Extract trace from prepare_and_convert and remove export_program

Differential Revision: D73440517

Pull Request resolved: pytorch/executorch#10493

Author: mcremon-meta

backends/cadence/aot/compiler.py

@@ -39,7 +39,6 @@
 from torch._inductor.decomposition import remove_decompositions
 from torch.ao.quantization.quantize_pt2e import convert_pt2e, prepare_pt2e
 
-from torch.export import export
 from torch.export.exported_program import ExportedProgram
 
 from .passes import get_cadence_passes
@@ -55,27 +54,24 @@
 # 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(
+def trace(
     model: torch.nn.Module,
     inputs: tuple[object, ...],
-    quantizer: CadenceQuantizer,
-    calibration_data: Optional[list[tuple[object, ...]]] = None,
     dump_graphs: bool = False,
-) -> torch.fx.GraphModule:
+) -> ExportedProgram:
     """
-    Prepare and convert a model using the given quantizer.
-    The quantizer must be supplied and be the same as the one used to
-    fuse the model later, if applicable. If you do not expect that behavior,
-    please use quantize_and_fuse_pt2 instead, which will instantiate a
-    default quantizer for you if needed.
-    If calibration data is provided, it will be used to calibrate the model. If
-    not, the inputs will be used for calibration instead, which is useful for
-    unit tests but should not be used for end-to-end use cases.
-    Returns a GraphModule with the converted model.
+    Trace the model with export_for_training and return an ExportedProgram.
     """
 
+    # Make the model inference mode by calling model.eval()
+    model.eval()
+
+    # Prevent mkldnn decompositions
+    torch._C._set_mkldnn_enabled(False)
+
     # Get default decompositions
     decomp_table = torch.export.default_decompositions()
+
     # Select ops to keep
     ops_to_keep = [
         torch.ops.aten.conv1d.default,
@@ -85,19 +81,46 @@ def prepare_and_convert_pt2(
         torch.ops.aten.matmul.default,
         torch.ops.aten.rms_norm.default,
     ]
+
     # Remove decompositions for the ops we want to keep
     # pyre-fixme[6]: For 1st argument expected `Dict[typing.Callable[..., typing.Any
     remove_decompositions(decomp_table, ops_to_keep)
+
     # Export with dynamo
-    model_gm = (
-        torch.export.export_for_training(model, inputs, strict=True)
-        .run_decompositions(decomp_table)
-        .module()
-    )
+    program = torch.export.export_for_training(
+        model, inputs, strict=True
+    ).run_decompositions(decomp_table)
 
     if dump_graphs:
         logging.info("Graph before quantization:")
-        logging.info(model_gm.graph.print_tabular())
+        logging.info(program.module().graph.print_tabular())
+
+    return program
+
+
+def prepare_and_convert_pt2(
+    program: ExportedProgram,
+    inputs: tuple[object, ...],
+    quantizer: CadenceQuantizer,
+    calibration_data: Optional[list[tuple[object, ...]]] = None,
+    dump_graphs: bool = False,
+) -> torch.fx.GraphModule:
+    """
+    Prepare and convert a model using the given quantizer.
+    The quantizer must be supplied and be the same as the one used to
+    fuse the model later, if applicable. If you do not expect that behavior,
+    please use quantize_and_fuse_pt2 instead, which will instantiate a
+    default quantizer for you if needed.
+    If calibration data is provided, it will be used to calibrate the model. If
+    not, the inputs will be used for calibration instead, which is useful for
+    unit tests but should not be used for end-to-end use cases.
+    Returns a GraphModule with the converted model.
+    """
+
+    # Get the graph module from the ExportedProgram
+    model_gm = program.module()
+
+    assert isinstance(model_gm, torch.fx.GraphModule)
 
     # Prepare
     prepared_model = prepare_pt2e(model_gm, quantizer)
@@ -121,10 +144,10 @@ def prepare_and_convert_pt2(
 
 
 # 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.
+# if the quantizer here is different from the quantizer used to prepare/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_pt2 instead.
 def fuse_pt2(
     converted_graph_module: torch.fx.GraphModule,
     quantizer: CadenceQuantizer,
@@ -167,9 +190,15 @@ def quantize_pt2(
     if not quantizer:
         quantizer = CadenceDefaultQuantizer()
 
+    program = trace(model, inputs, dump_graphs=dump_graphs)
+
+    if dump_graphs:
+        logging.info("Graph after trace:")
+        logging.info(program.graph.print_tabular())
+
     # Get converted graph module
     converted_gm = prepare_and_convert_pt2(
-        model, inputs, quantizer, calibration_data, dump_graphs=dump_graphs
+        program, inputs, quantizer, calibration_data, dump_graphs=dump_graphs
     )
 
     # Get fused model
@@ -184,22 +213,6 @@ def quantize_pt2(
     return program
 
 
-# Export the model and lower it to an ExportedProgram (in aten IR)
-def export_program(
-    model: torch.nn.Module,
-    inputs: tuple[object, ...],
-) -> ExportedProgram:
-    assert isinstance(model, torch.nn.Module), "model should be an nn.Module"
-
-    # Prevent mkldnn decompositions
-    torch._C._set_mkldnn_enabled(False)
-
-    # Export the model and return it.
-    expo_program = export(model, inputs, strict=True)
-
-    return expo_program
-
-
 def _lower_ep_to_edge(
     expo_program: ExportedProgram,
     dump_graphs: bool = False,
@@ -248,7 +261,7 @@ def export_to_edge(
     assert isinstance(model, torch.nn.Module), "model should be an nn.Module"
 
     # Export the model into an ExportedProgram.
-    expo_program = export_program(model, inputs)
+    expo_program = trace(model, inputs)
 
     # Lower the model to edge IR.
     edge_prog_manager = _lower_ep_to_edge(expo_program, dump_graphs, constant_methods)

backends/cadence/aot/export_example.py

@@ -18,6 +18,7 @@
     export_to_executorch_gen_etrecord,
     fuse_pt2,
     prepare_and_convert_pt2,
+    trace,
 )
 
 from executorch.backends.cadence.aot.quantizer.quantizer import CadenceDefaultQuantizer
@@ -48,8 +49,11 @@ def export_model(
     # Instantiate the quantizer
     quantizer = CadenceDefaultQuantizer()
 
+    # Trace the model
+    ep = trace(model, example_inputs)
+
     # Convert the model
-    converted_model = prepare_and_convert_pt2(model, example_inputs, quantizer)
+    converted_model = prepare_and_convert_pt2(ep, example_inputs, quantizer)
 
     # Get reference outputs from converted model
     ref_outputs = converted_model(*example_inputs)

backends/cadence/aot/tests/test_remove_ops_passes.py

@@ -16,10 +16,10 @@
 import torch.nn.functional as F
 from executorch.backends.cadence.aot import compiler
 from executorch.backends.cadence.aot.compiler import export_to_edge
+from executorch.backends.cadence.aot.fuse_ops import FuseQuantDequantToRequantizePass
 from executorch.backends.cadence.aot.graph_builder import GraphBuilder
 
 from executorch.backends.cadence.aot.pass_utils import count_node, op_counts_match
-from executorch.backends.cadence.aot.quantizer.quantizer import CadenceDefaultQuantizer
 from executorch.backends.cadence.aot.remove_ops import (
     RemoveAliasCopyOpPass,
     RemoveBranchedQuantDequant,
@@ -42,9 +42,6 @@
 from parameterized.parameterized import parameterized
 from pyre_extensions import none_throws
 
-from torch.ao.quantization.quantize_pt2e import convert_pt2e, prepare_pt2e
-
-from torch.export import export_for_training
 from torch.fx.passes.infra.pass_base import PassResult
 
 
@@ -459,44 +456,53 @@ def forward(self, x, y):
         )
 
     def test_remove_nop_quant_dequant(self):
-        class M(torch.nn.Module):
-            def __init__(self):
-                super(M, self).__init__()
-                self.linear = torch.nn.Linear(6, 12, bias=False)
+        builder = GraphBuilder()
+        x = builder.placeholder("x", torch.randn(8, 8))
+        q0 = builder.call_operator(
+            op=exir_ops.edge.cadence.quantize_per_tensor.default,
+            args=(x, 0.01662161760032177, -4, -128, 127, torch.int8),
+        )
+        dq0 = builder.call_operator(
+            op=exir_ops.edge.cadence.dequantize_per_tensor.default,
+            args=(q0, 0.01662161760032177, -4, -128, 127, torch.int8),
+        )
+        q1 = builder.call_operator(
+            op=exir_ops.edge.cadence.quantize_per_tensor.default,
+            args=(x, 0.012577153742313385, -9, -128, 127, torch.int8),
+        )
+        builder.output([dq0, q1])
+        graph_module = builder.get_graph_module()
 
-            def forward(self, x):
-                x = self.linear(x)
-                return x
+        # Expect the dq op to be removed by the pass
+        self.assertEqual(
+            count_node(
+                graph_module, exir_ops.edge.cadence.dequantize_per_tensor.default
+            ),
+            1,
+        )
 
-        inp = torch.randn(2, 8, 1, 6)
+        # Expect 1 quantize op left since it has no matching dequant
+        self.assertEqual(
+            count_node(graph_module, exir_ops.edge.cadence.quantize_per_tensor.default),
+            2,
+        )
 
-        # Run the standard quant/convert steps, but without fusing
-        # this leaves two redundant quant/dequant pairs to test with
-        quantizer = CadenceDefaultQuantizer()
-        model_exp = export_for_training(M(), (inp,), strict=True).module()
-        prepared_model = prepare_pt2e(model_exp, quantizer)
-        prepared_model(inp)
-        converted_model = convert_pt2e(prepared_model)
+        p = FuseQuantDequantToRequantizePass()
 
-        graph_module = (
-            compiler.export_to_cadence(
-                converted_model,
-                (inp,),
-            )
-            .exported_program()
-            .graph_module
-        )
+        graph_after_passes = cast(PassResult, p(graph_module)).graph_module
 
-        # Expect all quantize ops to be removed by the pass
+        # Expect the dq op to be removed by the pass
         self.assertEqual(
-            count_node(graph_module, exir_ops.edge.cadence.quantize_per_tensor.default),
+            count_node(
+                graph_after_passes, exir_ops.edge.cadence.dequantize_per_tensor.default
+            ),
             0,
         )
 
-        # Expect 1 dequantize op for the weights
+        # Expect 1 quantize op left since it has no matching dequant
         self.assertEqual(
             count_node(
-                graph_module, exir_ops.edge.cadence.dequantize_per_tensor.default
+                graph_after_passes, exir_ops.edge.cadence.quantize_per_tensor.default
             ),
             1,
         )