Make Resnet work with rexportable flow (#40)

Differential Revision: D47890878

fbshipit-source-id: c98f38ab90a6ff57c4dc047adcc3bdff044ae25d

Author: tugsbayasgalan

backends/xnnpack/test/test_xnnpack_utils.py

@@ -35,6 +35,7 @@
 
 from executorch.exir.passes.spec_prop_pass import SpecPropPass
 from executorch.exir.serialize import serialize_to_flatbuffer
+from executorch.exir.tracer import _default_decomposition_table
 
 # pyre-ignore[21]: Could not find module `executorch.extension.pybindings.portable`.
 from executorch.extension.pybindings.portable import (  # @manual
@@ -312,19 +313,22 @@ def quantize_and_test_model_with_quantizer(
     ):
         module.eval()
         # program capture
-        capture_config = exir.CaptureConfig(
-            pt2_mode=True, enable_functionalization=True
+        m = torch._export.capture_pre_autograd_graph(
+            module, example_inputs, decomp_table=_default_decomposition_table()
         )
-        captured_program = exir.capture(module, example_inputs, config=capture_config)
-        m = captured_program.exported_program.graph_module
 
         quantizer = XNNPACKQuantizer()
         quantization_config = get_symmetric_quantization_config()
         quantizer.set_global(quantization_config)
         prepared = prepare_pt2e(m, quantizer)
         converted = convert_pt2e(prepared)
 
-        captured_program.exported_program.graph_module = converted
+        captured_program = exir.capture(
+            converted,
+            example_inputs,
+            config=exir.CaptureConfig(enable_aot=True, _unlift=True),
+        )
+
         edge_program = captured_program.to_edge(get_xnnpack_edge_compile_config())
         delegated_module = self.lower_module_and_test_output(
             module=edge_program,

exir/capture/TARGETS

@@ -8,7 +8,6 @@ python_library(
     deps = [
         ":capture",
         ":config",
-        ":unlift",
     ],
 )
 
@@ -19,7 +18,6 @@ python_library(
     ],
     deps = [
         ":config",
-        ":unlift",
         "//caffe2:torch",
         "//executorch/exir:error",
         "//executorch/exir:tracer",
@@ -40,13 +38,3 @@ python_library(
         "//executorch/exir/passes:lib",
     ],
 )
-
-python_library(
-    name = "unlift",
-    srcs = [
-        "_unlift.py",
-    ],
-    deps = [
-        "//caffe2:torch",
-    ],
-)

exir/capture/__init__.py

@@ -12,13 +12,11 @@
     EdgeCompileConfig,
     ExecutorchBackendConfig,
 )
-from executorch.exir.capture._unlift import unlift_exported_program_lifted_states
 
 __all__ = [
     "capture",
     "capture_multiple",
     "CaptureConfig",
     "EdgeCompileConfig",
     "ExecutorchBackendConfig",
-    "unlift_exported_program_lifted_states",
 ]

exir/capture/_capture.py

@@ -13,7 +13,6 @@
 import torch
 import torch._export
 from executorch.exir.capture._config import CaptureConfig
-from executorch.exir.capture._unlift import unlift_exported_program_lifted_states
 from executorch.exir.error import ExportError, ExportErrorType, InternalError
 from executorch.exir.program import ExirExportedProgram, MultiMethodExirExportedProgram
 from executorch.exir.tracer import (
@@ -76,7 +75,7 @@ def capture(
             ep = ep.transform(ReplaceViewOpsWithViewCopyOpsPass())
             if not config._unlift:
                 return ExirExportedProgram(ep, False)
-            graph_module = unlift_exported_program_lifted_states(ep)
+            graph_module = ep.module()
 
         elif config.enable_dynamic_shape:
             graph_module, _ = dynamo_trace(

exir/delegate.py

@@ -107,8 +107,9 @@ def call_delegate_autograd(lowered_module, *args):
             def fake_requires_grad(var):
                 if var is not None:
                     var = var.detach()
-                    var.requires_grad = True
-                return err_fn(var)
+                    if torch.is_floating_point(var) or torch.is_complex(var):
+                        var.requires_grad = True
+                return var
 
             return pytree.tree_map(fake_requires_grad, res)
 

exir/pass_base.py

@@ -134,7 +134,11 @@ def make_val(
                     x = torch.dequantize(x)
 
                 try:
-                    fake_tensor = self.fake_tensor_mode.from_tensor(x)
+                    # TODO we should allocate static shapes
+                    # for param/buffer values
+                    fake_tensor = self.fake_tensor_mode.from_tensor(
+                        x, static_shapes=isinstance(x, torch.nn.Parameter)
+                    )
                 except UnsupportedFakeTensorException:
                     # TODO: This is just a workaround to get over the
                     # x.as_subclass error