[ExecuTorch][Core AI][3/x] PT2E quantizer front-end

backends/apple/coreai/quantizer/__init__.py

@@ -0,0 +1,11 @@
+# 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.
+
+from executorch.backends.apple.coreai.quantizer.quantizer import CoreAIQuantizer
+
+__all__ = [
+    "CoreAIQuantizer",
+]

backends/apple/coreai/quantizer/quantizer.py

@@ -0,0 +1,87 @@
+# 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.
+
+"""Core AI quantization front-end for ExecuTorch.
+
+Keeps the canonical ExecuTorch quantization pipeline::
+
+    prepare -> calibrate -> convert -> export -> to_edge_transform_and_lower
+
+``CoreAIQuantizer.convert()`` produces the export-ready, fully quantized graph
+(Core AI ``coreai::`` ops). It intentionally runs the entire ``coreai_opt``
+``finalize(CoreAI)`` (``convert_pt2e`` + conv/bn fold + Q/DQ to ``coreai::``
+rewrite + kv-cache relocation), because that whole rewrite must run before
+``torch.export``: a graph still carrying ``coreai_opt`` fake-quant cannot be
+strict-``torch.export``ed (its ``FakeQuantize.forward`` has a data-dependent
+guard). Nothing quant-related is left for the backend ``preprocess`` to do; it
+simply lowers the ``coreai::`` ops via ``add_exported_program``.
+
+``coreai_opt`` (the ``coreai-optimization`` package) is imported lazily so this
+module stays importable in environments without it installed.
+"""
+
+from __future__ import annotations
+
+from typing import Any, Optional, Sequence
+
+from torch import fx, nn
+
+
+class CoreAIQuantizer:
+    """PT2E quantization front-end for Core AI, wrapping ``coreai_opt``.
+
+    Standard PT2E lifecycle: :meth:`prepare` (exports + inserts fake-quant),
+    :meth:`calibration_mode` / :meth:`training_mode`, then :meth:`convert`.
+
+    Unlike stock PT2E (where ``convert`` yields a backend-agnostic Q/DQ graph),
+    :meth:`convert` here returns the Core AI quantized graph (``coreai::`` ops),
+    because ``coreai_opt``'s fake-quant has no meaningful generic Q/DQ stage and
+    the full rewrite must complete before ``torch.export``.
+
+    Args:
+        model: the ``nn.Module`` to quantize.
+        config: a ``coreai_opt.quantization.QuantizerConfig``. Defaults to
+            ``coreai_opt``'s default (int8 weight + activation) when ``None``.
+    """
+
+    def __init__(self, model: nn.Module, config: Optional[Any] = None) -> None:
+        from coreai_opt.quantization import Quantizer, QuantizerConfig
+
+        self._quantizer = Quantizer(model, config or QuantizerConfig())
+        self._prepared: Optional[fx.GraphModule] = None
+
+    def prepare(
+        self,
+        example_inputs: Sequence[Any],
+        dynamic_shapes: Any = None,
+    ) -> fx.GraphModule:
+        """PT2E prepare: export + insert fake-quant. Returns the prepared graph."""
+        self._prepared = self._quantizer.prepare(
+            tuple(example_inputs), dynamic_shapes=dynamic_shapes
+        )
+        return self._prepared
+
+    def calibration_mode(self):
+        """Context manager to collect activation statistics on calibration data."""
+        return self._quantizer.calibration_mode()
+
+    def training_mode(self):
+        """Context manager for quantization-aware training."""
+        return self._quantizer.training_mode()
+
+    def convert(self) -> fx.GraphModule:
+        """Produce the export-ready, fully quantized Core AI graph.
+
+        Runs the entire ``coreai_opt`` ``finalize(CoreAI)`` (``convert_pt2e`` +
+        conv/bn fold + Q/DQ -> ``coreai::`` rewrite + kv-cache relocation).  The
+        result has no fake-quant, so it can be strict-``torch.export``ed and then
+        lowered via ``to_edge_transform_and_lower([CoreAIPartitioner()])``.
+        """
+        from coreai_opt.common import ExportBackend
+
+        if self._prepared is None:
+            raise RuntimeError("Call prepare() before convert().")
+        return self._quantizer.finalize(backend=ExportBackend.CoreAI)

backends/apple/coreai/test/test_quantizer.py

@@ -0,0 +1,183 @@
+# 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.
+
+import unittest
+from collections import Counter
+
+import torch
+import torch.nn as nn
+
+from executorch.backends.apple.coreai.partition.partitioner import CoreAIPartitioner
+from executorch.backends.apple.coreai.quantizer.quantizer import CoreAIQuantizer
+from executorch.exir import to_edge_transform_and_lower
+from executorch.exir.lowered_backend_module import (
+    executorch_call_delegate,
+    get_lowered_backend_modules,
+)
+
+
+def _coreai_quant_op_names(graph_module):
+    """Names of call_function targets in a converted graph (test helper)."""
+    return [
+        getattr(node.target, "__name__", str(node.target))
+        for node in graph_module.graph.nodes
+        if node.op == "call_function"
+    ]
+
+
+def _model():
+    return nn.Sequential(nn.Linear(32, 32), nn.ReLU(), nn.Linear(32, 32)).eval()
+
+
+class _ConvBN(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.conv = nn.Conv2d(3, 8, 3)
+        self.bn = nn.BatchNorm2d(8)
+        self.relu = nn.ReLU()
+
+    def forward(self, x):
+        return self.relu(self.bn(self.conv(x)))
+
+
+def _full_finalize(model, example_inputs):
+    """coreai_opt standalone finalize(CoreAI) for equivalence checks."""
+    from coreai_opt.common import ExportBackend
+    from coreai_opt.quantization import Quantizer, QuantizerConfig
+
+    q = Quantizer(model, QuantizerConfig())
+    prepared = q.prepare(example_inputs)
+    with q.calibration_mode():
+        prepared(*example_inputs)
+    return q.finalize(backend=ExportBackend.CoreAI)
+
+
+class CoreAIQuantizerTest(unittest.TestCase):
+    def setUp(self):
+        self.example_inputs = (torch.randn(2, 32),)
+
+    def _convert(self, model):
+        q = CoreAIQuantizer(model)
+        prepared = q.prepare(self.example_inputs)
+        with q.calibration_mode():
+            prepared(*self.example_inputs)
+        return q.convert()
+
+    def test_convert_produces_coreai_quant_ops(self):
+        names = set(_coreai_quant_op_names(self._convert(_model())))
+        self.assertIn("quantize", names)
+        self.assertIn("dequantize", names)
+        self.assertIn("constexpr_blockwise_shift_scale", names)
+
+    def test_convert_matches_full_finalize(self):
+        # convert() IS the full coreai_opt finalize(CoreAI).
+        converted = self._convert(_model())
+        full = _full_finalize(_model(), self.example_inputs)
+        self.assertEqual(
+            Counter(_coreai_quant_op_names(converted)),
+            Counter(_coreai_quant_op_names(full)),
+        )
+
+    def test_convert_output_is_exportable(self):
+        # The key property: no fake-quant remains, so strict torch.export works.
+        converted = self._convert(_model())
+        ep = torch.export.export(converted, self.example_inputs)
+        namespaces = {
+            getattr(n.target, "namespace", None)
+            for n in ep.graph.nodes
+            if n.op == "call_function"
+        }
+        # Quant ops are lowered into the coreai namespace after export.
+        self.assertIn("coreai", namespaces)
+
+    def test_prepare_is_required_before_convert(self):
+        q = CoreAIQuantizer(_model())
+        with self.assertRaises(RuntimeError):
+            q.convert()
+
+
+class CoreAIQuantizerConvBNTest(unittest.TestCase):
+    def setUp(self):
+        self.example_inputs = (torch.randn(1, 3, 16, 16),)
+
+    def _convert(self, model):
+        q = CoreAIQuantizer(model)
+        prepared = q.prepare(self.example_inputs)
+        with q.calibration_mode():
+            prepared(*self.example_inputs)
+        return q.convert()
+
+    def test_convert_folds_conv_bn(self):
+        # convert() folds conv+bn (pre-export) so no batch_norm survives.
+        names = _coreai_quant_op_names(self._convert(_ConvBN().eval()))
+        self.assertTrue(any("conv" in n for n in names), names)
+        self.assertFalse(
+            any("batch_norm" in n for n in names),
+            f"batch_norm was not folded: {names}",
+        )
+
+    def test_convert_matches_full_finalize_conv_bn(self):
+        converted = self._convert(_ConvBN().eval())
+        full = _full_finalize(_ConvBN().eval(), self.example_inputs)
+        self.assertEqual(
+            Counter(_coreai_quant_op_names(converted)),
+            Counter(_coreai_quant_op_names(full)),
+        )
+
+
+class CoreAIQuantizedLoweringTest(unittest.TestCase):
+    """Full quantized lowering: quantize -> export -> to_edge_transform_and_lower."""
+
+    def setUp(self):
+        self.example_inputs = (torch.randn(2, 32),)
+
+    def _lower(self):
+        q = CoreAIQuantizer(_model())
+        prepared = q.prepare(self.example_inputs)
+        with q.calibration_mode():
+            prepared(*self.example_inputs)
+        ep = torch.export.export(q.convert(), self.example_inputs)
+        return to_edge_transform_and_lower(ep, partitioner=[CoreAIPartitioner()])
+
+    def test_quantized_linear_lowers_to_coreai(self):
+        lowered = self._lower()
+        gm = lowered.exported_program().graph_module
+
+        delegates = [
+            n
+            for n in gm.graph.nodes
+            if n.op == "call_function" and n.target is executorch_call_delegate
+        ]
+        leftover = [
+            n
+            for n in gm.graph.nodes
+            if n.op == "call_function"
+            and n.target is not executorch_call_delegate
+            and "getitem" not in str(n.target)
+        ]
+        # The quantized subgraph (incl. coreai:: quant ops) is delegated and
+        # converted by coreai-torch, leaving nothing outside the delegate.
+        self.assertGreaterEqual(len(delegates), 1)
+        self.assertEqual(
+            leftover,
+            [],
+            f"unexpected ops left outside the delegate: "
+            f"{[str(n.target) for n in leftover]}",
+        )
+
+    def test_quantized_lowers_through_to_executorch(self):
+        lowered = self._lower()
+        # A single CoreAI delegate, with its asset embedded (NDS present).
+        lbms = get_lowered_backend_modules(lowered.exported_program().graph_module)
+        self.assertEqual([lbm.backend_id for lbm in lbms], ["CoreAIBackend"])
+        self.assertIsNotNone(lbms[0].named_data_store_output)
+        # It lowers all the way to a non-empty .pte.
+        buffer = bytes(lowered.to_executorch().buffer)
+        self.assertGreater(len(buffer), 0)
+
+
+if __name__ == "__main__":
+    unittest.main()