Move SDPA decomp pass from Qualcomm's directory to be shareable and c…

…all it for Cadence backends (#4258)

Summary:
Pull Request resolved: #4258

Moving the pass to backends/transforms so that other backends can call it, and call it from the Cadence side so that we can quantize the bmm ops in SDPA.

Reviewed By: cccclai

Differential Revision: D59600486

backends/cadence/aot/TARGETS

@@ -34,6 +34,7 @@ python_library(
         "//caffe2:torch",
         "//executorch/backends/cadence/aot/quantizer:fusion_pass",
         "//executorch/backends/cadence/aot/quantizer:quantizer",
+        "//executorch/backends/transforms:decompose_sdpa",
         "//executorch/exir:lib",
     ],
 )

backends/cadence/aot/compiler.py

@@ -23,6 +23,9 @@
     CadenceQuantizer,
 )
 from executorch.backends.cadence.aot.utils import model_is_quantized
+from executorch.backends.transforms.decompose_sdpa import (
+    DecomposeScaledDotProductAttention,
+)
 from executorch.exir import EdgeCompileConfig, EdgeProgramManager, to_edge
 from pyre_extensions import assert_is_instance
 from torch._export import capture_pre_autograd_graph
@@ -47,6 +50,9 @@ def quantize_pt2(
     # Export with dynamo
     model_exp = capture_pre_autograd_graph(model, inputs)
 
+    # Decompose SDPA
+    DecomposeScaledDotProductAttention(False)(model_exp)
+
     # Prepare
     prepared_model = prepare_pt2e(model_exp, quantizer)
 

backends/qualcomm/quantizer/quantizer.py

@@ -7,16 +7,16 @@
 from typing import Callable, Dict, Optional, Sequence, Set
 
 import torch
-from executorch.backends.qualcomm.passes.decompose_scaled_dot_product_attention import (
-    DecomposeScaledDotProductAttention,
-)
 from executorch.backends.qualcomm.passes.decompose_silu import DecomposeSilu
 from executorch.backends.qualcomm.passes.recompose_pixel_unshuffle import (
     RecomposePixelUnshuffle,
 )
 from executorch.backends.qualcomm.passes.reduce_dynamic_range import ReduceDynamicRange
 from executorch.backends.qualcomm.passes.remove_redundancy import RemoveRedundancy
 from executorch.backends.qualcomm.passes.replace_inf_buffer import ReplaceInfBuffer
+from executorch.backends.transforms.decompose_sdpa import (
+    DecomposeScaledDotProductAttention,
+)
 
 from torch._ops import OpOverload
 from torch.ao.quantization.quantizer import Quantizer

backends/transforms/TARGETS

@@ -29,6 +29,19 @@ runtime.python_library(
     ],
 )
 
+runtime.python_library(
+    name = "decompose_sdpa",
+    srcs = ["decompose_sdpa.py"],
+    visibility = [
+        "//executorch/backends/...",
+        "@EXECUTORCH_CLIENTS",
+    ],
+    deps = [
+        "//caffe2:torch",
+        "//executorch/exir:pass_base",
+    ],
+)
+
 runtime.python_library(
     name = "fuse_batch_norm_with_conv",
     srcs = ["fuse_batch_norm_with_conv.py"],

backends/qualcomm/passes/decompose_scaled_dot_product_attention.py → backends/transforms/decompose_sdpa.py

@@ -1,8 +1,11 @@
-# Copyright (c) Qualcomm Innovation Center, Inc.
-# All rights reserved
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
+
+# pyre-strict
+
 import torch
 from executorch.exir.pass_base import ExportPass, PassResult
 from torch._decomp import get_decompositions
@@ -14,7 +17,15 @@ class DecomposeScaledDotProductAttention(ExportPass):
     Decompose from scaled_dot_product_attention to multiple nodes.
     """
 
-    def call(self, graph_module: torch.fx.GraphModule):
+    def __init__(self, allow_non_fake_inputs: bool = True) -> None:
+        super().__init__()
+        # With allow_non_fake_inputs=False, we don't get _unsafe_view ops
+        # in the graph, we allow disabling it here.
+        self._allow_non_fake_inputs = allow_non_fake_inputs
+
+    def call(
+        self, graph_module: torch.fx.GraphModule, allow_non_fake_inputs: bool = True
+    ) -> PassResult:
         graph = graph_module.graph
         for node in graph.nodes:
             if node.target == torch.ops.aten.scaled_dot_product_attention.default:
@@ -29,7 +40,7 @@ def call(self, graph_module: torch.fx.GraphModule):
                         ]
                     ),
                     tracing_mode="fake",
-                    _allow_non_fake_inputs=True,
+                    _allow_non_fake_inputs=allow_non_fake_inputs,
                 )(*input_tensors)
                 with graph.inserting_before(node):
                     name_to_input_tensor_map = {}