diff --git a/benchmarks/data.js b/benchmarks/data.js
index 446cd748..762ac288 100644
--- a/benchmarks/data.js
+++ b/benchmarks/data.js
@@ -1,5 +1,5 @@
 window.BENCHMARK_DATA = {
-  "lastUpdate": 1724443708207,
+  "lastUpdate": 1724454030302,
   "repoUrl": "https://github.com/LuxDL/LuxLib.jl",
   "entries": {
     "LuxLib Benchmarks": [
@@ -126974,6 +126974,9498 @@ window.BENCHMARK_DATA = {
             "extra": "gctime=0\nmemory=77624\nallocs=1666\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
           }
         ]
+      },
+      {
+        "commit": {
+          "author": {
+            "email": "avikpal@mit.edu",
+            "name": "Avik Pal",
+            "username": "avik-pal"
+          },
+          "committer": {
+            "email": "avik.pal.2017@gmail.com",
+            "name": "Avik Pal",
+            "username": "avik-pal"
+          },
+          "distinct": true,
+          "id": "45fee0a4ac07b4907d153d927a6bdc10b77f20bc",
+          "message": "test: simple enzyme forward test to check no crash",
+          "timestamp": "2024-08-23T14:57:51-07:00",
+          "tree_id": "0c0785d421a458f0b84b084dd4bf5a0971136f9e",
+          "url": "https://github.com/LuxDL/LuxLib.jl/commit/45fee0a4ac07b4907d153d927a6bdc10b77f20bc"
+        },
+        "date": 1724454029707,
+        "tool": "julia",
+        "benches": [
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 6645.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1632\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 5958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1632\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 7792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1632\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 5437,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1632\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(4 x 32)/forward/GPU/CUDA",
+            "value": 117530,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11984\nallocs=467\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(4 x 32)/forward/GPU/Metal",
+            "value": 851584,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=40528\nallocs=1247\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 424104,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17360\nallocs=532\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 9834,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15136\nallocs=102\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 10042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15136\nallocs=102\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 10000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15136\nallocs=102\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 9708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15136\nallocs=102\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(4 x 32)/zygote/GPU/CUDA",
+            "value": 538763,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=73824\nallocs=2886\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(4 x 32)/zygote/GPU/Metal",
+            "value": 2566416,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=233136\nallocs=7642\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 685807,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=54496\nallocs=1194\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=relu)(32 x 128)/forward/CPU/2 thread(s)",
+            "value": 1625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=relu)(32 x 128)/forward/CPU/4 thread(s)",
+            "value": 2416.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=relu)(32 x 128)/forward/CPU/8 thread(s)",
+            "value": 1895.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=relu)(32 x 128)/forward/CPU/1 thread(s)",
+            "value": 2958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=relu)(32 x 128)/forward/GPU/CUDA",
+            "value": 21687,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2384\nallocs=107\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=relu)(32 x 128)/forward/GPU/Metal",
+            "value": 212666.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7160\nallocs=233\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=relu)(32 x 128)/forward/GPU/AMDGPU",
+            "value": 37011,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1888\nallocs=70\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=relu)(32 x 128)/zygote/CPU/2 thread(s)",
+            "value": 3958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=50496\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=relu)(32 x 128)/zygote/CPU/4 thread(s)",
+            "value": 3854,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=50496\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=relu)(32 x 128)/zygote/CPU/8 thread(s)",
+            "value": 3937.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=50496\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=relu)(32 x 128)/zygote/CPU/1 thread(s)",
+            "value": 4354.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=50496\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=relu)(32 x 128)/zygote/GPU/CUDA",
+            "value": 143890,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16840\nallocs=631\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=relu)(32 x 128)/zygote/GPU/Metal",
+            "value": 1548375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=55984\nallocs=1868\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=relu)(32 x 128)/zygote/GPU/AMDGPU",
+            "value": 150252,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16344\nallocs=385\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 57792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 39959,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 46666,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 83333,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 37041,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3712\nallocs=128\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1028896,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13760\nallocs=447\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 82890.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7088\nallocs=208\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 2029520.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7355648\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 2096833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7355648\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 2092209,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7355648\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 1990709,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7355648\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 227911.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=21240\nallocs=777\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 7599333,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=211512\nallocs=6761\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1314332,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=94088\nallocs=2266\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 147417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1050128\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 150979,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1050128\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 152333,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1050128\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 152708.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1050128\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 165380,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17904\nallocs=612\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1531291.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=56176\nallocs=1803\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 174842,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18960\nallocs=437\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 1106333.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8400368\nallocs=142\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 1125708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8400368\nallocs=142\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 1115083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8400368\nallocs=142\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 1115937.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8400368\nallocs=142\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 695023,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=83920\nallocs=3063\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 5817541,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=280824\nallocs=8736\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1029500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=74560\nallocs=1556\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 5125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2400\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 4458.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2400\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 6083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2400\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 5000.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2400\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(4 x 32)/forward/GPU/CUDA",
+            "value": 91412.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13104\nallocs=537\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(4 x 32)/forward/GPU/Metal",
+            "value": 448979,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=37584\nallocs=1260\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 60800,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10896\nallocs=270\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 8666,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17152\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 8792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17152\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 9042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17152\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 8792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17152\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(4 x 32)/zygote/GPU/CUDA",
+            "value": 598490,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=75968\nallocs=3041\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(4 x 32)/zygote/GPU/Metal",
+            "value": 5903917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=244040\nallocs=8111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 389469,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=56560\nallocs=1238\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 20667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134288\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 19250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134288\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 19958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134288\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 18541,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134288\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 68484,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12544\nallocs=282\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 1311833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=34288\nallocs=1053\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 73831,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=333\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 212750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=813456\nallocs=127\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 220458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=813456\nallocs=127\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 220250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=813456\nallocs=127\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 221500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=813456\nallocs=127\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 354536,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=64176\nallocs=1466\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 5845458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=171040\nallocs=4725\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 474284,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=68112\nallocs=1148\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=relu)(2 x 128)/forward/CPU/2 thread(s)",
+            "value": 625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=relu)(2 x 128)/forward/CPU/4 thread(s)",
+            "value": 667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=relu)(2 x 128)/forward/CPU/8 thread(s)",
+            "value": 833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=relu)(2 x 128)/forward/CPU/1 thread(s)",
+            "value": 645.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=relu)(2 x 128)/forward/GPU/CUDA",
+            "value": 20000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2384\nallocs=107\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=relu)(2 x 128)/forward/GPU/Metal",
+            "value": 294458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6880\nallocs=231\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=relu)(2 x 128)/forward/GPU/AMDGPU",
+            "value": 32961,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1888\nallocs=70\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=relu)(2 x 128)/zygote/CPU/2 thread(s)",
+            "value": 1417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3712\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=relu)(2 x 128)/zygote/CPU/4 thread(s)",
+            "value": 1417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3712\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=relu)(2 x 128)/zygote/CPU/8 thread(s)",
+            "value": 1583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3712\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=relu)(2 x 128)/zygote/CPU/1 thread(s)",
+            "value": 1583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3712\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=relu)(2 x 128)/zygote/GPU/CUDA",
+            "value": 123534,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=14616\nallocs=562\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=relu)(2 x 128)/zygote/GPU/Metal",
+            "value": 1493874.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=54336\nallocs=1843\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=relu)(2 x 128)/zygote/GPU/AMDGPU",
+            "value": 126886,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13448\nallocs=319\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 7458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 5334,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 6250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 10292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 24129,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2800\nallocs=95\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 615729,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16032\nallocs=533\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 47500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6576\nallocs=166\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 230041,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=933072\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 237500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=933072\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 262792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=933072\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 233458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=933072\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 182530.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19384\nallocs=711\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 9080458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=247640\nallocs=8001\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 610315.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=79208\nallocs=1823\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=relu)(32 x 128)/forward/CPU/2 thread(s)",
+            "value": 4125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=relu)(32 x 128)/forward/CPU/4 thread(s)",
+            "value": 4125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=relu)(32 x 128)/forward/CPU/8 thread(s)",
+            "value": 4125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=relu)(32 x 128)/forward/CPU/1 thread(s)",
+            "value": 4125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=relu)(32 x 128)/forward/GPU/CUDA",
+            "value": 23152,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=960\nallocs=41\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=relu)(32 x 128)/forward/GPU/Metal",
+            "value": 224833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7384\nallocs=263\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=relu)(32 x 128)/forward/GPU/AMDGPU",
+            "value": 49110,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2064\nallocs=84\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=relu)(32 x 128)/zygote/CPU/2 thread(s)",
+            "value": 16833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=71920\nallocs=26\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=relu)(32 x 128)/zygote/CPU/4 thread(s)",
+            "value": 16584,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=71920\nallocs=26\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=relu)(32 x 128)/zygote/CPU/8 thread(s)",
+            "value": 17084,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=71920\nallocs=26\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=relu)(32 x 128)/zygote/CPU/1 thread(s)",
+            "value": 16916,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=71920\nallocs=26\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=relu)(32 x 128)/zygote/GPU/CUDA",
+            "value": 192479.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15672\nallocs=581\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=relu)(32 x 128)/zygote/GPU/Metal",
+            "value": 965958.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49448\nallocs=1643\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=relu)(32 x 128)/zygote/GPU/AMDGPU",
+            "value": 178772,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13768\nallocs=346\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=gelu)(512 x 128)/forward/CPU/2 thread(s)",
+            "value": 509792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=gelu)(512 x 128)/forward/CPU/4 thread(s)",
+            "value": 332375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=gelu)(512 x 128)/forward/CPU/8 thread(s)",
+            "value": 404500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=gelu)(512 x 128)/forward/CPU/1 thread(s)",
+            "value": 864583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=gelu)(512 x 128)/forward/GPU/CUDA",
+            "value": 113511.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=960\nallocs=41\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=gelu)(512 x 128)/forward/GPU/Metal",
+            "value": 474250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7408\nallocs=264\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=gelu)(512 x 128)/forward/GPU/AMDGPU",
+            "value": 248712,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8240\nallocs=328\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=gelu)(512 x 128)/zygote/CPU/2 thread(s)",
+            "value": 2322292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2360672\nallocs=33\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=gelu)(512 x 128)/zygote/CPU/4 thread(s)",
+            "value": 1755583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2360672\nallocs=33\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=gelu)(512 x 128)/zygote/CPU/8 thread(s)",
+            "value": 2033750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2360672\nallocs=33\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=gelu)(512 x 128)/zygote/CPU/1 thread(s)",
+            "value": 3280334,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2360672\nallocs=33\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=gelu)(512 x 128)/zygote/GPU/CUDA",
+            "value": 238729,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18568\nallocs=713\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=gelu)(512 x 128)/zygote/GPU/Metal",
+            "value": 1974104,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=58840\nallocs=1949\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=gelu)(512 x 128)/zygote/GPU/AMDGPU",
+            "value": 760717.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16760\nallocs=520\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 6520.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5408\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 6292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5408\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 7583.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5408\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 5875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5408\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(32 x 32)/forward/GPU/CUDA",
+            "value": 90935.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11984\nallocs=467\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(32 x 32)/forward/GPU/Metal",
+            "value": 743625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=36856\nallocs=1231\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 60411,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10944\nallocs=273\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 11625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=51120\nallocs=87\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 11958.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=51120\nallocs=87\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 12271,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=51120\nallocs=87\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 12396,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=51120\nallocs=87\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(32 x 32)/zygote/GPU/CUDA",
+            "value": 637206,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=72848\nallocs=2885\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(32 x 32)/zygote/GPU/Metal",
+            "value": 5519292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=231728\nallocs=7544\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 414184,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=54128\nallocs=1185\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=relu)(2 x 128)/forward/CPU/2 thread(s)",
+            "value": 541,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=relu)(2 x 128)/forward/CPU/4 thread(s)",
+            "value": 500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=relu)(2 x 128)/forward/CPU/8 thread(s)",
+            "value": 500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=relu)(2 x 128)/forward/CPU/1 thread(s)",
+            "value": 500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=relu)(2 x 128)/forward/GPU/CUDA",
+            "value": 23373,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1056\nallocs=47\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=relu)(2 x 128)/forward/GPU/Metal",
+            "value": 325562.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7696\nallocs=277\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=relu)(2 x 128)/forward/GPU/AMDGPU",
+            "value": 54120,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1888\nallocs=70\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=relu)(2 x 128)/zygote/CPU/2 thread(s)",
+            "value": 2084,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4880\nallocs=15\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=relu)(2 x 128)/zygote/CPU/4 thread(s)",
+            "value": 2083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4880\nallocs=15\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=relu)(2 x 128)/zygote/CPU/8 thread(s)",
+            "value": 2167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4880\nallocs=15\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=relu)(2 x 128)/zygote/CPU/1 thread(s)",
+            "value": 2125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4880\nallocs=15\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=relu)(2 x 128)/zygote/GPU/CUDA",
+            "value": 220940,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15592\nallocs=578\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=relu)(2 x 128)/zygote/GPU/Metal",
+            "value": 1999750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=59120\nallocs=2034\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=relu)(2 x 128)/zygote/GPU/AMDGPU",
+            "value": 182151,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=14408\nallocs=343\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 8959,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7344\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 8917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7344\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 10479.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7344\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 9458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7344\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(32 x 32)/forward/GPU/CUDA",
+            "value": 107673.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12752\nallocs=309\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(32 x 32)/forward/GPU/Metal",
+            "value": 816459,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=35736\nallocs=1143\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 75421,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16128\nallocs=351\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 18145.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=60944\nallocs=124\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 17834,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=60944\nallocs=124\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 18875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=60944\nallocs=124\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 18563,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=60944\nallocs=124\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(32 x 32)/zygote/GPU/CUDA",
+            "value": 617019,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=73832\nallocs=1708\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(32 x 32)/zygote/GPU/Metal",
+            "value": 5305458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=190776\nallocs=5371\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 387888.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=77832\nallocs=1299\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 541,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(32 x 32)/forward/GPU/CUDA",
+            "value": 35086,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4848\nallocs=166\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(32 x 32)/forward/GPU/Metal",
+            "value": 440000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13656\nallocs=447\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 45921,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6480\nallocs=173\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 10750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=40448\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 10479,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=40448\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 10291.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=40448\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 10250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=40448\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(32 x 32)/zygote/GPU/CUDA",
+            "value": 261459,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=20192\nallocs=724\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(32 x 32)/zygote/GPU/Metal",
+            "value": 4700125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=152672\nallocs=5001\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 373724,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49312\nallocs=1213\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=identity)(512 x 128)/forward/CPU/2 thread(s)",
+            "value": 396937.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=identity)(512 x 128)/forward/CPU/4 thread(s)",
+            "value": 215458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=identity)(512 x 128)/forward/CPU/8 thread(s)",
+            "value": 288000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=identity)(512 x 128)/forward/CPU/1 thread(s)",
+            "value": 756125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=identity)(512 x 128)/forward/GPU/CUDA",
+            "value": 112222,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1168\nallocs=48\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=identity)(512 x 128)/forward/GPU/Metal",
+            "value": 396166,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2736\nallocs=96\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=identity)(512 x 128)/forward/GPU/AMDGPU",
+            "value": 76981,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2064\nallocs=120\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=identity)(512 x 128)/zygote/CPU/2 thread(s)",
+            "value": 1455708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1836144\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=identity)(512 x 128)/zygote/CPU/4 thread(s)",
+            "value": 857833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1836144\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=identity)(512 x 128)/zygote/CPU/8 thread(s)",
+            "value": 1133208,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1836144\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=identity)(512 x 128)/zygote/CPU/1 thread(s)",
+            "value": 2441333.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1836144\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=identity)(512 x 128)/zygote/GPU/CUDA",
+            "value": 205916,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13808\nallocs=502\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=identity)(512 x 128)/zygote/GPU/Metal",
+            "value": 1638958.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=45152\nallocs=1518\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=identity)(512 x 128)/zygote/GPU/AMDGPU",
+            "value": 325533,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12944\nallocs=380\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 6916,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13728\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 7833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13728\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 8417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13728\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 7208,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13728\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(32 x 32)/forward/GPU/CUDA",
+            "value": 152876.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13104\nallocs=537\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(32 x 32)/forward/GPU/Metal",
+            "value": 451312.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=38232\nallocs=1254\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 62081,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10800\nallocs=264\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 15708.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=89120\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 14854,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=89120\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 14521,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=89120\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 16625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=89120\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(32 x 32)/zygote/GPU/CUDA",
+            "value": 1044876,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=78192\nallocs=3110\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(32 x 32)/zygote/GPU/Metal",
+            "value": 5855584,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=248432\nallocs=8060\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=true)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 433489.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=59328\nallocs=1293\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 31750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132624\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 28417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132624\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 28166,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132624\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 25875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132624\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 225622,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17904\nallocs=612\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 631917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=48184\nallocs=1530\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 118121,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18304\nallocs=396\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 112958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1060336\nallocs=142\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 103792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1060336\nallocs=142\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 151396,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1060336\nallocs=142\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 146291.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1060336\nallocs=142\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 1190197.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=83920\nallocs=3063\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 5968062.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=256368\nallocs=7902\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 597156,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=70336\nallocs=1439\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 88250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=394480\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 77750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=394480\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 78291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=394480\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 74166,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=394480\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 231353,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19024\nallocs=682\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 537791.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49976\nallocs=1562\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 127991.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18304\nallocs=396\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 279770.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2508432\nallocs=149\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 306041,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2508432\nallocs=149\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 319103.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2508432\nallocs=149\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 301792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2508432\nallocs=149\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 1223547,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=90592\nallocs=3393\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 6540125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=288696\nallocs=8884\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 698201.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=75808\nallocs=1556\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 17042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13728\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 17250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13728\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 17146,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13728\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 16771,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13728\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(32 x 32)/forward/GPU/CUDA",
+            "value": 164086,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13104\nallocs=537\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(32 x 32)/forward/GPU/Metal",
+            "value": 451167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=38232\nallocs=1254\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 238683,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13424\nallocs=428\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 27354.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=89120\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 24833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=89120\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 27312.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=89120\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 27749.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=89120\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(32 x 32)/zygote/GPU/CUDA",
+            "value": 1038807,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=79248\nallocs=3154\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(32 x 32)/zygote/GPU/Metal",
+            "value": 6017250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=248240\nallocs=8054\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 701206,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=59728\nallocs=1299\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 11395.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7344\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 12250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7344\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 12375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7344\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 11333,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7344\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(32 x 32)/forward/GPU/CUDA",
+            "value": 139180,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12496\nallocs=308\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(32 x 32)/forward/GPU/Metal",
+            "value": 842812.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=35736\nallocs=1143\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 241222,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=22416\nallocs=602\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 21687,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=52368\nallocs=104\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 21271,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=52368\nallocs=104\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 22125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=52368\nallocs=104\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 22625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=52368\nallocs=104\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(32 x 32)/zygote/GPU/CUDA",
+            "value": 749757,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=62104\nallocs=1758\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(32 x 32)/zygote/GPU/Metal",
+            "value": 5233187,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=179904\nallocs=5393\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 686992,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=66264\nallocs=1365\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 66999.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134288\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 63750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134288\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 65708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134288\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 63479,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134288\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 120818,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12544\nallocs=282\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 1336750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=34288\nallocs=1053\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 228922.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16816\nallocs=344\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 484146,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1064816\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 449000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1064816\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 445542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1064816\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 448916.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1064816\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 559738.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=56264\nallocs=1599\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 6432542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=169656\nallocs=4968\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 721207,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=58664\nallocs=1178\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 7624.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13728\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 7834,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13728\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 7750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13728\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 7458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13728\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(32 x 32)/forward/GPU/CUDA",
+            "value": 161448.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13104\nallocs=537\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(32 x 32)/forward/GPU/Metal",
+            "value": 449750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=38232\nallocs=1254\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 59581,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10912\nallocs=271\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 13708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=80352\nallocs=86\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 14625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=80352\nallocs=86\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 15667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=80352\nallocs=86\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 14541.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=80352\nallocs=86\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(32 x 32)/zygote/GPU/CUDA",
+            "value": 1019503,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=75120\nallocs=2979\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(32 x 32)/zygote/GPU/Metal",
+            "value": 5626021,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=235968\nallocs=7662\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 403264,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=57328\nallocs=1244\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=4)/forward/CPU/2 thread(s)",
+            "value": 6154458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4194368\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=4)/forward/CPU/4 thread(s)",
+            "value": 3223916.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4194368\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=4)/forward/CPU/8 thread(s)",
+            "value": 6375625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4194368\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=4)/forward/CPU/1 thread(s)",
+            "value": 11909375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4194368\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=4)/forward/GPU/CUDA",
+            "value": 301453,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=512\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=4)/forward/GPU/AMDGPU",
+            "value": 294773,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5472\nallocs=294\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=4)/zygote/CPU/2 thread(s)",
+            "value": 19093562.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16778320\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=4)/zygote/CPU/4 thread(s)",
+            "value": 11085604,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16778320\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=4)/zygote/CPU/8 thread(s)",
+            "value": 19938020.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16778320\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=4)/zygote/CPU/1 thread(s)",
+            "value": 36629791,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16778128\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=4)/zygote/GPU/CUDA",
+            "value": 1145132,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11824\nallocs=431\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=4)/zygote/GPU/AMDGPU",
+            "value": 1161766.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=20192\nallocs=583\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=gelu)(2 x 128)/forward/CPU/2 thread(s)",
+            "value": 958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=gelu)(2 x 128)/forward/CPU/4 thread(s)",
+            "value": 1000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=gelu)(2 x 128)/forward/CPU/8 thread(s)",
+            "value": 958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=gelu)(2 x 128)/forward/CPU/1 thread(s)",
+            "value": 958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=gelu)(2 x 128)/forward/GPU/CUDA",
+            "value": 23494,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1056\nallocs=47\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=gelu)(2 x 128)/forward/GPU/Metal",
+            "value": 329625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7696\nallocs=277\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=gelu)(2 x 128)/forward/GPU/AMDGPU",
+            "value": 216242,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8320\nallocs=330\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=gelu)(2 x 128)/zygote/CPU/2 thread(s)",
+            "value": 3667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5776\nallocs=12\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=gelu)(2 x 128)/zygote/CPU/4 thread(s)",
+            "value": 3708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5776\nallocs=12\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=gelu)(2 x 128)/zygote/CPU/8 thread(s)",
+            "value": 3750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5776\nallocs=12\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=gelu)(2 x 128)/zygote/CPU/1 thread(s)",
+            "value": 3667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5776\nallocs=12\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=gelu)(2 x 128)/zygote/GPU/CUDA",
+            "value": 296983,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17960\nallocs=683\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=gelu)(2 x 128)/zygote/GPU/Metal",
+            "value": 2127729,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=64992\nallocs=2231\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=gelu)(2 x 128)/zygote/GPU/AMDGPU",
+            "value": 647936,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17576\nallocs=495\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 8334,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3616\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 8041,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3616\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 9688,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3616\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 7917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3616\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(4 x 32)/forward/GPU/CUDA",
+            "value": 136939,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12368\nallocs=300\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(4 x 32)/forward/GPU/Metal",
+            "value": 738708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=35440\nallocs=1140\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 67671,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15920\nallocs=338\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 11584,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=34432\nallocs=124\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 11792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=34432\nallocs=124\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 12750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=34432\nallocs=124\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 12625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=34432\nallocs=124\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(4 x 32)/zygote/GPU/CUDA",
+            "value": 721402,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=71096\nallocs=1622\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(4 x 32)/zygote/GPU/Metal",
+            "value": 5162750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=188288\nallocs=5386\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=true)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 363024,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=74088\nallocs=1180\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=relu)(2 x 128)/forward/CPU/2 thread(s)",
+            "value": 292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=relu)(2 x 128)/forward/CPU/4 thread(s)",
+            "value": 292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=relu)(2 x 128)/forward/CPU/8 thread(s)",
+            "value": 292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=relu)(2 x 128)/forward/CPU/1 thread(s)",
+            "value": 250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=relu)(2 x 128)/forward/GPU/CUDA",
+            "value": 22847,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=960\nallocs=41\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=relu)(2 x 128)/forward/GPU/Metal",
+            "value": 317792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7104\nallocs=261\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=relu)(2 x 128)/forward/GPU/AMDGPU",
+            "value": 51281,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1968\nallocs=78\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=relu)(2 x 128)/zygote/CPU/2 thread(s)",
+            "value": 3000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=26\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=relu)(2 x 128)/zygote/CPU/4 thread(s)",
+            "value": 2875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=26\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=relu)(2 x 128)/zygote/CPU/8 thread(s)",
+            "value": 3125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=26\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=relu)(2 x 128)/zygote/CPU/1 thread(s)",
+            "value": 2875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=26\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=relu)(2 x 128)/zygote/GPU/CUDA",
+            "value": 210440.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13448\nallocs=512\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=relu)(2 x 128)/zygote/GPU/Metal",
+            "value": 1577208,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=46784\nallocs=1608\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=relu)(2 x 128)/zygote/GPU/AMDGPU",
+            "value": 160702,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10680\nallocs=268\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 10625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7088\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 11709,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7088\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 13167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7088\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 11167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7088\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(32 x 32)/forward/GPU/CUDA",
+            "value": 138633,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11664\nallocs=282\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(32 x 32)/forward/GPU/Metal",
+            "value": 907792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=33016\nallocs=1047\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 239227,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18720\nallocs=538\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 21062.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=48912\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 23604.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=48912\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 21416.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=48912\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 20958.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=48912\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(32 x 32)/zygote/GPU/CUDA",
+            "value": 652488.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=54152\nallocs=1599\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(32 x 32)/zygote/GPU/Metal",
+            "value": 4746874.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=153416\nallocs=4612\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 654136,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=55992\nallocs=1179\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=relu)(32 x 128)/forward/CPU/2 thread(s)",
+            "value": 4416,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=relu)(32 x 128)/forward/CPU/4 thread(s)",
+            "value": 4542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=relu)(32 x 128)/forward/CPU/8 thread(s)",
+            "value": 4375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=relu)(32 x 128)/forward/CPU/1 thread(s)",
+            "value": 4334,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=relu)(32 x 128)/forward/GPU/CUDA",
+            "value": 24791.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1056\nallocs=47\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=relu)(32 x 128)/forward/GPU/Metal",
+            "value": 227833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7976\nallocs=279\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=relu)(32 x 128)/forward/GPU/AMDGPU",
+            "value": 52640,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2176\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=relu)(32 x 128)/zygote/CPU/2 thread(s)",
+            "value": 16542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=71488\nallocs=15\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=relu)(32 x 128)/zygote/CPU/4 thread(s)",
+            "value": 16375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=71488\nallocs=15\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=relu)(32 x 128)/zygote/CPU/8 thread(s)",
+            "value": 16708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=71488\nallocs=15\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=relu)(32 x 128)/zygote/CPU/1 thread(s)",
+            "value": 16625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=71488\nallocs=15\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=relu)(32 x 128)/zygote/GPU/CUDA",
+            "value": 351658.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17816\nallocs=647\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=relu)(32 x 128)/zygote/GPU/Metal",
+            "value": 1485875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=61216\nallocs=2053\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=relu)(32 x 128)/zygote/GPU/AMDGPU",
+            "value": 212337,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17672\nallocs=432\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 1958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 2042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 2042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 2000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(32 x 32)/forward/GPU/CUDA",
+            "value": 36523,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4848\nallocs=166\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(32 x 32)/forward/GPU/Metal",
+            "value": 462667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13656\nallocs=447\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 207782,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12848\nallocs=429\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 19125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49152\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 18458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49152\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 18833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49152\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 18125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49152\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(32 x 32)/zygote/GPU/CUDA",
+            "value": 303670,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=23592\nallocs=890\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(32 x 32)/zygote/GPU/Metal",
+            "value": 4639708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=161880\nallocs=5261\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 702222,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=51960\nallocs=1325\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=512)/forward/CPU/2 thread(s)",
+            "value": 60208,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=524352\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=512)/forward/CPU/4 thread(s)",
+            "value": 60959,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=524352\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=512)/forward/CPU/8 thread(s)",
+            "value": 65604.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=524352\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=512)/forward/CPU/1 thread(s)",
+            "value": 51083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=524352\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=512)/forward/GPU/CUDA",
+            "value": 66579,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=512\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=512)/forward/GPU/AMDGPU",
+            "value": 95291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=14208\nallocs=66\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=512)/zygote/CPU/2 thread(s)",
+            "value": 199667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2098256\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=512)/zygote/CPU/4 thread(s)",
+            "value": 115166,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2098256\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=512)/zygote/CPU/8 thread(s)",
+            "value": 119292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2098256\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=512)/zygote/CPU/1 thread(s)",
+            "value": 295479,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2098064\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=512)/zygote/GPU/CUDA",
+            "value": 229950,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11808\nallocs=430\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=512)/zygote/GPU/AMDGPU",
+            "value": 584990.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=55040\nallocs=439\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 86146,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051792\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 81709,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051792\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 87667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051792\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 127312.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051792\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 192185,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12560\nallocs=283\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1983854.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=50976\nallocs=1614\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 169421,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=24240\nallocs=514\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 1920229,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6318480\nallocs=127\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 1925709,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6318480\nallocs=127\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 1919895.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6318480\nallocs=127\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 1914750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6318480\nallocs=127\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 573852,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=64208\nallocs=1468\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 9135542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=188160\nallocs=5313\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1085326,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=88160\nallocs=1694\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=identity)(2 x 128)/forward/CPU/2 thread(s)",
+            "value": 291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=identity)(2 x 128)/forward/CPU/4 thread(s)",
+            "value": 292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=identity)(2 x 128)/forward/CPU/8 thread(s)",
+            "value": 292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=identity)(2 x 128)/forward/CPU/1 thread(s)",
+            "value": 292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=identity)(2 x 128)/forward/GPU/CUDA",
+            "value": 21768,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1056\nallocs=47\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=identity)(2 x 128)/forward/GPU/Metal",
+            "value": 353875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7072\nallocs=261\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=identity)(2 x 128)/forward/GPU/AMDGPU",
+            "value": 46981,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1936\nallocs=76\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=identity)(2 x 128)/zygote/CPU/2 thread(s)",
+            "value": 1833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3584\nallocs=9\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=identity)(2 x 128)/zygote/CPU/4 thread(s)",
+            "value": 1875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3584\nallocs=9\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=identity)(2 x 128)/zygote/CPU/8 thread(s)",
+            "value": 1834,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3584\nallocs=9\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=identity)(2 x 128)/zygote/CPU/1 thread(s)",
+            "value": 1792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3584\nallocs=9\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=identity)(2 x 128)/zygote/GPU/CUDA",
+            "value": 267259,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13200\nallocs=472\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=identity)(2 x 128)/zygote/GPU/Metal",
+            "value": 1021146,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=52520\nallocs=1821\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=true, act=identity)(2 x 128)/zygote/GPU/AMDGPU",
+            "value": 185402,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13344\nallocs=329\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 10042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3360\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 8250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3360\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 10625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3360\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 9000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3360\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(4 x 32)/forward/GPU/CUDA",
+            "value": 134677.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11536\nallocs=274\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(4 x 32)/forward/GPU/Metal",
+            "value": 887250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=32720\nallocs=1044\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 239022,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=21872\nallocs=593\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 9500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18896\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 11709,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18896\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 9395.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18896\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 9208,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18896\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(4 x 32)/zygote/GPU/CUDA",
+            "value": 583672,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=51672\nallocs=1514\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(4 x 32)/zygote/GPU/Metal",
+            "value": 4182708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=150864\nallocs=4623\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=false)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 636841.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=52680\nallocs=1087\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 58000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 39791,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 46708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 82834,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 40507,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3536\nallocs=119\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1125958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16032\nallocs=533\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 76281,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6832\nallocs=182\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 1893542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6308192\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 1983291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6308192\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 1983291.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6308192\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 1902916,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6308192\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 236090,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=22144\nallocs=649\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 11366312.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=240632\nallocs=7788\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1031795,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=83456\nallocs=1981\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 417500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=394480\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 418750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=394480\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 421625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=394480\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 417042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=394480\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 235292.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19024\nallocs=682\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 542791.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49976\nallocs=1562\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 285283,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=21712\nallocs=609\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 772062.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2508432\nallocs=149\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 706208.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2508432\nallocs=149\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 697646.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2508432\nallocs=149\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 762583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2508432\nallocs=149\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 1149490,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=91648\nallocs=3437\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 6544625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=289016\nallocs=8896\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 922659,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=76128\nallocs=1558\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 3408770.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3146992\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 3443353.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3146992\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 3457250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3146992\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 3456792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3146992\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 171296,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19024\nallocs=682\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1419875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=58128\nallocs=1840\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 409534,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19824\nallocs=491\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 6173000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19997968\nallocs=151\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 6227062,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19997968\nallocs=151\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 6229500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19997968\nallocs=151\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 6220208,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19997968\nallocs=151\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 1077389,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=91648\nallocs=3437\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 7523396,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=314112\nallocs=9747\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1562795,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=79616\nallocs=1629\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=gelu)(512 x 128)/forward/CPU/2 thread(s)",
+            "value": 471979.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=gelu)(512 x 128)/forward/CPU/4 thread(s)",
+            "value": 254250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=gelu)(512 x 128)/forward/CPU/8 thread(s)",
+            "value": 341666,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=gelu)(512 x 128)/forward/CPU/1 thread(s)",
+            "value": 903750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=gelu)(512 x 128)/forward/GPU/CUDA",
+            "value": 47482,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1056\nallocs=47\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=gelu)(512 x 128)/forward/GPU/Metal",
+            "value": 444459,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8016\nallocs=281\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=gelu)(512 x 128)/forward/GPU/AMDGPU",
+            "value": 250912,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2016\nallocs=78\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=gelu)(512 x 128)/zygote/CPU/2 thread(s)",
+            "value": 2323333,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2362080\nallocs=18\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=gelu)(512 x 128)/zygote/CPU/4 thread(s)",
+            "value": 1760000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2362080\nallocs=18\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=gelu)(512 x 128)/zygote/CPU/8 thread(s)",
+            "value": 2038125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2362080\nallocs=18\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=gelu)(512 x 128)/zygote/CPU/1 thread(s)",
+            "value": 3288375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2362080\nallocs=18\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=gelu)(512 x 128)/zygote/GPU/CUDA",
+            "value": 256085,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=20504\nallocs=772\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=gelu)(512 x 128)/zygote/GPU/Metal",
+            "value": 2239750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=75576\nallocs=2533\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=gelu)(512 x 128)/zygote/GPU/AMDGPU",
+            "value": 788328,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=21320\nallocs=614\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 57770.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 39583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 46209,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 82958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 28615,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2800\nallocs=95\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1133833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16032\nallocs=533\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 76141,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7248\nallocs=208\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 2030708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7356832\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 2061250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7356832\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 2095541.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7356832\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 1953708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7356832\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 235210,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19384\nallocs=711\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 11813583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=247640\nallocs=8001\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1051380,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=83912\nallocs=1970\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 58041,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 39500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 46583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 83000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 49005,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4448\nallocs=152\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1119604.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13760\nallocs=447\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 74191,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6688\nallocs=183\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 1921625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6307008\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 1977958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6307008\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 1975834,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6307008\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 1880167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6307008\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 241693.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19648\nallocs=695\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 9953791,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=202400\nallocs=6495\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 942554.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=78624\nallocs=1628\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(4 x 32)/forward/GPU/CUDA",
+            "value": 34479,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4848\nallocs=166\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(4 x 32)/forward/GPU/Metal",
+            "value": 430437,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13344\nallocs=443\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 50680,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6576\nallocs=179\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 6792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16784\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 7958.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16784\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 7916.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16784\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 7750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16784\nallocs=92\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(4 x 32)/zygote/GPU/CUDA",
+            "value": 203416,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17968\nallocs=655\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(4 x 32)/zygote/GPU/Metal",
+            "value": 4793624.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=150536\nallocs=5007\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=false)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 372109,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=46208\nallocs=1134\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=identity)(2 x 128)/forward/CPU/2 thread(s)",
+            "value": 250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=identity)(2 x 128)/forward/CPU/4 thread(s)",
+            "value": 292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=identity)(2 x 128)/forward/CPU/8 thread(s)",
+            "value": 291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=identity)(2 x 128)/forward/CPU/1 thread(s)",
+            "value": 250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=identity)(2 x 128)/forward/GPU/CUDA",
+            "value": 31576,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1008\nallocs=38\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=identity)(2 x 128)/forward/GPU/Metal",
+            "value": 256375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2496\nallocs=97\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=identity)(2 x 128)/forward/GPU/AMDGPU",
+            "value": 41481,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1008\nallocs=54\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=identity)(2 x 128)/zygote/CPU/2 thread(s)",
+            "value": 2625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4288\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=identity)(2 x 128)/zygote/CPU/4 thread(s)",
+            "value": 2959,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4288\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=identity)(2 x 128)/zygote/CPU/8 thread(s)",
+            "value": 3708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4288\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=identity)(2 x 128)/zygote/CPU/1 thread(s)",
+            "value": 3083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4288\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=identity)(2 x 128)/zygote/GPU/CUDA",
+            "value": 195018,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11104\nallocs=403\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=identity)(2 x 128)/zygote/GPU/Metal",
+            "value": 948875.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=36200\nallocs=1247\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=identity)(2 x 128)/zygote/GPU/AMDGPU",
+            "value": 155612,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8640\nallocs=226\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 421875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1052048\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 441709,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1052048\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 429167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1052048\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 435458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1052048\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 141479,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13392\nallocs=309\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 2056458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=53696\nallocs=1710\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 371433.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=30608\nallocs=745\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 3799959,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8408304\nallocs=123\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 3818354,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8408304\nallocs=123\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 3807708.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8408304\nallocs=123\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 3804750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8408304\nallocs=123\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 767195,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=69016\nallocs=1898\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 10786853.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=230376\nallocs=6922\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1321443,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=96008\nallocs=2083\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=32)/forward/CPU/2 thread(s)",
+            "value": 49849770.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=33554496\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=32)/forward/CPU/4 thread(s)",
+            "value": 26028875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=33554496\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=32)/forward/CPU/8 thread(s)",
+            "value": 35533000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=33554496\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=32)/forward/CPU/1 thread(s)",
+            "value": 98215750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=33554496\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=32)/forward/GPU/CUDA",
+            "value": 1616583.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=528\nallocs=24\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=32)/forward/GPU/AMDGPU",
+            "value": 1008870,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=9328\nallocs=348\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=32)/zygote/CPU/2 thread(s)",
+            "value": 154650938,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134218832\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=32)/zygote/CPU/4 thread(s)",
+            "value": 88783312.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134218832\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=32)/zygote/CPU/8 thread(s)",
+            "value": 112429209,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134218832\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=32)/zygote/CPU/1 thread(s)",
+            "value": 294910479,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134218640\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=32)/zygote/GPU/CUDA",
+            "value": 6458815,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11840\nallocs=432\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=32)/zygote/GPU/AMDGPU",
+            "value": 5555575,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=27072\nallocs=736\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=tanh)(32 x 128)/forward/CPU/2 thread(s)",
+            "value": 19000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=tanh)(32 x 128)/forward/CPU/4 thread(s)",
+            "value": 15250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=tanh)(32 x 128)/forward/CPU/8 thread(s)",
+            "value": 17333,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=tanh)(32 x 128)/forward/CPU/1 thread(s)",
+            "value": 15833.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=tanh)(32 x 128)/forward/GPU/CUDA",
+            "value": 19992,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2384\nallocs=107\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=tanh)(32 x 128)/forward/GPU/Metal",
+            "value": 220583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7176\nallocs=234\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=tanh)(32 x 128)/forward/GPU/AMDGPU",
+            "value": 26010,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1712\nallocs=59\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=tanh)(32 x 128)/zygote/CPU/2 thread(s)",
+            "value": 11083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=50496\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=tanh)(32 x 128)/zygote/CPU/4 thread(s)",
+            "value": 7874.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=50496\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=tanh)(32 x 128)/zygote/CPU/8 thread(s)",
+            "value": 9250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=50496\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=tanh)(32 x 128)/zygote/CPU/1 thread(s)",
+            "value": 17209,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=50496\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=tanh)(32 x 128)/zygote/GPU/CUDA",
+            "value": 289103,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16840\nallocs=631\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=tanh)(32 x 128)/zygote/GPU/Metal",
+            "value": 1568584,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=56552\nallocs=1881\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=tanh)(32 x 128)/zygote/GPU/AMDGPU",
+            "value": 152592,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15736\nallocs=347\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 9000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3616\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 8625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3616\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 10541,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3616\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 7500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3616\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(4 x 32)/forward/GPU/CUDA",
+            "value": 139112,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12368\nallocs=300\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(4 x 32)/forward/GPU/Metal",
+            "value": 697750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=35440\nallocs=1140\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 238993,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18992\nallocs=530\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 9354.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=22128\nallocs=104\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 9833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=22128\nallocs=104\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 10083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=22128\nallocs=104\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 9166.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=22128\nallocs=104\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(4 x 32)/zygote/GPU/CUDA",
+            "value": 687570,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=59624\nallocs=1673\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(4 x 32)/zygote/GPU/Metal",
+            "value": 5031875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=177168\nallocs=5408\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=gelu, affine=true)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 662346.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=62456\nallocs=1242\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 9917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7088\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 9541,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7088\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 10542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7088\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 8749.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7088\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(32 x 32)/forward/GPU/CUDA",
+            "value": 134423,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11664\nallocs=282\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(32 x 32)/forward/GPU/Metal",
+            "value": 923917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=33016\nallocs=1047\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 70081,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15568\nallocs=341\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 13458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=44512\nallocs=90\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 14542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=44512\nallocs=90\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 18166,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=44512\nallocs=90\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 13353.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=44512\nallocs=90\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(32 x 32)/zygote/GPU/CUDA",
+            "value": 641312,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=52312\nallocs=1515\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(32 x 32)/zygote/GPU/Metal",
+            "value": 4618500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=151056\nallocs=4562\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 358499,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=54296\nallocs=1083\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(4 x 32)/forward/GPU/CUDA",
+            "value": 34342,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4848\nallocs=166\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(4 x 32)/forward/GPU/Metal",
+            "value": 431187.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13344\nallocs=443\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 207942,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=9520\nallocs=363\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 7979.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17936\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 9167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17936\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 10042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17936\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 7958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17936\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(4 x 32)/zygote/GPU/CUDA",
+            "value": 228270.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=21368\nallocs=821\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(4 x 32)/zygote/GPU/Metal",
+            "value": 4859333,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=159360\nallocs=5274\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=false)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 663502,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=48104\nallocs=1199\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=gelu)(32 x 128)/forward/CPU/2 thread(s)",
+            "value": 16417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=gelu)(32 x 128)/forward/CPU/4 thread(s)",
+            "value": 14500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=gelu)(32 x 128)/forward/CPU/8 thread(s)",
+            "value": 14625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=gelu)(32 x 128)/forward/CPU/1 thread(s)",
+            "value": 10688,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=gelu)(32 x 128)/forward/GPU/CUDA",
+            "value": 21203,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2384\nallocs=107\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=gelu)(32 x 128)/forward/GPU/Metal",
+            "value": 211000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7176\nallocs=234\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=gelu)(32 x 128)/forward/GPU/AMDGPU",
+            "value": 190002,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8272\nallocs=327\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=gelu)(32 x 128)/zygote/CPU/2 thread(s)",
+            "value": 32021,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=67136\nallocs=14\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=gelu)(32 x 128)/zygote/CPU/4 thread(s)",
+            "value": 32333.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=67136\nallocs=14\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=gelu)(32 x 128)/zygote/CPU/8 thread(s)",
+            "value": 32000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=67136\nallocs=14\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=gelu)(32 x 128)/zygote/CPU/1 thread(s)",
+            "value": 31958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=67136\nallocs=14\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=gelu)(32 x 128)/zygote/GPU/CUDA",
+            "value": 306523.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19208\nallocs=736\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=gelu)(32 x 128)/zygote/GPU/Metal",
+            "value": 1710500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=62696\nallocs=2079\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(32, act=gelu)(32 x 128)/zygote/GPU/AMDGPU",
+            "value": 607906,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19768\nallocs=550\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 440750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051792\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 447459,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051792\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 462000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051792\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 440958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051792\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 195262,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12560\nallocs=283\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1977271,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=50976\nallocs=1614\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 352533,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=30112\nallocs=739\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 3835270.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8404848\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 3836854,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8404848\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 3833000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8404848\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 3792291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8404848\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 575749.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=56296\nallocs=1601\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 9765667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=187400\nallocs=5572\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1392133,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=80344\nallocs=1826\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=512)/forward/CPU/2 thread(s)",
+            "value": 832773604.5,
+            "unit": "ns",
+            "extra": "gctime=49778812.5\nmemory=536870976\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=512)/forward/CPU/4 thread(s)",
+            "value": 416178167,
+            "unit": "ns",
+            "extra": "gctime=415666\nmemory=536870976\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=512)/forward/CPU/8 thread(s)",
+            "value": 542118041,
+            "unit": "ns",
+            "extra": "gctime=437292\nmemory=536870976\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=512)/forward/CPU/1 thread(s)",
+            "value": 1579553104.5,
+            "unit": "ns",
+            "extra": "gctime=49996770.5\nmemory=536870976\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=512)/forward/GPU/CUDA",
+            "value": 22761017,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=528\nallocs=24\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=512)/forward/GPU/AMDGPU",
+            "value": 14155420,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=20992\nallocs=348\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=512)/zygote/CPU/2 thread(s)",
+            "value": 3970317583,
+            "unit": "ns",
+            "extra": "gctime=128399168\nmemory=2147484752\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=512)/zygote/CPU/4 thread(s)",
+            "value": 1513852416,
+            "unit": "ns",
+            "extra": "gctime=36266459\nmemory=2147484752\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=512)/zygote/CPU/8 thread(s)",
+            "value": 1789856541,
+            "unit": "ns",
+            "extra": "gctime=30905416\nmemory=2147484752\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=512)/zygote/CPU/1 thread(s)",
+            "value": 5283712875,
+            "unit": "ns",
+            "extra": "gctime=44651958\nmemory=2147484560\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=512)/zygote/GPU/CUDA",
+            "value": 365579607,
+            "unit": "ns",
+            "extra": "gctime=295145420\nmemory=12048\nallocs=437\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=512)/zygote/GPU/AMDGPU",
+            "value": 88621175,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=65536\nallocs=806\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 76438,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=394480\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 77000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=394480\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 79229,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=394480\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 76708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=394480\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 229997.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19024\nallocs=682\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 543458.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49976\nallocs=1562\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 110081,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17840\nallocs=367\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 260208,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2246000\nallocs=143\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 203667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2246000\nallocs=143\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 196833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2246000\nallocs=143\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 191729,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2246000\nallocs=143\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 1114396,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=87408\nallocs=3262\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 6038000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=274440\nallocs=8464\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 642801.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=73792\nallocs=1514\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=128)/forward/CPU/2 thread(s)",
+            "value": 200090500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134217792\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=128)/forward/CPU/4 thread(s)",
+            "value": 103769271,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134217792\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=128)/forward/CPU/8 thread(s)",
+            "value": 138991458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134217792\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=128)/forward/CPU/1 thread(s)",
+            "value": 388799084,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134217792\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=128)/forward/GPU/CUDA",
+            "value": 5837129,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=528\nallocs=24\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=128)/forward/GPU/AMDGPU",
+            "value": 3417993,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11584\nallocs=348\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=128)/zygote/CPU/2 thread(s)",
+            "value": 619060979.5,
+            "unit": "ns",
+            "extra": "gctime=2591062.5\nmemory=536872016\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=128)/zygote/CPU/4 thread(s)",
+            "value": 352556416.5,
+            "unit": "ns",
+            "extra": "gctime=2853291.5\nmemory=536872016\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=128)/zygote/CPU/8 thread(s)",
+            "value": 439711625,
+            "unit": "ns",
+            "extra": "gctime=2695687.5\nmemory=536872016\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=128)/zygote/CPU/1 thread(s)",
+            "value": 1186861083,
+            "unit": "ns",
+            "extra": "gctime=3098000\nmemory=536871824\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=128)/zygote/GPU/CUDA",
+            "value": 26566336,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11840\nallocs=432\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(512, Bsize=128)/zygote/GPU/AMDGPU",
+            "value": 21956286,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=37312\nallocs=806\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 7458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 5500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 6125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 10417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 27251,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3536\nallocs=119\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 562916,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16032\nallocs=533\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 50400,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6544\nallocs=164\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 213500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=801936\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 222020.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=801936\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 223812.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=801936\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 207000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=801936\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 236231,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=22144\nallocs=649\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 9332834,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=240568\nallocs=7784\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 535116,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=77808\nallocs=1780\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 7937.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3360\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 8542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3360\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 10209,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3360\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 8063,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3360\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(4 x 32)/forward/GPU/CUDA",
+            "value": 131595.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11536\nallocs=274\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(4 x 32)/forward/GPU/Metal",
+            "value": 853687.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=32720\nallocs=1044\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 76050,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15552\nallocs=340\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 7458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18224\nallocs=90\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 8792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18224\nallocs=90\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 11791,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18224\nallocs=90\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 7292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18224\nallocs=90\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(4 x 32)/zygote/GPU/CUDA",
+            "value": 567254,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49832\nallocs=1430\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(4 x 32)/zygote/GPU/Metal",
+            "value": 4301958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=148704\nallocs=4570\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=relu, affine=false)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 325113,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=51160\nallocs=1002\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 459,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(32 x 32)/forward/GPU/CUDA",
+            "value": 25819,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3808\nallocs=128\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(32 x 32)/forward/GPU/Metal",
+            "value": 434312.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15928\nallocs=533\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 49671,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6816\nallocs=184\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 10312.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=40912\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 10083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=40912\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 10750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=40912\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 9750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=40912\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(32 x 32)/zygote/GPU/CUDA",
+            "value": 265547,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=22304\nallocs=678\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(32 x 32)/zygote/GPU/Metal",
+            "value": 5398313,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=174672\nallocs=5728\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 396234,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=55536\nallocs=1354\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=gelu)(512 x 128)/forward/CPU/2 thread(s)",
+            "value": 109687.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=gelu)(512 x 128)/forward/CPU/4 thread(s)",
+            "value": 84812.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=gelu)(512 x 128)/forward/CPU/8 thread(s)",
+            "value": 99979.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=gelu)(512 x 128)/forward/CPU/1 thread(s)",
+            "value": 146708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=gelu)(512 x 128)/forward/GPU/CUDA",
+            "value": 24233,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2384\nallocs=107\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=gelu)(512 x 128)/forward/GPU/Metal",
+            "value": 264270.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7184\nallocs=234\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=gelu)(512 x 128)/forward/GPU/AMDGPU",
+            "value": 190652,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3808\nallocs=190\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=gelu)(512 x 128)/zygote/CPU/2 thread(s)",
+            "value": 514729,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051456\nallocs=18\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=gelu)(512 x 128)/zygote/CPU/4 thread(s)",
+            "value": 478375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051456\nallocs=18\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=gelu)(512 x 128)/zygote/CPU/8 thread(s)",
+            "value": 478291.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051456\nallocs=18\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=gelu)(512 x 128)/zygote/CPU/1 thread(s)",
+            "value": 477791,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051456\nallocs=18\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=gelu)(512 x 128)/zygote/GPU/CUDA",
+            "value": 248436.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19208\nallocs=736\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=gelu)(512 x 128)/zygote/GPU/Metal",
+            "value": 2180959,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=70856\nallocs=2355\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=gelu)(512 x 128)/zygote/GPU/AMDGPU",
+            "value": 625596.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=20008\nallocs=558\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=32)/forward/CPU/2 thread(s)",
+            "value": 5312,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=32832\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=32)/forward/CPU/4 thread(s)",
+            "value": 6666.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=32832\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=32)/forward/CPU/8 thread(s)",
+            "value": 7125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=32832\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=32)/forward/CPU/1 thread(s)",
+            "value": 4250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=32832\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=32)/forward/GPU/CUDA",
+            "value": 15903,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=512\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=32)/forward/GPU/AMDGPU",
+            "value": 73410,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2832\nallocs=84\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=32)/zygote/CPU/2 thread(s)",
+            "value": 11604.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132176\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=32)/zygote/CPU/4 thread(s)",
+            "value": 10333,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132176\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=32)/zygote/CPU/8 thread(s)",
+            "value": 10833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132176\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=32)/zygote/CPU/1 thread(s)",
+            "value": 16375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=131984\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=32)/zygote/GPU/CUDA",
+            "value": 226653,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11808\nallocs=430\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=32)/zygote/GPU/AMDGPU",
+            "value": 376134,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17328\nallocs=402\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=128)/forward/CPU/2 thread(s)",
+            "value": 39791.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=131136\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=128)/forward/CPU/4 thread(s)",
+            "value": 51167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=131136\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=128)/forward/CPU/8 thread(s)",
+            "value": 52125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=131136\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=128)/forward/CPU/1 thread(s)",
+            "value": 13416,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=131136\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=128)/forward/GPU/CUDA",
+            "value": 21786,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=512\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=128)/forward/GPU/AMDGPU",
+            "value": 85321,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5104\nallocs=85\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=128)/zygote/CPU/2 thread(s)",
+            "value": 36271,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=525392\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=128)/zygote/CPU/4 thread(s)",
+            "value": 27895.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=525392\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=128)/zygote/CPU/8 thread(s)",
+            "value": 31083.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=525392\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=128)/zygote/CPU/1 thread(s)",
+            "value": 57167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=525200\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=128)/zygote/GPU/CUDA",
+            "value": 202344,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11808\nallocs=430\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=128)/zygote/GPU/AMDGPU",
+            "value": 421284,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=24016\nallocs=397\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=tanh)(2 x 128)/forward/CPU/2 thread(s)",
+            "value": 1958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=tanh)(2 x 128)/forward/CPU/4 thread(s)",
+            "value": 1917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=tanh)(2 x 128)/forward/CPU/8 thread(s)",
+            "value": 2125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=tanh)(2 x 128)/forward/CPU/1 thread(s)",
+            "value": 1792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=tanh)(2 x 128)/forward/GPU/CUDA",
+            "value": 20196,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2384\nallocs=107\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=tanh)(2 x 128)/forward/GPU/Metal",
+            "value": 299167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6896\nallocs=232\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=tanh)(2 x 128)/forward/GPU/AMDGPU",
+            "value": 28950,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1776\nallocs=63\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=tanh)(2 x 128)/zygote/CPU/2 thread(s)",
+            "value": 2250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3712\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=tanh)(2 x 128)/zygote/CPU/4 thread(s)",
+            "value": 2292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3712\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=tanh)(2 x 128)/zygote/CPU/8 thread(s)",
+            "value": 2417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3712\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=tanh)(2 x 128)/zygote/CPU/1 thread(s)",
+            "value": 2042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3712\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=tanh)(2 x 128)/zygote/GPU/CUDA",
+            "value": 217371,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=14616\nallocs=562\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=tanh)(2 x 128)/zygote/GPU/Metal",
+            "value": 1508312.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=54968\nallocs=1860\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=tanh)(2 x 128)/zygote/GPU/AMDGPU",
+            "value": 138781,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13512\nallocs=323\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 5542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2400\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 4771,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2400\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 6458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2400\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 3979.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2400\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(4 x 32)/forward/GPU/CUDA",
+            "value": 157473.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13104\nallocs=537\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(4 x 32)/forward/GPU/Metal",
+            "value": 631417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=37632\nallocs=1263\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 61190,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10912\nallocs=271\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 8167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15936\nallocs=86\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 8459,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15936\nallocs=86\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 8521,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15936\nallocs=86\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 8458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15936\nallocs=86\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(4 x 32)/zygote/GPU/CUDA",
+            "value": 937391.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=72896\nallocs=2910\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(4 x 32)/zygote/GPU/Metal",
+            "value": 5777041.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=231432\nallocs=7696\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=true)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 389174,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=54352\nallocs=1173\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 57250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 57000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 57833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 58417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 37214,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4448\nallocs=152\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 547062.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13760\nallocs=447\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 208472,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12816\nallocs=428\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 459083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1063248\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 465000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1063248\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 465625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1063248\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 433750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1063248\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 272599,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=23080\nallocs=861\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 8129896,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=211704\nallocs=6760\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 806418,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=77896\nallocs=1676\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=128)/forward/CPU/2 thread(s)",
+            "value": 3328417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8388672\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=128)/forward/CPU/4 thread(s)",
+            "value": 1771625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8388672\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=128)/forward/CPU/8 thread(s)",
+            "value": 2335146.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8388672\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=128)/forward/CPU/1 thread(s)",
+            "value": 6300271,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8388672\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=128)/forward/GPU/CUDA",
+            "value": 204810,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=512\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=128)/forward/GPU/AMDGPU",
+            "value": 203992,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6704\nallocs=185\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=128)/zygote/CPU/2 thread(s)",
+            "value": 11439854.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=33555536\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=128)/zygote/CPU/4 thread(s)",
+            "value": 6573354.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=33555536\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=128)/zygote/CPU/8 thread(s)",
+            "value": 8339250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=33555536\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=128)/zygote/CPU/1 thread(s)",
+            "value": 21058146,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=33555344\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=128)/zygote/GPU/CUDA",
+            "value": 743384,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11808\nallocs=430\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=128)/zygote/GPU/AMDGPU",
+            "value": 1068820.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=28944\nallocs=572\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 5167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1632\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 5042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1632\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 6167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1632\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 4500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1632\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(4 x 32)/forward/GPU/CUDA",
+            "value": 151540.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11984\nallocs=467\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(4 x 32)/forward/GPU/Metal",
+            "value": 722604.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=36672\nallocs=1235\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 56651,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10816\nallocs=265\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 7083.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12128\nallocs=85\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 7500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12128\nallocs=85\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 7334,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12128\nallocs=85\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 7125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12128\nallocs=85\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(4 x 32)/zygote/GPU/CUDA",
+            "value": 791278.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=67568\nallocs=2685\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(4 x 32)/zygote/GPU/Metal",
+            "value": 5188146,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=216304\nallocs=7156\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 384994,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49584\nallocs=1088\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 97604,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1052048\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 95667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1052048\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 128792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1052048\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 101375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1052048\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 158164,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13392\nallocs=309\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 2055833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=53696\nallocs=1710\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 182992,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=24480\nallocs=504\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 2022875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7359616\nallocs=120\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 2032584,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7359616\nallocs=120\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 2028770.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7359616\nallocs=120\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 2017771,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7359616\nallocs=120\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 765556,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=67176\nallocs=1814\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 11016875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=227400\nallocs=6857\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1122166,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=92760\nallocs=1890\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=4)/forward/CPU/2 thread(s)",
+            "value": 34604.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=128\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=4)/forward/CPU/4 thread(s)",
+            "value": 34646,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=128\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=4)/forward/CPU/8 thread(s)",
+            "value": 34625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=128\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=4)/forward/CPU/1 thread(s)",
+            "value": 625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=128\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=4)/forward/GPU/CUDA",
+            "value": 15150,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=512\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=4)/forward/GPU/AMDGPU",
+            "value": 81540,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2368\nallocs=100\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=4)/zygote/CPU/2 thread(s)",
+            "value": 2562.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1360\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=4)/zygote/CPU/4 thread(s)",
+            "value": 3042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1360\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=4)/zygote/CPU/8 thread(s)",
+            "value": 3000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1360\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=4)/zygote/CPU/1 thread(s)",
+            "value": 2084,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1168\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=4)/zygote/GPU/CUDA",
+            "value": 143347.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=9456\nallocs=360\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=4)/zygote/GPU/AMDGPU",
+            "value": 346678.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12032\nallocs=335\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 7166,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 5416,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 6083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 10000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 36135,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4448\nallocs=152\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 551291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13760\nallocs=447\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 49200,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5936\nallocs=136\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 223083.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=800976\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 221187.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=800976\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 221792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=800976\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 242208,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=800976\nallocs=111\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 248462,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19648\nallocs=695\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 8070728.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=202400\nallocs=6495\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=identity, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 522185,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=74352\nallocs=1508\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=identity)(32 x 128)/forward/CPU/2 thread(s)",
+            "value": 3917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=identity)(32 x 128)/forward/CPU/4 thread(s)",
+            "value": 4000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=identity)(32 x 128)/forward/CPU/8 thread(s)",
+            "value": 3917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=identity)(32 x 128)/forward/CPU/1 thread(s)",
+            "value": 3917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=identity)(32 x 128)/forward/GPU/CUDA",
+            "value": 21588,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1056\nallocs=47\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=identity)(32 x 128)/forward/GPU/Metal",
+            "value": 246125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7352\nallocs=263\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=identity)(32 x 128)/forward/GPU/AMDGPU",
+            "value": 46230,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1920\nallocs=75\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=identity)(32 x 128)/zygote/CPU/2 thread(s)",
+            "value": 14917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=54640\nallocs=9\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=identity)(32 x 128)/zygote/CPU/4 thread(s)",
+            "value": 14916,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=54640\nallocs=9\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=identity)(32 x 128)/zygote/CPU/8 thread(s)",
+            "value": 14917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=54640\nallocs=9\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=identity)(32 x 128)/zygote/CPU/1 thread(s)",
+            "value": 14917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=54640\nallocs=9\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=identity)(32 x 128)/zygote/GPU/CUDA",
+            "value": 332170,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15424\nallocs=541\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=identity)(32 x 128)/zygote/GPU/Metal",
+            "value": 1019208,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=54272\nallocs=1837\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=identity)(32 x 128)/zygote/GPU/AMDGPU",
+            "value": 196351.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16048\nallocs=383\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 120916.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1052048\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 127979,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1052048\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 128958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1052048\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 117208,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1052048\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 151353,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13392\nallocs=309\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1994500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=53696\nallocs=1710\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 169312,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=24784\nallocs=523\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 1914625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6310928\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 1908250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6310928\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 1923666,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6310928\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 1920583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6310928\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 750665,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=65584\nallocs=1732\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 10544021,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=219904\nallocs=6637\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1234013,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=96032\nallocs=1998\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 18833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134288\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 18292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134288\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 20291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134288\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 18125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134288\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 121910,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12544\nallocs=282\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 1357604,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=34288\nallocs=1053\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 78121,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16144\nallocs=302\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 229792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=933632\nallocs=108\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 226249.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=933632\nallocs=108\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 217104.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=933632\nallocs=108\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 216500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=933632\nallocs=108\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 554792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=54424\nallocs=1515\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 6233583.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=167336\nallocs=4921\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 487889,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=57784\nallocs=1133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=4)/forward/CPU/2 thread(s)",
+            "value": 26458.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4352\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=4)/forward/CPU/4 thread(s)",
+            "value": 29375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4352\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=4)/forward/CPU/8 thread(s)",
+            "value": 28895.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4352\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=4)/forward/CPU/1 thread(s)",
+            "value": 1209,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4352\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=4)/forward/GPU/CUDA",
+            "value": 16098,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=512\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=4)/forward/GPU/AMDGPU",
+            "value": 83481,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2368\nallocs=100\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=4)/zygote/CPU/2 thread(s)",
+            "value": 4541.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18256\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=4)/zygote/CPU/4 thread(s)",
+            "value": 5083.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18256\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=4)/zygote/CPU/8 thread(s)",
+            "value": 5312.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18256\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=4)/zygote/CPU/1 thread(s)",
+            "value": 4625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18064\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=4)/zygote/GPU/CUDA",
+            "value": 220627,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11808\nallocs=430\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(16, Bsize=4)/zygote/GPU/AMDGPU",
+            "value": 386633.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15360\nallocs=414\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 308417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132624\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 309792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132624\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 307375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132624\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 304375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132624\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 251986,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17904\nallocs=612\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 1131042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=48184\nallocs=1530\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 279722,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=24752\nallocs=661\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 542542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1320960\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 530625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1320960\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 532167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1320960\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 552625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1320960\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 1165833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=85904\nallocs=3212\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 6317687.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=268496\nallocs=8303\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 872538,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=70688\nallocs=1474\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 22083.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134544\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 23292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134544\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 22583.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134544\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 19833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134544\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 127624.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13376\nallocs=308\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 1486562.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=37008\nallocs=1149\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 76640,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16864\nallocs=322\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 226271,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=805680\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 212542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=805680\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 214229.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=805680\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 212500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=805680\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 857148,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=65552\nallocs=1730\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 7438083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=202848\nallocs=6053\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=identity, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 550605,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=73344\nallocs=1429\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 7417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5408\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 6583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5408\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 8000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5408\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 5958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5408\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(32 x 32)/forward/GPU/CUDA",
+            "value": 153413,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11984\nallocs=467\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(32 x 32)/forward/GPU/Metal",
+            "value": 792104,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=36872\nallocs=1232\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 69611,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10864\nallocs=268\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 10209,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=42336\nallocs=85\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 10729,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=42336\nallocs=85\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 10709,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=42336\nallocs=85\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 10917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=42336\nallocs=85\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(32 x 32)/zygote/GPU/CUDA",
+            "value": 866679,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=69792\nallocs=2754\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(32 x 32)/zygote/GPU/Metal",
+            "value": 5386000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=218432\nallocs=7134\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=identity, affine=false)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 393064,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=52160\nallocs=1135\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 6708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1632\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 5312.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1632\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 6646,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1632\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 5042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1632\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(4 x 32)/forward/GPU/CUDA",
+            "value": 156782.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11984\nallocs=467\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(4 x 32)/forward/GPU/Metal",
+            "value": 767937.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=36672\nallocs=1235\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 59941,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10928\nallocs=272\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 7250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13360\nallocs=87\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 7334,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13360\nallocs=87\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 7604.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13360\nallocs=87\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 7542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13360\nallocs=87\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(4 x 32)/zygote/GPU/CUDA",
+            "value": 815076,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=70624\nallocs=2816\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(4 x 32)/zygote/GPU/Metal",
+            "value": 5485417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=228896\nallocs=7571\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=relu, affine=false)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 400604,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=51264\nallocs=1122\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=512)/forward/CPU/2 thread(s)",
+            "value": 14539750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=33554496\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=512)/forward/CPU/4 thread(s)",
+            "value": 7703625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=33554496\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=512)/forward/CPU/8 thread(s)",
+            "value": 10129458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=33554496\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=512)/forward/CPU/1 thread(s)",
+            "value": 27745125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=33554496\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=512)/forward/GPU/CUDA",
+            "value": 531314,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=528\nallocs=24\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=512)/forward/GPU/AMDGPU",
+            "value": 395439,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=20992\nallocs=348\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=512)/zygote/CPU/2 thread(s)",
+            "value": 46281666,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134218832\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=512)/zygote/CPU/4 thread(s)",
+            "value": 26486646,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134218832\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=512)/zygote/CPU/8 thread(s)",
+            "value": 33555291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134218832\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=512)/zygote/CPU/1 thread(s)",
+            "value": 85896125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134218640\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=512)/zygote/GPU/CUDA",
+            "value": 2671865,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11840\nallocs=432\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=512)/zygote/GPU/AMDGPU",
+            "value": 3267562,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=62064\nallocs=736\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 69667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134544\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 66916.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134544\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 67667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134544\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 67000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134544\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 134478.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13376\nallocs=308\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 1454896,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=37008\nallocs=1149\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 228332,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17904\nallocs=387\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 477062,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1068048\nallocs=123\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 441208,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1068048\nallocs=123\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 441375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1068048\nallocs=123\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 440792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1068048\nallocs=123\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 789089,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=68984\nallocs=1896\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 7688417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=212712\nallocs=6323\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 805468,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=76920\nallocs=1597\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 584,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(32 x 32)/forward/GPU/CUDA",
+            "value": 32040,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4112\nallocs=142\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(32 x 32)/forward/GPU/Metal",
+            "value": 444854,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13656\nallocs=447\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 49641,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6496\nallocs=174\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 9834,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=44800\nallocs=93\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 10687,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=44800\nallocs=93\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 11083.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=44800\nallocs=93\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 9042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=44800\nallocs=93\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(32 x 32)/zygote/GPU/CUDA",
+            "value": 291258,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=21768\nallocs=806\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(32 x 32)/zygote/GPU/Metal",
+            "value": 4894917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=159704\nallocs=5216\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 392374,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49992\nallocs=1211\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=gelu)(32 x 128)/forward/CPU/2 thread(s)",
+            "value": 9916,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=gelu)(32 x 128)/forward/CPU/4 thread(s)",
+            "value": 9875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=gelu)(32 x 128)/forward/CPU/8 thread(s)",
+            "value": 9791,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=gelu)(32 x 128)/forward/CPU/1 thread(s)",
+            "value": 9792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=gelu)(32 x 128)/forward/GPU/CUDA",
+            "value": 22712,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=960\nallocs=41\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=gelu)(32 x 128)/forward/GPU/Metal",
+            "value": 227500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7384\nallocs=263\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=gelu)(32 x 128)/forward/GPU/AMDGPU",
+            "value": 217302,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8240\nallocs=328\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=gelu)(32 x 128)/zygote/CPU/2 thread(s)",
+            "value": 45958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=88640\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=gelu)(32 x 128)/zygote/CPU/4 thread(s)",
+            "value": 46083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=88640\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=gelu)(32 x 128)/zygote/CPU/8 thread(s)",
+            "value": 46417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=88640\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=gelu)(32 x 128)/zygote/CPU/1 thread(s)",
+            "value": 46042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=88640\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=gelu)(32 x 128)/zygote/GPU/CUDA",
+            "value": 299669,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18088\nallocs=683\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=gelu)(32 x 128)/zygote/GPU/Metal",
+            "value": 956875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=50896\nallocs=1672\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=gelu)(32 x 128)/zygote/GPU/AMDGPU",
+            "value": 627331.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15992\nallocs=472\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 56459,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 56459,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 57042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 57791,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 29020,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3536\nallocs=119\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 571042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16032\nallocs=533\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 206262,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13008\nallocs=430\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 499291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1064208\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 466208,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1064208\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 466374.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1064208\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 483250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1064208\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 250980,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=21224\nallocs=795\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 9451750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=249920\nallocs=8052\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 850639,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=80120\nallocs=1871\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 659313,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3146992\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 657750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3146992\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 664375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3146992\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 613687.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3146992\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 221566,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19024\nallocs=682\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1372416,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=58128\nallocs=1840\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 238043,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18976\nallocs=438\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 2210125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17900528\nallocs=145\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 2247917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17900528\nallocs=145\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 2238417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17900528\nallocs=145\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 2232375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17900528\nallocs=145\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 1036904,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=87408\nallocs=3262\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 8492166,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=298448\nallocs=9286\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=identity, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1379048.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=78144\nallocs=1626\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 20604.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134544\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 20500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134544\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 21375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134544\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 21291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=134544\nallocs=43\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 125525.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13376\nallocs=308\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 1450958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=37008\nallocs=1149\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 76440.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16864\nallocs=322\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 233020.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=936864\nallocs=120\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 255209,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=936864\nallocs=120\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 220583.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=936864\nallocs=120\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 224854.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=936864\nallocs=120\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 823179,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=67144\nallocs=1812\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 7635209,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=210344\nallocs=6273\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=true)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 561210.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=74232\nallocs=1434\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(32 x 32)/forward/GPU/CUDA",
+            "value": 22882,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3072\nallocs=104\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(32 x 32)/forward/GPU/Metal",
+            "value": 439812,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15928\nallocs=533\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 50421,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6880\nallocs=188\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 10604,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=45264\nallocs=95\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 11416,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=45264\nallocs=95\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 10791.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=45264\nallocs=95\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 10375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=45264\nallocs=95\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(32 x 32)/zygote/GPU/CUDA",
+            "value": 270980,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19960\nallocs=740\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(32 x 32)/zygote/GPU/Metal",
+            "value": 5356520.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=181672\nallocs=5941\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 411824,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=56360\nallocs=1361\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 10792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3360\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 10250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3360\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 10167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3360\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 8750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3360\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(4 x 32)/forward/GPU/CUDA",
+            "value": 134053,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11536\nallocs=274\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(4 x 32)/forward/GPU/Metal",
+            "value": 834416,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=32720\nallocs=1044\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 67971,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15040\nallocs=308\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 7542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17552\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 7792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17552\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 8042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17552\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 7500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17552\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(4 x 32)/zygote/GPU/CUDA",
+            "value": 543489,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=48240\nallocs=1348\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(4 x 32)/zygote/GPU/Metal",
+            "value": 4166125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=141496\nallocs=4352\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 329703,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=50112\nallocs=982\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=gelu)(2 x 128)/forward/CPU/2 thread(s)",
+            "value": 1792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=gelu)(2 x 128)/forward/CPU/4 thread(s)",
+            "value": 1625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=gelu)(2 x 128)/forward/CPU/8 thread(s)",
+            "value": 1834,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=gelu)(2 x 128)/forward/CPU/1 thread(s)",
+            "value": 1542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1280\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=gelu)(2 x 128)/forward/GPU/CUDA",
+            "value": 20840,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2384\nallocs=107\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=gelu)(2 x 128)/forward/GPU/Metal",
+            "value": 300875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6896\nallocs=232\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=gelu)(2 x 128)/forward/GPU/AMDGPU",
+            "value": 192452,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4912\nallocs=259\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=gelu)(2 x 128)/zygote/CPU/2 thread(s)",
+            "value": 3375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4800\nallocs=14\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=gelu)(2 x 128)/zygote/CPU/4 thread(s)",
+            "value": 3500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4800\nallocs=14\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=gelu)(2 x 128)/zygote/CPU/8 thread(s)",
+            "value": 3583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4800\nallocs=14\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=gelu)(2 x 128)/zygote/CPU/1 thread(s)",
+            "value": 3166,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4800\nallocs=14\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=gelu)(2 x 128)/zygote/GPU/CUDA",
+            "value": 231108,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16984\nallocs=667\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=gelu)(2 x 128)/zygote/GPU/Metal",
+            "value": 1624666.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=60832\nallocs=2056\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(2, act=gelu)(2 x 128)/zygote/GPU/AMDGPU",
+            "value": 597176,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16648\nallocs=470\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=tanh)(512 x 128)/forward/CPU/2 thread(s)",
+            "value": 151791.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=tanh)(512 x 128)/forward/CPU/4 thread(s)",
+            "value": 106625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=tanh)(512 x 128)/forward/CPU/8 thread(s)",
+            "value": 129417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=tanh)(512 x 128)/forward/CPU/1 thread(s)",
+            "value": 225312,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=tanh)(512 x 128)/forward/GPU/CUDA",
+            "value": 23800,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2384\nallocs=107\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=tanh)(512 x 128)/forward/GPU/Metal",
+            "value": 298375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7184\nallocs=234\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=tanh)(512 x 128)/forward/GPU/AMDGPU",
+            "value": 36681,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1232\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=tanh)(512 x 128)/zygote/CPU/2 thread(s)",
+            "value": 176833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=789264\nallocs=16\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=tanh)(512 x 128)/zygote/CPU/4 thread(s)",
+            "value": 87833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=789264\nallocs=16\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=tanh)(512 x 128)/zygote/CPU/8 thread(s)",
+            "value": 110583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=789264\nallocs=16\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=tanh)(512 x 128)/zygote/CPU/1 thread(s)",
+            "value": 251000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=789264\nallocs=16\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=tanh)(512 x 128)/zygote/GPU/CUDA",
+            "value": 234463.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16840\nallocs=631\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=tanh)(512 x 128)/zygote/GPU/Metal",
+            "value": 1991770.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=64072\nallocs=2140\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=tanh)(512 x 128)/zygote/GPU/AMDGPU",
+            "value": 221652,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15976\nallocs=362\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 7291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 5416,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 5959,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 10167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132096\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 32586,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3712\nallocs=128\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 559708.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13760\nallocs=447\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 50471,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6016\nallocs=141\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 223250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=932112\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 265500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=932112\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 229334,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=932112\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 213562.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=932112\nallocs=113\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 264008.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=21240\nallocs=777\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 8366500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=209488\nallocs=6713\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 536235,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=75080\nallocs=1509\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 15417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5408\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 14917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5408\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 16667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5408\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 15750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5408\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(32 x 32)/forward/GPU/CUDA",
+            "value": 153470.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11984\nallocs=467\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(32 x 32)/forward/GPU/Metal",
+            "value": 758125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=36872\nallocs=1232\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 239772,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17360\nallocs=532\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 23958.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=52896\nallocs=102\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 24417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=52896\nallocs=102\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 23291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=52896\nallocs=102\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 22833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=52896\nallocs=102\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(32 x 32)/zygote/GPU/CUDA",
+            "value": 908275,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=76048\nallocs=2955\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(32 x 32)/zygote/GPU/Metal",
+            "value": 5574500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=233472\nallocs=7569\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=false)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 696222.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=56624\nallocs=1212\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 9625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7344\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 9792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7344\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 10875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7344\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 9625.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7344\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(32 x 32)/forward/GPU/CUDA",
+            "value": 138347.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12496\nallocs=308\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(32 x 32)/forward/GPU/Metal",
+            "value": 828291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=35736\nallocs=1143\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 71601,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16064\nallocs=347\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 14333,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=43568\nallocs=100\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 14625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=43568\nallocs=100\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 14354.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=43568\nallocs=100\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 13875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=43568\nallocs=100\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(32 x 32)/zygote/GPU/CUDA",
+            "value": 750912,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=58672\nallocs=1592\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(32 x 32)/zygote/GPU/Metal",
+            "value": 5349500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=170152\nallocs=5129\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 372023.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=62704\nallocs=1198\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 10458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7088\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 9917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7088\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 11291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7088\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 9958.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7088\nallocs=32\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(32 x 32)/forward/GPU/CUDA",
+            "value": 136667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11664\nallocs=282\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(32 x 32)/forward/GPU/Metal",
+            "value": 897645.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=33016\nallocs=1047\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 70471,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15360\nallocs=328\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 12708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=40112\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 12958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=40112\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 12834,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=40112\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 12249.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=40112\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(32 x 32)/zygote/GPU/CUDA",
+            "value": 610096,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=50720\nallocs=1433\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(32 x 32)/zygote/GPU/Metal",
+            "value": 4478395.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=143648\nallocs=4347\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=false)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 343163,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=53136\nallocs=1056\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=128)/forward/CPU/2 thread(s)",
+            "value": 31500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2304\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=128)/forward/CPU/4 thread(s)",
+            "value": 32334,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2304\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=128)/forward/CPU/8 thread(s)",
+            "value": 31063,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2304\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=128)/forward/CPU/1 thread(s)",
+            "value": 1667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2304\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=128)/forward/GPU/CUDA",
+            "value": 16053,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=512\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=128)/forward/GPU/AMDGPU",
+            "value": 74040,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5120\nallocs=86\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=128)/zygote/CPU/2 thread(s)",
+            "value": 5104,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10064\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=128)/zygote/CPU/4 thread(s)",
+            "value": 5000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10064\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=128)/zygote/CPU/8 thread(s)",
+            "value": 5291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10064\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=128)/zygote/CPU/1 thread(s)",
+            "value": 6521,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=9872\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=128)/zygote/GPU/CUDA",
+            "value": 143200,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=9456\nallocs=360\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=128)/zygote/GPU/AMDGPU",
+            "value": 371674,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=24192\nallocs=408\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(4 x 32)/forward/GPU/CUDA",
+            "value": 25340,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3808\nallocs=128\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(4 x 32)/forward/GPU/Metal",
+            "value": 455083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15616\nallocs=529\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 48400,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6736\nallocs=179\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 7000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17024\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 7500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17024\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 7792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17024\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 7062.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17024\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(4 x 32)/zygote/GPU/CUDA",
+            "value": 193241,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=20080\nallocs=609\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(4 x 32)/zygote/GPU/Metal",
+            "value": 5200291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=172320\nallocs=5736\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=identity, affine=true)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 392484,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=52480\nallocs=1278\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(32 x 32)/forward/CPU/2 thread(s)",
+            "value": 2042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(32 x 32)/forward/CPU/4 thread(s)",
+            "value": 2042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(32 x 32)/forward/CPU/8 thread(s)",
+            "value": 2125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(32 x 32)/forward/CPU/1 thread(s)",
+            "value": 1959,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=5440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(32 x 32)/forward/GPU/CUDA",
+            "value": 26289,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3808\nallocs=128\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(32 x 32)/forward/GPU/Metal",
+            "value": 430104.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15928\nallocs=533\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(32 x 32)/forward/GPU/AMDGPU",
+            "value": 209682,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12960\nallocs=430\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 17604,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49616\nallocs=96\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 17791.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49616\nallocs=96\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 17646,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49616\nallocs=96\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 17396,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=49616\nallocs=96\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(32 x 32)/zygote/GPU/CUDA",
+            "value": 277617,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=21784\nallocs=824\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(32 x 32)/zygote/GPU/Metal",
+            "value": 5801541,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=183896\nallocs=5989\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(32 x 32)/zygote/GPU/AMDGPU",
+            "value": 717947,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=57880\nallocs=1441\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 177042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1050128\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 174792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1050128\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 154291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1050128\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 148125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1050128\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 216713,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17904\nallocs=612\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1455354,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=56336\nallocs=1808\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 198202,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18928\nallocs=435\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 1277542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10496000\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 1327937,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10496000\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 1325666.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10496000\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 1327375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10496000\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 968973,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=84848\nallocs=3168\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 6943375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=293688\nallocs=9154\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1113911,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=74880\nallocs=1607\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/2 thread(s)",
+            "value": 24542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132624\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/4 thread(s)",
+            "value": 26041,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132624\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/8 thread(s)",
+            "value": 25666,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132624\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/CPU/1 thread(s)",
+            "value": 24771,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=132624\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/CUDA",
+            "value": 260001,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17904\nallocs=612\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/Metal",
+            "value": 1053562.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=48184\nallocs=1530\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/forward/GPU/AMDGPU",
+            "value": 119771,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18608\nallocs=415\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 128854.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1320960\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 173624.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1320960\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 118334,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1320960\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 117125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1320960\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/CUDA",
+            "value": 1179575,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=84848\nallocs=3168\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/Metal",
+            "value": 6306396,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=268592\nallocs=8303\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=false)(16 x 16 x 4 x 32)/zygote/GPU/AMDGPU",
+            "value": 607237,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=70224\nallocs=1446\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 334,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 334,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(4 x 32)/forward/GPU/CUDA",
+            "value": 22595,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3072\nallocs=104\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(4 x 32)/forward/GPU/Metal",
+            "value": 442041.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15616\nallocs=529\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 49271,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6784\nallocs=182\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 7458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17600\nallocs=95\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 7667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17600\nallocs=95\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 7834,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17600\nallocs=95\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 7187.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17600\nallocs=95\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(4 x 32)/zygote/GPU/CUDA",
+            "value": 198802.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17736\nallocs=671\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(4 x 32)/zygote/GPU/Metal",
+            "value": 5362999.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=179120\nallocs=5952\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=true)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 396804,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=53640\nallocs=1306\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 7125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2400\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 6500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2400\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 7292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2400\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 5041.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2400\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(4 x 32)/forward/GPU/CUDA",
+            "value": 164192.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13104\nallocs=537\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(4 x 32)/forward/GPU/Metal",
+            "value": 455646,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=37632\nallocs=1263\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 237732,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17360\nallocs=532\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 9875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17152\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 10417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17152\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 10208,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17152\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 10042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17152\nallocs=88\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(4 x 32)/zygote/GPU/CUDA",
+            "value": 959197,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=77024\nallocs=3085\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(4 x 32)/zygote/GPU/Metal",
+            "value": 6029979,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=243912\nallocs=8109\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(2, act=gelu, affine=true)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 680937,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=56608\nallocs=1219\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=gelu)(2 x 128)/forward/CPU/2 thread(s)",
+            "value": 667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=gelu)(2 x 128)/forward/CPU/4 thread(s)",
+            "value": 666,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=gelu)(2 x 128)/forward/CPU/8 thread(s)",
+            "value": 625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=gelu)(2 x 128)/forward/CPU/1 thread(s)",
+            "value": 625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1088\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=gelu)(2 x 128)/forward/GPU/CUDA",
+            "value": 22283,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=960\nallocs=41\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=gelu)(2 x 128)/forward/GPU/Metal",
+            "value": 328083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7104\nallocs=261\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=gelu)(2 x 128)/forward/GPU/AMDGPU",
+            "value": 215392,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4960\nallocs=265\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=gelu)(2 x 128)/zygote/CPU/2 thread(s)",
+            "value": 4583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6608\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=gelu)(2 x 128)/zygote/CPU/4 thread(s)",
+            "value": 4583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6608\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=gelu)(2 x 128)/zygote/CPU/8 thread(s)",
+            "value": 4792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6608\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=gelu)(2 x 128)/zygote/CPU/1 thread(s)",
+            "value": 4583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6608\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=gelu)(2 x 128)/zygote/GPU/CUDA",
+            "value": 228018.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15864\nallocs=614\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=gelu)(2 x 128)/zygote/GPU/Metal",
+            "value": 1580083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=48672\nallocs=1657\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(2, bias=false, act=gelu)(2 x 128)/zygote/GPU/AMDGPU",
+            "value": 600937,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11288\nallocs=293\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 8729.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3616\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 8416,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3616\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 9958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3616\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 8041.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3616\nallocs=36\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(4 x 32)/forward/GPU/CUDA",
+            "value": 136239,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12368\nallocs=300\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(4 x 32)/forward/GPU/Metal",
+            "value": 802291.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=35440\nallocs=1140\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 70015.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15968\nallocs=341\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 8750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=20784\nallocs=100\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 8625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=20784\nallocs=100\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 8959,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=20784\nallocs=100\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 8229.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=20784\nallocs=100\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(4 x 32)/zygote/GPU/CUDA",
+            "value": 658040,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=56192\nallocs=1507\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(4 x 32)/zygote/GPU/Metal",
+            "value": 4729875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=167800\nallocs=5137\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(2, act=identity, affine=true)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 352293.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=59616\nallocs=1120\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=4)/forward/CPU/2 thread(s)",
+            "value": 126416,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262208\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=4)/forward/CPU/4 thread(s)",
+            "value": 95833,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262208\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=4)/forward/CPU/8 thread(s)",
+            "value": 128834,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262208\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=4)/forward/CPU/1 thread(s)",
+            "value": 182979,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262208\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=4)/forward/GPU/CUDA",
+            "value": 45872,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=512\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=4)/forward/GPU/AMDGPU",
+            "value": 94981,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2112\nallocs=84\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=4)/zygote/CPU/2 thread(s)",
+            "value": 324875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049680\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=4)/zygote/CPU/4 thread(s)",
+            "value": 187937.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049680\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=4)/zygote/CPU/8 thread(s)",
+            "value": 313667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049680\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=4)/zygote/CPU/1 thread(s)",
+            "value": 567334,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049488\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=4)/zygote/GPU/CUDA",
+            "value": 202221.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11808\nallocs=430\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=4)/zygote/GPU/AMDGPU",
+            "value": 486324.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15088\nallocs=397\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=identity)(512 x 128)/forward/CPU/2 thread(s)",
+            "value": 397167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=identity)(512 x 128)/forward/CPU/4 thread(s)",
+            "value": 215333,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=identity)(512 x 128)/forward/CPU/8 thread(s)",
+            "value": 288334,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=identity)(512 x 128)/forward/CPU/1 thread(s)",
+            "value": 756250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=identity)(512 x 128)/forward/GPU/CUDA",
+            "value": 43363,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1056\nallocs=47\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=identity)(512 x 128)/forward/GPU/Metal",
+            "value": 434250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7392\nallocs=265\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=identity)(512 x 128)/forward/GPU/AMDGPU",
+            "value": 83831,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2720\nallocs=125\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=identity)(512 x 128)/zygote/CPU/2 thread(s)",
+            "value": 1452083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1837680\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=identity)(512 x 128)/zygote/CPU/4 thread(s)",
+            "value": 860875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1837680\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=identity)(512 x 128)/zygote/CPU/8 thread(s)",
+            "value": 1133979.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1837680\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=identity)(512 x 128)/zygote/CPU/1 thread(s)",
+            "value": 2441959,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1837680\nallocs=13\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=identity)(512 x 128)/zygote/GPU/CUDA",
+            "value": 257428,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15744\nallocs=561\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=identity)(512 x 128)/zygote/GPU/Metal",
+            "value": 1874458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=62528\nallocs=2119\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=identity)(512 x 128)/zygote/GPU/AMDGPU",
+            "value": 354814,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16816\nallocs=431\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 653292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3146992\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 672583.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3146992\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 659166,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3146992\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 630917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3146992\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 215216.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19024\nallocs=682\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1367729,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=58128\nallocs=1840\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 250902.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18432\nallocs=404\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 2450458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19997968\nallocs=151\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 2465792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19997968\nallocs=151\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 2418771,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19997968\nallocs=151\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 2409583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19997968\nallocs=151\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 1051964,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=90592\nallocs=3393\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 12445958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=313568\nallocs=9730\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=relu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1482025,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=80224\nallocs=1685\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=32)/forward/CPU/2 thread(s)",
+            "value": 33917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=576\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=32)/forward/CPU/4 thread(s)",
+            "value": 34417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=576\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=32)/forward/CPU/8 thread(s)",
+            "value": 34209,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=576\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=32)/forward/CPU/1 thread(s)",
+            "value": 750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=576\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=32)/forward/GPU/CUDA",
+            "value": 15426,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=512\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=32)/forward/GPU/AMDGPU",
+            "value": 73461,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2768\nallocs=80\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=32)/zygote/CPU/2 thread(s)",
+            "value": 3125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3152\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=32)/zygote/CPU/4 thread(s)",
+            "value": 3333,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3152\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=32)/zygote/CPU/8 thread(s)",
+            "value": 3354.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3152\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=32)/zygote/CPU/1 thread(s)",
+            "value": 3125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2960\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=32)/zygote/GPU/CUDA",
+            "value": 142341,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=9456\nallocs=360\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=32)/zygote/GPU/AMDGPU",
+            "value": 346283,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=14224\nallocs=337\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 407041,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 402250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 408917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 421250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 43456,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3536\nallocs=119\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1157209,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16032\nallocs=533\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 240343,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13088\nallocs=431\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 3870917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8405472\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 3997208.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8405472\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 3992125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8405472\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 3774916.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8405472\nallocs=117\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 245622,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=21224\nallocs=795\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 11639916.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=249920\nallocs=8052\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=true)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1244322,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=85128\nallocs=2037\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=identity)(32 x 128)/forward/CPU/2 thread(s)",
+            "value": 3917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=identity)(32 x 128)/forward/CPU/4 thread(s)",
+            "value": 3959,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=identity)(32 x 128)/forward/CPU/8 thread(s)",
+            "value": 3916,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=identity)(32 x 128)/forward/CPU/1 thread(s)",
+            "value": 3875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16640\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=identity)(32 x 128)/forward/GPU/CUDA",
+            "value": 33107,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1008\nallocs=38\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=identity)(32 x 128)/forward/GPU/Metal",
+            "value": 264209,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2696\nallocs=94\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=identity)(32 x 128)/forward/GPU/AMDGPU",
+            "value": 41200,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1072\nallocs=58\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=identity)(32 x 128)/zygote/CPU/2 thread(s)",
+            "value": 15708,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=55216\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=identity)(32 x 128)/zygote/CPU/4 thread(s)",
+            "value": 15792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=55216\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=identity)(32 x 128)/zygote/CPU/8 thread(s)",
+            "value": 15958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=55216\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=identity)(32 x 128)/zygote/CPU/1 thread(s)",
+            "value": 15667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=55216\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=identity)(32 x 128)/zygote/GPU/CUDA",
+            "value": 266369,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13328\nallocs=472\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=identity)(32 x 128)/zygote/GPU/Metal",
+            "value": 881458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=37864\nallocs=1258\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=false, act=identity)(32 x 128)/zygote/GPU/AMDGPU",
+            "value": 166502,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11776\nallocs=307\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=relu)(512 x 128)/forward/CPU/2 thread(s)",
+            "value": 404166,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=relu)(512 x 128)/forward/CPU/4 thread(s)",
+            "value": 221375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=relu)(512 x 128)/forward/CPU/8 thread(s)",
+            "value": 295292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=relu)(512 x 128)/forward/CPU/1 thread(s)",
+            "value": 760167,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262192\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=relu)(512 x 128)/forward/GPU/CUDA",
+            "value": 113190,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=960\nallocs=41\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=relu)(512 x 128)/forward/GPU/Metal",
+            "value": 496062.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7424\nallocs=265\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=relu)(512 x 128)/forward/GPU/AMDGPU",
+            "value": 90120,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3024\nallocs=144\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=relu)(512 x 128)/zygote/CPU/2 thread(s)",
+            "value": 1467208,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2098400\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=relu)(512 x 128)/zygote/CPU/4 thread(s)",
+            "value": 886541,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2098400\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=relu)(512 x 128)/zygote/CPU/8 thread(s)",
+            "value": 1159125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2098400\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=relu)(512 x 128)/zygote/CPU/1 thread(s)",
+            "value": 2466583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2098400\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=relu)(512 x 128)/zygote/GPU/CUDA",
+            "value": 240367,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15992\nallocs=601\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=relu)(512 x 128)/zygote/GPU/Metal",
+            "value": 1925500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=57384\nallocs=1920\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=false, act=relu)(512 x 128)/zygote/GPU/AMDGPU",
+            "value": 356978.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=14984\nallocs=422\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 583,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(4 x 32)/forward/GPU/CUDA",
+            "value": 25613,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3808\nallocs=128\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(4 x 32)/forward/GPU/Metal",
+            "value": 450208.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=15616\nallocs=529\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 209312,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12976\nallocs=427\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 8333.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18176\nallocs=96\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 8500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18176\nallocs=96\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 9250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18176\nallocs=96\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 8208,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18176\nallocs=96\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(4 x 32)/zygote/GPU/CUDA",
+            "value": 210905.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19560\nallocs=755\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(4 x 32)/zygote/GPU/Metal",
+            "value": 5731708.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=181144\nallocs=6003\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=gelu, affine=true)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 707187.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=53704\nallocs=1301\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=32)/forward/CPU/2 thread(s)",
+            "value": 834541.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2097216\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=32)/forward/CPU/4 thread(s)",
+            "value": 468500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2097216\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=32)/forward/CPU/8 thread(s)",
+            "value": 621917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2097216\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=32)/forward/CPU/1 thread(s)",
+            "value": 1560125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2097216\nallocs=2\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=32)/forward/GPU/CUDA",
+            "value": 134548.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=512\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=32)/forward/GPU/AMDGPU",
+            "value": 178651,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2832\nallocs=84\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=32)/zygote/CPU/2 thread(s)",
+            "value": 2687229.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8389712\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=32)/zygote/CPU/4 thread(s)",
+            "value": 1538104.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8389712\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=32)/zygote/CPU/8 thread(s)",
+            "value": 2006000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8389712\nallocs=31\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=32)/zygote/CPU/1 thread(s)",
+            "value": 4939208.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8389520\nallocs=29\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=32)/zygote/GPU/CUDA",
+            "value": 259430,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11808\nallocs=430\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(128, Bsize=32)/zygote/GPU/AMDGPU",
+            "value": 775478,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=20368\nallocs=459\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(4 x 32)/forward/CPU/2 thread(s)",
+            "value": 334,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(4 x 32)/forward/CPU/4 thread(s)",
+            "value": 375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(4 x 32)/forward/CPU/8 thread(s)",
+            "value": 375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(4 x 32)/forward/CPU/1 thread(s)",
+            "value": 292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1440\nallocs=19\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(4 x 32)/forward/GPU/CUDA",
+            "value": 31695,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4112\nallocs=142\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(4 x 32)/forward/GPU/Metal",
+            "value": 443479.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13344\nallocs=443\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(4 x 32)/forward/GPU/AMDGPU",
+            "value": 48751,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=6496\nallocs=174\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(4 x 32)/zygote/CPU/2 thread(s)",
+            "value": 7375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17360\nallocs=93\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(4 x 32)/zygote/CPU/4 thread(s)",
+            "value": 7292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17360\nallocs=93\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(4 x 32)/zygote/CPU/8 thread(s)",
+            "value": 8000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17360\nallocs=93\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(4 x 32)/zygote/CPU/1 thread(s)",
+            "value": 7333.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17360\nallocs=93\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(4 x 32)/zygote/GPU/CUDA",
+            "value": 221109,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=19544\nallocs=737\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(4 x 32)/zygote/GPU/Metal",
+            "value": 4885229,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=157352\nallocs=5224\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(2, act=relu, affine=false)(4 x 32)/zygote/GPU/AMDGPU",
+            "value": 371974,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=47240\nallocs=1154\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 2377333,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1050128\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 2416875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1050128\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 2384125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1050128\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 2389542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1050128\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 214311,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=17904\nallocs=612\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1592375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=56336\nallocs=1808\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 360673,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=20128\nallocs=510\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 4636270.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10496000\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 4666250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10496000\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 4654917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10496000\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 4646958,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=10496000\nallocs=133\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 959765,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=85904\nallocs=3212\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 6792729.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=293592\nallocs=9154\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "layernorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1254372,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=74512\nallocs=1551\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=relu)(512 x 128)/forward/CPU/2 thread(s)",
+            "value": 7729,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=relu)(512 x 128)/forward/CPU/4 thread(s)",
+            "value": 7021,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=relu)(512 x 128)/forward/CPU/8 thread(s)",
+            "value": 7250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=relu)(512 x 128)/forward/CPU/1 thread(s)",
+            "value": 9458,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=relu)(512 x 128)/forward/GPU/CUDA",
+            "value": 23256,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2384\nallocs=107\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=relu)(512 x 128)/forward/GPU/Metal",
+            "value": 265042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7184\nallocs=234\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=relu)(512 x 128)/forward/GPU/AMDGPU",
+            "value": 38891,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1184\nallocs=26\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=relu)(512 x 128)/zygote/CPU/2 thread(s)",
+            "value": 52312.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=789264\nallocs=16\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=relu)(512 x 128)/zygote/CPU/4 thread(s)",
+            "value": 52104,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=789264\nallocs=16\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=relu)(512 x 128)/zygote/CPU/8 thread(s)",
+            "value": 33229,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=789264\nallocs=16\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=relu)(512 x 128)/zygote/CPU/1 thread(s)",
+            "value": 63750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=789264\nallocs=16\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=relu)(512 x 128)/zygote/GPU/CUDA",
+            "value": 233595.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16840\nallocs=631\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=relu)(512 x 128)/zygote/GPU/Metal",
+            "value": 2003500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=64072\nallocs=2140\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "bias_activation(512, act=relu)(512 x 128)/zygote/GPU/AMDGPU",
+            "value": 225372,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16056\nallocs=367\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=512)/forward/CPU/2 thread(s)",
+            "value": 21541,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8448\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=512)/forward/CPU/4 thread(s)",
+            "value": 24916,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8448\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=512)/forward/CPU/8 thread(s)",
+            "value": 24437.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8448\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=512)/forward/CPU/1 thread(s)",
+            "value": 5292,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8448\nallocs=1\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=512)/forward/GPU/CUDA",
+            "value": 17600,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=512\nallocs=23\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=512)/forward/GPU/AMDGPU",
+            "value": 86991,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=14768\nallocs=101\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=512)/zygote/CPU/2 thread(s)",
+            "value": 11875,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=34640\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=512)/zygote/CPU/4 thread(s)",
+            "value": 9250,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=34640\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=512)/zygote/CPU/8 thread(s)",
+            "value": 10479.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=34640\nallocs=27\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=512)/zygote/CPU/1 thread(s)",
+            "value": 17979,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=34448\nallocs=25\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=512)/zygote/GPU/CUDA",
+            "value": 242050.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=11808\nallocs=430\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchedmm(2, Bsize=512)/zygote/GPU/AMDGPU",
+            "value": 379203,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=52528\nallocs=415\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=relu)(512 x 128)/forward/CPU/2 thread(s)",
+            "value": 406209,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=relu)(512 x 128)/forward/CPU/4 thread(s)",
+            "value": 223645.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=relu)(512 x 128)/forward/CPU/8 thread(s)",
+            "value": 297291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=relu)(512 x 128)/forward/CPU/1 thread(s)",
+            "value": 762770.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=262384\nallocs=6\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=relu)(512 x 128)/forward/GPU/CUDA",
+            "value": 46561,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1056\nallocs=47\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=relu)(512 x 128)/forward/GPU/Metal",
+            "value": 507959,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8016\nallocs=281\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=relu)(512 x 128)/forward/GPU/AMDGPU",
+            "value": 89681,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=3152\nallocs=149\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=relu)(512 x 128)/zygote/CPU/2 thread(s)",
+            "value": 1479041,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2100080\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=relu)(512 x 128)/zygote/CPU/4 thread(s)",
+            "value": 892812.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2100080\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=relu)(512 x 128)/zygote/CPU/8 thread(s)",
+            "value": 1165500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2100080\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=relu)(512 x 128)/zygote/CPU/1 thread(s)",
+            "value": 2475375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=2100080\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=relu)(512 x 128)/zygote/GPU/CUDA",
+            "value": 302356,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18136\nallocs=667\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=relu)(512 x 128)/zygote/GPU/Metal",
+            "value": 2142375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=70056\nallocs=2351\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(512, bias=true, act=relu)(512 x 128)/zygote/GPU/AMDGPU",
+            "value": 379453,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=18200\nallocs=465\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 435625,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 429834,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 437541,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 448000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1049824\nallocs=20\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 54368,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4448\nallocs=152\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1064291,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=13760\nallocs=447\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 237093,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12896\nallocs=429\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 3904000,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8404288\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 4036083,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8404288\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 4027208.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8404288\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 3808500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=8404288\nallocs=115\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 265452,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=23080\nallocs=861\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 10355209,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=211720\nallocs=6761\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "batchnorm(4, act=gelu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1238912.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=81272\nallocs=1740\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=gelu)(32 x 128)/forward/CPU/2 thread(s)",
+            "value": 8792,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=gelu)(32 x 128)/forward/CPU/4 thread(s)",
+            "value": 6917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=gelu)(32 x 128)/forward/CPU/8 thread(s)",
+            "value": 7667,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=gelu)(32 x 128)/forward/CPU/1 thread(s)",
+            "value": 12375,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=16832\nallocs=5\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=gelu)(32 x 128)/forward/GPU/CUDA",
+            "value": 23715,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1056\nallocs=47\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=gelu)(32 x 128)/forward/GPU/Metal",
+            "value": 228854.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7976\nallocs=279\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=gelu)(32 x 128)/forward/GPU/AMDGPU",
+            "value": 217862,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=4720\nallocs=247\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=gelu)(32 x 128)/zygote/CPU/2 thread(s)",
+            "value": 45125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=87936\nallocs=12\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=gelu)(32 x 128)/zygote/CPU/4 thread(s)",
+            "value": 45334,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=87936\nallocs=12\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=gelu)(32 x 128)/zygote/CPU/8 thread(s)",
+            "value": 45166,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=87936\nallocs=12\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=gelu)(32 x 128)/zygote/CPU/1 thread(s)",
+            "value": 45042,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=87936\nallocs=12\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=gelu)(32 x 128)/zygote/GPU/CUDA",
+            "value": 356362,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=20184\nallocs=752\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=gelu)(32 x 128)/zygote/GPU/Metal",
+            "value": 1233500,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=67304\nallocs=2251\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "dense(32, bias=true, act=gelu)(32 x 128)/zygote/GPU/AMDGPU",
+            "value": 675687,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=20520\nallocs=564\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/2 thread(s)",
+            "value": 85542,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051792\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/4 thread(s)",
+            "value": 123125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051792\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/8 thread(s)",
+            "value": 86770.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051792\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/CPU/1 thread(s)",
+            "value": 83125,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=1051792\nallocs=39\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/CUDA",
+            "value": 190402,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=12560\nallocs=283\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/Metal",
+            "value": 1991958.5,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=50976\nallocs=1614\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/forward/GPU/AMDGPU",
+            "value": 203932,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=23424\nallocs=463\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/2 thread(s)",
+            "value": 2017750,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7356160\nallocs=108\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/4 thread(s)",
+            "value": 2018541,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7356160\nallocs=108\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/8 thread(s)",
+            "value": 2017917,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7356160\nallocs=108\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/CPU/1 thread(s)",
+            "value": 1986854,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=7356160\nallocs=108\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/CUDA",
+            "value": 571148,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=54456\nallocs=1517\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/Metal",
+            "value": 9347417,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=184408\nallocs=5506\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          },
+          {
+            "name": "groupnorm(4, act=relu, affine=false)(16 x 16 x 32 x 32)/zygote/GPU/AMDGPU",
+            "value": 1095271,
+            "unit": "ns",
+            "extra": "gctime=0\nmemory=78616\nallocs=1724\nparams={\"gctrial\":true,\"time_tolerance\":0.05,\"evals_set\":false,\"samples\":10000,\"evals\":1,\"gcsample\":false,\"seconds\":5,\"overhead\":0,\"memory_tolerance\":0.01}"
+          }
+        ]
       }
     ]
   }