Move XNNPACKQuantizer from PyTorch to ExecuTorch (#7804)

Summary:
Pull Request resolved: #7804

X-link: pytorch/pytorch#144940

This migrates XNNPACKQuantizer from PyTorch to ExecuTorch.

Rationale:
Main motivation is to avoid pytorch pin update in OSS after updating XNNPACKQuantizer, which can be rather frequent.

Other impact and considerations:
- PT2e flow (which lives in PyTorch) relies havily on XNNPACKQuantizer for a "example" implementation for quantizer and more importantly tests. Fow now, we will keep the torch.ao.quantization.xnnpack_quantizer as is but mark is as not BC, and deprecated to discourace future new dependencies on it.

- Other OSS repository using XNNPACKQuantizer from PyTorch now have to take an additional dependency on ExecuTorch.

Reviewed By: mcr229

Differential Revision: D68191752

backends/cadence/aot/TARGETS

@@ -65,6 +65,7 @@ python_library(
         "//executorch/backends/cadence/aot/quantizer:fusion_pass",
         "//executorch/backends/cadence/runtime:runtime",
         "//executorch/backends/cadence/aot/quantizer:quantizer",
+        "//executorch/backends/xnnpack/quantizer:xnnpack_quantizer",
         "//executorch/backends/transforms:decompose_sdpa",
         "//executorch/backends/transforms:remove_clone_ops",
         "//executorch/exir:lib",

backends/cadence/aot/export_example.py

@@ -23,13 +23,13 @@
 from executorch.backends.cadence.aot.quantizer.quantizer import CadenceDefaultQuantizer
 from executorch.backends.cadence.runtime import runtime
 from executorch.backends.cadence.runtime.executor import BundledProgramManager
-from executorch.exir import ExecutorchProgramManager
-from torch import nn
-from torch.ao.quantization.observer import HistogramObserver, MinMaxObserver
-from torch.ao.quantization.quantizer.xnnpack_quantizer_utils import (
+from executorch.backends.xnnpack.quantizer.xnnpack_quantizer_utils import (
     QuantizationConfig,
     QuantizationSpec,
 )
+from executorch.exir import ExecutorchProgramManager
+from torch import nn
+from torch.ao.quantization.observer import HistogramObserver, MinMaxObserver
 
 from .utils import save_bpte_program, save_pte_program
 

backends/cadence/aot/quantizer/TARGETS

@@ -34,6 +34,7 @@ python_library(
         ":patterns",
         ":utils",
         "//caffe2:torch",
+        "//executorch/backends/xnnpack/quantizer:xnnpack_quantizer_utils",
     ],
 )
 

backends/cadence/aot/quantizer/quantizer.py

@@ -26,18 +26,18 @@
     is_annotated,
     no_outside_users,
 )
+from executorch.backends.xnnpack.quantizer.xnnpack_quantizer_utils import (
+    OperatorConfig,
+    QuantizationAnnotation,
+    QuantizationConfig,
+    QuantizationSpec,
+)
 
 from torch import fx
 
 from torch.ao.quantization.observer import HistogramObserver, MinMaxObserver
 from torch.ao.quantization.quantizer import DerivedQuantizationSpec, Quantizer
 from torch.ao.quantization.quantizer.composable_quantizer import ComposableQuantizer
-from torch.ao.quantization.quantizer.xnnpack_quantizer_utils import (
-    OperatorConfig,
-    QuantizationAnnotation,
-    QuantizationConfig,
-    QuantizationSpec,
-)
 
 
 act_qspec = QuantizationSpec(

backends/example/TARGETS

@@ -11,6 +11,7 @@ python_library(
     deps = [
         "//caffe2:torch",
         "//executorch/backends/example/example_operators:example_operators_lib",
+        "//executorch/backends/xnnpack/quantizer:xnnpack_quantizer",
     ],
 )
 

backends/example/example_quantizer.py

@@ -9,11 +9,11 @@
 
 import torch
 from executorch.backends.example.example_operators.ops import module_to_annotator
+from executorch.backends.xnnpack.quantizer.xnnpack_quantizer_utils import OperatorConfig
 from torch import fx
 from torch.ao.quantization.observer import HistogramObserver, MinMaxObserver
 from torch.ao.quantization.pt2e.graph_utils import find_sequential_partitions
 from torch.ao.quantization.quantizer import QuantizationSpec, Quantizer
-from torch.ao.quantization.quantizer.xnnpack_quantizer_utils import OperatorConfig
 
 
 def get_uint8_tensor_spec(observer_or_fake_quant_ctr):

backends/transforms/targets.bzl

@@ -195,6 +195,7 @@ def define_common_targets():
         deps = [
             "fbsource//third-party/pypi/expecttest:expecttest",  # @manual
             ":duplicate_dynamic_quant_chain",
+            "//executorch/backends/xnnpack/quantizer:xnnpack_quantizer",
             "//caffe2:torch",
             "//executorch/exir:lib",
         ],

backends/transforms/test/test_duplicate_dynamic_quant_chain.py

@@ -11,11 +11,11 @@
 from executorch.backends.transforms.duplicate_dynamic_quant_chain import (
     DuplicateDynamicQuantChainPass,
 )
-from torch.ao.quantization.quantize_pt2e import convert_pt2e, prepare_pt2e
-from torch.ao.quantization.quantizer.xnnpack_quantizer import (
+from executorch.backends.xnnpack.quantizer.xnnpack_quantizer import (
     get_symmetric_quantization_config,
     XNNPACKQuantizer,
 )
+from torch.ao.quantization.quantize_pt2e import convert_pt2e, prepare_pt2e
 
 # TODO: Move away from using torch's internal testing utils
 from torch.testing._internal.common_quantization import (

backends/vulkan/quantizer/TARGETS

@@ -9,5 +9,6 @@ python_library(
     ],
     deps = [
         "//caffe2:torch",
+        "//executorch/backends/xnnpack/quantizer:xnnpack_quantizer_utils",
     ],
 )

backends/vulkan/quantizer/vulkan_quantizer.py

@@ -12,15 +12,15 @@
 from typing import Any, Callable, Dict, Optional
 
 import torch
-from torch.ao.quantization.observer import MinMaxObserver, PerChannelMinMaxObserver
-from torch.ao.quantization.qconfig import _ObserverOrFakeQuantizeConstructor
-from torch.ao.quantization.quantizer import QuantizationSpec, Quantizer
-from torch.ao.quantization.quantizer.xnnpack_quantizer_utils import (
+from executorch.backends.xnnpack.quantizer.xnnpack_quantizer_utils import (
     _convert_scalars_to_attrs,
     OP_TO_ANNOTATOR,
     propagate_annotation,
     QuantizationConfig,
 )
+from torch.ao.quantization.observer import MinMaxObserver, PerChannelMinMaxObserver
+from torch.ao.quantization.qconfig import _ObserverOrFakeQuantizeConstructor
+from torch.ao.quantization.quantizer import QuantizationSpec, Quantizer
 from torch.fx import Node
 
 

backends/xnnpack/quantizer/TARGETS

@@ -0,0 +1,20 @@
+load("@fbcode_macros//build_defs:python_library.bzl", "python_library")
+
+python_library(
+    name = "xnnpack_quantizer",
+    srcs = ["xnnpack_quantizer.py"],
+    deps = [
+        ":xnnpack_quantizer_utils",
+        "//caffe2:torch",
+        "//executorch/exir:lib",
+    ],
+)
+
+python_library(
+    name = "xnnpack_quantizer_utils",
+    srcs = ["xnnpack_quantizer_utils.py"],
+    deps = [
+        "//caffe2:torch",
+        "//executorch/exir:lib",
+    ],
+)