Improve _safe_softmax lowering

PiperOrigin-RevId: 704838624

Author: chunnienc

ai_edge_torch/odml_torch/lowerings/__init__.py

@@ -21,6 +21,6 @@
 from . import context
 from . import registry
 from . import utils
-from .registry import decompositions
+from .decomp import decompositions
 from .registry import lookup
 from .registry import lower

ai_edge_torch/odml_torch/lowerings/decomp.py

@@ -0,0 +1,59 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Torch export decompositions to run before lowering."""
+
+import functools
+
+import torch
+
+
+@functools.cache
+def decompositions():
+  # Base: Core ATen decompositions
+  decompositions = torch._decomp.core_aten_decompositions()
+
+  decompositions.update(
+      torch._decomp.get_decompositions([
+          torch.ops.aten.upsample_nearest2d,
+          torch.ops.aten._native_batch_norm_legit.no_stats,
+          torch.ops.aten._native_batch_norm_legit_functional,
+          torch.ops.aten._adaptive_avg_pool2d,
+          torch.ops.aten._adaptive_avg_pool3d,
+          torch.ops.aten.grid_sampler_2d,
+          torch.ops.aten.native_group_norm,
+          torch.ops.aten.native_dropout,
+          torch.ops.aten.reflection_pad1d,
+          torch.ops.aten.reflection_pad2d,
+          torch.ops.aten.reflection_pad3d,
+          torch.ops.aten.replication_pad1d,
+          torch.ops.aten.replication_pad2d,
+          torch.ops.aten.replication_pad3d,
+          torch.ops.aten.addmm,
+      ])
+  )
+
+  torch._decomp.remove_decompositions(
+      decompositions,
+      [torch.ops.aten.roll],
+  )
+
+  # Override _safe_softmax decompositions with regular softmax.
+  # _safe_softmax introduces additional check-select ops to guard extreme
+  # input values to softmax, which could make the converted model inefficient
+  # on-device.
+  if hasattr(torch.ops.aten, "_safe_softmax"):
+    decompositions[torch.ops.aten._safe_softmax.default] = torch.softmax
+
+  return decompositions

ai_edge_torch/odml_torch/lowerings/registry.py

@@ -26,7 +26,6 @@ class LoweringRegistry:
 
   def __init__(self):
     self.registered_ops = {}
-    self.decompositions = {}
 
   def lookup(self, op_or_name):
     candidate = self._get_lowering(op_or_name)
@@ -52,33 +51,6 @@ def register(self, op, lowering):
 
 
 global_registry = LoweringRegistry()
-global_registry.decompositions.update(torch._decomp.core_aten_decompositions())
-global_registry.decompositions.update(
-    torch._decomp.get_decompositions([
-        torch.ops.aten.upsample_nearest2d,
-        torch.ops.aten._native_batch_norm_legit.no_stats,
-        torch.ops.aten._native_batch_norm_legit_functional,
-        torch.ops.aten._adaptive_avg_pool2d,
-        torch.ops.aten._adaptive_avg_pool3d,
-        torch.ops.aten.grid_sampler_2d,
-        torch.ops.aten.native_group_norm,
-        torch.ops.aten.native_dropout,
-        torch.ops.aten.reflection_pad1d,
-        torch.ops.aten.reflection_pad2d,
-        torch.ops.aten.reflection_pad3d,
-        torch.ops.aten.replication_pad1d,
-        torch.ops.aten.replication_pad2d,
-        torch.ops.aten.replication_pad3d,
-        torch.ops.aten.addmm,
-    ])
-)
-
-torch._decomp.remove_decompositions(
-    global_registry.decompositions,
-    [
-        torch.ops.aten.roll,
-    ],
-)
 
 
 def lookup(op):
@@ -91,7 +63,3 @@ def inner(lowering: Callable[[context.LoweringContext, ...], Any]):
     return lowering
 
   return inner
-
-
-def decompositions():
-  return global_registry.decompositions