Improve accuracy for models using shuffle, unshuffle, cat ops (#19159)

abhinaykukkadapu · facebook-github-bot · commit 7fdec04c77e4 · 2026-04-29T09:36:14.000-07:00
Summary:

Replace the Qualcomm concat observer path with an explicit same-domain-or-requantize model for `aten.cat`. Preserve shared qparams for `pixel_shuffle` and `pixel_unshuffle`, extend `split_with_sizes_copy` coverage, and add regressions for mismatched `cat` branches plus value-preserving ops that must use `SharedQuantizationSpec`.

Differential Revision: D102626539
diff --git a/backends/qualcomm/_passes/__init__.py b/backends/qualcomm/_passes/__init__.py
@@ -5,6 +5,7 @@
 # LICENSE file in the root directory of this source tree.
 
 from .annotate_avg_pool1d import AnnotateAvgPool1D
+from .annotate_concat_requant import AnnotateConcatRequant
 from .annotate_quant_attrs import AnnotateQuantAttrs
 from .annotate_stack import AnnotateStack
 from .annotate_unbind import AnnotateUnbind
@@ -60,6 +61,7 @@
 
 __all__ = [
     AnnotateAvgPool1D,
+    AnnotateConcatRequant,
     AnnotateQuantAttrs,
     AnnotateStack,
     AnnotateUnbind,
diff --git a/backends/qualcomm/_passes/annotate_concat_requant.py b/backends/qualcomm/_passes/annotate_concat_requant.py
@@ -0,0 +1,97 @@
+# Copyright (c) Qualcomm Innovation Center, Inc.
+# 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.
+
+from typing import Any, Dict
+
+import torch
+from executorch.backends.qualcomm.builders.node_visitor import dq_ops, q_ops
+from executorch.backends.qualcomm.utils.constants import (
+    QCOM_DTYPE,
+    QCOM_ENCODING,
+    QCOM_QUANT_MAX,
+    QCOM_QUANT_MIN,
+    QCOM_REQUANTIZE,
+    QCOM_SCALE,
+    QCOM_ZERO_POINT,
+)
+from executorch.exir.dialects._ops import ops as exir_ops
+from executorch.exir.pass_base import ExportPass, PassResult
+
+from .utils import get_quant_attrs
+
+
+EDGE_CAT_OPS = {
+    exir_ops.edge.aten.cat.default,
+    exir_ops.edge.aten.concat.default,
+}
+
+
+class AnnotateConcatRequant(ExportPass):
+    """
+    Record explicit requantization needs for concat inputs whose concrete
+    post-calibration qparams do not match concat's output domain.
+    """
+
+    def __init__(
+        self,
+        edge_program: torch.export.ExportedProgram,
+        skip_advanced_requant: bool = False,
+    ):
+        super(AnnotateConcatRequant, self).__init__()
+        self.edge_program = edge_program
+        self.skip_advanced_requant = skip_advanced_requant
+
+    def _is_requant_needed(self, src_attrs: Dict[str, Any], dst_attrs: Dict[str, Any]):
+        if self.skip_advanced_requant:
+            return src_attrs[QCOM_DTYPE] != dst_attrs[QCOM_DTYPE]
+
+        return any(
+            src_attrs[attr] != dst_attrs[attr]
+            for attr in [
+                QCOM_SCALE,
+                QCOM_ZERO_POINT,
+                QCOM_QUANT_MIN,
+                QCOM_QUANT_MAX,
+                QCOM_DTYPE,
+            ]
+        )
+
+    def _annotate_concat_input_requant(self, quant_node: torch.fx.Node) -> None:
+        cat_node = quant_node.args[0]
+        if cat_node.target not in EDGE_CAT_OPS:
+            return
+
+        output_q_attrs = get_quant_attrs(self.edge_program, quant_node)
+        for input_node in cat_node.args[0]:
+            if input_node.target not in dq_ops:
+                continue
+
+            source_q_node = input_node.args[0]
+            if source_q_node.target not in q_ops:
+                continue
+
+            source_q_attrs = get_quant_attrs(self.edge_program, source_q_node)
+            if not self._is_requant_needed(source_q_attrs, output_q_attrs):
+                continue
+
+            source_node = source_q_node.args[0]
+            if not isinstance(source_node, torch.fx.Node):
+                continue
+
+            requant_attrs = output_q_attrs.copy()
+            requant_attrs[QCOM_ENCODING] = source_q_attrs[QCOM_ENCODING]
+            source_node.meta.setdefault(QCOM_REQUANTIZE, {})
+            source_node.meta[QCOM_REQUANTIZE][cat_node.name] = requant_attrs
+
+    def call(self, graph_module: torch.fx.GraphModule):
+        for node in graph_module.graph.nodes:
+            if (
+                node.target in q_ops
+                and isinstance(node.args[0], torch.fx.Node)
+                and node.args[0].target in EDGE_CAT_OPS
+            ):
+                self._annotate_concat_input_requant(node)
+        return PassResult(graph_module, True)
diff --git a/backends/qualcomm/_passes/annotate_quant_attrs.py b/backends/qualcomm/_passes/annotate_quant_attrs.py
@@ -79,6 +79,21 @@ def _find_last_dq_nodes(self, node: torch.fx.node.Node) -> torch.fx.node.Node:
 
         return last_dq_nodes
 
+    def _is_requant_needed(self, src_attrs: Dict[str, Any], dst_attrs: Dict[str, Any]):
+        if self.skip_advanced_requant:
+            return src_attrs[QCOM_DTYPE] != dst_attrs[QCOM_DTYPE]
+
+        return any(
+            src_attrs[attr] != dst_attrs[attr]
+            for attr in [
+                QCOM_SCALE,
+                QCOM_ZERO_POINT,
+                QCOM_QUANT_MIN,
+                QCOM_QUANT_MAX,
+                QCOM_DTYPE,
+            ]
+        )
+
     def _annotate_requant(self, n):
         # Record requant attributes:
         # node1 -> q_ui8 (n) -> dq_ui8 -> q_int32 -> dq_int32 -> node2 -> ....
@@ -96,28 +111,7 @@ def _annotate_requant(self, n):
                 # that has multiple outputs that requires quant attributes.
 
                 # Determine if requantization is needed based on configuration and attribute mismatch.
-                is_requant_needed = False
-                if self.skip_advanced_requant:
-                    # In skip_advanced_requant mode, only consider requant if dtypes differ.
-                    if q_attrs[QCOM_DTYPE] != dq_attrs[QCOM_DTYPE]:
-                        is_requant_needed = True
-                else:
-                    # In full requant mode, consider requant if any key attribute differs.
-                    # This aims to improve accuracy by adjusting scale, zero_point, etc.
-                    # Users can disable this if it causes regressions.
-                    if any(
-                        q_attrs[attr] != dq_attrs[attr]
-                        for attr in [
-                            QCOM_SCALE,
-                            QCOM_ZERO_POINT,
-                            QCOM_QUANT_MIN,
-                            QCOM_QUANT_MAX,
-                            QCOM_DTYPE,
-                        ]
-                    ):
-                        is_requant_needed = True
-
-                if is_requant_needed:
+                if self._is_requant_needed(q_attrs, dq_attrs):
                     dq_attrs[QCOM_ENCODING] = q_attrs[QCOM_ENCODING]
                     user_node = list(dq_node.users)[0]
                     n.args[0].meta.setdefault(QCOM_REQUANTIZE, {})
diff --git a/backends/qualcomm/_passes/qnn_pass_manager.py b/backends/qualcomm/_passes/qnn_pass_manager.py
@@ -10,6 +10,7 @@
 
 from executorch.backends.qualcomm._passes import (
     AnnotateAvgPool1D,
+    AnnotateConcatRequant,
     AnnotateQuantAttrs,
     AnnotateStack,
     AnnotateUnbind,
@@ -99,6 +100,7 @@ def get_capture_program_passes():
     default_passes_and_setting = [
         (AnnotateAvgPool1D, True),
         (AnnotateQuantAttrs, True),
+        (AnnotateConcatRequant, True),
         (AnnotateStack, True),
         (AnnotateUnbind, True),
         (ConvertBmmToMatmul, False),
diff --git a/backends/qualcomm/_passes/utils.py b/backends/qualcomm/_passes/utils.py
@@ -61,6 +61,7 @@ def get_passes_dependency_for_capture_program():
     """
     from executorch.backends.qualcomm._passes import (
         AnnotateAvgPool1D,
+        AnnotateConcatRequant,
         AnnotateQuantAttrs,
         AnnotateStack,
         AnnotateUnbind,
@@ -89,6 +90,7 @@ def get_passes_dependency_for_capture_program():
 
     return {
         AnnotateAvgPool1D: [RemoveRedundancy],
+        AnnotateConcatRequant: [AnnotateQuantAttrs],
         AnnotateQuantAttrs: [
             ConvertBmmToMatmul,
             RecomposePixelUnshuffle,
@@ -108,9 +110,15 @@ def get_passes_dependency_for_capture_program():
         DecomposeTrunc: [RemoveRedundancy],
         ExpandBroadcastTensorShape: [FoldQDQ],
         FixedLinearKeepDim: [FoldQDQ],
-        FoldQDQ: [AnnotateQuantAttrs, AnnotateStack, AnnotateUnbind],
+        FoldQDQ: [
+            AnnotateConcatRequant,
+            AnnotateQuantAttrs,
+            AnnotateStack,
+            AnnotateUnbind,
+        ],
         I64toI32: [RemoveRedundancy],
         LayoutTransform: [
+            AnnotateConcatRequant,
             AnnotateQuantAttrs,
             ExpandBroadcastTensorShape,
             FixedLinearKeepDim,
diff --git a/backends/qualcomm/quantizer/annotators/htp_rules.py b/backends/qualcomm/quantizer/annotators/htp_rules.py
@@ -12,7 +12,6 @@
 
 import executorch.backends.qualcomm.builders.qnn_constants as QnnConstants
 import torch
-
 from executorch.backends.qualcomm.quantizer.observers.concat_observer import (
     ConcatObserver,
 )
@@ -235,31 +234,28 @@ def annotate(node: Node, quantization_config: QuantizationConfig) -> None:
             return
 
         input_qspec_map, input_nodes = {}, node.args[0]
-        for input in input_nodes:
-            input_qspec = input.meta.get(Q_ANNOTATION_KEY, None)
+        for input_node in input_nodes:
+            assert isinstance(input_node, Node)
+            input_qspec = input_node.meta.get(Q_ANNOTATION_KEY, None)
             qspec = getattr(input_qspec, "output_qspec", None)
-            # keep shared qspec here for propagation the data range
-            # without introducing extra requantizations
+            # Preserve shared upstream qspecs, but derive concat's output domain
+            # from the merged output range to avoid clipping wider branches.
             if isinstance(qspec, SharedQuantizationSpec):
-                input_qspec_map[input] = SharedQuantizationSpec(input)
+                input_qspec_map[input_node] = SharedQuantizationSpec(input_node)
             else:
-                input_qspec_map[input] = quantization_config.input_activation
+                input_qspec_map[input_node] = quantization_config.input_activation
 
         output_qspec = QuantizationSpec(
             dtype=quantization_config.output_activation.dtype,
             qscheme=quantization_config.output_activation.qscheme,
             quant_max=quantization_config.output_activation.quant_max,
             quant_min=quantization_config.output_activation.quant_min,
             observer_or_fake_quant_ctr=ConcatObserver.with_args(
-                # we need to know the concat node in order to hack all the input observers' data range
-                # since deep copy of fake tensor (node.meta["val"]) is inhibited
-                # we could only ship grap & node name and perform postprocess inside observer currently
-                **{
-                    "node_name": node.name,
-                    "graph": node.graph,
-                }
+                node_name=node.name,
+                graph=node.graph,
             ),
         )
+
         node.meta[Q_ANNOTATION_KEY] = QuantizationAnnotation(
             input_qspec_map=input_qspec_map,
             output_qspec=output_qspec,
@@ -295,6 +291,7 @@ def annotate(node: Node, quantization_config: QuantizationConfig) -> None:
 @register_annotator(
     [
         torch.ops.aten.split_with_sizes.default,
+        torch.ops.aten.split_with_sizes_copy.default,
         torch.ops.aten.split.Tensor,
         torch.ops.aten.chunk.default,
     ],
@@ -1203,14 +1200,22 @@ def annotate(node: Node, quantization_config: QuantizationConfig) -> None:
     [torch.ops.aten.pixel_shuffle.default], QnnConstants.OpDepthToSpace.op_name
 )
 class PixelShuffle(GeneralOpDef):
-    pass
+    @staticmethod
+    def annotate(node: Node, quantization_config: QuantizationConfig) -> None:
+        annotate_in_out_obs_sharing_op(node, quantization_config)
+        if not _is_annotated([node]):
+            annotate_single_in_share_out(node, quantization_config)
 
 
 @register_annotator(
     [torch.ops.aten.pixel_unshuffle.default], QnnConstants.OpSpaceToDepth.op_name
 )
 class PixelUnshuffle(GeneralOpDef):
-    pass
+    @staticmethod
+    def annotate(node: Node, quantization_config: QuantizationConfig) -> None:
+        annotate_in_out_obs_sharing_op(node, quantization_config)
+        if not _is_annotated([node]):
+            annotate_single_in_share_out(node, quantization_config)
 
 
 @register_annotator(
diff --git a/backends/qualcomm/quantizer/annotators/lpai_rules.py b/backends/qualcomm/quantizer/annotators/lpai_rules.py
@@ -11,7 +11,6 @@
 
 import executorch.backends.qualcomm.builders.qnn_constants as QnnConstants
 import torch
-
 from executorch.backends.qualcomm.quantizer.observers.concat_observer import (
     ConcatObserver,
 )
@@ -181,31 +180,26 @@ def annotate(node: Node, quantization_config: QuantizationConfig) -> None:
             return
 
         input_qspec_map, input_nodes = {}, node.args[0]
-        for input in input_nodes:
-            input_qspec = input.meta.get(Q_ANNOTATION_KEY, None)
+        for input_node in input_nodes:
+            assert isinstance(input_node, Node)
+            input_qspec = input_node.meta.get(Q_ANNOTATION_KEY, None)
             qspec = getattr(input_qspec, "output_qspec", None)
-            # keep shared qspec here for propagation the data range
-            # without introducing extra requantizations
             if isinstance(qspec, SharedQuantizationSpec):
-                input_qspec_map[input] = SharedQuantizationSpec(input)
+                input_qspec_map[input_node] = SharedQuantizationSpec(input_node)
             else:
-                input_qspec_map[input] = quantization_config.input_activation
+                input_qspec_map[input_node] = quantization_config.input_activation
 
         output_qspec = QuantizationSpec(
             dtype=quantization_config.output_activation.dtype,
             qscheme=quantization_config.output_activation.qscheme,
             quant_max=quantization_config.output_activation.quant_max,
             quant_min=quantization_config.output_activation.quant_min,
             observer_or_fake_quant_ctr=ConcatObserver.with_args(
-                # we need to know the concat node in order to hack all the input observers' data range
-                # since deep copy of fake tensor (node.meta["val"]) is inhibited
-                # we could only ship grap & node name and perform postprocess inside observer currently
-                **{
-                    "node_name": node.name,
-                    "graph": node.graph,
-                }
+                node_name=node.name,
+                graph=node.graph,
             ),
         )
+
         node.meta[Q_ANNOTATION_KEY] = QuantizationAnnotation(
             input_qspec_map=input_qspec_map,
             output_qspec=output_qspec,
@@ -223,6 +217,7 @@ def annotate(node: Node, quantization_config: QuantizationConfig) -> None:
 @register_annotator(
     [
         torch.ops.aten.split_with_sizes.default,
+        torch.ops.aten.split_with_sizes_copy.default,
         torch.ops.aten.split.Tensor,
         torch.ops.aten.chunk.default,
     ],
@@ -705,14 +700,22 @@ def annotate(node: Node, quantization_config: QuantizationConfig) -> None:
     [torch.ops.aten.pixel_shuffle.default], QnnConstants.OpDepthToSpace.op_name
 )
 class PixelShuffle(GeneralOpDef):
-    pass
+    @staticmethod
+    def annotate(node: Node, quantization_config: QuantizationConfig) -> None:
+        annotate_in_out_obs_sharing_op(node, quantization_config)
+        if not _is_annotated([node]):
+            annotate_single_in_share_out(node, quantization_config)
 
 
 @register_annotator(
     [torch.ops.aten.pixel_unshuffle.default], QnnConstants.OpSpaceToDepth.op_name
 )
 class PixelUnshuffle(GeneralOpDef):
-    pass
+    @staticmethod
+    def annotate(node: Node, quantization_config: QuantizationConfig) -> None:
+        annotate_in_out_obs_sharing_op(node, quantization_config)
+        if not _is_annotated([node]):
+            annotate_single_in_share_out(node, quantization_config)
 
 
 @register_annotator(
diff --git a/backends/qualcomm/tests/test_passes.py b/backends/qualcomm/tests/test_passes.py
diff --git a/backends/qualcomm/tests/test_qnn_delegate.py b/backends/qualcomm/tests/test_qnn_delegate.py
diff --git a/examples/qualcomm/executor_runner/qnn_executor_runner.cpp b/examples/qualcomm/executor_runner/qnn_executor_runner.cpp
