Fix quantized max_pool2d output observer for tuple-returning ops (pytorch#19259)

Summary:

prepare_pt2e skips the output observer for max_pool2d_with_indices because its output is a tuple, not a single tensor. This caused the quantized_max_pool2d fusion to silently fail.

Fix: annotate getitem[0] (single tensor) instead of the tuple-returning op. Also bail out if indices are consumed or the graph structure is unexpected.

Reviewed By: mcremon-meta

Differential Revision: D103436172

Author: abeakkas

backends/cadence/aot/quantizer/patterns.py

@@ -6,6 +6,7 @@
 
 # pyre-strict
 
+import operator
 from abc import ABC, abstractmethod
 from dataclasses import dataclass, field
 from typing import List, Tuple, Union
@@ -493,7 +494,16 @@ def get_anchors(
         # pyre-fixme[29]: `Union[BoundMethod[typing.Callable(torch._C.TensorBase.__ge...
         max_pool_node = fused_partition[0].nodes[-1]
 
-        # Input and output share quantization parameters since max is order-preserving
+        # Since max_pool2d_with_indices returns a tuple, the output observer must be
+        # placed on getitem[0] rather than the tuple-returning op. Otherwise
+        # prepare_pt2e silently skips it.
+        # Expect exactly one user: getitem[0] extracting the values tensor. If indices
+        # are also used or the structure is unexpected, bail out.
+        users = list(max_pool_node.users)
+        if len(users) != 1 or users[0].target is not operator.getitem:
+            return PartitionAnchors(empty=True), max_pool_node
+        getitem_0 = users[0]
+
         return (
             PartitionAnchors(
                 inputs=[(max_pool_node, 0)],
@@ -505,7 +515,7 @@ def get_anchors(
                 ],
                 output=[
                     (
-                        max_pool_node,
+                        getitem_0,
                         SharedQuantizationSpec((max_pool_node.args[0], max_pool_node)),
                     )
                 ],

backends/cadence/aot/tests/test_quantizer_ops.py

@@ -7,6 +7,7 @@
 # pyre-strict
 
 import inspect
+import operator
 import unittest
 from typing import Callable
 
@@ -483,12 +484,18 @@ def _build_addmm_graph(self) -> tuple[torch.fx.GraphModule, torch.fx.Node]:
         self.assertEqual(len(addmm_nodes), 1, "Should find exactly one addmm node")
         return gm, addmm_nodes[0]
 
-    def _build_max_pool2d_graph(self) -> tuple[torch.fx.GraphModule, torch.fx.Node]:
-        """Build a simple graph with a max_pool2d_with_indices operation."""
+    def _build_max_pool2d_graph(
+        self,
+    ) -> tuple[torch.fx.GraphModule, torch.fx.Node, torch.fx.Node]:
+        """Build a graph with max_pool2d_with_indices followed by getitem[0].
+
+        Returns:
+            A tuple of (graph_module, getitem_node, max_pool_node).
+            The getitem_node is where the output annotation is placed.
+            The max_pool_node is where the input annotation is placed.
+        """
         builder = GraphBuilder()
-        # Input shape: (batch, channels, height, width)
         x = builder.placeholder("x", torch.randn(1, 3, 8, 8))
-        # max_pool2d_with_indices args: (input, kernel_size, stride, padding, dilation, ceil_mode)
         max_pool = builder.call_operator(
             op=torch.ops.aten.max_pool2d_with_indices.default,
             args=(x, [2, 2], [2, 2], [0, 0], [1, 1], False),
@@ -503,19 +510,24 @@ def _build_max_pool2d_graph(self) -> tuple[torch.fx.GraphModule, torch.fx.Node]:
                 }
             ),
         )
-        builder.output([max_pool])
+        getitem = builder.call_operator(
+            op=operator.getitem,
+            args=(max_pool, 0),
+        )
+        builder.output([getitem])
         gm = builder.get_graph_module()
 
         max_pool_nodes = gm.graph.find_nodes(
             op="call_function",
             target=torch.ops.aten.max_pool2d_with_indices.default,
         )
-        self.assertEqual(
-            len(max_pool_nodes),
-            1,
-            "Should find exactly one max_pool2d_with_indices node",
+        self.assertEqual(len(max_pool_nodes), 1)
+        getitem_nodes = gm.graph.find_nodes(
+            op="call_function",
+            target=operator.getitem,
         )
-        return gm, max_pool_nodes[0]
+        self.assertEqual(len(getitem_nodes), 1)
+        return gm, getitem_nodes[0], max_pool_nodes[0]
 
     def _build_add_relu_graph(
         self,