[inductor] Clean up use_mixed_mm and mixed_mm_choice usage inside pytorch

henrylhtsang · henrylhtsang · commit 8f3a77f783b7 · 2025-04-23T16:56:23.000-07:00
Differential Revision: [D73551912](https://our.internmc.facebook.com/intern/diff/D73551912/) [ghstack-poisoned]
diff --git a/test/inductor/test_max_autotune.py b/test/inductor/test_max_autotune.py
@@ -1498,8 +1498,6 @@ def foo(x, y):
         {
             "max_autotune_gemm_backends": "Triton",
             "benchmark_epilogue_fusion": True,
-            "use_mixed_mm": False,
-            "mixed_mm_choice": "default",
             "max_epilogue_benchmarked_choices": 3,
         }
     )
@@ -1534,8 +1532,6 @@ def resolve_pending(x):
         {
             "max_autotune_gemm_backends": "Triton",
             "benchmark_epilogue_fusion": True,
-            "use_mixed_mm": False,
-            "mixed_mm_choice": "default",
             "max_epilogue_benchmarked_choices": 3,
         }
     )
diff --git a/test/inductor/test_torchinductor.py b/test/inductor/test_torchinductor.py
@@ -3724,7 +3724,6 @@ def fn(a, b):
 
     @skipIfPy312  # segfaults
     @skipCUDAIf(not SM80OrLater, "Requires sm80")
-    @config.patch(mixed_mm_choice="triton")
     def test_mixed_mm(self):
         def fn(a, b):
             return torch.mm(a, b.to(a.dtype))
@@ -3740,7 +3739,6 @@ def fn(a, b):
 
     @skipIfPy312  # segfaults
     @skipCUDAIf(not SM80OrLater, "Requires sm80")
-    @config.patch(mixed_mm_choice="triton")
     def test_mixed_mm2(self):
         def fn(a, b, scale, bias):
             return torch.mm(a, b.to(a.dtype)) * scale + bias
@@ -3758,7 +3756,6 @@ def fn(a, b, scale, bias):
 
     @skipIfPy312  # segfaults
     @skipCUDAIf(not SM80OrLater, "Requires sm80")
-    @config.patch(mixed_mm_choice="triton")
     def test_mixed_mm3(self):
         def fn(a, b):
             return torch.mm(a, b.to(a.dtype))
@@ -3777,7 +3774,6 @@ def fn(a, b):
         )
 
     @with_tf32_off
-    @config.patch(use_mixed_mm=True)
     def test_uint4x2_mixed_mm(self):
         def fn(a, b):
             return torch.mm(
diff --git a/torch/_inductor/config.py b/torch/_inductor/config.py
@@ -294,9 +294,7 @@ def prologue_fusion_enabled() -> bool:
 # but the mul gets fused with other pointwise ops instead.
 force_fuse_int_mm_with_mul = False
 
-# for pattern torch.mm(a, b.to(dtype)) with cuda tensors,
-# enable torch._inductor.kernel.mm.tuned_mixed_mm fused kernel.
-# Autotune will compare perf with normal cast->then->mm option
+# DEPRECATED. This setting is ignored.
 use_mixed_mm = True
 
 # enable runtime numeric check for pre/post grad fx passes
@@ -311,19 +309,7 @@ def prologue_fusion_enabled() -> bool:
     "requires_optimizer": True,
 }
 
-# mixed_mm_choice can be used to control the behaviour for pattern torch.mm(a, b.to(dtype)) with cuda tensors.
-# The fallback aten implementation is normal cast->then->mm option.
-# If mixed_mm_choice is "default": this flag will be ignored.
-# If mixed_mm_choice is "triton":
-# - Always use torch._inductor.kernel.mm.tuned_mixed_mm's fused kernel.
-# - Autotune will not compare with fallback.
-# If mixed_mm_choice is "aten": always use the fallback aten implementation.
-# If mixed_mm_choice is "heuristic":
-# - Enables the heuristic.
-# - If the heuristic decides to add a config, it will add the config as the first choice.
-# - If autotune is disabled, this config will always be chosen.
-# - If autotune is enabled, it will also compare with fallback aten implementation and fused kernel.
-# The use_mixed_mm flag will be ignored if mixed_mm_choice != "default".
+# DEPRECATED. This setting is ignored.
 mixed_mm_choice: Literal["default", "triton", "aten", "heuristic"] = "heuristic"
 
 # enable reordering pass for increasing overlap between compute and communication
diff --git a/torch/_inductor/kernel/mm.py b/torch/_inductor/kernel/mm.py
@@ -1171,7 +1171,6 @@ def get_context(m, k, n, mat1, mat2, mat1_stride, mat2_stride):
             "mat2_iscontig", mat2.layout.is_contiguous(), is_categorical=True
         )
         if name == "mm":
-            # for mixed_mm, we only consider fp16
             context_add_using_tf32(context, mat1.layout.dtype)
         return context
 
