Merge branch 'main' into toupstream/linear_int16

Author: zingo

.github/workflows/cuda.yml

@@ -89,6 +89,8 @@ jobs:
 
   export-voxtral-cuda-artifact:
     name: export-voxtral-cuda-${{ matrix.quant.name }}
+    # Skip this job if the pull request is from a fork (HuggingFace secrets are not available)
+    if: github.event.pull_request.head.repo.full_name == github.repository || github.event_name != 'pull_request'
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     permissions:
       id-token: write
@@ -166,6 +168,8 @@ jobs:
 
   export-gemma3-cuda-artifact:
     name: export-gemma3-cuda-${{ matrix.quant.name }}
+    # Skip this job if the pull request is from a fork (HuggingFace secrets are not available)
+    if: github.event.pull_request.head.repo.full_name == github.repository || github.event_name != 'pull_request'
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     permissions:
       id-token: write

backends/arm/_passes/__init__.py

@@ -88,8 +88,7 @@
 from .mm_to_bmm_pass import ConvertMmToBmmPass  # noqa
 from .remove_noop_pass import RemoveNoopPass  # noqa
 from .replace_scalar_with_tensor_pass import (  # noqa
-    ReplaceScalarWithTensorArgPassTOSABI,
-    ReplaceScalarWithTensorArgPassTOSAMI,
+    ReplaceScalarWithTensorByProfilePass,
 )
 from .rewrite_conv2d_pass import RewriteConv2dPass  # noqa
 from .rewrite_matmul import RewriteMatmulPass  # noqa

backends/arm/_passes/arm_pass_manager.py

@@ -87,8 +87,7 @@
     QuantizeOperatorArguments,
     RemoveNoopPass,
     ReplaceInfValues,
-    ReplaceScalarWithTensorArgPassTOSABI,
-    ReplaceScalarWithTensorArgPassTOSAMI,
+    ReplaceScalarWithTensorByProfilePass,
     RetraceFoldedDtypesPass,
     RewriteConv2dPass,
     RewriteMatmulPass,
@@ -154,15 +153,15 @@ def _transform(self, graph_module: GraphModule):
         with TosaLoweringContext(self.tosa_spec):
             return self(graph_module).graph_module
 
-    def _tosa_INT_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
+    def _tosa_INT_pipeline(
+        self, exported_program: ExportedProgram, graph_module: GraphModule
+    ) -> GraphModule:
         self.add_pass(AnnotateOutputDimOrderPass())
         self.add_pass(FuseQuantizedActivationPass())
         self.add_pass(RemoveGetItemPass())
         self.add_pass(ConvertSplitToSlicePass())
         self.add_pass(ConvertMmToBmmPass())
-        self.add_pass(
-            DecomposeMeanDimPass(exported_program.graph_module, self.tosa_spec)
-        )
+        self.add_pass(DecomposeMeanDimPass(graph_module, self.tosa_spec))
         self.add_pass(ConvertFullLikeToFullPass())
         self.add_pass(ConvertToClampPass())
         self.add_pass(ConvertMinMaxPass())
@@ -172,7 +171,7 @@ def _tosa_INT_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
             self.add_pass(CastToInt32Pass())
 
         self.add_pass(CastBoolToInt8Pass())
-        self.add_pass(ReplaceScalarWithTensorArgPassTOSABI())
+        self.add_pass(ReplaceScalarWithTensorByProfilePass())
         self.add_pass(AnnotateDecomposedMatmulPass())
         self.add_pass(QuantizeOperatorArguments())
         self.add_pass(ConvertELUParamsPass())
@@ -219,9 +218,11 @@ def _tosa_INT_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
         self.add_pass(InsertRescalePass())
 
         self.validate_constraints_mandatory()
-        return self._transform(exported_program.graph_module)
+        return self._transform(graph_module)
 
-    def _tosa_FP_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
+    def _tosa_FP_pipeline(
+        self, exported_program: ExportedProgram, graph_module: GraphModule
+    ) -> GraphModule:
         self.add_pass(AnnotateOutputDimOrderPass())
         self.add_pass(DecomposeExpm1Pass())
         self.add_pass(DecomposeLogitPass())
@@ -242,7 +243,7 @@ def _tosa_FP_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
         self.add_pass(DecomposeSinhPass())
         self.add_pass(DecomposeSignPass())
         self.add_pass(DecomposeDivTensorModePass())
-        self.add_pass(ReplaceScalarWithTensorArgPassTOSAMI())
+        self.add_pass(ReplaceScalarWithTensorByProfilePass())
         self.add_pass(DecomposeEmbeddingPass())
         self.add_pass(FuseQuantizedActivationPass())
         self.add_pass(RemoveGetItemPass())
@@ -256,9 +257,7 @@ def _tosa_FP_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
         self.add_pass(DecomposeLayerNormPass())
         self.add_pass(DecomposeBatchNormNoStatsPass())
         self.add_pass(DecomposeVarPass())
-        self.add_pass(
-            DecomposeMeanDimPass(exported_program.graph_module, self.tosa_spec)
-        )
+        self.add_pass(DecomposeMeanDimPass(graph_module, self.tosa_spec))
         self.add_pass(DecomposeNotEqualPass())
         self.add_pass(DecomposeDivPass())
         self.add_pass(DecomposeAddSubAlphaPass())
@@ -306,14 +305,16 @@ def _tosa_FP_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
         self.add_pass(InsertRescalePass())
 
         self.validate_constraints_mandatory()
-        return self._transform(exported_program.graph_module)
+        return self._transform(graph_module)
 
-    def transform_to_backend_pipeline(self, exported_program: ExportedProgram):
+    def transform_to_backend_pipeline(
+        self, exported_program: ExportedProgram, graph_module: GraphModule
+    ):
         """Apply passes before transforming program to backend"""
         if self.tosa_spec == TosaSpecification.create_from_string("TOSA-1.0+FP"):
-            return self._tosa_FP_pipeline(exported_program)
+            return self._tosa_FP_pipeline(exported_program, graph_module)
         elif self.tosa_spec == TosaSpecification.create_from_string("TOSA-1.0+INT"):
-            return self._tosa_INT_pipeline(exported_program)
+            return self._tosa_INT_pipeline(exported_program, graph_module)
         else:
             raise NotImplementedError(
                 f"No pass pipeline implemented for {self.tosa_spec=}"
@@ -335,7 +336,7 @@ def transform_for_annotation_pipeline(self, graph_module: GraphModule):
         self.add_pass(DecomposeAddmmPass())
         self.add_pass(DecomposeDivTensorModePass())
         self.add_pass(DecomposeAddSubAlphaPass())
-        self.add_pass(ReplaceScalarWithTensorArgPassTOSABI())
+        self.add_pass(ReplaceScalarWithTensorByProfilePass())
         self.add_pass(ScalarsToAttributePass())
         self.add_pass(DecomposeGroupNormPass())
         self.add_pass(DecomposeLayerNormPass())

backends/arm/_passes/decompose_acosh_pass.py

@@ -12,7 +12,7 @@
 from executorch.backends.arm._passes.match_arg_dtype_pass import MatchArgDtypePass
 from executorch.backends.arm._passes.match_arg_ranks_pass import MatchArgRanksPass
 from executorch.backends.arm._passes.replace_scalar_with_tensor_pass import (
-    ReplaceScalarWithTensorArgPassTOSAMI,
+    ReplaceScalarWithTensorByProfilePass,
 )
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass
@@ -32,7 +32,7 @@ class DecomposeAcoshPass(ArmPass):
         DecomposeSqrtPass,
         InsertTableOpsPass,
         MatchArgRanksPass,
-        ReplaceScalarWithTensorArgPassTOSAMI,
+        ReplaceScalarWithTensorByProfilePass,
         MatchArgDtypePass,
     }
 

backends/arm/_passes/decompose_asin_and_acos_pass.py

@@ -19,7 +19,7 @@
 from executorch.backends.arm._passes.match_arg_dtype_pass import MatchArgDtypePass
 from executorch.backends.arm._passes.match_arg_ranks_pass import MatchArgRanksPass
 from executorch.backends.arm._passes.replace_scalar_with_tensor_pass import (
-    ReplaceScalarWithTensorArgPassTOSAMI,
+    ReplaceScalarWithTensorByProfilePass,
 )
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass
@@ -71,7 +71,7 @@ class DecomposeAsinAndAcosPass(ArmPass):
         ConvertFullLikeToFullPass,
         MatchArgRanksPass,
         MatchArgDtypePass,
-        ReplaceScalarWithTensorArgPassTOSAMI,
+        ReplaceScalarWithTensorByProfilePass,
     }
 
     def _build_polynomial(

backends/arm/_passes/decompose_asinh_pass.py

@@ -12,7 +12,7 @@
 from executorch.backends.arm._passes.match_arg_dtype_pass import MatchArgDtypePass
 from executorch.backends.arm._passes.match_arg_ranks_pass import MatchArgRanksPass
 from executorch.backends.arm._passes.replace_scalar_with_tensor_pass import (
-    ReplaceScalarWithTensorArgPassTOSAMI,
+    ReplaceScalarWithTensorByProfilePass,
 )
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass
@@ -32,7 +32,7 @@ class DecomposeAsinhPass(ArmPass):
         DecomposeSqrtPass,
         InsertTableOpsPass,
         MatchArgRanksPass,
-        ReplaceScalarWithTensorArgPassTOSAMI,
+        ReplaceScalarWithTensorByProfilePass,
         MatchArgDtypePass,
     }
 

backends/arm/_passes/decompose_atan_pass.py

@@ -12,7 +12,7 @@
 from executorch.backends.arm._passes.match_arg_dtype_pass import MatchArgDtypePass
 from executorch.backends.arm._passes.match_arg_ranks_pass import MatchArgRanksPass
 from executorch.backends.arm._passes.replace_scalar_with_tensor_pass import (
-    ReplaceScalarWithTensorArgPassTOSAMI,
+    ReplaceScalarWithTensorByProfilePass,
 )
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass
@@ -47,7 +47,7 @@ class DecomposeAtanPass(ArmPass):
         InsertTableOpsPass,
         MatchArgRanksPass,
         MatchArgDtypePass,
-        ReplaceScalarWithTensorArgPassTOSAMI,
+        ReplaceScalarWithTensorByProfilePass,
     }
 
     def _rational_approximation(self, z, ops, meta):

backends/arm/_passes/decompose_atanh_pass.py

@@ -10,7 +10,7 @@
 from executorch.backends.arm._passes.match_arg_dtype_pass import MatchArgDtypePass
 from executorch.backends.arm._passes.match_arg_ranks_pass import MatchArgRanksPass
 from executorch.backends.arm._passes.replace_scalar_with_tensor_pass import (
-    ReplaceScalarWithTensorArgPassTOSAMI,
+    ReplaceScalarWithTensorByProfilePass,
 )
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass
@@ -43,7 +43,7 @@ class DecomposeAtanhPass(ArmPass):
         InsertTableOpsPass,
         MatchArgRanksPass,
         MatchArgDtypePass,
-        ReplaceScalarWithTensorArgPassTOSAMI,
+        ReplaceScalarWithTensorByProfilePass,
     }
 
     def call_operator(self, op, args, kwargs, meta):

backends/arm/_passes/decompose_cosh_pass.py

@@ -10,7 +10,7 @@
 from executorch.backends.arm._passes.match_arg_dtype_pass import MatchArgDtypePass
 from executorch.backends.arm._passes.match_arg_ranks_pass import MatchArgRanksPass
 from executorch.backends.arm._passes.replace_scalar_with_tensor_pass import (
-    ReplaceScalarWithTensorArgPassTOSAMI,
+    ReplaceScalarWithTensorByProfilePass,
 )
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass
@@ -31,7 +31,7 @@ class DecomposeCoshPass(ArmPass):
     _passes_required_after: Set[Type[ExportPass]] = {
         InsertTableOpsPass,
         MatchArgRanksPass,
-        ReplaceScalarWithTensorArgPassTOSAMI,
+        ReplaceScalarWithTensorByProfilePass,
         MatchArgDtypePass,
     }
 

backends/arm/_passes/decompose_expm1_pass.py

@@ -12,7 +12,7 @@
 from executorch.backends.arm._passes.match_arg_dtype_pass import MatchArgDtypePass
 from executorch.backends.arm._passes.match_arg_ranks_pass import MatchArgRanksPass
 from executorch.backends.arm._passes.replace_scalar_with_tensor_pass import (
-    ReplaceScalarWithTensorArgPassTOSAMI,
+    ReplaceScalarWithTensorByProfilePass,
 )
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass
@@ -83,7 +83,7 @@ class DecomposeExpm1Pass(ArmPass):
         ConvertIntPowToMuls,
         InsertTableOpsPass,
         DecomposeDivPass,
-        ReplaceScalarWithTensorArgPassTOSAMI,
+        ReplaceScalarWithTensorByProfilePass,
         MatchArgDtypePass,
         MatchArgRanksPass,
     }