Update

[ghstack-poisoned]

Author: swolchok

.ci/scripts/check_c10_sync.sh

@@ -5,7 +5,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-set -exu
+set -eux
 ls pytorch/.git || git clone https://github.com/pytorch/pytorch.git
 pytorch_pin="$(< .ci/docker/ci_commit_pins/pytorch.txt)"
 pushd pytorch

.github/workflows/check-c10-sync.yml

@@ -5,7 +5,7 @@ on:
     paths:
       - .ci/docker/ci_commit_pins/pytorch.txt
       - .ci/scripts/compare_dirs.sh
-      - .ci/scripts/diff_c10_mirror_with_pytorch.sh
+      - runtime/core/portable_type/c10/**
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}

backends/cadence/aot/compiler.py

@@ -47,6 +47,14 @@
 from .utils import print_ops_info
 
 
+default_quantizer = CadenceDefaultQuantizer()
+
+
+# Note: this is not meant as a primary API since it can create inconsistencies
+# if the quantizer here is different from the quantizer used to convert. It is
+# however useful for unit tests to separate the converted model from the fused
+# model, to be able to get reference numerics.
+# If this does not apply, please use quantize_and_fuse_pt2 instead.
 def prepare_and_convert_pt2(
     model: torch.nn.Module,
     inputs: tuple[object, ...],
@@ -245,6 +253,28 @@ def export_to_edge(
     return edge_prog_manager
 
 
+def quantize_and_export_to_edge(
+    model: torch.nn.Module,
+    inputs: tuple[object, ...],
+    quantizer: Optional[CadenceQuantizer] = None,
+    dump_graphs: bool = False,
+    constant_methods: Optional[dict[str, object]] = None,
+) -> EdgeProgramManager:
+    quantized_model = quantize_pt2(
+        model,
+        inputs,
+        quantizer=quantizer,
+        dump_graphs=dump_graphs,
+    )
+
+    return export_to_edge(
+        quantized_model,
+        inputs,
+        dump_graphs=dump_graphs,
+        constant_methods=constant_methods,
+    )
+
+
 def export_to_cadence(
     model: torch.nn.Module,
     inputs: tuple[object, ...],

backends/cadence/aot/tests/test_fusion_ops_passes.py

@@ -12,7 +12,10 @@
 import executorch.backends.cadence.aot.ops_registrations  # noqa
 import torch
 from executorch.backends.cadence.aot import compiler
-from executorch.backends.cadence.aot.compiler import export_to_edge, quantize_pt2
+from executorch.backends.cadence.aot.compiler import (
+    export_to_edge,
+    quantize_and_export_to_edge,
+)
 from executorch.backends.cadence.aot.fuse_ops import (
     FuseFullThenReshapePass,
     FuseMulIntoDequantPass,
@@ -414,9 +417,10 @@ def forward(self, x):
 
         inputs = torch.randn(2, 12, 1, 6)
         model = M()
-        quantized_model = quantize_pt2(model, (inputs,))
         graph_module = (
-            export_to_edge(quantized_model, (inputs,)).exported_program().graph_module
+            quantize_and_export_to_edge(model, (inputs,))
+            .exported_program()
+            .graph_module
         )
         graph_module = FuseQuantDequantToRequantizePass()(graph_module).graph_module
         self.check_op_counts(

backends/cadence/aot/tests/test_replace_ops_passes.py

@@ -13,7 +13,10 @@
 import torch
 import torch.nn.functional as F
 from executorch.backends.cadence.aot import compiler
-from executorch.backends.cadence.aot.compiler import export_to_edge, quantize_pt2
+from executorch.backends.cadence.aot.compiler import (
+    export_to_edge,
+    quantize_and_export_to_edge,
+)
 from executorch.backends.cadence.aot.graph_builder import (
     GraphBuilder,
     single_op_builder,
@@ -851,9 +854,8 @@ def test_replace_single_element_tensor_arguments_from_full_op_with_scalar(
 
         inputs = (x,)
         model = torch.nn.Linear(in_features=in_features, out_features=out_features)
-        quantized_model = quantize_pt2(model, inputs)
 
-        exported_program = export_to_edge(quantized_model, inputs).exported_program()
+        exported_program = quantize_and_export_to_edge(model, inputs).exported_program()
 
         # By default, the quantized linear op should have constant scalar attributes.
         self.assertTargetCountsEqual(
@@ -898,9 +900,8 @@ def test_replace_single_element_tensor_arguments_from_full_op_with_scalar_tuple_
 
         inputs = (x,)
         model = torch.nn.Linear(in_features=in_features, out_features=out_features)
-        quantized_model = quantize_pt2(model, inputs)
 
-        exported_program = export_to_edge(quantized_model, inputs).exported_program()
+        exported_program = quantize_and_export_to_edge(model, inputs).exported_program()
 
         # By default, the quantized linear op should have constant scalar attributes.
         self.assertTargetCountsEqual(