fp8 benchmark testing

benchmarks/profile_linear_float8.py

@@ -10,6 +10,7 @@
 from typing import Callable, Optional
 
 import fire
+import functools
 
 import torch
 from float8_experimental.float8_linear_utils import (
@@ -19,6 +20,7 @@
     sync_float8_amax_and_scale_history,
 )
 from torch.profiler import profile, ProfilerActivity, record_function
+from torch._inductor.utils import do_bench_using_profiling
 
 
 @dataclass
@@ -73,6 +75,10 @@ def profile_function(
     if config.file_path is None:
         print(prof.key_averages().table(sort_by="cpu_time_total", row_limit=10))
 
+    full_func = functools.partial(func, *args, **kwargs)
+    latency = do_bench_using_profiling(full_func)
+    print(f"{func=}, {latency=}")
+
     return prof
 
 
@@ -134,24 +140,24 @@ def main(profile_path: Path, compile: bool, linear_type: str):
 
     def ref_forw_backward(x):
         if params.layer_norm:
-            with record_function("layer_norm"):
-                x = ln(x)
-        with record_function("forward"):
-            out = linear_ref(x)
-        with record_function("backward"):
-            out.sum().backward()
+            #with record_function("layer_norm"):
+            x = ln(x)
+        #with record_function("forward"):
+        out = linear_ref(x)
+        #with record_function("backward"):
+        out.sum().backward()
 
     def float8_forw_backward(x):
         if linear_requires_sync(linear_type):
-            with record_function("scale_amax_and_scales"):
-                sync_float8_amax_and_scale_history(linear_float8)
+            # with record_function("scale_amax_and_scales"):
+            sync_float8_amax_and_scale_history(linear_float8)
         if params.layer_norm:
-            with record_function("layer_norm"):
-                x = ln(x)
-        with record_function("forward"):
-            out = linear_float8(x)
-        with record_function("backward"):
-            out.sum().backward()
+            # with record_function("layer_norm"):
+            x = ln(x)
+        # with record_function("forward"):
+        out = linear_float8(x)
+        # with record_function("backward"):
+        out.sum().backward()
 
     if transformer_engine_installed:
         # Create an FP8 recipe. Note: All input args are optional.
@@ -170,15 +176,20 @@ def te_forw_backward(x):
                     out.sum().backward()
 
     if params.torch_compile:
-        ref_forw_backward = torch.compile(ref_forw_backward)
+        #ref_forw_backward = torch.compile(ref_forw_backward)
         float8_forw_backward = torch.compile(float8_forw_backward)
         # Compiling TE_linear fails but they are already compiling under the hood
         # if transformer_engine_installed:
         #     te_forw_backward = torch.compile(te_forw_backward)
 
+    def wrapper_float8(x):
+        if linear_requires_sync(linear_type):
+            sync_float8_amax_and_scale_history(linear_float8)
+        float8_forw_backward(x)
+
     for _ in range(5):
         ref_forw_backward(input_tensor)
-        float8_forw_backward(input_tensor)
+        wrapper_float8(input_tensor)
         if transformer_engine_installed:
             te_forw_backward(input_tensor)
 
@@ -189,7 +200,7 @@ def te_forw_backward(x):
     )
     profile_function(profile_config, ref_forw_backward, input_tensor)
 
-    # # Profile Float8 Linear
+    # Profile Float8 Linear
     float8_string = f"linear_float8_M_{params.M}_K_{params.K}_N_{params.N}_input_bias_{params.input_bias}_compile_{params.torch_compile}_{linear_type}.json"
     profile_config = ProfileConfig(
         str(profile_path / float8_string),
@@ -198,7 +209,7 @@ def te_forw_backward(x):
         warmup_iters=5,
         sync=True,
     )
-    profile_function(profile_config, float8_forw_backward, input_tensor)
+    profile_function(profile_config, wrapper_float8, input_tensor)
 
     te_string = f"linear_transformer_engine_M_{params.M}_K_{params.K}_N_{params.N}_input_bias_{params.input_bias}.json"
     if transformer_engine_installed: