fix_qat_quantization

python/paddle/quantization/imperative/qat.py

@@ -294,9 +294,10 @@ def quantize(self, model):
         return model
 
     def save_quantized_model(self, layer, path, input_spec=None, **config):
-        self._quantize_outputs.save_quantized_model(
-            layer, path, input_spec, **config
-        )
+        with paddle.pir_utils.OldIrGuard():
+            self._quantize_outputs.save_quantized_model(
+                layer, path, input_spec, **config
+            )
 
 
 class ImperativeQuantizeInputs:

test/deprecated/quantization/CMakeLists.txt

@@ -188,15 +188,15 @@ endfunction()
 
 if(WIN32)
   list(REMOVE_ITEM TEST_OPS test_light_nas)
-  list(REMOVE_ITEM TEST_OPS test_post_training_quantization_while)
+  list(REMOVE_ITEM TEST_OPS test_post_training_quantization_while_deprecated)
   list(REMOVE_ITEM TEST_OPS test_weight_quantization_mobilenetv1)
   list(REMOVE_ITEM TEST_OPS test_imperative_qat_amp)
   list(REMOVE_ITEM TEST_OPS test_weight_only_linear)
   list(REMOVE_ITEM TEST_OPS test_llm_int8_linear)
   list(REMOVE_ITEM TEST_OPS test_quant_aware_deprecated)
   list(REMOVE_ITEM TEST_OPS test_quant_post_quant_aware)
   list(REMOVE_ITEM TEST_OPS test_quant_aware_user_defined)
-  list(REMOVE_ITEM TEST_OPS test_quant_amp)
+  list(REMOVE_ITEM TEST_OPS test_quant_amp_deprecated)
   list(REMOVE_ITEM TEST_OPS test_apply_per_channel_scale)
 
 endif()
@@ -237,13 +237,8 @@ list(REMOVE_ITEM TEST_OPS test_filter_pruning)
 if(WIN32)
   set(SINGLE_CARD_TEST_OPS
       test_user_defined_quantization_deprecated
-      test_quantization_scale_pass_deprecated
-      test_quantization_pass_deprecated
-      test_moving_average_abs_max_scale_op_deprecated
-      test_imperative_qat_channelwise
-      test_imperative_qat
-      test_imperative_out_scale
-      test_graph_deprecated)
+      test_quantization_scale_pass_deprecated test_quantization_pass_deprecated
+      test_moving_average_abs_max_scale_op_deprecated test_graph_deprecated)
   list(REMOVE_ITEM TEST_OPS ${SINGLE_CARD_TEST_OPS})
   foreach(src ${SINGLE_CARD_TEST_OPS})
     py_test(${src} SRCS ${src}.py ENVS CUDA_VISIBLE_DEVICES=0)
@@ -256,29 +251,24 @@ endforeach()
 
 # setting timeout value for old unittests
 if(NOT WIN32)
-  set_tests_properties(test_post_training_quantization_while PROPERTIES TIMEOUT
-                                                                        120)
+  set_tests_properties(test_post_training_quantization_while_deprecated
+                       PROPERTIES TIMEOUT 120)
   set_tests_properties(test_weight_quantization_mobilenetv1 PROPERTIES TIMEOUT
                                                                        120)
   set_tests_properties(test_quant_aware_deprecated PROPERTIES TIMEOUT 200)
   set_tests_properties(test_quant_post_quant_aware PROPERTIES TIMEOUT 200)
   set_tests_properties(test_quant_aware_user_defined PROPERTIES TIMEOUT 200)
-  set_tests_properties(test_quant_amp PROPERTIES TIMEOUT 200)
+  set_tests_properties(test_quant_amp_deprecated PROPERTIES TIMEOUT 200)
 endif()
 
 set_tests_properties(test_graph_deprecated PROPERTIES TIMEOUT 120)
 set_tests_properties(test_quantization_pass_deprecated PROPERTIES TIMEOUT 120)
-set_tests_properties(test_imperative_qat_channelwise PROPERTIES TIMEOUT 200)
 set_tests_properties(test_user_defined_quantization_deprecated
                      PROPERTIES TIMEOUT 200)
-set_tests_properties(test_imperative_qat PROPERTIES TIMEOUT 200)
-set_tests_properties(test_imperative_qat_fuse PROPERTIES TIMEOUT 200)
-set_tests_properties(test_imperative_out_scale PROPERTIES TIMEOUT 200)
 
 if(APPLE)
-  set_tests_properties(test_post_training_quantization_while PROPERTIES TIMEOUT
-                                                                        300)
-  set_tests_properties(test_imperative_skip_op PROPERTIES TIMEOUT 300)
+  set_tests_properties(test_post_training_quantization_while_deprecated
+                       PROPERTIES TIMEOUT 300)
 endif()
 
 set_tests_properties(test_quantization_scale_pass_deprecated PROPERTIES TIMEOUT

test/quantization/CMakeLists.txt

@@ -462,8 +462,8 @@ list(REMOVE_ITEM TEST_OPS test_filter_pruning)
 if(WIN32)
   set(SINGLE_CARD_TEST_OPS
       test_imperative_qat_channelwise test_imperative_qat
-      test_imperative_qat_lsq test_imperative_qat_matmul
-      test_imperative_out_scale)
+      test_imperative_qat_fuse test_imperative_qat_lsq
+      test_imperative_qat_matmul test_imperative_out_scale)
   list(REMOVE_ITEM TEST_OPS ${SINGLE_CARD_TEST_OPS})
   foreach(src ${SINGLE_CARD_TEST_OPS})
     py_test(${src} SRCS ${src}.py ENVS CUDA_VISIBLE_DEVICES=0)
@@ -493,7 +493,11 @@ endif()
 set_tests_properties(test_imperative_qat_user_defined PROPERTIES TIMEOUT 200)
 set_tests_properties(test_imperative_qat_lsq PROPERTIES TIMEOUT 300)
 set_tests_properties(test_imperative_qat_matmul PROPERTIES TIMEOUT 300)
-
+set_tests_properties(test_imperative_qat PROPERTIES TIMEOUT 200)
+set_tests_properties(test_imperative_qat_fuse PROPERTIES TIMEOUT 200)
+set_tests_properties(test_imperative_qat_channelwise PROPERTIES TIMEOUT 200)
+set_tests_properties(test_imperative_out_scale PROPERTIES TIMEOUT 200)
+set_tests_properties(test_imperative_skip_op PROPERTIES TIMEOUT 300)
 if(LINUX AND WITH_ONEDNN)
   set_tests_properties(test_quant2_int8_mobilenetv1_mkldnn PROPERTIES TIMEOUT
                                                                       120)
@@ -517,4 +521,5 @@ if(APPLE)
   set_tests_properties(test_post_training_quantization_mnist PROPERTIES TIMEOUT
                                                                         300)
   set_tests_properties(test_imperative_ptq PROPERTIES TIMEOUT 300)
+  set_tests_properties(test_imperative_skip_op PROPERTIES TIMEOUT 300)
 endif()

test/deprecated/quantization/test_imperative_qat.py → test/quantization/test_imperative_qat.py

@@ -214,21 +214,22 @@ def test_qat(self):
             else:
                 place = core.CPUPlace()
             exe = paddle.static.Executor(place)
-            [
-                inference_program,
-                feed_target_names,
-                fetch_targets,
-            ] = paddle.static.load_inference_model(
-                tmpdir,
-                executor=exe,
-                model_filename="lenet" + INFER_MODEL_SUFFIX,
-                params_filename="lenet" + INFER_PARAMS_SUFFIX,
-            )
-            (quant_out,) = exe.run(
-                inference_program,
-                feed={feed_target_names[0]: test_data},
-                fetch_list=fetch_targets,
-            )
+            with paddle.pir_utils.OldIrGuard():
+                [
+                    inference_program,
+                    feed_target_names,
+                    fetch_targets,
+                ] = paddle.static.load_inference_model(
+                    tmpdir,
+                    executor=exe,
+                    model_filename="lenet" + INFER_MODEL_SUFFIX,
+                    params_filename="lenet" + INFER_PARAMS_SUFFIX,
+                )
+                (quant_out,) = exe.run(
+                    inference_program,
+                    feed={feed_target_names[0]: test_data},
+                    fetch_list=fetch_targets,
+                )
             paddle.disable_static()
             quant_out = paddle.to_tensor(quant_out)
             quant_acc = paddle.metric.accuracy(quant_out, label).numpy()

test/quantization/test_imperative_qat_amp.py

@@ -225,7 +225,10 @@ def test_ptq(self):
         input_spec = [
             paddle.static.InputSpec(shape=[None, 1, 28, 28], dtype='float32')
         ]
-        paddle.jit.save(layer=model, path=self.save_path, input_spec=input_spec)
+        with paddle.pir_utils.OldIrGuard():
+            paddle.jit.save(
+                layer=model, path=self.save_path, input_spec=input_spec
+            )
         print(f'Quantized model saved in {{{self.save_path}}}')
 
         end_time = time.time()

test/deprecated/quantization/test_imperative_skip_op.py → test/quantization/test_imperative_skip_op.py

@@ -81,17 +81,17 @@ def test_out_scale_acc(self):
         else:
             place = core.CPUPlace()
         exe = paddle.static.Executor(place)
-
-        [
-            inference_program,
-            feed_target_names,
-            fetch_targets,
-        ] = paddle.static.load_inference_model(
-            save_dir,
-            executor=exe,
-            model_filename="lenet" + INFER_MODEL_SUFFIX,
-            params_filename="lenet" + INFER_PARAMS_SUFFIX,
-        )
+        with paddle.pir_utils.OldIrGuard():
+            [
+                inference_program,
+                feed_target_names,
+                fetch_targets,
+            ] = paddle.static.load_inference_model(
+                save_dir,
+                executor=exe,
+                model_filename="lenet" + INFER_MODEL_SUFFIX,
+                params_filename="lenet" + INFER_PARAMS_SUFFIX,
+            )
         model_ops = inference_program.global_block().ops
 
         conv2d_count, matmul_count = 0, 0