Remove unnecessary graph rewrites for both cuda and whole ET pipeline

backends/aoti/aoti_backend.py

@@ -20,6 +20,7 @@
 from executorch.exir._warnings import experimental
 from executorch.exir.backend.backend_details import ExportedProgram, PreprocessResult
 from executorch.exir.backend.compile_spec_schema import CompileSpec
+from executorch.exir.graph_module import contains_any_op
 from torch._inductor.codegen.cpp_wrapper_cpu import CppWrapperCpu
 from torch.export.passes import move_to_device_pass
 
@@ -248,8 +249,12 @@ def preprocess(
             else:
                 custom_pass(device_edge_program.graph_module)
 
-        # Run decompositions if any
-        if decomposition_table:
+        # ``run_decompositions`` retraces the complete ExportedProgram even
+        # when none of the table's operators occur in the graph. Large CUDA
+        # models make that no-op expensive, so only run it when it can apply.
+        if contains_any_op(
+            device_edge_program.graph_module, decomposition_table.keys()
+        ):
             device_edge_program = device_edge_program.run_decompositions(
                 decomposition_table
             )

backends/cuda/BUCK

@@ -77,7 +77,6 @@ fbcode_target(
     srcs = [
         "passes/__init__.py",
         "passes/move_cond_predicate_to_cpu.py",
-        "passes/replace_int64_floordiv.py",
     ],
     visibility = [
         "//executorch/backends/cuda/...",

backends/cuda/cuda_backend.py

@@ -23,9 +23,6 @@
 from executorch.backends.cuda.passes.move_cond_predicate_to_cpu import (
     MoveCondPredicateToCpuPass,
 )
-from executorch.backends.cuda.passes.replace_int64_floordiv import (
-    ReplaceInt64FloorDivWithFloatPass,
-)
 from executorch.backends.cuda.triton.replacement_pass import (
     ReplaceEdgeOpWithTritonOpPass,
 )
@@ -568,7 +565,7 @@ def get_custom_passes(cls, compile_specs: List[CompileSpec]) -> List[typing.Any]
                         f"Invalid triton_kernel_mode: {mode}. Expected 'ON' or 'OFF'."
                     )
                 triton_kernel_mode = mode
-        passes = [MoveCondPredicateToCpuPass(), ReplaceInt64FloorDivWithFloatPass()]
+        passes = [MoveCondPredicateToCpuPass()]
         if triton_kernel_mode == "ON":
             passes.append(ReplaceEdgeOpWithTritonOpPass())
         return passes