Enable pickling model prepared with QAT qconfig (#109288)

sindish · facebook-github-bot · commit 8834d057e033 · 2023-09-26T15:12:32.000-07:00
Summary:

Resolving error:

AttributeError: Can't pickle local object '_add_module_to_qconfig_obs_ctr.&lt;locals&gt;.get_factory_kwargs_based_on_module_device'

by moving nested function out to the main module

Test Plan: Added test to CI

Reviewed By: andrewor14

Differential Revision: D49187352
diff --git a/test/quantization/fx/test_quantize_fx.py b/test/quantization/fx/test_quantize_fx.py
@@ -1,5 +1,5 @@
 # Owner(s): ["oncall: quantization"]
-
+import pickle
 from collections import OrderedDict
 import contextlib
 import torch
@@ -9714,6 +9714,18 @@ def forward(self, x):
             out_ref = converted_ref(inp)
 
             torch.testing.assert_close(out, out_ref)
+
+    @override_qengines
+    def test_qat_pickle(self):
+        model = torch.nn.Sequential(nn.Conv2d(3, 32, 5), nn.BatchNorm2d(32), nn.ReLU())
+        example_inputs = torch.randn(1, 3, 128, 128)
+
+        qengine = torch.backends.quantized.engine
+        qconfig_dict = {'': torch.ao.quantization.get_default_qat_qconfig(qengine)}
+        model = prepare_qat_fx(model, qconfig_dict, example_inputs)
+        pickle.dumps(model)
+
+
 if __name__ == '__main__':
     raise RuntimeError("This test file is not meant to be run directly, use:\n\n"
                        "\tpython test/test_quantization.py TESTNAME\n\n"
diff --git a/torch/ao/quantization/qconfig.py b/torch/ao/quantization/qconfig.py
@@ -1,4 +1,5 @@
 from collections import namedtuple
+from functools import partial
 from typing import Optional, Any, Union, Type
 
 import torch
@@ -458,6 +459,13 @@ def _assert_valid_qconfig(qconfig: Optional[QConfig],
 QConfigAny = Optional[QConfig]
 QConfigAny.__module__ = "torch.ao.quantization.qconfig"
 
+def _get_factory_kwargs_based_on_module_device(module):
+    assert isinstance(module, torch.nn.Module)
+    devices = {p.device for p in module.parameters()} | \
+        {p.device for p in module.buffers()}
+    device = next(iter(devices)) if len(devices) > 0 else None
+    return None if device is None else {'device': device}
+
 def _add_module_to_qconfig_obs_ctr(
         qconfig: QConfigAny,
         module: Optional[nn.Module]) -> Any:
@@ -477,19 +485,14 @@ def _add_module_to_qconfig_obs_ctr(
     if module is None or qconfig is None or qconfig._fields != ('activation', 'weight'):
         return qconfig
 
-    def get_factory_kwargs_based_on_module_device():
-        assert isinstance(module, torch.nn.Module)
-        devices = {p.device for p in module.parameters()} | \
-            {p.device for p in module.buffers()}
-        device = next(iter(devices)) if len(devices) > 0 else None
-        return None if device is None else {'device': device}
 
     def configure_constructor_to_put_obs_on_module_device(original_constructor):
+        get_factory_kwargs_based_on_module_device_with_model = partial(_get_factory_kwargs_based_on_module_device, module)
         try:
             # check if constructor can accept factory_kwargs
             check = original_constructor.with_args(factory_kwargs=None)
             check()
-            return original_constructor.with_callable_args(factory_kwargs=get_factory_kwargs_based_on_module_device)
+            return original_constructor.with_callable_args(factory_kwargs=get_factory_kwargs_based_on_module_device_with_model)
         except AttributeError:  # qconfig doesn't have activation or weight
             return original_constructor
         except TypeError:  # the class doesn't accept factory_kwargs argument
