[Inference] Fix CI Tests (#68887)

vivienfanghuagood · web-flow · commit 9825f4703df4 · 2024-10-28T23:20:28.000-07:00
* fix_pir_ci_test

* fix_pir_ci_test

* update

* add

* add

* update

* update

* update
diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -386,10 +386,6 @@ AnalysisPredictor::AnalysisPredictor(const AnalysisConfig &config)
   if (config_.shape_range_info_collected()) {
     config_.SwitchIrOptim(false);
   }
-  if (FLAGS_enable_pir_api) {
-    config_.EnableNewExecutor(true);
-    config_.EnableNewIR(true);
-  }
   if (config_.new_executor_enabled()) {
     config_.EnableMemoryOptim(false);
     if (config_.new_ir_enabled()) {
@@ -447,13 +443,8 @@ bool AnalysisPredictor::Init(
   load_pir_model_ =
       model_path.substr(model_path.find_last_of(".") + 1) == "json";
   if (load_pir_model_) {
-    PADDLE_ENFORCE_EQ(
-        FLAGS_enable_pir_api || (config_.use_pir_ && config_.use_new_executor_),
-        true,
-        common::errors::InvalidArgument(
-            "Models with a .json suffix can only run in PIR mode. Please set "
-            "export FLAGS_enable_pir_api=True or "
-            "config.EnableNewExecutor(true)) and config.EnableNewIR(true)"));
+    config_.use_pir_ = true;
+    config_.use_new_executor_ = true;
   }
 
   // Use Optimized model to inference
diff --git a/python/paddle/incubate/nn/functional/block_multihead_attention.py b/python/paddle/incubate/nn/functional/block_multihead_attention.py
@@ -21,7 +21,6 @@
 from paddle import _C_ops
 from paddle.framework import (
     LayerHelper,
-    in_dynamic_mode,
     in_dynamic_or_pir_mode,
 )
 
@@ -451,7 +450,7 @@ def block_multihead_attention_xpu(
     compute_dtype: str = "default",
     rope_theta: float = 10000.0,
 ) -> tuple[Tensor, Tensor, Tensor, Tensor]:
-    if in_dynamic_mode():
+    if in_dynamic_or_pir_mode():
         return _C_ops.block_multihead_attention_xpu(
             qkv,
             key_cache,
diff --git a/test/cpp/inference/api/CMakeLists.txt b/test/cpp/inference/api/CMakeLists.txt
@@ -746,6 +746,17 @@ if(WITH_TESTING AND WITH_INFERENCE_API_TEST)
         "conv_bn_swish_split_gelu.tar.gz" 2a5e8791e47b221b4f782151d76da9c6)
     endif()
 
+    inference_analysis_test(
+      trt_dynamic_shape_test
+      SRCS
+      trt_dynamic_shape_test.cc
+      EXTRA_DEPS
+      common
+      paddle_inference_shared
+      ARGS
+      --infer_model=${TRT_MODEL_INSTALL_DIR})
+    set_tests_properties(trt_dynamic_shape_test PROPERTIES TIMEOUT 300)
+
     set(TEST_TRT_ERNIE_MODEL "${TRT_MODEL_INSTALL_DIR}/ernie_test")
     if(NOT EXISTS ${TEST_TRT_ERNIE_MODEL}/ernie_model_4.tar.gz)
       inference_download_and_uncompress(
@@ -960,6 +971,31 @@ if(WITH_TESTING AND WITH_INFERENCE_API_TEST)
     SRCS paddle_infer_api_errors_tester.cc
     DEPS ${inference_api_tester_deps} common)
 
+  if(WITH_GPU)
+    inference_analysis_test(
+      paddle_infer_api_test
+      SRCS
+      paddle_infer_api_test.cc
+      EXTRA_DEPS
+      common
+      paddle_inference_shared
+      ARGS
+      --infer_model=${RESNET50_MODEL_DIR})
+    set_tests_properties(paddle_infer_api_test PROPERTIES TIMEOUT 300)
+
+    inference_analysis_test(
+      paddle_infer_api_copy_tensor_tester
+      SRCS
+      paddle_infer_api_copy_tensor_tester.cc
+      EXTRA_DEPS
+      common
+      paddle_inference_shared
+      ARGS
+      --infer_model=${RESNET50_MODEL_DIR})
+    set_tests_properties(paddle_infer_api_copy_tensor_tester PROPERTIES TIMEOUT
+                                                                        300)
+  endif()
+
   if(WITH_GPU AND TENSORRT_FOUND)
     set_tests_properties(test_trt_dynamic_shape_ernie_ser_deser
                          PROPERTIES TIMEOUT 300)
diff --git a/test/cpp/inference/api/paddle_infer_api_copy_tensor_tester.cc b/test/cpp/inference/api/paddle_infer_api_copy_tensor_tester.cc
@@ -48,6 +48,7 @@ TEST(Tensor, copy_to_cpu_async_stream) {
   UpdateDllFlag("conv_workspace_size_limit", "4000");
   std::string model_dir = FLAGS_infer_model + "/model";
   Config config;
+  config.EnableNewIR(false);
   config.SetModel(model_dir + "/model", model_dir + "/params");
   config.EnableUseGpu(100, 0);
 
@@ -113,6 +114,8 @@ TEST(Tensor, copy_to_cpu_async_callback) {
   UpdateDllFlag("conv_workspace_size_limit", "4000");
   std::string model_dir = FLAGS_infer_model + "/model";
   Config config;
+  config.SwitchIrOptim(false);
+  config.EnableNewIR(false);
   config.SetModel(model_dir + "/model", model_dir + "/params");
   config.EnableUseGpu(100, 0);
 
@@ -148,6 +151,7 @@ TEST(Tensor, copy_to_cpu_async_callback) {
   for (int i = 0; i < 100; i++) {
     predictor->Run();
   }
+  cudaDeviceSynchronize();
 
   output_tensor->CopyToCpuAsync(
       out_data,
diff --git a/test/cpp/inference/api/paddle_infer_api_test.cc b/test/cpp/inference/api/paddle_infer_api_test.cc
@@ -25,6 +25,7 @@ TEST(Predictor, use_gpu) {
   UpdateDllFlag("conv_workspace_size_limit", "4000");
   std::string model_dir = FLAGS_infer_model + "/model";
   Config config;
+  config.EnableNewIR(false);
   config.SetModel(model_dir + "/model", model_dir + "/params");
   config.EnableUseGpu(100, 0);
 
@@ -63,6 +64,7 @@ TEST(PredictorPool, basic) {
   UpdateDllFlag("conv_workspace_size_limit", "4000");
   std::string model_dir = FLAGS_infer_model + "/model";
   Config config;
+  config.EnableNewIR(false);
   config.SetModel(model_dir + "/model", model_dir + "/params");
   config.EnableUseGpu(100, 0);
 
diff --git a/test/cpp/inference/api/trt_cascade_rcnn_test.cc b/test/cpp/inference/api/trt_cascade_rcnn_test.cc
@@ -25,6 +25,7 @@ TEST(TensorRT, cascade_rcnn) {
   std::string model_dir = FLAGS_infer_model + "/cascade_rcnn";
   AnalysisConfig config;
   int batch_size = 1;
+  config.EnableNewIR(false);
   config.EnableUseGpu(100, 0);
   config.SetModel(model_dir + "/model", model_dir + "/params");
   config.EnableTensorRtEngine(
diff --git a/test/cpp/inference/api/trt_dynamic_shape_test.cc b/test/cpp/inference/api/trt_dynamic_shape_test.cc
@@ -33,6 +33,7 @@ void TestDynamic(bool with_dynamic = true,
   }
 
   AnalysisConfig config;
+  config.EnableNewIR(false);
   config.EnableUseGpu(100, 0);
   std::string buffer_prog, buffer_param;
   ReadBinaryFile(model_dir + "/model", &buffer_prog);
diff --git a/test/cpp/inference/api/trt_quant_int8_yolov3_r50_test.cc b/test/cpp/inference/api/trt_quant_int8_yolov3_r50_test.cc
@@ -22,6 +22,7 @@ namespace inference {
 
 TEST(quant_int8, yolov3_resnet50) {
   AnalysisConfig config;
+  config.EnableNewIR(false);
   config.EnableUseGpu(100, 0);
   config.SetModel(FLAGS_infer_model + "/model", FLAGS_infer_model + "/params");
   config.EnableTensorRtEngine(
diff --git a/test/ir/CMakeLists.txt b/test/ir/CMakeLists.txt
@@ -4,8 +4,6 @@ file(
   "test_*.py")
 string(REPLACE ".py" "" TEST_IR_PASSES "${TEST_IR_PASSES}")
 
-list(REMOVE_ITEM TEST_IR_PASSES test_inference_datatype)
-
 list(REMOVE_ITEM TEST_IR_PASSES test_ir_fusion_group_pass)
 
 if(WIN32 AND WIN_UNITTEST_LEVEL LESS 2)
diff --git a/test/ir/inference/CMakeLists.txt b/test/ir/inference/CMakeLists.txt
@@ -28,12 +28,7 @@ list(REMOVE_ITEM TEST_INFERENCE_IR_PASSES
 list(REMOVE_ITEM TEST_INFERENCE_IR_PASSES "test_onednn_multi_gru_fuse_pass")
 list(REMOVE_ITEM TEST_INFERENCE_IR_PASSES
      "test_onednn_reshape_transpose_matmul_fuse_pass")
-list(REMOVE_ITEM TEST_INFERENCE_IR_PASSES "test_forbid_dynamic_op_api")
 list(REMOVE_ITEM TEST_INFERENCE_IR_PASSES "test_save_optimized_model_pass")
-list(REMOVE_ITEM TEST_INFERENCE_IR_PASSES
-     "test_trt_explicit_quantization_mobilenet")
-list(REMOVE_ITEM TEST_INFERENCE_IR_PASSES
-     "test_trt_explicit_quantization_resnet")
 list(REMOVE_ITEM TEST_INFERENCE_IR_PASSES
      "test_trt_remove_amp_strategy_op_pass")
 
diff --git a/test/ir/inference/test_forbid_dynamic_op_api.py b/test/ir/inference/test_forbid_dynamic_op_api.py
@@ -93,7 +93,7 @@ def build_model(self):
 
     def init_predictor(self):
         config = Config(
-            self.model_prefix + '.pdmodel', self.model_prefix + '.pdiparams'
+            self.model_prefix + '.json', self.model_prefix + '.pdiparams'
         )
         config.enable_use_gpu(256, 0, PrecisionType.Half)
         config.enable_tensorrt_engine(
diff --git a/test/ir/inference/test_save_optimized_model_pass.py b/test/ir/inference/test_save_optimized_model_pass.py
@@ -80,7 +80,7 @@ class TestSaveOptimizedModelPassWithGPU(
     def init_predictor(self, save_optimized_model: bool):
         if save_optimized_model is True:
             config = Config(
-                os.path.join(self.temp_dir.name, 'alexnet/inference.pdmodel'),
+                os.path.join(self.temp_dir.name, 'alexnet/inference.json'),
                 os.path.join(self.temp_dir.name, 'alexnet/inference.pdiparams'),
             )
             config.enable_use_gpu(256, 0, PrecisionType.Half)
@@ -92,7 +92,7 @@ def init_predictor(self, save_optimized_model: bool):
             config.enable_save_optim_model(True)
         else:
             config = Config(
-                os.path.join(self.temp_dir.name, 'alexnet/_optimized.pdmodel'),
+                os.path.join(self.temp_dir.name, 'alexnet/_optimized.json'),
                 os.path.join(
                     self.temp_dir.name, 'alexnet/_optimized.pdiparams'
                 ),
@@ -186,7 +186,7 @@ class TestSaveOptimizedModelPassWithTRT(
     def init_predictor(self, save_optimized_model: bool):
         if save_optimized_model is True:
             config = Config(
-                os.path.join(self.temp_dir.name, 'alexnet/inference.pdmodel'),
+                os.path.join(self.temp_dir.name, 'alexnet/inference.json'),
                 os.path.join(self.temp_dir.name, 'alexnet/inference.pdiparams'),
             )
             config.enable_use_gpu(256, 0)
@@ -212,7 +212,7 @@ def init_predictor(self, save_optimized_model: bool):
             config.enable_save_optim_model(True)
         else:
             config = Config(
-                os.path.join(self.temp_dir.name, 'alexnet/_optimized.pdmodel'),
+                os.path.join(self.temp_dir.name, 'alexnet/_optimized.json'),
                 os.path.join(
                     self.temp_dir.name, 'alexnet/_optimized.pdiparams'
                 ),
diff --git a/test/ir/inference/test_trt_explicit_quantization_model.py b/test/ir/inference/test_trt_explicit_quantization_model.py
@@ -118,12 +118,12 @@ def insert_qdq(program, scope, place, for_test=False):
             path_prefix, [image], [out], exe, program=quant_val_prog
         )
 
-    def infer_program(self, trt_int8=False):
+    def infer_program(self, trt_int8=False, collect_shape=False):
         config = Config(
             os.path.join(self.path, 'inference.pdmodel'),
             os.path.join(self.path, 'inference.pdiparams'),
         )
-        config.enable_use_gpu(256, 0, PrecisionType.Half)
+        config.enable_use_gpu(256, 0, PrecisionType.Float32)
         config.enable_memory_optim()
         if trt_int8:
             precision_mode = PrecisionType.Int8
@@ -139,11 +139,11 @@ def infer_program(self, trt_int8=False):
         )
         if trt_int8:
             config.enable_tensorrt_explicit_quantization()
-        config.set_trt_dynamic_shape_info(
-            {"image": (1, 1, 28, 28)},
-            {"image": (128, 1, 28, 28)},
-            {"image": (64, 1, 28, 28)},
-        )
+        shape_path = self.path + ".shape.txt"
+        if collect_shape:
+            config.collect_shape_range_info(shape_path)
+        else:
+            config.enable_tuned_tensorrt_dynamic_shape(shape_path)
         config.disable_glog_info()
         predictor = create_predictor(config)
         input_names = predictor.get_input_names()
@@ -159,8 +159,12 @@ def infer_program(self, trt_int8=False):
     def test_model(self):
         with paddle.pir_utils.OldIrGuard():
             self.build_program()
-            baseline_output = self.infer_program(trt_int8=False)
-            trt_output = self.infer_program(trt_int8=True)
+            self.infer_program(trt_int8=False, collect_shape=True)
+            baseline_output = self.infer_program(
+                trt_int8=False, collect_shape=False
+            )
+            self.infer_program(trt_int8=True, collect_shape=True)
+            trt_output = self.infer_program(trt_int8=True, collect_shape=False)
             trt_predict = np.argmax(trt_output, axis=1)
             baseline_predict = np.argmax(baseline_output, axis=1)
             same = (trt_predict == baseline_predict).sum() / len(trt_predict)
diff --git a/test/ir/inference/test_trt_remove_amp_strategy_op_pass.py b/test/ir/inference/test_trt_remove_amp_strategy_op_pass.py
@@ -91,7 +91,7 @@ def infer_program(self, use_trt=False):
         )
         config.enable_use_gpu(256, 0, PrecisionType.Half)
         config.enable_memory_optim()
-        config.disable_glog_info()
+        # config.disable_glog_info()
         if use_trt:
             config.enable_tensorrt_engine(
                 workspace_size=1 << 30,
@@ -192,8 +192,8 @@ def build_program(self):
 
 
 @unittest.skipIf(
-    paddle.inference.get_trt_compile_version() < (8, 5, 1),
-    "Quantization axis is consistent with Paddle after TRT 8.5.2.",
+    paddle.inference.get_trt_compile_version() < (8, 6, 1),
+    "Quantization axis is consistent with Paddle after TRT 8.6.1.",
 )
 class TestRemoveStrategyOpAMPQAT(TestRemoveStrategyOpBase, unittest.TestCase):
     def build_program(self):
diff --git a/test/ir/inference/test_use_optimized_model_api.py b/test/ir/inference/test_use_optimized_model_api.py
@@ -79,7 +79,7 @@ def test_check_output(self):
     def inference(self):
         # Config
         config = Config(
-            self.path_prefix + ".pdmodel", self.path_prefix + ".pdiparams"
+            self.path_prefix + ".json", self.path_prefix + ".pdiparams"
         )
         config.enable_use_gpu(100, 0)
         config.enable_tensorrt_engine(
diff --git a/test/ir/test_inference_datatype.py b/test/ir/test_inference_datatype.py
@@ -65,7 +65,7 @@ def init_predictor(self):
         config = Config(
             os.path.join(
                 self.temp_dir.name,
-                'test_inference_datatype_model/inference.pdmodel',
+                'test_inference_datatype_model/inference.json',
             ),
             os.path.join(
                 self.temp_dir.name,

Original file line number	Diff line number	Diff line change
`@@ -33,6 +33,7 @@ void TestDynamic(bool with_dynamic = true,`
`33`	`33`	`}`
`34`	`34`
`35`	`35`	`AnalysisConfig config;`
	`36`	`+ config.EnableNewIR(false);`
`36`	`37`	`config.EnableUseGpu(100, 0);`
`37`	`38`	`std::string buffer_prog, buffer_param;`
`38`	`39`	`ReadBinaryFile(model_dir + "/model", &buffer_prog);`