Disable donated_buffer for all ops's backward benchmarking (#104)

Summary:
It is still a temporary fix for backward benchmarking. Related discussion #40

Pull Request resolved: #104

Reviewed By: xuzhao9

Differential Revision: D66911331

Pulled By: FindHao

fbshipit-source-id: 6b3e5188fb6c929d6fe34aaf3a141bafa92c33f3

tritonbench/operators/geglu/operator.py

@@ -8,6 +8,7 @@
 
 from tritonbench.utils.triton_op import (
     BenchmarkOperator,
+    Mode,
     register_benchmark,
     register_x_val,
 )
@@ -59,6 +60,12 @@ def liger_geglu(self, input) -> Callable:
 
     @register_benchmark()
     def inductor_geglu(self, input) -> Callable:
+        # TODO: remove this once we have a better way to handle backward benchmarking
+        # We need to run backward multiple times for proper benchmarking
+        # so donated buffer have to be disabled
+        if self.mode == Mode.BWD or self.mode == Mode.FWD_BWD:
+            import torch._functorch.config
+
         compiled = torch.compile(self.baseline_model)
         return lambda: compiled(input)
 

tritonbench/operators/layer_norm/operator.py

@@ -34,6 +34,7 @@ def torch_layer_norm(self, *args):
 
     @register_benchmark()
     def torch_compile_layer_norm(self, *args):
+        # TODO: remove this once we have a better way to handle backward benchmarking
         # We need to run backward multiple times for proper benchmarking
         # so donated buffer have to be disabled
         if self.mode == Mode.BWD or self.mode == Mode.FWD_BWD:

tritonbench/operators/swiglu/operator.py

@@ -7,6 +7,7 @@
 
 from tritonbench.utils.triton_op import (
     BenchmarkOperator,
+    Mode,
     register_benchmark,
     register_x_val,
 )
@@ -59,6 +60,12 @@ def liger_swiglu(self, input) -> Callable:
 
     @register_benchmark()
     def inductor_swiglu(self, input) -> Callable:
+        # TODO: remove this once we have a better way to handle backward benchmarking
+        # We need to run backward multiple times for proper benchmarking
+        # so donated buffer have to be disabled
+        if self.mode == Mode.BWD or self.mode == Mode.FWD_BWD:
+            import torch._functorch.config
+
         compiled = torch.compile(self.baseline_op)
         return lambda: compiled(input)