Enable quant fusion and const propagation by default (pytorch#10394)

Summary: This diff enables quant fusion and constant propagation by
default in ExecuTorch. It occurs after all to_edge passes, but before
memory planning.

Differential Revision: D73513516

Author: metascroy

exir/capture/_config.py

@@ -102,3 +102,6 @@ class ExecutorchBackendConfig:
     # serialized in the PTE file. Its value is ignored if mutable buffers are not
     # memory planned as the names must be serialized in that case.
     emit_mutable_buffer_names: bool = False
+
+    # If set to true, we run quant fusion and constant propagation passes
+    do_quant_fusion_and_const_prop: bool = False

exir/passes/TARGETS

@@ -154,6 +154,8 @@ python_library(
         "//caffe2:torch",
         "//executorch/exir:pass_base",
         "//executorch/exir/dialects:lib",
+        "//pytorch/ao:torchao",
+        "//executorch/exir/passes:constant_prop_pass",
     ],
 )
 

exir/passes/constant_prop_pass.py

@@ -6,6 +6,7 @@
 
 # pyre-unsafe
 
+import logging
 from collections import OrderedDict
 from typing import cast, Mapping, Optional
 
@@ -29,6 +30,32 @@
 # Propagating aten.full can significantly increase compiled model size.
 _DEFAULT_SKIP_TARGETS = {exir_ops.edge.aten.full.default}
 
+# Do not const prop quantization primitives
+_QDQ_OPS = [
+    exir_ops.edge.quantized_decomposed.dequantize_per_channel.default,
+    exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default,
+    exir_ops.edge.quantized_decomposed.dequantize_per_tensor.tensor,
+    exir_ops.edge.quantized_decomposed.convert_element_type.no_fuse,
+    exir_ops.edge.quantized_decomposed.quantize_per_tensor.default,
+    exir_ops.edge.quantized_decomposed.quantize_per_tensor.tensor,
+    exir_ops.edge.quantized_decomposed.quantize_per_channel.default,
+    exir_ops.edge.quantized_decomposed.choose_qparams.tensor,
+]
+try:
+    import torchao  # noqa: F401
+
+    _QDQ_OPS.extend(
+        [
+            exir_ops.edge.torchao.dequantize_affine.default,
+            exir_ops.edge.torchao.quantize_affine.default,
+            exir_ops.edge.torchao.choose_qparams_affine.default,
+        ]
+    )
+except ImportError:
+    pass
+_DEFAULT_SKIP_TARGETS.update(set(_QDQ_OPS))
+
+
 _PRIMITIVE_TYPES = (
     float,
     int,
@@ -308,7 +335,9 @@ def constant_prop_pass(
         if node.target == torch.ops.higher_order.cond
     ]
     if len(has_control_flow) > 0:
-        raise RuntimeError("constant_prop_pass for control flow is not supported yet.")
+        logging.warning(
+            "constant_prop_pass does not constant propagate in control flow modules"
+        )
 
     const_node_to_tensor = get_propagated_const_tensor_dict(
         exported_program, custom_skip_targets

exir/passes/quant_fusion_pass.py

@@ -7,6 +7,8 @@
 import torch
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass
+from executorch.exir.passes.constant_prop_pass import constant_prop_pass
+from torch.export import ExportedProgram
 from torch.fx import GraphModule, subgraph_rewriter
 from torch.fx.passes.infra.pass_base import PassResult
 from torch.utils import _pytree as pytree
@@ -139,3 +141,13 @@ def call(self, graph_module: GraphModule) -> PassResult:
         graph_module.graph.lint()
         graph_module.graph.eliminate_dead_code()
         return PassResult(graph_module, True)
+
+
+def quant_fusion_and_const_prop_pass(program: ExportedProgram) -> ExportedProgram:
+    gm = program.graph_module
+    gm_res = QuantFusionPass(_fix_node_meta_val=True)(gm)
+    gm = gm_res.graph_module
+
+    # Do const prop pass to remove packing/dtype conversion ops
+    program = constant_prop_pass(program)
+    return program

exir/program/_program.py

@@ -52,6 +52,7 @@
 from executorch.exir.passes.normalize_view_copy_base_pass import (
     NormalizeViewCopyBasePass,
 )
+from executorch.exir.passes.quant_fusion_pass import quant_fusion_and_const_prop_pass
 from executorch.exir.passes.remove_graph_asserts_pass import (
     RemoveGraphAssertsPass,
     RemoveNonCoreAtenOpGraphAssertsPass,
@@ -1524,9 +1525,15 @@ def to_executorch(
             after it has been transformed to the ExecuTorch backend.
         """
         config = config if config else ExecutorchBackendConfig()
-
         execution_programs: Dict[str, ExportedProgram] = {}
         for name, program in self._edge_programs.items():
+            if config.do_quant_fusion_and_const_prop:
+                if program.graph_signature.backward_signature is not None:
+                    raise Exception(
+                        "Cannot run do_quant_fusion_and_const_prop on a graph with a backward signature intended for on-device training."
+                        " Please set do_quant_fusion_and_const_prop to False in the ExecutorchBackendConfig."
+                    )
+                program = quant_fusion_and_const_prop_pass(program)
             program = weights_to_outputs_pass(program)
             program = unsafe_remove_auto_functionalized_pass(program)
             gm, new_signature = insert_write_back_for_buffers_pass(program)

exir/tests/test_passes.py

@@ -1279,6 +1279,7 @@ class Module(torch.nn.Module):
             def __init__(self):
                 super().__init__()
                 self.linear = torch.nn.Linear(3, 3)
+                self.w = torch.randn(3, 3)
 
             def t(self, val):
                 return val + 1
@@ -1293,8 +1294,11 @@ def false_fn(self, val):
                 return self.linear(val) - self.f(val)
 
             def forward(self, pred, x):
+                out = torch.nn.functional.linear(
+                    x, self.w.to(torch.float16).to(torch.float32)
+                )
                 return torch.ops.higher_order.cond(
-                    pred, self.true_fn, self.false_fn, [x]
+                    pred, self.true_fn, self.false_fn, [out]
                 )
 
         mod = Module()
@@ -1304,14 +1308,48 @@ def forward(self, pred, x):
             export(mod, (pred, x), strict=True),
             compile_config=exir.EdgeCompileConfig(_check_ir_validity=False),
         )
-        error_msg = r"constant_prop_pass for control flow is not supported yet."
-
-        # TODO(chenlai): enable constant prop pass for control flow
-        with self.assertRaisesRegex(
-            RuntimeError,
-            error_msg,
-        ):
-            _ = constant_prop_pass(edge.exported_program())
+        expected_out = edge.exported_program().module()(pred, x)
+
+        warn_log = (
+            "constant_prop_pass does not constant propagate in control flow modules"
+        )
+        with self.assertLogs(level="WARNING") as log:
+            program = constant_prop_pass(edge.exported_program())
+            self.assertIn(warn_log, log.output[0])
+
+        out = program.module()(pred, x)
+        self.assertTrue(torch.allclose(expected_out, out))
+
+        # dtype casts in parent module are const propagated
+        FileCheck().check(
+            "executorch_exir_dialects_edge__ops_aten_mm_default(x, _prop_tensor_constant"
+        ).run(program.graph_module.code)
+
+    def test_constant_prop_pass_quant_primitives(self) -> None:
+        class M(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.w_int = torch.ones(3, 3, dtype=torch.int8)
+                self.w_scale = 3.0
+                self.w_zero_point = 3
+
+            def forward(self, x):
+                w_dq = torch.ops.quantized_decomposed.dequantize_per_tensor.default(
+                    self.w_int, self.w_scale, self.w_zero_point, -127, 128, torch.int8
+                )
+                return torch.nn.functional.linear(x, w_dq)
+
+        mod = M()
+        x = torch.randn([3])
+        mod(x)
+        edge = to_edge(
+            export(mod, (x,), strict=True),
+            compile_config=exir.EdgeCompileConfig(_check_ir_validity=False),
+        )
+        constant_prop_pass(edge.exported_program())
+        FileCheck().check(
+            "executorch_exir_dialects_edge__ops_quantized_decomposed_dequantize_per_tensor_default"
+        ).run(edge.exported_program().graph_module.code)
 
     def test_mutable_buffers(self) -> None:
         def count_copies(gm: torch.fx.GraphModule) -> int:

exir/tests/test_quant_fusion_pass.py

@@ -6,13 +6,17 @@
 
 # pyre-strict
 
+import copy
 import unittest
 
 import torch
 from executorch import exir
 from executorch.exir import EdgeCompileConfig, to_edge
 from executorch.exir.passes.constant_prop_pass import constant_prop_pass
-from executorch.exir.passes.quant_fusion_pass import QuantFusionPass
+from executorch.exir.passes.quant_fusion_pass import (
+    quant_fusion_and_const_prop_pass,
+    QuantFusionPass,
+)
 from executorch.exir.tests.common import register_additional_test_aten_ops
 from torch.ao.quantization import (  # @manual
     float_qparams_weight_only_qconfig,
@@ -419,6 +423,7 @@ def _test_embedding_torchao(
         m = to_edge(
             export(model, example_inputs, strict=True), compile_config=compile_config
         )
+        m_copy = copy.deepcopy(m)
 
         # Before pass, we see torchao dequantize and embedding ops
         FileCheck().check_count(
@@ -468,4 +473,22 @@ def _test_embedding_torchao(
         self.assertTrue(torch.allclose(expected_outputs, actual_outputs))
 
         # Can lower to executorch
-        exec_prog = m.to_executorch()  # noqa: F841
+        exec_prog = m.to_executorch()  # noqa
+
+        # Alternative flow 2 using quant_fusion_pass on exported program
+        quant_fusion_and_const_prop_pass(m_copy.exported_program())
+        FileCheck().check_count(
+            f"executorch_exir_dialects_edge__ops_quantized_decomposed_embedding_{embedding_suffix}",
+            1,
+            exactly=True,
+        ).check_not(
+            "executorch_exir_dialects_edge__ops_quant_fusion__pack_embedding_weight_default",
+        ).run(
+            m_copy.exported_program().graph_module.code
+        )
+
+        actual_outputs2 = m_copy.exported_program().module()(*example_inputs)
+        self.assertTrue(torch.allclose(expected_outputs, actual_outputs2))
+
+        # Can lower to executorch
+        exec_prog2 = m_copy.to_executorch()  # noqa

extension/llm/export/builder.py

@@ -29,7 +29,6 @@
 
 from executorch.exir.pass_base import ExportPass
 from executorch.exir.passes import MemoryPlanningPass
-from executorch.exir.passes.quant_fusion_pass import QuantFusionPass
 from executorch.exir.passes.sym_shape_eval_pass import ConstraintBasedSymShapeEvalPass
 
 from executorch.extension.export_util.utils import export_to_edge, save_pte_program
@@ -504,20 +503,23 @@ def to_executorch(
         """
         Lower the model to executorch and get an ExecutorchProgram.
         """
-        to_executorch_passes = [
-            # If there are Linear operations left in the graph, let's execute
-            # them with the optimized op_linear rather than materializing a
-            # transpose followed by a regular op_mm.
-            ConvertToLinearPass(),
-            QuantFusionPass(),
-        ]
+        to_executorch_passes = []
         if passes:
             # pyre-fixme[6]: In call `list.extend`, for 1st positional argument,
             # expected `Iterable[Union[ConvertToLinearPass, QuantFusionPass]]` but
             # got `List[ExportPass]
             to_executorch_passes.extend(passes)
 
         assert self.edge_manager, "Need to run export_to_edge() first"
+
+        # If there are Linear operations left in the graph, let's execute
+        # them with the optimized op_linear rather than materializing a
+        # transpose followed by a regular op_mm.
+        # TODO: ConvertToLinearPass is not a sound pass and must be called before
+        # const propagation.  It requires fixing:
+        # https://github.com/pytorch/executorch/issues/10499
+        self.edge_manager.transform([ConvertToLinearPass()])
+
         self.export_program = self.edge_manager.to_executorch(
             ExecutorchBackendConfig(
                 extract_delegate_segments=True,
@@ -526,6 +528,7 @@ def to_executorch(
                 # Optional[PassResult]]]` but got `List[Union[ConvertToLinearPass,
                 # QuantFusionPass]]`.
                 passes=to_executorch_passes,
+                do_quant_fusion_and_const_prop=True,
                 memory_planning_pass=MemoryPlanningPass(alloc_graph_input=False),
                 sym_shape_eval_pass=ConstraintBasedSymShapeEvalPass(),
             )