[GH-PAGES] Updated website

triton-lang · Jun 12, 2024 · 22f4a18 · 22f4a18
1 parent acda248
commit 22f4a18
Show file tree

Hide file tree

Showing 62 changed files with 445 additions and 442 deletions.
diff --git a/main/.doctrees/environment.pickle b/main/.doctrees/environment.pickle
diff --git a/main/.doctrees/getting-started/tutorials/01-vector-add.doctree b/main/.doctrees/getting-started/tutorials/01-vector-add.doctree
diff --git a/main/.doctrees/getting-started/tutorials/02-fused-softmax.doctree b/main/.doctrees/getting-started/tutorials/02-fused-softmax.doctree
diff --git a/main/.doctrees/getting-started/tutorials/03-matrix-multiplication.doctree b/main/.doctrees/getting-started/tutorials/03-matrix-multiplication.doctree
diff --git a/main/.doctrees/getting-started/tutorials/04-low-memory-dropout.doctree b/main/.doctrees/getting-started/tutorials/04-low-memory-dropout.doctree
diff --git a/main/.doctrees/getting-started/tutorials/05-layer-norm.doctree b/main/.doctrees/getting-started/tutorials/05-layer-norm.doctree
diff --git a/main/.doctrees/getting-started/tutorials/06-fused-attention.doctree b/main/.doctrees/getting-started/tutorials/06-fused-attention.doctree
diff --git a/main/.doctrees/getting-started/tutorials/07-extern-functions.doctree b/main/.doctrees/getting-started/tutorials/07-extern-functions.doctree
diff --git a/main/.doctrees/getting-started/tutorials/08-grouped-gemm.doctree b/main/.doctrees/getting-started/tutorials/08-grouped-gemm.doctree
diff --git a/main/.doctrees/getting-started/tutorials/09-persistent-fp8-matmul.doctree b/main/.doctrees/getting-started/tutorials/09-persistent-fp8-matmul.doctree
diff --git a/main/.doctrees/getting-started/tutorials/sg_execution_times.doctree b/main/.doctrees/getting-started/tutorials/sg_execution_times.doctree
diff --git a/main/.doctrees/python-api/generated/triton.language.div_rn.doctree b/main/.doctrees/python-api/generated/triton.language.div_rn.doctree
diff --git a/main/.doctrees/python-api/generated/triton.language.sqrt_rn.doctree b/main/.doctrees/python-api/generated/triton.language.sqrt_rn.doctree
diff --git a/main/.doctrees/python-api/triton.language.doctree b/main/.doctrees/python-api/triton.language.doctree
diff --git a/main/.doctrees/sg_execution_times.doctree b/main/.doctrees/sg_execution_times.doctree
diff --git a/main/_downloads/3176accb6c7288b0e45f41d94eebacb9/06-fused-attention.ipynb b/main/_downloads/3176accb6c7288b0e45f41d94eebacb9/06-fused-attention.ipynb
diff --git a/main/_downloads/54a35f6ec55f9746935b9566fb6bb1df/06-fused-attention.py b/main/_downloads/54a35f6ec55f9746935b9566fb6bb1df/06-fused-attention.py
@@ -95,17 +95,17 @@ def keep(conf):
     return True
 
 
-@triton.autotune(list(filter(keep, configs)), key=["N_CTX"])
+@triton.autotune(list(filter(keep, configs)), key=["N_CTX", "HEAD_DIM"])
 @triton.jit
 def _attn_fwd(Q, K, V, sm_scale, M, Out,  #
               stride_qz, stride_qh, stride_qm, stride_qk,  #
               stride_kz, stride_kh, stride_kn, stride_kk,  #
               stride_vz, stride_vh, stride_vk, stride_vn,  #
               stride_oz, stride_oh, stride_om, stride_on,  #
               Z, H, N_CTX,  #
+              HEAD_DIM: tl.constexpr,  #
               BLOCK_M: tl.constexpr,  #
               BLOCK_N: tl.constexpr,  #
-              HEAD_DIM: tl.constexpr,  #
               STAGE: tl.constexpr  #
               ):
     tl.static_assert(BLOCK_N <= HEAD_DIM)
@@ -442,7 +442,7 @@ def forward(ctx, q, k, v, causal, sm_scale):
         # shape constraints
         HEAD_DIM_Q, HEAD_DIM_K = q.shape[-1], k.shape[-1]
         # when v is in float8_e5m2 it is transposed.
-        HEAD_DIM_V = v.shape[-2] if v.dtype == torch.float8_e5m2 else v.shape[-1]
+        HEAD_DIM_V = v.shape[-1]
         assert HEAD_DIM_Q == HEAD_DIM_K and HEAD_DIM_K == HEAD_DIM_V
         assert HEAD_DIM_K in {16, 32, 64, 128, 256}
         o = torch.empty_like(q)
@@ -609,6 +609,7 @@ def bench_flash_attention(BATCH, H, N_CTX, HEAD_DIM, causal, mode, provider, dev
         if mode == "fwd" and "fp8" in provider:
             q = q.to(torch.float8_e5m2)
             k = k.to(torch.float8_e5m2)
+            v = v.permute(0, 1, 3, 2).contiguous()
             v = v.permute(0, 1, 3, 2)
             v = v.to(torch.float8_e5m2)
         sm_scale = 1.3

diff --git a/main/_downloads/662999063954282841dc90b8945f85ce/tutorials_jupyter.zip b/main/_downloads/662999063954282841dc90b8945f85ce/tutorials_jupyter.zip
diff --git a/main/_downloads/763344228ae6bc253ed1a6cf586aa30d/tutorials_python.zip b/main/_downloads/763344228ae6bc253ed1a6cf586aa30d/tutorials_python.zip
diff --git a/main/_downloads/b9ce5a39c8d43007681c288566b27368/09-persistent-fp8-matmul.py b/main/_downloads/b9ce5a39c8d43007681c288566b27368/09-persistent-fp8-matmul.py
@@ -2,7 +2,7 @@
 Persistent FP8 Matmul
 =====================
 This script demonstrates persistent kernel implementations of matrix multiplication using Triton.
-It includes various matmul methods, such as naive, persistent, and TMA (Tile Matrix Accumulation) based approaches, and only supports GPUs with compute capability >= 9.0.
+It includes various matmul methods, such as naive, persistent, and TMA (Tensor Memory Accelerator) based approaches, and only supports GPUs with compute capability >= 9.0.
 Triton and CuBLAS implementations are benchmarked under different configurations and evaluated using the proton profiler.
 Users can pass command-line arguments to specify matrix dimensions and iteration steps flexibly.
 """

diff --git a/main/_downloads/c640f91b8ed671656d9d9ec8f7d29d34/09-persistent-fp8-matmul.ipynb b/main/_downloads/c640f91b8ed671656d9d9ec8f7d29d34/09-persistent-fp8-matmul.ipynb
@@ -4,7 +4,7 @@
       "cell_type": "markdown",
       "metadata": {},
       "source": [
-        "\n# Persistent FP8 Matmul\nThis script demonstrates persistent kernel implementations of matrix multiplication using Triton.\nIt includes various matmul methods, such as naive, persistent, and TMA (Tile Matrix Accumulation) based approaches, and only supports GPUs with compute capability >= 9.0.\nTriton and CuBLAS implementations are benchmarked under different configurations and evaluated using the proton profiler.\nUsers can pass command-line arguments to specify matrix dimensions and iteration steps flexibly.\n"
+        "\n# Persistent FP8 Matmul\nThis script demonstrates persistent kernel implementations of matrix multiplication using Triton.\nIt includes various matmul methods, such as naive, persistent, and TMA (Tensor Memory Accelerator) based approaches, and only supports GPUs with compute capability >= 9.0.\nTriton and CuBLAS implementations are benchmarked under different configurations and evaluated using the proton profiler.\nUsers can pass command-line arguments to specify matrix dimensions and iteration steps flexibly.\n"
       ]
     },
     {

diff --git a/main/_images/sphx_glr_01-vector-add_001.png b/main/_images/sphx_glr_01-vector-add_001.png
diff --git a/main/_images/sphx_glr_01-vector-add_thumb.png b/main/_images/sphx_glr_01-vector-add_thumb.png
diff --git a/main/_images/sphx_glr_02-fused-softmax_001.png b/main/_images/sphx_glr_02-fused-softmax_001.png
diff --git a/main/_images/sphx_glr_02-fused-softmax_thumb.png b/main/_images/sphx_glr_02-fused-softmax_thumb.png
diff --git a/main/_images/sphx_glr_03-matrix-multiplication_001.png b/main/_images/sphx_glr_03-matrix-multiplication_001.png
diff --git a/main/_images/sphx_glr_03-matrix-multiplication_002.png b/main/_images/sphx_glr_03-matrix-multiplication_002.png
diff --git a/main/_images/sphx_glr_03-matrix-multiplication_thumb.png b/main/_images/sphx_glr_03-matrix-multiplication_thumb.png
diff --git a/main/_images/sphx_glr_05-layer-norm_001.png b/main/_images/sphx_glr_05-layer-norm_001.png
diff --git a/main/_images/sphx_glr_05-layer-norm_thumb.png b/main/_images/sphx_glr_05-layer-norm_thumb.png
diff --git a/main/_images/sphx_glr_06-fused-attention_001.png b/main/_images/sphx_glr_06-fused-attention_001.png
diff --git a/main/_images/sphx_glr_06-fused-attention_002.png b/main/_images/sphx_glr_06-fused-attention_002.png
diff --git a/main/_images/sphx_glr_06-fused-attention_003.png b/main/_images/sphx_glr_06-fused-attention_003.png
diff --git a/main/_images/sphx_glr_06-fused-attention_thumb.png b/main/_images/sphx_glr_06-fused-attention_thumb.png
diff --git a/main/_images/sphx_glr_08-grouped-gemm_001.png b/main/_images/sphx_glr_08-grouped-gemm_001.png
diff --git a/main/_images/sphx_glr_08-grouped-gemm_thumb.png b/main/_images/sphx_glr_08-grouped-gemm_thumb.png
diff --git a/main/_sources/getting-started/tutorials/01-vector-add.rst.txt b/main/_sources/getting-started/tutorials/01-vector-add.rst.txt
@@ -230,17 +230,17 @@ We can now run the decorated function above. Pass `print_data=True` to see the p
 
     vector-add-performance:
                size       Triton        Torch
-    0        4096.0     8.000000     8.000000
-    1        8192.0    15.999999    15.999999
-    2       16384.0    31.999999    31.999999
+    0        4096.0     8.000000     9.600000
+    1        8192.0    19.200000    15.999999
+    2       16384.0    38.400001    31.999999
     3       32768.0    63.999998    63.999998
     4       65536.0   127.999995   127.999995
     5      131072.0   219.428568   219.428568
-    6      262144.0   384.000001   341.333321
-    7      524288.0   614.400016   558.545450
+    6      262144.0   384.000001   384.000001
+    7      524288.0   614.400016   614.400016
     8     1048576.0   819.200021   819.200021
     9     2097152.0  1068.521715  1023.999964
-    10    4194304.0  1260.307736  1260.307736
+    10    4194304.0  1260.307736  1228.800031
     11    8388608.0  1424.695621  1424.695621
     12   16777216.0  1560.380965  1560.380965
     13   33554432.0  1624.859540  1624.859540
@@ -253,7 +253,7 @@ We can now run the decorated function above. Pass `print_data=True` to see the p
 
 .. rst-class:: sphx-glr-timing
 
-   **Total running time of the script:** (0 minutes 8.965 seconds)
+   **Total running time of the script:** (0 minutes 8.340 seconds)
 
 
 .. _sphx_glr_download_getting-started_tutorials_01-vector-add.py:

diff --git a/main/_sources/getting-started/tutorials/02-fused-softmax.rst.txt b/main/_sources/getting-started/tutorials/02-fused-softmax.rst.txt
@@ -303,104 +303,104 @@ We will then compare its performance against (1) :code:`torch.softmax` and (2) t
 
     softmax-performance:
               N       Triton        Torch
-    0     256.0   474.886872   694.483260
-    1     384.0   608.275337   793.865707
-    2     512.0   750.325069   907.709196
-    3     640.0   786.230049   967.184122
-    4     768.0   882.550621  1019.516965
-    5     896.0   933.082975  1070.506037
-    6    1024.0   990.964158  1116.796272
-    7    1152.0  1110.548913   613.568651
-    8    1280.0  1139.226810   669.627030
-    9    1408.0  1161.877551   724.977307
-    10   1536.0  1192.678787   780.087866
-    11   1664.0  1208.494960   813.990785
-    12   1792.0  1228.799871   855.423724
-    13   1920.0  1253.318071   910.135827
-    14   2048.0  1272.176961   958.212585
-    15   2176.0  1262.729082   977.668960
-    16   2304.0  1263.769744  1009.549848
-    17   2432.0  1300.221255  1058.137453
-    18   2560.0  1304.930348  1083.903210
-    19   2688.0  1310.316031  1101.638060
-    20   2816.0  1333.607448  1130.354944
-    21   2944.0  1326.281560  1170.386713
-    22   3072.0  1354.066553  1182.816857
-    23   3200.0  1354.196012  1191.882725
-    24   3328.0  1362.253764  1226.216284
-    25   3456.0  1372.578742  1249.475639
-    26   3584.0  1375.513688  1257.543915
-    27   3712.0  1383.660024  1266.022294
-    28   3840.0  1384.542849  1301.957450
-    29   3968.0  1387.246937  1313.791403
-    30   4096.0  1401.801736  1323.349043
-    31   4224.0  1336.502774  1162.260414
-    32   4352.0  1339.787153  1172.044541
-    33   4480.0  1351.870184  1183.736530
-    34   4608.0  1357.423399  1196.479290
-    35   4736.0  1361.989884  1198.912098
-    36   4864.0  1380.536616  1222.557769
-    37   4992.0  1369.355432  1238.439229
-    38   5120.0  1373.097897  1251.131265
-    39   5248.0  1377.856986  1258.540361
-    40   5376.0  1380.042163  1286.206289
-    41   5504.0  1383.007577  1297.681281
-    42   5632.0  1387.751447  1310.564525
-    43   5760.0  1393.654050  1328.675932
-    44   5888.0  1393.712112  1339.936464
-    45   6016.0  1403.099460  1354.114534
-    46   6144.0  1409.151442  1373.076994
-    47   6272.0  1412.647808  1378.095969
-    48   6400.0  1416.736394  1388.316354
-    49   6528.0  1415.977502  1395.751557
-    50   6656.0  1418.974179  1404.295022
-    51   6784.0  1417.375807  1416.872608
-    52   6912.0  1426.498550  1424.602149
-    53   7040.0  1418.695818  1430.935232
-    54   7168.0  1430.265776  1433.229186
-    55   7296.0  1434.456799  1443.744323
-    56   7424.0  1430.059198  1446.000678
-    57   7552.0  1427.527213  1454.072801
-    58   7680.0  1438.971950  1461.880615
-    59   7808.0  1431.538789  1463.922739
-    60   7936.0  1433.513264  1469.388283
-    61   8064.0  1438.438301  1475.123808
-    62   8192.0  1436.675086  1485.034777
-    63   8320.0  1384.416885  1403.418539
-    64   8448.0  1381.969366  1402.997531
-    65   8576.0  1396.067673  1392.623468
-    66   8704.0  1387.553880  1402.493073
-    67   8832.0  1384.059893  1403.281876
-    68   8960.0  1395.940239  1411.576239
-    69   9088.0  1408.235295  1415.618970
-    70   9216.0  1405.069496  1427.334371
-    71   9344.0  1397.409535  1422.342505
-    72   9472.0  1398.859515  1433.090827
-    73   9600.0  1397.536414  1430.981887
-    74   9728.0  1399.649646  1444.275164
-    75   9856.0  1418.279890  1440.370402
-    76   9984.0  1403.428461  1452.395129
-    77  10112.0  1410.351708  1452.974939
-    78  10240.0  1414.819754  1466.205557
-    79  10368.0  1411.204824  1462.887115
-    80  10496.0  1414.606195  1466.766495
-    81  10624.0  1409.348141  1467.196829
-    82  10752.0  1402.620501  1473.049479
-    83  10880.0  1400.636832  1481.012938
-    84  11008.0  1417.088863  1478.179239
-    85  11136.0  1423.359373  1486.956090
-    86  11264.0  1430.935728  1486.185796
-    87  11392.0  1414.594401  1490.294891
-    88  11520.0  1425.302497  1494.486333
-    89  11648.0  1431.407108  1497.469564
-    90  11776.0  1428.276153  1501.569828
-    91  11904.0  1439.345217  1507.349950
-    92  12032.0  1420.063148  1506.856766
-    93  12160.0  1418.473769  1509.935419
-    94  12288.0  1438.239656  1392.164615
-    95  12416.0  1445.141343  1392.158306
-    96  12544.0  1442.222511  1390.996328
-    97  12672.0  1446.057806  1389.808823
+    0     256.0   483.373814   691.410530
+    1     384.0   618.284228   807.811239
+    2     512.0   758.376010   913.515445
+    3     640.0   796.139488   952.027425
+    4     768.0   884.106445  1017.410462
+    5     896.0   931.881621  1060.687548
+    6    1024.0   987.890277  1124.106325
+    7    1152.0  1109.174905   614.814920
+    8    1280.0  1155.811618   669.919401
+    9    1408.0  1159.883220   725.626002
+    10   1536.0  1195.718508   779.164094
+    11   1664.0  1211.134824   814.654418
+    12   1792.0  1234.185297   856.258949
+    13   1920.0  1254.848750   909.547141
+    14   2048.0  1280.527072   960.181979
+    15   2176.0  1266.130989   977.655676
+    16   2304.0  1269.216999  1008.152209
+    17   2432.0  1301.518687  1054.155940
+    18   2560.0  1305.882814  1083.215684
+    19   2688.0  1309.052134  1103.331214
+    20   2816.0  1326.479814  1133.798566
+    21   2944.0  1327.003204  1164.515965
+    22   3072.0  1355.254207  1182.967957
+    23   3200.0  1354.312383  1196.631655
+    24   3328.0  1363.235926  1222.992633
+    25   3456.0  1374.813321  1247.572000
+    26   3584.0  1379.471259  1261.300215
+    27   3712.0  1383.468547  1269.207431
+    28   3840.0  1390.274965  1303.357204
+    29   3968.0  1392.744551  1313.293799
+    30   4096.0  1395.432489  1324.162252
+    31   4224.0  1330.762484  1160.325788
+    32   4352.0  1337.949192  1173.518168
+    33   4480.0  1354.147132  1180.916725
+    34   4608.0  1364.426067  1194.316176
+    35   4736.0  1358.579305  1196.606959
+    36   4864.0  1377.283327  1220.686468
+    37   4992.0  1373.385271  1235.483630
+    38   5120.0  1372.565674  1248.511059
+    39   5248.0  1373.629334  1257.727135
+    40   5376.0  1374.573949  1286.122332
+    41   5504.0  1383.845804  1297.084539
+    42   5632.0  1383.685094  1312.995012
+    43   5760.0  1395.206666  1324.070267
+    44   5888.0  1388.315574  1342.564480
+    45   6016.0  1398.865022  1352.200223
+    46   6144.0  1412.620414  1373.401229
+    47   6272.0  1412.668209  1374.286825
+    48   6400.0  1419.615232  1385.639329
+    49   6528.0  1412.254639  1396.029543
+    50   6656.0  1420.756083  1403.343240
+    51   6784.0  1414.896258  1413.407734
+    52   6912.0  1428.907998  1423.411192
+    53   7040.0  1420.585163  1432.429243
+    54   7168.0  1427.360996  1432.997128
+    55   7296.0  1431.118521  1444.424889
+    56   7424.0  1428.966318  1445.439665
+    57   7552.0  1423.730409  1454.009783
+    58   7680.0  1433.413558  1459.313934
+    59   7808.0  1432.827777  1467.103305
+    60   7936.0  1434.701316  1467.759825
+    61   8064.0  1440.947239  1472.594700
+    62   8192.0  1436.460859  1483.730067
+    63   8320.0  1389.977920  1403.779373
+    64   8448.0  1378.195253  1405.602653
+    65   8576.0  1398.062328  1394.030166
+    66   8704.0  1390.322779  1401.275831
+    67   8832.0  1385.130096  1402.587790
+    68   8960.0  1399.843525  1411.445897
+    69   9088.0  1410.506406  1416.167489
+    70   9216.0  1405.039396  1425.971433
+    71   9344.0  1400.395161  1423.667293
+    72   9472.0  1396.965321  1432.095267
+    73   9600.0  1393.987492  1436.089884
+    74   9728.0  1398.630681  1442.419996
+    75   9856.0  1415.951782  1442.786485
+    76   9984.0  1400.951074  1451.195595
+    77  10112.0  1413.151543  1455.469015
+    78  10240.0  1419.068719  1466.755375
+    79  10368.0  1415.493789  1460.576378
+    80  10496.0  1415.119991  1467.882429
+    81  10624.0  1412.048253  1467.768703
+    82  10752.0  1413.444281  1472.473186
+    83  10880.0  1402.165427  1477.500992
+    84  11008.0  1422.362282  1479.161425
+    85  11136.0  1422.191503  1484.824616
+    86  11264.0  1431.432382  1485.935090
+    87  11392.0  1417.856347  1491.204361
+    88  11520.0  1422.069283  1492.275406
+    89  11648.0  1427.668803  1499.136712
+    90  11776.0  1433.812357  1501.546145
+    91  11904.0  1445.187994  1506.729085
+    92  12032.0  1426.405254  1508.512115
+    93  12160.0  1419.359808  1512.522930
+    94  12288.0  1435.564739  1391.805939
+    95  12416.0  1448.499886  1391.218675
+    96  12544.0  1441.379527  1393.247778
+    97  12672.0  1449.838003  1391.809115
 
 
 
@@ -415,7 +415,7 @@ In the above plot, we can see that:
 
 .. rst-class:: sphx-glr-timing
 
-   **Total running time of the script:** (0 minutes 34.483 seconds)
+   **Total running time of the script:** (0 minutes 28.602 seconds)
 
 
 .. _sphx_glr_download_getting-started_tutorials_02-fused-softmax.py: