From 242d26d45e8fd12dfa3cc04320959f2c1e2e0155 Mon Sep 17 00:00:00 2001 From: Frank Lin Date: Tue, 8 Aug 2023 20:28:11 -0700 Subject: [PATCH] Set Input Type for TensorRT convert tests (#55823) * set input type to fp16 --- test/ir/inference/auto_scan_test.py | 154 +++++++++--------- test/ir/inference/program_config.py | 6 +- .../inference/test_trt_convert_activation.py | 4 + .../test_trt_convert_affine_channel.py | 4 + .../test_trt_convert_anchor_generator.py | 4 + test/ir/inference/test_trt_convert_arg_max.py | 4 + test/ir/inference/test_trt_convert_arg_min.py | 4 + test/ir/inference/test_trt_convert_assign.py | 4 + .../inference/test_trt_convert_batch_norm.py | 4 + .../test_trt_convert_bilinear_interp_v2.py | 4 + .../inference/test_trt_convert_bitwise_not.py | 4 + test/ir/inference/test_trt_convert_bmm.py | 4 + test/ir/inference/test_trt_convert_cast.py | 4 + test/ir/inference/test_trt_convert_clip.py | 4 + .../test_trt_convert_compare_and_logical.py | 20 +++ test/ir/inference/test_trt_convert_concat.py | 4 + test/ir/inference/test_trt_convert_conv2d.py | 8 + .../test_trt_convert_conv2d_fusion.py | 6 + .../test_trt_convert_conv2d_transpose.py | 8 + .../test_trt_convert_conv3d_transpose.py | 2 + ...test_trt_convert_cross_multihead_matmul.py | 4 + test/ir/inference/test_trt_convert_cumsum.py | 2 + .../test_trt_convert_deformable_conv.py | 2 + .../test_trt_convert_depthwise_conv2d.py | 6 + ..._trt_convert_depthwise_conv2d_transpose.py | 4 + test/ir/inference/test_trt_convert_dropout.py | 4 + test/ir/inference/test_trt_convert_einsum.py | 12 ++ .../inference/test_trt_convert_elementwise.py | 36 ++++ ...st_trt_convert_elementwiseadd_transpose.py | 2 + .../test_trt_convert_emb_eltwise_layernorm.py | 4 + test/ir/inference/test_trt_convert_equal.py | 4 + .../test_trt_convert_expand_as_v2.py | 4 + .../inference/test_trt_convert_expand_v2.py | 8 + .../test_trt_convert_fill_any_like.py | 4 + .../test_trt_convert_fill_constant.py | 2 + ...test_trt_convert_flash_multihead_matmul.py | 4 + test/ir/inference/test_trt_convert_flatten.py | 16 ++ ...st_trt_convert_flatten_contiguous_range.py | 4 + test/ir/inference/test_trt_convert_flip.py | 2 + .../test_trt_convert_fused_token_prune.py | 2 + test/ir/inference/test_trt_convert_gather.py | 4 + .../inference/test_trt_convert_gather_nd.py | 24 +++ test/ir/inference/test_trt_convert_gelu.py | 4 + .../test_trt_convert_grid_sampler.py | 2 + .../inference/test_trt_convert_group_norm.py | 4 + .../test_trt_convert_hard_sigmoid.py | 4 + .../inference/test_trt_convert_hard_swish.py | 4 + .../test_trt_convert_index_select.py | 4 + .../test_trt_convert_instance_norm.py | 4 + test/ir/inference/test_trt_convert_inverse.py | 4 + .../inference/test_trt_convert_layer_norm.py | 8 + .../inference/test_trt_convert_leaky_relu.py | 6 + .../test_trt_convert_lookup_table.py | 2 + .../test_trt_convert_lookup_table_v2.py | 2 + test/ir/inference/test_trt_convert_matmul.py | 4 + .../inference/test_trt_convert_matmul_v2.py | 6 + test/ir/inference/test_trt_convert_mish.py | 4 + .../test_trt_convert_multiclass_nms.py | 3 + .../test_trt_convert_multiclass_nms3.py | 3 + .../test_trt_convert_multihead_matmul.py | 7 + ...t_trt_convert_multihead_matmul_roformer.py | 2 + .../test_trt_convert_nearest_interp.py | 4 + .../test_trt_convert_nearest_interp_v2.py | 8 + test/ir/inference/test_trt_convert_one_hot.py | 4 + test/ir/inference/test_trt_convert_p_norm.py | 2 + test/ir/inference/test_trt_convert_pad.py | 4 + test/ir/inference/test_trt_convert_pad3d.py | 8 + test/ir/inference/test_trt_convert_pool2d.py | 4 + .../test_trt_convert_preln_residual_bias.py | 4 + ...test_trt_convert_preln_residual_no_bias.py | 4 + test/ir/inference/test_trt_convert_prelu.py | 4 + .../test_trt_convert_qk_multihead_matmul.py | 4 + test/ir/inference/test_trt_convert_range.py | 4 + test/ir/inference/test_trt_convert_reduce.py | 4 + test/ir/inference/test_trt_convert_reshape.py | 10 ++ test/ir/inference/test_trt_convert_rnn.py | 2 + .../inference/test_trt_convert_roi_align.py | 4 + test/ir/inference/test_trt_convert_roll.py | 4 + test/ir/inference/test_trt_convert_scale.py | 4 + .../test_trt_convert_scatter_nd_add.py | 4 + .../inference/test_trt_convert_set_value.py | 2 + test/ir/inference/test_trt_convert_shape.py | 4 + .../test_trt_convert_shuffle_channel.py | 4 + test/ir/inference/test_trt_convert_slice.py | 4 + test/ir/inference/test_trt_convert_softmax.py | 4 + test/ir/inference/test_trt_convert_split.py | 6 + test/ir/inference/test_trt_convert_square.py | 4 + .../ir/inference/test_trt_convert_squeeze2.py | 4 + test/ir/inference/test_trt_convert_stack.py | 4 + .../test_trt_convert_strided_slice.py | 4 + test/ir/inference/test_trt_convert_sum.py | 8 + test/ir/inference/test_trt_convert_swish.py | 4 + .../test_trt_convert_take_along_axis.py | 4 + .../test_trt_convert_temporal_shift.py | 4 + test/ir/inference/test_trt_convert_tile.py | 8 + test/ir/inference/test_trt_convert_top_k.py | 4 + .../ir/inference/test_trt_convert_top_k_v2.py | 4 + .../test_trt_convert_trans_layernorm.py | 2 + .../inference/test_trt_convert_transpose.py | 4 + test/ir/inference/test_trt_convert_unary.py | 8 + test/ir/inference/test_trt_convert_unbind.py | 2 + test/ir/inference/test_trt_convert_unfold.py | 4 + .../inference/test_trt_convert_unsqueeze2.py | 4 + test/ir/inference/test_trt_convert_where.py | 4 + .../ir/inference/test_trt_convert_yolo_box.py | 4 + .../test_trt_convert_yolo_box_head.py | 1 + 106 files changed, 601 insertions(+), 81 deletions(-) diff --git a/test/ir/inference/auto_scan_test.py b/test/ir/inference/auto_scan_test.py index 9f21b94027a..225318b5f86 100755 --- a/test/ir/inference/auto_scan_test.py +++ b/test/ir/inference/auto_scan_test.py @@ -14,7 +14,6 @@ import abc import enum -import logging import os import shutil import time @@ -35,9 +34,12 @@ from program_config import ( import paddle import paddle.inference as paddle_infer from paddle.fluid.core import PassVersionChecker +from paddle.static.log_helper import get_logger LOGLEVEL = os.environ.get("PADDLE_TEST_LOGLEVEL", "INFO").upper() -logging.basicConfig(level=LOGLEVEL, format="%(message)s") +logging = get_logger( + __name__, LOGLEVEL, fmt='%(asctime)s-%(levelname)s: %(message)s' +) settings.register_profile( "ci", @@ -662,9 +664,6 @@ class TrtLayerAutoScanTest(AutoScanTest): ) return config - def get_avalible_input_type(self) -> List[np.dtype]: - return [np.float32] - def assert_tensors_near( self, atol: float, @@ -747,10 +746,13 @@ class TrtLayerAutoScanTest(AutoScanTest): if not skip_baseline: # baseline: gpu run, we only test float32 gpu_config = self.create_inference_config(use_trt=False) + prog_config = prog_config.set_input_type( + np.float16 + ).set_input_type(np.float32) baseline_result = self.run_test_config( model, params, - prog_config.set_input_type(np.float32), + prog_config, gpu_config, feed_data, ) @@ -761,82 +763,78 @@ class TrtLayerAutoScanTest(AutoScanTest): nodes_num, threshold, ) in self.sample_predictor_configs(prog_config): - for input_type in self.get_avalible_input_type(): - prog_config = prog_config.set_input_type(input_type) - if os.path.exists(self.cache_dir): - shutil.rmtree(self.cache_dir) + if os.path.exists(self.cache_dir): + shutil.rmtree(self.cache_dir) - if isinstance(threshold, float): - atol = threshold - rtol = 1e-8 - elif isinstance(threshold, list) or isinstance( - threshold, tuple - ): - atol = threshold[0] - rtol = threshold[1] - else: - raise NotImplementedError - - is_fp8 = ( - pred_config.tensorrt_precision_mode() - == paddle_infer.PrecisionType.Int8 - ) - if (not is_fp8 and quant) or (is_fp8 and not quant): - continue - - ignore_flag = False - for teller, reason, note in self.ignore_cases: - if teller(prog_config, pred_config): - ignore_flag = True - if reason == IgnoreReasons.TRT_NOT_IMPLEMENTED: - self.ignore_log( - f"[TRT_NOT_IMPLEMENTED] {note} vs {self.inference_config_str(pred_config)}" - ) - elif reason == IgnoreReasons.TRT_NOT_SUPPORT: - self.ignore_log( - f"[TRT_NOT_SUPPORT] {note} vs {self.inference_config_str(pred_config)}" - ) - else: - raise NotImplementedError - break - - if ignore_flag: - continue - - try: - pred_config_deserialize = paddle_infer.Config( - pred_config - ) - trt_result = self.run_test_config( - model, params, prog_config, pred_config, feed_data - ) - self.assert_tensors_near( - atol, rtol, trt_result, baseline_result - ) - trt_engine_num, paddle_op_num = nodes_num - self.assert_op_size(trt_engine_num, paddle_op_num) - - # deserialize test - if trt_engine_num > 0: - self.run_test_config( - model, - params, - prog_config, - pred_config_deserialize, - feed_data, + if isinstance(threshold, float): + atol = threshold + rtol = 1e-8 + elif isinstance(threshold, list) or isinstance( + threshold, tuple + ): + atol = threshold[0] + rtol = threshold[1] + else: + raise NotImplementedError + + is_fp8 = ( + pred_config.tensorrt_precision_mode() + == paddle_infer.PrecisionType.Int8 + ) + if (not is_fp8 and quant) or (is_fp8 and not quant): + continue + + ignore_flag = False + for teller, reason, note in self.ignore_cases: + if teller(prog_config, pred_config): + ignore_flag = True + if reason == IgnoreReasons.TRT_NOT_IMPLEMENTED: + self.ignore_log( + f"[TRT_NOT_IMPLEMENTED] {note} vs {self.inference_config_str(pred_config)}" ) + elif reason == IgnoreReasons.TRT_NOT_SUPPORT: + self.ignore_log( + f"[TRT_NOT_SUPPORT] {note} vs {self.inference_config_str(pred_config)}" + ) + else: + raise NotImplementedError + break - self.success_log(f"program_config: {prog_config}") - self.success_log( - f"predictor_config: {self.inference_config_str(pred_config)}" - ) - except Exception as e: - self.fail_log(f"program_config: {prog_config}") - self.fail_log( - f"predictor_config: {self.inference_config_str(pred_config)}" + if ignore_flag: + continue + + try: + pred_config_deserialize = paddle_infer.Config(pred_config) + trt_result = self.run_test_config( + model, params, prog_config, pred_config, feed_data + ) + self.assert_tensors_near( + atol, rtol, trt_result, baseline_result + ) + trt_engine_num, paddle_op_num = nodes_num + self.assert_op_size(trt_engine_num, paddle_op_num) + + # deserialize test + if trt_engine_num > 0: + self.run_test_config( + model, + params, + prog_config, + pred_config_deserialize, + feed_data, ) - self.fail_log(f"\033[1;31m ERROR INFO: {e}\033[0m") - all_passes = False + + self.success_log(f"program_config: {prog_config}") + self.success_log( + f"predictor_config: {self.inference_config_str(pred_config)}" + ) + except Exception as e: + self.fail_log(f"program_config: {prog_config}") + self.fail_log( + f"predictor_config: {self.inference_config_str(pred_config)}" + ) + self.fail_log(f"\033[1;31m ERROR INFO: {e}\033[0m") + all_passes = False self.assertTrue(all_passes) diff --git a/test/ir/inference/program_config.py b/test/ir/inference/program_config.py index 3dfd9b3f958..601ddfdc238 100644 --- a/test/ir/inference/program_config.py +++ b/test/ir/inference/program_config.py @@ -67,7 +67,7 @@ class TensorConfig: def __repr__(self): return str({'shape': self.shape, 'lod': self.lod, 'dtype': self.dtype}) - def astype(self, type: np.dtype): + def convert_type_inplace(self, type: np.dtype): self.data = self.data.astype(type) self.dtype = self.data.dtype return self @@ -277,9 +277,9 @@ class ProgramConfig: def set_input_type(self, type: np.dtype): for inp in self.inputs.values(): - inp.astype(type) + inp.convert_type_inplace(type) for weight in self.weights.values(): - weight.astype(type) + weight.convert_type_inplace(type) return self def get_input_type(self) -> np.dtype: diff --git a/test/ir/inference/test_trt_convert_activation.py b/test/ir/inference/test_trt_convert_activation.py index cec7e624b08..85f2729f351 100644 --- a/test/ir/inference/test_trt_convert_activation.py +++ b/test/ir/inference/test_trt_convert_activation.py @@ -178,10 +178,12 @@ class TrtConvertActivationTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -189,10 +191,12 @@ class TrtConvertActivationTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_affine_channel.py b/test/ir/inference/test_trt_convert_affine_channel.py index b326f14ed2f..b30a07fc1b5 100644 --- a/test/ir/inference/test_trt_convert_affine_channel.py +++ b/test/ir/inference/test_trt_convert_affine_channel.py @@ -133,10 +133,12 @@ class TrtConvertAffineChannelTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -144,10 +146,12 @@ class TrtConvertAffineChannelTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_anchor_generator.py b/test/ir/inference/test_trt_convert_anchor_generator.py index 1e7c380b40f..fd91a9945b6 100644 --- a/test/ir/inference/test_trt_convert_anchor_generator.py +++ b/test/ir/inference/test_trt_convert_anchor_generator.py @@ -107,10 +107,12 @@ class TrtConvertAnchorGeneratorTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -118,10 +120,12 @@ class TrtConvertAnchorGeneratorTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_arg_max.py b/test/ir/inference/test_trt_convert_arg_max.py index 081a58eadf4..f7e8185d3ad 100644 --- a/test/ir/inference/test_trt_convert_arg_max.py +++ b/test/ir/inference/test_trt_convert_arg_max.py @@ -118,10 +118,12 @@ class TrtConvertArgMaxTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -129,10 +131,12 @@ class TrtConvertArgMaxTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_arg_min.py b/test/ir/inference/test_trt_convert_arg_min.py index da448142d4f..c1e330c1c2b 100644 --- a/test/ir/inference/test_trt_convert_arg_min.py +++ b/test/ir/inference/test_trt_convert_arg_min.py @@ -118,10 +118,12 @@ class TrtConvertArgMinTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -129,10 +131,12 @@ class TrtConvertArgMinTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_assign.py b/test/ir/inference/test_trt_convert_assign.py index 64dbf16064e..55939982d5e 100644 --- a/test/ir/inference/test_trt_convert_assign.py +++ b/test/ir/inference/test_trt_convert_assign.py @@ -133,10 +133,12 @@ class TrtConvertAssignTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 @@ -144,10 +146,12 @@ class TrtConvertAssignTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 diff --git a/test/ir/inference/test_trt_convert_batch_norm.py b/test/ir/inference/test_trt_convert_batch_norm.py index 277eb132602..2cd98ad6998 100644 --- a/test/ir/inference/test_trt_convert_batch_norm.py +++ b/test/ir/inference/test_trt_convert_batch_norm.py @@ -232,10 +232,12 @@ class TrtConvertBatchNormTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -243,10 +245,12 @@ class TrtConvertBatchNormTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_bilinear_interp_v2.py b/test/ir/inference/test_trt_convert_bilinear_interp_v2.py index 148776365c0..5fb36deec1a 100644 --- a/test/ir/inference/test_trt_convert_bilinear_interp_v2.py +++ b/test/ir/inference/test_trt_convert_bilinear_interp_v2.py @@ -120,10 +120,12 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 @@ -131,10 +133,12 @@ class TrtConvertBilinearInterpV2Test(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 diff --git a/test/ir/inference/test_trt_convert_bitwise_not.py b/test/ir/inference/test_trt_convert_bitwise_not.py index ab8da867362..738ef5af23a 100644 --- a/test/ir/inference/test_trt_convert_bitwise_not.py +++ b/test/ir/inference/test_trt_convert_bitwise_not.py @@ -133,10 +133,12 @@ class TrtConvertActivationTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 @@ -144,10 +146,12 @@ class TrtConvertActivationTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 diff --git a/test/ir/inference/test_trt_convert_bmm.py b/test/ir/inference/test_trt_convert_bmm.py index ebd25646a18..0bb85f9abf9 100644 --- a/test/ir/inference/test_trt_convert_bmm.py +++ b/test/ir/inference/test_trt_convert_bmm.py @@ -98,10 +98,12 @@ class TrtConvertBmmTest_dynamic(TrtLayerAutoScanTest): clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -115,10 +117,12 @@ class TrtConvertBmmTest_dynamic(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), tol_fp32 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), tol_half diff --git a/test/ir/inference/test_trt_convert_cast.py b/test/ir/inference/test_trt_convert_cast.py index 026abc57105..da181e79225 100644 --- a/test/ir/inference/test_trt_convert_cast.py +++ b/test/ir/inference/test_trt_convert_cast.py @@ -164,10 +164,12 @@ class TrtConvertCastTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 @@ -175,10 +177,12 @@ class TrtConvertCastTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 diff --git a/test/ir/inference/test_trt_convert_clip.py b/test/ir/inference/test_trt_convert_clip.py index c6fd638c2f9..dc46c2a775d 100644 --- a/test/ir/inference/test_trt_convert_clip.py +++ b/test/ir/inference/test_trt_convert_clip.py @@ -143,10 +143,12 @@ class TrtConvertClipTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -154,10 +156,12 @@ class TrtConvertClipTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_compare_and_logical.py b/test/ir/inference/test_trt_convert_compare_and_logical.py index 79d157e9fe7..fde4b2b3587 100755 --- a/test/ir/inference/test_trt_convert_compare_and_logical.py +++ b/test/ir/inference/test_trt_convert_compare_and_logical.py @@ -153,10 +153,12 @@ class TrtConvertLogicalTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -164,10 +166,12 @@ class TrtConvertLogicalTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) @@ -295,10 +299,12 @@ class TrtConvertCompareTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -306,10 +312,12 @@ class TrtConvertCompareTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) @@ -454,10 +462,12 @@ class TrtConvertLessEqualTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -465,10 +475,12 @@ class TrtConvertLessEqualTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) @@ -613,10 +625,12 @@ class TrtConvertGreaterEqualTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -624,10 +638,12 @@ class TrtConvertGreaterEqualTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) @@ -750,10 +766,12 @@ class TrtConvertCompareSkipTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -761,10 +779,12 @@ class TrtConvertCompareSkipTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_concat.py b/test/ir/inference/test_trt_convert_concat.py index afe1851759e..fcf68b74816 100644 --- a/test/ir/inference/test_trt_convert_concat.py +++ b/test/ir/inference/test_trt_convert_concat.py @@ -320,10 +320,12 @@ class TrtConvertConcatTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -331,10 +333,12 @@ class TrtConvertConcatTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_conv2d.py b/test/ir/inference/test_trt_convert_conv2d.py index 432ad97f76a..8d3ea2e28ea 100644 --- a/test/ir/inference/test_trt_convert_conv2d.py +++ b/test/ir/inference/test_trt_convert_conv2d.py @@ -171,14 +171,17 @@ class TrtConvertConv2dTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) self.trt_param.precision = paddle_infer.PrecisionType.Int8 + program_config.set_input_type(np.int8) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-2, 1e-2) @@ -186,14 +189,17 @@ class TrtConvertConv2dTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) self.trt_param.precision = paddle_infer.PrecisionType.Int8 + program_config.set_input_type(np.int8) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-2, 1e-2) @@ -356,11 +362,13 @@ class TrtConvertConv2dNotPersistableTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-2, 1e-2) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-2, 1e-2) diff --git a/test/ir/inference/test_trt_convert_conv2d_fusion.py b/test/ir/inference/test_trt_convert_conv2d_fusion.py index dd6eb895b98..00d6a8678a9 100644 --- a/test/ir/inference/test_trt_convert_conv2d_fusion.py +++ b/test/ir/inference/test_trt_convert_conv2d_fusion.py @@ -179,14 +179,17 @@ class TrtConvertConv2dFusionTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) self.trt_param.precision = paddle_infer.PrecisionType.Int8 + program_config.set_input_type(np.int8) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -194,14 +197,17 @@ class TrtConvertConv2dFusionTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) self.trt_param.precision = paddle_infer.PrecisionType.Int8 + program_config.set_input_type(np.int8) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_conv2d_transpose.py b/test/ir/inference/test_trt_convert_conv2d_transpose.py index 42057072f70..57c973b4667 100644 --- a/test/ir/inference/test_trt_convert_conv2d_transpose.py +++ b/test/ir/inference/test_trt_convert_conv2d_transpose.py @@ -194,10 +194,12 @@ class TrtConvertConv2dTransposeTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -208,10 +210,12 @@ class TrtConvertConv2dTransposeTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) @@ -338,10 +342,12 @@ class TrtConvertConv2dTransposeTest2(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-4 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e0, 1e-3) @@ -349,10 +355,12 @@ class TrtConvertConv2dTransposeTest2(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-4 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e0, 1e-3) diff --git a/test/ir/inference/test_trt_convert_conv3d_transpose.py b/test/ir/inference/test_trt_convert_conv3d_transpose.py index 4134460d7d3..ba545275d80 100644 --- a/test/ir/inference/test_trt_convert_conv3d_transpose.py +++ b/test/ir/inference/test_trt_convert_conv3d_transpose.py @@ -118,6 +118,7 @@ class TrtConvertConv3dTransposeTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -125,6 +126,7 @@ class TrtConvertConv3dTransposeTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_cross_multihead_matmul.py b/test/ir/inference/test_trt_convert_cross_multihead_matmul.py index a8ce9ee0a59..d50875e9a93 100644 --- a/test/ir/inference/test_trt_convert_cross_multihead_matmul.py +++ b/test/ir/inference/test_trt_convert_cross_multihead_matmul.py @@ -270,17 +270,21 @@ class TrtConvertCrossMultiHeadMatmulTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) self.trt_param.workspace_size = 2013265920 yield self.create_inference_config(), (1, 4), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 4), (1e-2, 1e-3) # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) self.trt_param.workspace_size = 2013265920 yield self.create_inference_config(), (1, 3), (1e-5, 1e-4) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), (1e-2, 1e-2) def add_skip_trt_case(self): diff --git a/test/ir/inference/test_trt_convert_cumsum.py b/test/ir/inference/test_trt_convert_cumsum.py index 50c73afcc06..8f58f6b2ada 100644 --- a/test/ir/inference/test_trt_convert_cumsum.py +++ b/test/ir/inference/test_trt_convert_cumsum.py @@ -175,10 +175,12 @@ class TrtConvertCumsum(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 diff --git a/test/ir/inference/test_trt_convert_deformable_conv.py b/test/ir/inference/test_trt_convert_deformable_conv.py index ff32472a9d0..55682a287c2 100644 --- a/test/ir/inference/test_trt_convert_deformable_conv.py +++ b/test/ir/inference/test_trt_convert_deformable_conv.py @@ -219,11 +219,13 @@ class TrtConvertDeformableConvTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-5, 1e-5) diff --git a/test/ir/inference/test_trt_convert_depthwise_conv2d.py b/test/ir/inference/test_trt_convert_depthwise_conv2d.py index bdad374e5a7..066a517d3f1 100644 --- a/test/ir/inference/test_trt_convert_depthwise_conv2d.py +++ b/test/ir/inference/test_trt_convert_depthwise_conv2d.py @@ -151,13 +151,16 @@ class TrtConvertDepthwiseConv2dTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num(), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num(), ( 1e-3, 1e-3, ) self.trt_param.precision = paddle_infer.PrecisionType.Int8 + program_config.set_input_type(np.int8) yield self.create_inference_config(), generate_trt_nodes_num(), ( 1e-3, 1e-3, @@ -166,13 +169,16 @@ class TrtConvertDepthwiseConv2dTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num(), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num(), ( 1e-3, 1e-3, ) self.trt_param.precision = paddle_infer.PrecisionType.Int8 + program_config.set_input_type(np.int8) yield self.create_inference_config(), generate_trt_nodes_num(), ( 1e-3, 1e-3, diff --git a/test/ir/inference/test_trt_convert_depthwise_conv2d_transpose.py b/test/ir/inference/test_trt_convert_depthwise_conv2d_transpose.py index d346e001d09..ce97776d292 100644 --- a/test/ir/inference/test_trt_convert_depthwise_conv2d_transpose.py +++ b/test/ir/inference/test_trt_convert_depthwise_conv2d_transpose.py @@ -154,10 +154,12 @@ class TrtConvertDepthwiseConv2dTransposeTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -168,10 +170,12 @@ class TrtConvertDepthwiseConv2dTransposeTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_dropout.py b/test/ir/inference/test_trt_convert_dropout.py index 6d2509cefd7..7fcae9e6cfa 100644 --- a/test/ir/inference/test_trt_convert_dropout.py +++ b/test/ir/inference/test_trt_convert_dropout.py @@ -137,10 +137,12 @@ class TrtConvertDropoutTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -148,10 +150,12 @@ class TrtConvertDropoutTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_einsum.py b/test/ir/inference/test_trt_convert_einsum.py index 6f1fb5ebdd4..593d126aa7e 100644 --- a/test/ir/inference/test_trt_convert_einsum.py +++ b/test/ir/inference/test_trt_convert_einsum.py @@ -146,10 +146,12 @@ class TrtConvertEinsumTest_SingleOperand(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 @@ -157,10 +159,12 @@ class TrtConvertEinsumTest_SingleOperand(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 @@ -326,10 +330,12 @@ class TrtConvertEinsumTest_DoubuleOperand_Vector_Matrix(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 @@ -337,10 +343,12 @@ class TrtConvertEinsumTest_DoubuleOperand_Vector_Matrix(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 @@ -456,10 +464,12 @@ class TrtConvertEinsumTest_DoubuleOperand_Matrix_Matrix(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 @@ -467,10 +477,12 @@ class TrtConvertEinsumTest_DoubuleOperand_Matrix_Matrix(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 diff --git a/test/ir/inference/test_trt_convert_elementwise.py b/test/ir/inference/test_trt_convert_elementwise.py index 0ac4a2ba462..b5ca0d05d47 100644 --- a/test/ir/inference/test_trt_convert_elementwise.py +++ b/test/ir/inference/test_trt_convert_elementwise.py @@ -139,10 +139,12 @@ class TrtConvertElementwiseTestOneInputSpecialCase0(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -150,10 +152,12 @@ class TrtConvertElementwiseTestOneInputSpecialCase0(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) @@ -270,10 +274,12 @@ class TrtConvertElementwiseTestOneInputSpecialCase1(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -281,10 +287,12 @@ class TrtConvertElementwiseTestOneInputSpecialCase1(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) @@ -431,10 +439,12 @@ class TrtConvertElementwiseTestOneInput(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -442,10 +452,12 @@ class TrtConvertElementwiseTestOneInput(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) @@ -599,10 +611,12 @@ class TrtConvertElementwiseTestTwoInputWithoutBroadcast(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -610,8 +624,10 @@ class TrtConvertElementwiseTestTwoInputWithoutBroadcast(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 3), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), (1e-3, 1e-3) def add_skip_trt_case(self): @@ -770,15 +786,19 @@ class TrtConvertElementwiseTestTwoInputWithBroadcast(TrtLayerAutoScanTest): clear_dynamic_shape() if self.shape1[0] == self.shape2[0]: self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 3), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), (1e-3, 1e-3) # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 3), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), (1e-3, 1e-3) def add_skip_trt_case(self): @@ -921,15 +941,19 @@ class TrtConvertElementwiseTestOneInputCornerCase(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (0, 3), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (0, 3), (1e-3, 1e-3) # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 2), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 2), (1e-3, 1e-3) def add_skip_trt_case(self): @@ -1069,10 +1093,12 @@ class TrtConvertElementwiseTestTwoInputSkipCase(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -1080,8 +1106,10 @@ class TrtConvertElementwiseTestTwoInputSkipCase(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (0, 4), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (0, 4), (1e-3, 1e-3) def add_skip_trt_case(self): @@ -1187,10 +1215,12 @@ class TrtConvertPowOp(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -1198,10 +1228,12 @@ class TrtConvertPowOp(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) @@ -1347,10 +1379,12 @@ class TrtConvertElementwise0D(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -1358,10 +1392,12 @@ class TrtConvertElementwise0D(TrtLayerAutoScanTest): # # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_elementwiseadd_transpose.py b/test/ir/inference/test_trt_convert_elementwiseadd_transpose.py index 83d5855464f..dd9c495d497 100644 --- a/test/ir/inference/test_trt_convert_elementwiseadd_transpose.py +++ b/test/ir/inference/test_trt_convert_elementwiseadd_transpose.py @@ -162,10 +162,12 @@ class TrtConvertElementwiseaddTransposeTest(TrtLayerAutoScanTest): # just support dynamic_shape generate_dynamic_shape(attrs, inputs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), ( diff --git a/test/ir/inference/test_trt_convert_emb_eltwise_layernorm.py b/test/ir/inference/test_trt_convert_emb_eltwise_layernorm.py index cfaf3b9843e..5e619a1f5b8 100644 --- a/test/ir/inference/test_trt_convert_emb_eltwise_layernorm.py +++ b/test/ir/inference/test_trt_convert_emb_eltwise_layernorm.py @@ -293,15 +293,19 @@ class TrtConvertEmbEltwiseLayernormTest1(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (0, 5), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (0, 5), 2e-2 # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 4), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 4), 2e-2 def test(self): diff --git a/test/ir/inference/test_trt_convert_equal.py b/test/ir/inference/test_trt_convert_equal.py index 5879a003d95..72e8a1574e7 100644 --- a/test/ir/inference/test_trt_convert_equal.py +++ b/test/ir/inference/test_trt_convert_equal.py @@ -157,10 +157,12 @@ class TrtConvertEqualOneInputCornerCase(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -168,10 +170,12 @@ class TrtConvertEqualOneInputCornerCase(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_expand_as_v2.py b/test/ir/inference/test_trt_convert_expand_as_v2.py index be5458cac07..05b3cade8b4 100644 --- a/test/ir/inference/test_trt_convert_expand_as_v2.py +++ b/test/ir/inference/test_trt_convert_expand_as_v2.py @@ -154,10 +154,12 @@ class TrtConvertExpandASV2Test(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 @@ -244,9 +246,11 @@ class TrtConvertExpandV2Test2(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) # fill_constant will be folded by constnt folding pass! yield self.create_inference_config(), (1, 2), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 2), 1e-3 def add_skip_trt_case(self): diff --git a/test/ir/inference/test_trt_convert_expand_v2.py b/test/ir/inference/test_trt_convert_expand_v2.py index 49330b1fffc..5b6cc1c885b 100644 --- a/test/ir/inference/test_trt_convert_expand_v2.py +++ b/test/ir/inference/test_trt_convert_expand_v2.py @@ -147,10 +147,12 @@ class TrtConvertExpandV2Test(TrtLayerAutoScanTest): clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -158,10 +160,12 @@ class TrtConvertExpandV2Test(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 @@ -248,9 +252,11 @@ class TrtConvertExpandV2Test2(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) # fill_constant will be folded by constnt folding pass! yield self.create_inference_config(), (1, 2), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 2), 1e-3 def add_skip_trt_case(self): @@ -392,9 +398,11 @@ class TrtConvertExpandV2Test3(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) # fill_constant will be folded by constnt folding pass! yield self.create_inference_config(), (1, 2), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 2), 1e-3 def add_skip_trt_case(self): diff --git a/test/ir/inference/test_trt_convert_fill_any_like.py b/test/ir/inference/test_trt_convert_fill_any_like.py index 33871a58676..1c3dddf2c36 100644 --- a/test/ir/inference/test_trt_convert_fill_any_like.py +++ b/test/ir/inference/test_trt_convert_fill_any_like.py @@ -167,10 +167,12 @@ class TrtConvertExpandV2Test(TrtLayerAutoScanTest): clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 @@ -178,10 +180,12 @@ class TrtConvertExpandV2Test(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 diff --git a/test/ir/inference/test_trt_convert_fill_constant.py b/test/ir/inference/test_trt_convert_fill_constant.py index 767f460d4ca..73d15d26ab8 100644 --- a/test/ir/inference/test_trt_convert_fill_constant.py +++ b/test/ir/inference/test_trt_convert_fill_constant.py @@ -142,10 +142,12 @@ class TrtConvertFillConstantTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_flash_multihead_matmul.py b/test/ir/inference/test_trt_convert_flash_multihead_matmul.py index ebede2dce89..4d06e4764f2 100644 --- a/test/ir/inference/test_trt_convert_flash_multihead_matmul.py +++ b/test/ir/inference/test_trt_convert_flash_multihead_matmul.py @@ -265,17 +265,21 @@ class TrtConvertFlashMultiHeadMatmulTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) self.trt_param.workspace_size = 2013265920 yield self.create_inference_config(), (1, 2), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 2), (1e-3, 1e-3) # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) self.trt_param.workspace_size = 2013265920 yield self.create_inference_config(), (1, 2), (1e-5, 1e-4) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 2), (1e-2, 1e-3) def add_skip_trt_case(self): diff --git a/test/ir/inference/test_trt_convert_flatten.py b/test/ir/inference/test_trt_convert_flatten.py index 40546a7fbd3..636502aa231 100644 --- a/test/ir/inference/test_trt_convert_flatten.py +++ b/test/ir/inference/test_trt_convert_flatten.py @@ -101,10 +101,12 @@ class TrtConvertFlattenTest_dim_2(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -112,10 +114,12 @@ class TrtConvertFlattenTest_dim_2(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) @@ -202,10 +206,12 @@ class TrtConvertFlattenTest_dim_3(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -213,10 +219,12 @@ class TrtConvertFlattenTest_dim_3(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) @@ -303,10 +311,12 @@ class TrtConvertFlattenTest_dim_4(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -314,10 +324,12 @@ class TrtConvertFlattenTest_dim_4(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) @@ -404,10 +416,12 @@ class TrtConvertFlattenTest_dim_5(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -415,10 +429,12 @@ class TrtConvertFlattenTest_dim_5(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_flatten_contiguous_range.py b/test/ir/inference/test_trt_convert_flatten_contiguous_range.py index f6d5130ed07..df7d6102710 100644 --- a/test/ir/inference/test_trt_convert_flatten_contiguous_range.py +++ b/test/ir/inference/test_trt_convert_flatten_contiguous_range.py @@ -130,10 +130,12 @@ class TrtConvertFlattenContiguousRangeTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -141,10 +143,12 @@ class TrtConvertFlattenContiguousRangeTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_flip.py b/test/ir/inference/test_trt_convert_flip.py index 442a3381e38..d56cf50f7c9 100644 --- a/test/ir/inference/test_trt_convert_flip.py +++ b/test/ir/inference/test_trt_convert_flip.py @@ -124,10 +124,12 @@ class TrtConvertFlipTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_fused_token_prune.py b/test/ir/inference/test_trt_convert_fused_token_prune.py index 93a5253aad7..672812edb1d 100644 --- a/test/ir/inference/test_trt_convert_fused_token_prune.py +++ b/test/ir/inference/test_trt_convert_fused_token_prune.py @@ -121,10 +121,12 @@ class TrtConvertFusedTokenPruneTest(TrtLayerAutoScanTest): generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-2, 1e-2) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-1, 1e-2) diff --git a/test/ir/inference/test_trt_convert_gather.py b/test/ir/inference/test_trt_convert_gather.py index 69a2624b77e..fbf5c818b01 100644 --- a/test/ir/inference/test_trt_convert_gather.py +++ b/test/ir/inference/test_trt_convert_gather.py @@ -194,10 +194,12 @@ class TrtConvertGatherTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( False ), 1e-3 @@ -205,8 +207,10 @@ class TrtConvertGatherTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num(True), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num(True), 1e-3 def add_skip_trt_case(self): diff --git a/test/ir/inference/test_trt_convert_gather_nd.py b/test/ir/inference/test_trt_convert_gather_nd.py index 8b32e5516b9..ee0d0917c72 100644 --- a/test/ir/inference/test_trt_convert_gather_nd.py +++ b/test/ir/inference/test_trt_convert_gather_nd.py @@ -93,15 +93,19 @@ class TrtConvertGatherNdTest_dim_4_1(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (0, 4), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (0, 4), 1e-3 # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 3), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), 1e-3 def add_skip_trt_case(self): @@ -183,15 +187,19 @@ class TrtConvertGatherNdTest_dim_4_1_2(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (0, 4), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (0, 4), 1e-3 # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 3), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), 1e-3 def add_skip_trt_case(self): @@ -273,15 +281,19 @@ class TrtConvertGatherNdTest_dim_4_2(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (0, 4), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (0, 4), 1e-3 # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 3), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), 1e-3 def add_skip_trt_case(self): @@ -363,15 +375,19 @@ class TrtConvertGatherNdTest_dim_4_3(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (0, 4), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (0, 4), 1e-3 # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 3), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), 1e-3 def add_skip_trt_case(self): @@ -453,15 +469,19 @@ class TrtConvertGatherNdTest_dim_2_2(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (0, 4), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (0, 4), 1e-3 # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 3), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), 1e-3 def add_skip_trt_case(self): @@ -545,15 +565,19 @@ class TrtConvertGatherNdTest_dim_3_3(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (0, 4), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (0, 4), 1e-3 # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 3), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), 1e-3 def test(self): diff --git a/test/ir/inference/test_trt_convert_gelu.py b/test/ir/inference/test_trt_convert_gelu.py index 1f3847ff207..aa64994ad4c 100644 --- a/test/ir/inference/test_trt_convert_gelu.py +++ b/test/ir/inference/test_trt_convert_gelu.py @@ -127,10 +127,12 @@ class TrtConvertGeluTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -138,10 +140,12 @@ class TrtConvertGeluTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_grid_sampler.py b/test/ir/inference/test_trt_convert_grid_sampler.py index 49d2248a502..5c40eb6e5f7 100644 --- a/test/ir/inference/test_trt_convert_grid_sampler.py +++ b/test/ir/inference/test_trt_convert_grid_sampler.py @@ -138,8 +138,10 @@ class TrtConvertGridSampler(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 3), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), 1e-3 def test(self): diff --git a/test/ir/inference/test_trt_convert_group_norm.py b/test/ir/inference/test_trt_convert_group_norm.py index bd4652c4bbd..49993c43f14 100644 --- a/test/ir/inference/test_trt_convert_group_norm.py +++ b/test/ir/inference/test_trt_convert_group_norm.py @@ -125,11 +125,13 @@ class TrtConvertGroupNormTest(TrtLayerAutoScanTest): clear_dynamic_shape() self.trt_param.workspace_size = 2013265920 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 @@ -138,11 +140,13 @@ class TrtConvertGroupNormTest(TrtLayerAutoScanTest): self.trt_param.workspace_size = 2013265920 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 diff --git a/test/ir/inference/test_trt_convert_hard_sigmoid.py b/test/ir/inference/test_trt_convert_hard_sigmoid.py index e65b1e1e32b..69d7594fd43 100644 --- a/test/ir/inference/test_trt_convert_hard_sigmoid.py +++ b/test/ir/inference/test_trt_convert_hard_sigmoid.py @@ -97,15 +97,19 @@ class TrtConvertHardSigmoidTest_dim_2(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 2), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 2), 1e-3 # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 2), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 2), 1e-3 def test(self): diff --git a/test/ir/inference/test_trt_convert_hard_swish.py b/test/ir/inference/test_trt_convert_hard_swish.py index 3c3a98ee1e6..381712bb596 100644 --- a/test/ir/inference/test_trt_convert_hard_swish.py +++ b/test/ir/inference/test_trt_convert_hard_swish.py @@ -97,10 +97,12 @@ class TrtConvertHardSwishTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -108,10 +110,12 @@ class TrtConvertHardSwishTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_index_select.py b/test/ir/inference/test_trt_convert_index_select.py index 9d1d9fb26ba..c43347ba19a 100644 --- a/test/ir/inference/test_trt_convert_index_select.py +++ b/test/ir/inference/test_trt_convert_index_select.py @@ -171,10 +171,12 @@ class TrtConvertIndexSelectTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( False ), 1e-3 @@ -182,8 +184,10 @@ class TrtConvertIndexSelectTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num(True), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num(True), 1e-3 def test(self): diff --git a/test/ir/inference/test_trt_convert_instance_norm.py b/test/ir/inference/test_trt_convert_instance_norm.py index 72b728d5cc3..3200003821d 100644 --- a/test/ir/inference/test_trt_convert_instance_norm.py +++ b/test/ir/inference/test_trt_convert_instance_norm.py @@ -145,10 +145,12 @@ class TrtConvertInstanceNormTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -156,10 +158,12 @@ class TrtConvertInstanceNormTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_inverse.py b/test/ir/inference/test_trt_convert_inverse.py index 87adffe106d..e36d7429409 100644 --- a/test/ir/inference/test_trt_convert_inverse.py +++ b/test/ir/inference/test_trt_convert_inverse.py @@ -82,15 +82,19 @@ class TrtConvertInverse(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (0, 3), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (0, 3), 1e-3 # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 2), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 2), 1e-3 def test(self): diff --git a/test/ir/inference/test_trt_convert_layer_norm.py b/test/ir/inference/test_trt_convert_layer_norm.py index f3c2cbfeff3..63a7c9ac627 100644 --- a/test/ir/inference/test_trt_convert_layer_norm.py +++ b/test/ir/inference/test_trt_convert_layer_norm.py @@ -125,10 +125,12 @@ class TrtConvertLayerNormTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 @@ -136,10 +138,12 @@ class TrtConvertLayerNormTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 @@ -249,10 +253,12 @@ class TrtConvertLayerNormTest_2(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 @@ -260,10 +266,12 @@ class TrtConvertLayerNormTest_2(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 diff --git a/test/ir/inference/test_trt_convert_leaky_relu.py b/test/ir/inference/test_trt_convert_leaky_relu.py index 683bace7b00..4fac037674d 100644 --- a/test/ir/inference/test_trt_convert_leaky_relu.py +++ b/test/ir/inference/test_trt_convert_leaky_relu.py @@ -103,14 +103,17 @@ class TrtConvertLeakyReluTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) self.trt_param.precision = paddle_infer.PrecisionType.Int8 + program_config.set_input_type(np.int8) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -118,14 +121,17 @@ class TrtConvertLeakyReluTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) self.trt_param.precision = paddle_infer.PrecisionType.Int8 + program_config.set_input_type(np.int8) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_lookup_table.py b/test/ir/inference/test_trt_convert_lookup_table.py index dd5ee6d8fce..e1fb64bcdf5 100644 --- a/test/ir/inference/test_trt_convert_lookup_table.py +++ b/test/ir/inference/test_trt_convert_lookup_table.py @@ -138,10 +138,12 @@ class TrtConvertLookupTableV2Test(TrtLayerAutoScanTest): # for dynamic_shape mode generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_lookup_table_v2.py b/test/ir/inference/test_trt_convert_lookup_table_v2.py index ecc97fc3b94..4f2722a27f9 100644 --- a/test/ir/inference/test_trt_convert_lookup_table_v2.py +++ b/test/ir/inference/test_trt_convert_lookup_table_v2.py @@ -128,10 +128,12 @@ class TrtConvertLookupTableV2Test(TrtLayerAutoScanTest): # for dynamic_shape mode generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_matmul.py b/test/ir/inference/test_trt_convert_matmul.py index 51654282729..a529648c65c 100644 --- a/test/ir/inference/test_trt_convert_matmul.py +++ b/test/ir/inference/test_trt_convert_matmul.py @@ -101,8 +101,10 @@ class TrtConvertMatmulTest_static(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 3), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), 1e-3 def test(self): @@ -192,8 +194,10 @@ class TrtConvertMatmulTest_dynamic(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 3), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), 1e-3 def add_skip_trt_case(self): diff --git a/test/ir/inference/test_trt_convert_matmul_v2.py b/test/ir/inference/test_trt_convert_matmul_v2.py index f3dae2b3160..b511d4a0abf 100644 --- a/test/ir/inference/test_trt_convert_matmul_v2.py +++ b/test/ir/inference/test_trt_convert_matmul_v2.py @@ -96,8 +96,10 @@ class TrtConvertMatmulTest_dynamic(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 3), (tol_fp32, tol_fp32) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), (tol_half, tol_half) def add_skip_trt_case(self): @@ -182,8 +184,10 @@ class TrtConvertMatmulTest_dynamic2(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 3), (tol_fp32, tol_fp32) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), (tol_half, tol_half) def add_skip_trt_case(self): @@ -311,8 +315,10 @@ class TrtConvertMatmulTest_dynamic3(TrtLayerAutoScanTest): generate_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 3), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), 1e-3 def add_skip_trt_case(self): diff --git a/test/ir/inference/test_trt_convert_mish.py b/test/ir/inference/test_trt_convert_mish.py index 36375b67031..8206b089c66 100644 --- a/test/ir/inference/test_trt_convert_mish.py +++ b/test/ir/inference/test_trt_convert_mish.py @@ -139,10 +139,12 @@ class TrtConvertMishTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -150,10 +152,12 @@ class TrtConvertMishTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_multiclass_nms.py b/test/ir/inference/test_trt_convert_multiclass_nms.py index 20950d2564b..3e40c669935 100644 --- a/test/ir/inference/test_trt_convert_multiclass_nms.py +++ b/test/ir/inference/test_trt_convert_multiclass_nms.py @@ -160,10 +160,12 @@ class TrtConvertMulticlassNMSTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 @@ -171,6 +173,7 @@ class TrtConvertMulticlassNMSTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 diff --git a/test/ir/inference/test_trt_convert_multiclass_nms3.py b/test/ir/inference/test_trt_convert_multiclass_nms3.py index 89f40072986..60f2b0a68a4 100644 --- a/test/ir/inference/test_trt_convert_multiclass_nms3.py +++ b/test/ir/inference/test_trt_convert_multiclass_nms3.py @@ -167,10 +167,12 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 @@ -178,6 +180,7 @@ class TrtConvertMulticlassNMS3Test(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 diff --git a/test/ir/inference/test_trt_convert_multihead_matmul.py b/test/ir/inference/test_trt_convert_multihead_matmul.py index 4a29253f2c8..3d3100a70e7 100644 --- a/test/ir/inference/test_trt_convert_multihead_matmul.py +++ b/test/ir/inference/test_trt_convert_multihead_matmul.py @@ -385,9 +385,11 @@ class TrtConvertMultiHeadMatmulTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) self.trt_param.workspace_size = 2013265920 yield self.create_inference_config(), (1, 3), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), (1e-2, 1e-2) def test(self): @@ -983,16 +985,19 @@ class TrtConvertVitToMultiHeadMatmulTest(TrtLayerAutoScanTest): generate_dynamic_shape(attrs) self.trt_param.workspace_size = 2013265920 self.trt_param.precision = paddle_infer.PrecisionType.Int8 + program_config.set_input_type(np.int8) yield self.create_inference_config(), generate_trt_nodes_num(), ( 1e-3, 1e-3, ) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num(), ( 1e-3, 2e-2, ) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num(), ( 1e-5, 1e-5, @@ -1364,9 +1369,11 @@ class TrtConvertMultiHeadMatmulTest_biasqk_seqseq(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) self.trt_param.workspace_size = 2013265920 yield self.create_inference_config(), (1, 3), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), (1e-2, 1e-2) def test(self): diff --git a/test/ir/inference/test_trt_convert_multihead_matmul_roformer.py b/test/ir/inference/test_trt_convert_multihead_matmul_roformer.py index 2de959a6ee7..f1f8ecdc374 100644 --- a/test/ir/inference/test_trt_convert_multihead_matmul_roformer.py +++ b/test/ir/inference/test_trt_convert_multihead_matmul_roformer.py @@ -540,9 +540,11 @@ class TrtConvertMultiHeadMatmulRoformerTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) self.trt_param.workspace_size = 2013265920 yield self.create_inference_config(), (1, 5), (1e-3, 1e-3) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 5), (1e-3, 1e-3) def test(self): diff --git a/test/ir/inference/test_trt_convert_nearest_interp.py b/test/ir/inference/test_trt_convert_nearest_interp.py index e3dd3c74740..a96cbb9f81a 100644 --- a/test/ir/inference/test_trt_convert_nearest_interp.py +++ b/test/ir/inference/test_trt_convert_nearest_interp.py @@ -113,10 +113,12 @@ class TrtConvertNearestInterpTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 @@ -124,10 +126,12 @@ class TrtConvertNearestInterpTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 diff --git a/test/ir/inference/test_trt_convert_nearest_interp_v2.py b/test/ir/inference/test_trt_convert_nearest_interp_v2.py index 7a942168176..d83ec7ecbe9 100644 --- a/test/ir/inference/test_trt_convert_nearest_interp_v2.py +++ b/test/ir/inference/test_trt_convert_nearest_interp_v2.py @@ -83,10 +83,12 @@ class TrtConvertNearestInterpV2Test(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 @@ -94,10 +96,12 @@ class TrtConvertNearestInterpV2Test(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 @@ -174,10 +178,12 @@ class TrtConvertNearestInterpV2ShapeTensorTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 @@ -185,10 +191,12 @@ class TrtConvertNearestInterpV2ShapeTensorTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 diff --git a/test/ir/inference/test_trt_convert_one_hot.py b/test/ir/inference/test_trt_convert_one_hot.py index 5332cb950ae..72ca495ef8b 100644 --- a/test/ir/inference/test_trt_convert_one_hot.py +++ b/test/ir/inference/test_trt_convert_one_hot.py @@ -141,10 +141,12 @@ class TrtConvertOneHotTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 @@ -152,10 +154,12 @@ class TrtConvertOneHotTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 diff --git a/test/ir/inference/test_trt_convert_p_norm.py b/test/ir/inference/test_trt_convert_p_norm.py index 6fc8e5ec87c..bf5adf9715b 100644 --- a/test/ir/inference/test_trt_convert_p_norm.py +++ b/test/ir/inference/test_trt_convert_p_norm.py @@ -118,10 +118,12 @@ class TrtConvertPNormTest(TrtLayerAutoScanTest): # for dynamic_shape mode generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_pad.py b/test/ir/inference/test_trt_convert_pad.py index ecfc09090b7..f20b915c2a4 100644 --- a/test/ir/inference/test_trt_convert_pad.py +++ b/test/ir/inference/test_trt_convert_pad.py @@ -104,10 +104,12 @@ class TrtConvertPadTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 @@ -115,10 +117,12 @@ class TrtConvertPadTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 diff --git a/test/ir/inference/test_trt_convert_pad3d.py b/test/ir/inference/test_trt_convert_pad3d.py index df0426b2349..1237c26c474 100644 --- a/test/ir/inference/test_trt_convert_pad3d.py +++ b/test/ir/inference/test_trt_convert_pad3d.py @@ -120,10 +120,12 @@ class TrtConvertPad3dTensorPadding(TrtLayerAutoScanTest): clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -131,10 +133,12 @@ class TrtConvertPad3dTensorPadding(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 @@ -230,10 +234,12 @@ class TrtConvertPad3dListPadding(TrtLayerAutoScanTest): clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -241,10 +247,12 @@ class TrtConvertPad3dListPadding(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_pool2d.py b/test/ir/inference/test_trt_convert_pool2d.py index a44273c069b..669a83b7f2d 100644 --- a/test/ir/inference/test_trt_convert_pool2d.py +++ b/test/ir/inference/test_trt_convert_pool2d.py @@ -153,10 +153,12 @@ class TrtConvertPool2dTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -164,10 +166,12 @@ class TrtConvertPool2dTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_preln_residual_bias.py b/test/ir/inference/test_trt_convert_preln_residual_bias.py index a45ddfcae18..f9431e07511 100644 --- a/test/ir/inference/test_trt_convert_preln_residual_bias.py +++ b/test/ir/inference/test_trt_convert_preln_residual_bias.py @@ -169,10 +169,12 @@ class TrtConvertSkipLayernormTest(TrtLayerAutoScanTest): # for static_shape, fall back to fluid fused op clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 # atol=1e-2 while rtol is 1e-8 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 # atol=1e-2 while rtol is 1e-8 @@ -180,10 +182,12 @@ class TrtConvertSkipLayernormTest(TrtLayerAutoScanTest): # just support dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 # atol=1e-2 while rtol is 1e-8 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 # atol=1e-2 while rtol is 1e-8 diff --git a/test/ir/inference/test_trt_convert_preln_residual_no_bias.py b/test/ir/inference/test_trt_convert_preln_residual_no_bias.py index fd3bdb64c7e..7be194a953a 100644 --- a/test/ir/inference/test_trt_convert_preln_residual_no_bias.py +++ b/test/ir/inference/test_trt_convert_preln_residual_no_bias.py @@ -158,10 +158,12 @@ class TrtConvertSkipLayernormTest(TrtLayerAutoScanTest): # for static_shape, fall back to fluid fused op clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 # atol=1e-2 while rtol is 1e-8 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 # atol=1e-2 while rtol is 1e-8 @@ -169,10 +171,12 @@ class TrtConvertSkipLayernormTest(TrtLayerAutoScanTest): # just support dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 # atol=1e-2 while rtol is 1e-8 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 # atol=1e-2 while rtol is 1e-8 diff --git a/test/ir/inference/test_trt_convert_prelu.py b/test/ir/inference/test_trt_convert_prelu.py index ae53f10ad57..7c627f957ab 100644 --- a/test/ir/inference/test_trt_convert_prelu.py +++ b/test/ir/inference/test_trt_convert_prelu.py @@ -205,10 +205,12 @@ class TrtConvertPreluTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -216,10 +218,12 @@ class TrtConvertPreluTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_qk_multihead_matmul.py b/test/ir/inference/test_trt_convert_qk_multihead_matmul.py index 548f0486e12..ece654687fe 100644 --- a/test/ir/inference/test_trt_convert_qk_multihead_matmul.py +++ b/test/ir/inference/test_trt_convert_qk_multihead_matmul.py @@ -329,17 +329,21 @@ class TrtConvertQkAttentionTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) self.trt_param.workspace_size = 2013265920 yield self.create_inference_config(), (1, 3), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), (1e-3, 1e-3) # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) self.trt_param.workspace_size = 2013265920 yield self.create_inference_config(), (1, 3), (1e-5, 1e-4) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 3), (1e-2, 1e-3) def add_skip_trt_case(self): diff --git a/test/ir/inference/test_trt_convert_range.py b/test/ir/inference/test_trt_convert_range.py index 5ef6b8b9052..d8b139925b0 100644 --- a/test/ir/inference/test_trt_convert_range.py +++ b/test/ir/inference/test_trt_convert_range.py @@ -130,10 +130,12 @@ class TrtConvertRangeDynamicTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-2 @@ -214,10 +216,12 @@ class TrtConvertRangeStaticTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-2 diff --git a/test/ir/inference/test_trt_convert_reduce.py b/test/ir/inference/test_trt_convert_reduce.py index 59acef17964..57de3c5603b 100644 --- a/test/ir/inference/test_trt_convert_reduce.py +++ b/test/ir/inference/test_trt_convert_reduce.py @@ -164,10 +164,12 @@ class TrtConvertReduceTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -175,10 +177,12 @@ class TrtConvertReduceTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-5, 1e-5) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_reshape.py b/test/ir/inference/test_trt_convert_reshape.py index c30d973651b..5762d51c348 100644 --- a/test/ir/inference/test_trt_convert_reshape.py +++ b/test/ir/inference/test_trt_convert_reshape.py @@ -172,10 +172,12 @@ class TrtConvertReshapeTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -183,10 +185,12 @@ class TrtConvertReshapeTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 @@ -308,8 +312,10 @@ class TrtConvertReshapeTest2(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 2), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 2), 1e-3 def add_skip_trt_case(self): @@ -419,8 +425,10 @@ class TrtConvertReshapeTest3(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 2), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 2), 1e-3 def add_skip_trt_case(self): @@ -509,10 +517,12 @@ class TrtConvertReshapeZeroDimsTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_rnn.py b/test/ir/inference/test_trt_convert_rnn.py index 296da4db148..b00ac38ac0e 100644 --- a/test/ir/inference/test_trt_convert_rnn.py +++ b/test/ir/inference/test_trt_convert_rnn.py @@ -264,10 +264,12 @@ class TrtConvertSliceTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), tol_fp32 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), tol_half diff --git a/test/ir/inference/test_trt_convert_roi_align.py b/test/ir/inference/test_trt_convert_roi_align.py index 1e1a83a40e4..4a3558ac461 100644 --- a/test/ir/inference/test_trt_convert_roi_align.py +++ b/test/ir/inference/test_trt_convert_roi_align.py @@ -190,10 +190,12 @@ class TrtConvertRoiAlignTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -201,10 +203,12 @@ class TrtConvertRoiAlignTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_roll.py b/test/ir/inference/test_trt_convert_roll.py index e5267de3e59..21e265a4679 100644 --- a/test/ir/inference/test_trt_convert_roll.py +++ b/test/ir/inference/test_trt_convert_roll.py @@ -107,10 +107,12 @@ class TrtConvertRollTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -118,10 +120,12 @@ class TrtConvertRollTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_scale.py b/test/ir/inference/test_trt_convert_scale.py index 55d1752d499..c3916a65aa6 100644 --- a/test/ir/inference/test_trt_convert_scale.py +++ b/test/ir/inference/test_trt_convert_scale.py @@ -165,10 +165,12 @@ class TrtConvertScaleTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -176,10 +178,12 @@ class TrtConvertScaleTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_scatter_nd_add.py b/test/ir/inference/test_trt_convert_scatter_nd_add.py index 66f25781e6b..7667b40331e 100644 --- a/test/ir/inference/test_trt_convert_scatter_nd_add.py +++ b/test/ir/inference/test_trt_convert_scatter_nd_add.py @@ -102,15 +102,19 @@ class TrtConvertScatterNd(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (0, 5), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (0, 5), 1e-3 # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 4), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 4), 1e-3 def test(self): diff --git a/test/ir/inference/test_trt_convert_set_value.py b/test/ir/inference/test_trt_convert_set_value.py index 569a515e9ad..39f2d1c228b 100644 --- a/test/ir/inference/test_trt_convert_set_value.py +++ b/test/ir/inference/test_trt_convert_set_value.py @@ -154,11 +154,13 @@ class TrtConvertSetValue(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) self.trt_param.workspace_size = 2013265920 yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-5, 1e-4) self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_shape.py b/test/ir/inference/test_trt_convert_shape.py index 1f2a97018e2..f29473d9ded 100644 --- a/test/ir/inference/test_trt_convert_shape.py +++ b/test/ir/inference/test_trt_convert_shape.py @@ -99,10 +99,12 @@ class TrtConvertSumTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( False ), 1e-3 @@ -110,8 +112,10 @@ class TrtConvertSumTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num(True), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num(True), 1e-3 def test(self): diff --git a/test/ir/inference/test_trt_convert_shuffle_channel.py b/test/ir/inference/test_trt_convert_shuffle_channel.py index f61ecfbeea0..725a6c67c37 100644 --- a/test/ir/inference/test_trt_convert_shuffle_channel.py +++ b/test/ir/inference/test_trt_convert_shuffle_channel.py @@ -92,10 +92,12 @@ class TrtConvertShuffleChannelTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -103,10 +105,12 @@ class TrtConvertShuffleChannelTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_slice.py b/test/ir/inference/test_trt_convert_slice.py index ac5b6439030..66458d53a12 100644 --- a/test/ir/inference/test_trt_convert_slice.py +++ b/test/ir/inference/test_trt_convert_slice.py @@ -131,10 +131,12 @@ class TrtConvertSliceTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -142,10 +144,12 @@ class TrtConvertSliceTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_softmax.py b/test/ir/inference/test_trt_convert_softmax.py index 7fb9a69bb63..71334094686 100644 --- a/test/ir/inference/test_trt_convert_softmax.py +++ b/test/ir/inference/test_trt_convert_softmax.py @@ -135,10 +135,12 @@ class TrtConvertSoftmaxTest(TrtLayerAutoScanTest): pass else: self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -146,10 +148,12 @@ class TrtConvertSoftmaxTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_split.py b/test/ir/inference/test_trt_convert_split.py index bfb7099f16f..e29cb2ddef1 100644 --- a/test/ir/inference/test_trt_convert_split.py +++ b/test/ir/inference/test_trt_convert_split.py @@ -250,10 +250,12 @@ class TrtConvertSplitTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -261,10 +263,12 @@ class TrtConvertSplitTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 @@ -400,10 +404,12 @@ class TrtConvertSplitTest2(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_square.py b/test/ir/inference/test_trt_convert_square.py index 4dd30b4016f..a082726998b 100644 --- a/test/ir/inference/test_trt_convert_square.py +++ b/test/ir/inference/test_trt_convert_square.py @@ -117,10 +117,12 @@ class TrtConvertSquareTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -128,10 +130,12 @@ class TrtConvertSquareTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_squeeze2.py b/test/ir/inference/test_trt_convert_squeeze2.py index a24465428f6..3b73040c881 100644 --- a/test/ir/inference/test_trt_convert_squeeze2.py +++ b/test/ir/inference/test_trt_convert_squeeze2.py @@ -117,10 +117,12 @@ class TrtConvertSplitTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -128,10 +130,12 @@ class TrtConvertSplitTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_stack.py b/test/ir/inference/test_trt_convert_stack.py index 39f729e2be7..f83368e325f 100644 --- a/test/ir/inference/test_trt_convert_stack.py +++ b/test/ir/inference/test_trt_convert_stack.py @@ -218,10 +218,12 @@ class TrtConvertStackTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -229,10 +231,12 @@ class TrtConvertStackTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_strided_slice.py b/test/ir/inference/test_trt_convert_strided_slice.py index b814c9f9b57..e0e6198a16f 100644 --- a/test/ir/inference/test_trt_convert_strided_slice.py +++ b/test/ir/inference/test_trt_convert_strided_slice.py @@ -122,6 +122,7 @@ class TrtConvertStridedSliceTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 @@ -129,6 +130,7 @@ class TrtConvertStridedSliceTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 @@ -217,11 +219,13 @@ class TrtConvertStridedSliceTest2(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 2), 1e-5 # for dynamic_shape generate_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 2), 1e-5 def test(self): diff --git a/test/ir/inference/test_trt_convert_sum.py b/test/ir/inference/test_trt_convert_sum.py index 231921d641b..abcd0f8ed66 100644 --- a/test/ir/inference/test_trt_convert_sum.py +++ b/test/ir/inference/test_trt_convert_sum.py @@ -193,10 +193,12 @@ class TrtConvertSumTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( False ), 1e-3 @@ -204,8 +206,10 @@ class TrtConvertSumTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num(True), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num(True), 1e-3 def test(self): @@ -311,10 +315,12 @@ class TrtConvertSumTest1(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( False ), 1e-3 @@ -322,8 +328,10 @@ class TrtConvertSumTest1(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num(True), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num(True), 1e-3 def test(self): diff --git a/test/ir/inference/test_trt_convert_swish.py b/test/ir/inference/test_trt_convert_swish.py index 3db16d47cda..623cada1d46 100755 --- a/test/ir/inference/test_trt_convert_swish.py +++ b/test/ir/inference/test_trt_convert_swish.py @@ -121,10 +121,12 @@ class TrtConvertSwishTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -132,10 +134,12 @@ class TrtConvertSwishTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_take_along_axis.py b/test/ir/inference/test_trt_convert_take_along_axis.py index 63b340f8807..3d269ea6da4 100644 --- a/test/ir/inference/test_trt_convert_take_along_axis.py +++ b/test/ir/inference/test_trt_convert_take_along_axis.py @@ -160,10 +160,12 @@ class TrtConvertTakeAlongAxisTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( False ), 1e-5 @@ -171,8 +173,10 @@ class TrtConvertTakeAlongAxisTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num(True), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num(True), 1e-3 def add_skip_trt_case(self): diff --git a/test/ir/inference/test_trt_convert_temporal_shift.py b/test/ir/inference/test_trt_convert_temporal_shift.py index b0b2ce51062..9ac5369264a 100755 --- a/test/ir/inference/test_trt_convert_temporal_shift.py +++ b/test/ir/inference/test_trt_convert_temporal_shift.py @@ -108,10 +108,12 @@ class TrtConvertTemporalShiftTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -119,10 +121,12 @@ class TrtConvertTemporalShiftTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_tile.py b/test/ir/inference/test_trt_convert_tile.py index 240fd65b839..ba512438e24 100644 --- a/test/ir/inference/test_trt_convert_tile.py +++ b/test/ir/inference/test_trt_convert_tile.py @@ -93,10 +93,12 @@ class TrtConvertTileTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -104,10 +106,12 @@ class TrtConvertTileTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 @@ -184,10 +188,12 @@ class TrtConvertTileTest2(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 @@ -283,10 +289,12 @@ class TrtConvertTileTest3(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_top_k.py b/test/ir/inference/test_trt_convert_top_k.py index 8c8bcbde993..239e361ebd3 100644 --- a/test/ir/inference/test_trt_convert_top_k.py +++ b/test/ir/inference/test_trt_convert_top_k.py @@ -118,10 +118,12 @@ class TrtConvertActivationTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -129,10 +131,12 @@ class TrtConvertActivationTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_top_k_v2.py b/test/ir/inference/test_trt_convert_top_k_v2.py index 29a04c5d049..9f20b6260ac 100644 --- a/test/ir/inference/test_trt_convert_top_k_v2.py +++ b/test/ir/inference/test_trt_convert_top_k_v2.py @@ -147,10 +147,12 @@ class TrtConvertTopKV2Test(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -158,10 +160,12 @@ class TrtConvertTopKV2Test(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_trans_layernorm.py b/test/ir/inference/test_trt_convert_trans_layernorm.py index 3d7236d2bba..e8623b771db 100644 --- a/test/ir/inference/test_trt_convert_trans_layernorm.py +++ b/test/ir/inference/test_trt_convert_trans_layernorm.py @@ -226,10 +226,12 @@ class TrtConvertTransLayernormTest(TrtLayerAutoScanTest): # just support dynamic_shape generate_dynamic_shape(attrs, inputs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), ( diff --git a/test/ir/inference/test_trt_convert_transpose.py b/test/ir/inference/test_trt_convert_transpose.py index 9a583537569..d0d3aa98313 100644 --- a/test/ir/inference/test_trt_convert_transpose.py +++ b/test/ir/inference/test_trt_convert_transpose.py @@ -139,10 +139,12 @@ class TrtConvertTransposeTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -150,10 +152,12 @@ class TrtConvertTransposeTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_unary.py b/test/ir/inference/test_trt_convert_unary.py index 1d83978f2f0..ba70dfb085d 100644 --- a/test/ir/inference/test_trt_convert_unary.py +++ b/test/ir/inference/test_trt_convert_unary.py @@ -208,10 +208,12 @@ class TrtConvertActivationTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-4 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -219,10 +221,12 @@ class TrtConvertActivationTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-4 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 @@ -342,10 +346,12 @@ class TrtConvertLogicalNotTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), (1e-3, 1e-3) @@ -353,10 +359,12 @@ class TrtConvertLogicalNotTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), (1e-3, 1e-3) diff --git a/test/ir/inference/test_trt_convert_unbind.py b/test/ir/inference/test_trt_convert_unbind.py index 875f37ba54a..91464aabeef 100644 --- a/test/ir/inference/test_trt_convert_unbind.py +++ b/test/ir/inference/test_trt_convert_unbind.py @@ -106,10 +106,12 @@ class TrtConvertUnbind(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_unfold.py b/test/ir/inference/test_trt_convert_unfold.py index 1f63a2eb97e..ea18d50e1a4 100644 --- a/test/ir/inference/test_trt_convert_unfold.py +++ b/test/ir/inference/test_trt_convert_unfold.py @@ -87,15 +87,19 @@ class TrtConvertUnfold(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (0, 3), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (0, 3), 1e-3 # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), (1, 2), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), (1, 2), 1e-3 def test(self): diff --git a/test/ir/inference/test_trt_convert_unsqueeze2.py b/test/ir/inference/test_trt_convert_unsqueeze2.py index 7e92c52b553..d1f27ace147 100644 --- a/test/ir/inference/test_trt_convert_unsqueeze2.py +++ b/test/ir/inference/test_trt_convert_unsqueeze2.py @@ -100,10 +100,12 @@ class TrtConvertSplitTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -111,10 +113,12 @@ class TrtConvertSplitTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_where.py b/test/ir/inference/test_trt_convert_where.py index 17c20e56de2..fafa1a137fa 100644 --- a/test/ir/inference/test_trt_convert_where.py +++ b/test/ir/inference/test_trt_convert_where.py @@ -206,10 +206,12 @@ class TrtConvertActivationTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 @@ -217,10 +219,12 @@ class TrtConvertActivationTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 diff --git a/test/ir/inference/test_trt_convert_yolo_box.py b/test/ir/inference/test_trt_convert_yolo_box.py index 0bc410cc81f..dc112ba02d7 100644 --- a/test/ir/inference/test_trt_convert_yolo_box.py +++ b/test/ir/inference/test_trt_convert_yolo_box.py @@ -153,10 +153,12 @@ class TrtConvertYoloBoxTest(TrtLayerAutoScanTest): # for static_shape clear_dynamic_shape() self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, False ), 1e-3 @@ -164,10 +166,12 @@ class TrtConvertYoloBoxTest(TrtLayerAutoScanTest): # for dynamic_shape generate_dynamic_shape(attrs) self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-5 self.trt_param.precision = paddle_infer.PrecisionType.Half + program_config.set_input_type(np.float16) yield self.create_inference_config(), generate_trt_nodes_num( attrs, True ), 1e-3 diff --git a/test/ir/inference/test_trt_convert_yolo_box_head.py b/test/ir/inference/test_trt_convert_yolo_box_head.py index cf31681b1ab..df53e2fec65 100644 --- a/test/ir/inference/test_trt_convert_yolo_box_head.py +++ b/test/ir/inference/test_trt_convert_yolo_box_head.py @@ -79,6 +79,7 @@ class TrtConvertYoloBoxHeadTest(TrtLayerAutoScanTest): ) -> (paddle_infer.Config, List[int], float): # for static_shape self.trt_param.precision = paddle_infer.PrecisionType.Float32 + program_config.set_input_type(np.float32) yield self.create_inference_config(), [1, 2], 1e-5 def test(self): -- GitLab