[Paddle Inference]fix some transformer unitest (#48929)

* fix some transformer unitest

paddle/fluid/inference/tensorrt/convert/matmul_op.cc

@@ -67,7 +67,8 @@ class MatMulOpConverter : public OpConverter {
     if (op_desc.HasAttr("support_int8") &&
         PADDLE_GET_CONST(bool, op_desc.GetAttr("support_int8")) &&
         engine_->precision() == AnalysisConfig::Precision::kInt8 &&
-        platform::GetGPUComputeCapability(0) >= 75) {
+        platform::GetGPUComputeCapability(platform::GetCurrentDeviceId()) >=
+            75) {
       if (engine_->with_dynamic_shape()) {
         VLOG(3) << "Convert a fluid matmul_op_int8_dynamic to TensorRT "
                    "MatmulPluginLayer";

paddle/fluid/inference/tensorrt/convert/multihead_matmul_op.cc

@@ -88,11 +88,10 @@ class MultiheadMatMulOpConverter : public OpConverter {
                           engine_->tensorrt_transformer_posid() != "" &&
                           engine_->tensorrt_transformer_maskid() != "";
     if (engine_->with_dynamic_shape()) {
-      if (engine_->tensorrt_transformer_maskid() != "") {
-        if (engine_->precision() == AnalysisConfig::Precision::kFloat32) {
-          PADDLE_THROW(platform::errors::Fatal(
-              "use use_varseqlen must be int8 or half, not float32."));
-        }
+      if (engine_->tensorrt_transformer_maskid() != "" &&
+          engine_->precision() != AnalysisConfig::Precision::kFloat32 &&
+          platform::GetGPUComputeCapability(platform::GetCurrentDeviceId()) >=
+              75) {
         nvinfer1::Weights weight{nvinfer1::DataType::kFLOAT,
                                  static_cast<void*>(weight_data),
                                  static_cast<int32_t>(weight_t->numel())};
@@ -401,7 +400,8 @@ class MultiheadMatMulOpConverter : public OpConverter {
       } else {
         if (input_dims.d[1] <= 384 && !bias_qk_attr &&
             engine_->precision() != AnalysisConfig::Precision::kFloat32 &&
-            platform::GetGPUComputeCapability(0) >= 75) {
+            platform::GetGPUComputeCapability(platform::GetCurrentDeviceId()) >=
+                75) {
           /*
             * input_dims.d[0]: batch(-1)
             * input_dims.d[1]: length:256

paddle/fluid/inference/tensorrt/plugin/CMakeLists.txt

@@ -39,12 +39,12 @@ list(
   generic_plugin.cu
   lookup_table.cu
   many_emb_layernorm_plugin.cu
-  many_emb_Layernorm_kernel.cu)
+  many_emb_layernorm_kernel.cu)
 
 if(${TENSORRT_MAJOR_VERSION} GREATER_EQUAL 7)
   list(APPEND TRT_FILES many_emb_layernorm_varseqlen_plugin.cu
-       many_emb_Layernorm_varseqlen_kernel_mtron.cu
-       many_emb_Layernorm_varseqlen_kernel_hface.cu)
+       many_emb_layernorm_varseqlen_kernel_mtron.cu
+       many_emb_layernorm_varseqlen_kernel_hface.cu)
 endif()
 
 if(CUSPARSELT_FOUND AND ${TENSORRT_MAJOR_VERSION} GREATER_EQUAL 8)

paddle/fluid/inference/tests/api/trt_dynamic_shape_ernie_serialize_deserialize_test.h

@@ -139,6 +139,8 @@ static void trt_ernie(bool with_fp16, std::vector<float> result) {
   config.EnableTensorRtEngine(1 << 30, 1, 5, precision, true, false);
   config.SetTRTDynamicShapeInfo(
       min_input_shape, max_input_shape, opt_input_shape);
+  paddle_infer::experimental::InternalUtils::SetTransformerMaskid(
+      &config, "read_file_0.tmp_4");
   AnalysisConfig* config_deser = new AnalysisConfig(config);
 
   std::vector<float> out_data;

paddle/fluid/inference/tests/api/trt_dynamic_shape_ernie_test.cc

@@ -133,6 +133,8 @@ void trt_ernie(bool with_fp16,
   config.EnableTensorRtEngine(1 << 30, 1, 5, precision, false, false);
   config.SetTRTDynamicShapeInfo(
       min_input_shape, max_input_shape, opt_input_shape);
+  paddle_infer::experimental::InternalUtils::SetTransformerMaskid(
+      &config, "read_file_0.tmp_4");
   std::vector<float> out_data;
   run(config, &out_data, batch_size);
 
@@ -423,7 +425,7 @@ void run(paddle_infer::Predictor* predictor, std::vector<float>* out_data) {
 
 TEST(AnalysisPredictor, ernie_varlen) {
 #if IS_TRT_VERSION_GE(7234)
-  if (platform::GetGPUComputeCapability(0) >= 75) {
+  if (platform::GetGPUComputeCapability(platform::GetCurrentDeviceId()) >= 75) {
     auto predictor = InitPredictor();
     std::vector<float> out_data;
     run(predictor.get(), &out_data);