算子黑名单批量删除-ModelZoo-PyTorch-AtomGit

MPyTorch/built-in/cv/classification/Shufflenetv2_for_PyTorch/8p_main_med.py+0-3

	@@ -631,7 +631,4 @@ def get_pytorch_val_loader(data_path, batch_size, workers=5, _worker_init_fn=Non
631		631
632		632
633	if __name__ == '__main__':	633	if __name__ == '__main__':
634	- option = {}
635	- option["NPU_FUZZY_COMPILE_BLACKLIST"] = "BNTrainingUpdate"
636	- torch.npu.set_option(option)
637	main()	634	main()

MPyTorch/built-in/cv/detection/ABINet_for_PyTorch/tools/train.py+0-3

	@@ -35,9 +35,6 @@ from mmocr.utils import (collect_env, get_root_logger, is_2dlist,
35	import torch_npu	35	import torch_npu
36	from torch_npu.contrib import transfer_to_npu	36	from torch_npu.contrib import transfer_to_npu
37		37
38	-option = {}
39	-option["NPU_FUZZY_COMPILE_BLACKLIST"] = "MaskedFill"
40	-torch.npu.set_option(option)
41	torch.npu.set_compile_mode(jit_compile=False)	38	torch.npu.set_compile_mode(jit_compile=False)
42		39
43		40

MPyTorch/built-in/cv/detection/Faster_Mask_RCNN_for_PyTorch/tools/train_net.py+0-2

	@@ -199,8 +199,6 @@ if __name__ == "__main__":
199	option = {'ACL_PRECISION_MODE': 'allow_fp32_to_fp16'}	199	option = {'ACL_PRECISION_MODE': 'allow_fp32_to_fp16'}
200	if os.getenv('ALLOW_HF32', False):	200	if os.getenv('ALLOW_HF32', False):
201	option = {'ACL_PRECISION_MODE': 'must_keep_origin_dtype'}	201	option = {'ACL_PRECISION_MODE': 'must_keep_origin_dtype'}
202	- # LogSoftmaxV2算子在二进制下有精度问题，添加至黑名单规避
203	- option['NPU_FUZZY_COMPILE_BLACKLIST'] = 'LogSoftmaxV2'
204	torch_npu.npu.set_option(option)	202	torch_npu.npu.set_option(option)
205		203
206	launch(	204	launch(

MPyTorch/built-in/cv/detection/YOLOV4_ID0396_for_PyTorch/test.py+0-3

	@@ -313,9 +313,6 @@ if __name__ == '__main__':
313	print(opt)	313	print(opt)
314		314
315	torch.npu.set_compile_mode(jit_compile=False)	315	torch.npu.set_compile_mode(jit_compile=False)
316	- option = dict()
317	- option["NPU_FUZZY_COMPILE_BLACKLIST"] = "Identity"
318	- torch.npu.set_option(option)
319	if opt.task in ['val', 'test']: # run normally	316	if opt.task in ['val', 'test']: # run normally
320	test(opt.data,	317	test(opt.data,
321	opt.weights,	318	opt.weights,

MPyTorch/built-in/foundation/CodeGeeX2/ptuning/main.py+0-3

	@@ -34,9 +34,6 @@ import deepspeed_npu
34	from torch_npu.contrib import transfer_to_npu	34	from torch_npu.contrib import transfer_to_npu
35		35
36	torch.npu.set_compile_mode(jit_compile=True)	36	torch.npu.set_compile_mode(jit_compile=True)
37	-option = {"NPU_FUZZY_COMPILE_BLACKLIST":"Tril,LayerNormGrad"}
38	-torch.npu.set_option(option)
39	-
40		37
41	import transformers	38	import transformers
42	from transformers import (	39	from transformers import (

MPyTorch/built-in/foundation/CodeGeeX2/ptuning/main_without_tokenizer.py+0-2

	@@ -76,8 +76,6 @@ def main():
76	model_args, data_args, training_args = parser.parse_args_into_dataclasses()	76	model_args, data_args, training_args = parser.parse_args_into_dataclasses()
77	if training_args.do_train:	77	if training_args.do_train:
78	torch.npu.set_compile_mode(jit_compile=True)	78	torch.npu.set_compile_mode(jit_compile=True)
79	- option = {"NPU_FUZZY_COMPILE_BLACKLIST": "Tril,LayerNormGrad"}
80	- torch.npu.set_option(option)
81		79
82	if training_args.do_predict:	80	if training_args.do_predict:
83	training_args.local_rank = -1	81	training_args.local_rank = -1

MPyTorch/built-in/foundation/GPT-NeoX/train.py+0-2

	@@ -28,8 +28,6 @@ if __name__ == "__main__":
28		28
29	use_jit_compile = os.getenv('JIT_COMPILE', 'False').lower() in ['true', '1']	29	use_jit_compile = os.getenv('JIT_COMPILE', 'False').lower() in ['true', '1']
30	torch.npu.set_compile_mode(jit_compile=use_jit_compile)	30	torch.npu.set_compile_mode(jit_compile=use_jit_compile)
31	- option = {"NPU_FUZZY_COMPILE_BLACKLIST": "Tril,SoftmaxV2,LayerNormGrad", "MM_BMM_ND_ENABLE": 'enable'}
32	- torch.npu.set_option(option)
33	neox_args = NeoXArgs.consume_neox_args()	31	neox_args = NeoXArgs.consume_neox_args()
34	neox_args.configure_distributed_args()	32	neox_args.configure_distributed_args()
35	neox_args.build_tokenizer() # tokenizer needs to be build in training in order to set the padding vocab	33	neox_args.build_tokenizer() # tokenizer needs to be build in training in order to set the padding vocab

MPyTorch/built-in/nlp/mBART_ID2372_for_PyTorch/fairseq_cli/generate.py+0-1

	@@ -60,7 +60,6 @@ def _main(args, output_file):
60	option={}	60	option={}
61	option['ACL_OP_SELECT_IMPL_MODE'] = 'high_performance'	61	option['ACL_OP_SELECT_IMPL_MODE'] = 'high_performance'
62	option['ACL_OPTYPELIST_FOR_IMPLMODE'] = 'LayerNorm'	62	option['ACL_OPTYPELIST_FOR_IMPLMODE'] = 'LayerNorm'
63	- option['NPU_FUZZY_COMPILE_BLACKLIST'] = 'MultiHeadAttention,MultiHeadAttentionGrad'
64	torch.npu.set_option(option)	63	torch.npu.set_option(option)
65	utils.import_user_module(args)	64	utils.import_user_module(args)
66		65

MPyTorch/built-in/nlp/mBART_ID2372_for_PyTorch/fairseq_cli/train.py+0-1

	@@ -61,7 +61,6 @@ def main(args):
61	option = {}	61	option = {}
62	option['ACL_OP_SELECT_IMPL_MODE'] = 'high_performance'	62	option['ACL_OP_SELECT_IMPL_MODE'] = 'high_performance'
63	option['ACL_OPTYPELIST_FOR_IMPLMODE'] = 'LayerNorm'	63	option['ACL_OPTYPELIST_FOR_IMPLMODE'] = 'LayerNorm'
64	- option['NPU_FUZZY_COMPILE_BLACKLIST'] = 'MultiHeadAttention,MultiHeadAttentionGrad'
65	torch.npu.set_option(option)	64	torch.npu.set_option(option)
66	if args.distributed_world_size == 1:	65	if args.distributed_world_size == 1:
67	torch.npu.set_device('npu:{}'.format(args.npu_id))	66	torch.npu.set_device('npu:{}'.format(args.npu_id))

MPyTorch/built-in/others/LayoutXLM_for_PyTorch/examples/run_xfun_re.py+0-2

	@@ -76,8 +76,6 @@ elif os.getenv('ALLOW_FP32', False):
76	torch.npu.matmul.allow_hf32 = False	76	torch.npu.matmul.allow_hf32 = False
77	torch.npu.set_compile_mode(jit_compile=False)	77	torch.npu.set_compile_mode(jit_compile=False)
78	option = {}	78	option = {}
79	-# 二进制场景下，这三个算子前后存在大量aicpu上的transdata，性能较差；加入黑名单后性能提升明显
80	-option["NPU_FUZZY_COMPILE_BLACKLIST"] = "Conv2DBackpropFilter,Conv2DBackpropInput,Conv2D"
81	option["MM_BMM_ND_ENABLE"] = 'disable'	79	option["MM_BMM_ND_ENABLE"] = 'disable'
82	torch.npu.set_option(option)	80	torch.npu.set_option(option)
83		81

MPyTorch/built-in/others/LayoutXLM_for_PyTorch/examples/run_xfun_ser.py+0-1

	@@ -84,7 +84,6 @@ elif os.getenv('ALLOW_FP32', False):
84	torch.npu.matmul.allow_hf32 = False	84	torch.npu.matmul.allow_hf32 = False
85	torch.npu.set_compile_mode(jit_compile=False)	85	torch.npu.set_compile_mode(jit_compile=False)
86	option = {}	86	option = {}
87	-option["NPU_FUZZY_COMPILE_BLACKLIST"] = "Conv2DBackpropFilter,Conv2DBackpropInput,Conv2D"
88	option["MM_BMM_ND_ENABLE"] = 'disable'	87	option["MM_BMM_ND_ENABLE"] = 'disable'
89	torch.npu.set_option(option)	88	torch.npu.set_option(option)
90		89

MPyTorch/built-in/rl/MAPPO_for_PyTorch/onpolicy/scripts/train/train_mpe.py+0-1

	@@ -191,7 +191,6 @@ if __name__ == "__main__":
191	torch_npu.npu.set_compile_mode(jit_compile=False)	191	torch_npu.npu.set_compile_mode(jit_compile=False)
192	# TODO: DynamicGRUV2 now not support fp32	192	# TODO: DynamicGRUV2 now not support fp32
193	option = {}	193	option = {}
194	- option["NPU_FUZZY_COMPILE_BLACKLIST"] = "DynamicGRUV2"
195	option["ACL_PRECISION_MODE"] = "allow_fp32_to_fp16"	194	option["ACL_PRECISION_MODE"] = "allow_fp32_to_fp16"
196	torch.npu.set_option(option)	195	torch.npu.set_option(option)
197	main(sys.argv[1:])	196	main(sys.argv[1:])

MPyTorch/contrib/audio/Tacotron2_for_PyTorch/train.py+0-3

	@@ -601,7 +601,4 @@ def load_tensor(name, device):
601		601
602	if __name__ == '__main__':	602	if __name__ == '__main__':
603	torch.npu.set_compile_mode(jit_compile=False)	603	torch.npu.set_compile_mode(jit_compile=False)
604	- option = {}
605	- option['NPU_FUZZY_COMPILE_BLACKLIST'] = 'DynamicRNN,DynamicRNNV2'
606	- torch.npu.set_option(option)
607	main()	604	main()

MPyTorch/contrib/cv/detection/SSD-Resnet/train.py+0-4

	@@ -496,8 +496,4 @@ def main():
496		496
497		497
498	if __name__ == "__main__":	498	if __name__ == "__main__":
499	- option = {}
500	- option["NPU_FUZZY_COMPILE_BLACKLIST"] = '''BNTrainingReduce,BNTrainingReduceGrad,
501	- BNTrainingUpdate,BNTrainingUpdateGrad'''
502	- torch.npu.set_option(option)
503	main()	499	main()

MPyTorch/contrib/cv/detection/SSD/tools/test.py+0-3

	@@ -215,7 +215,4 @@ def main():
215		215
216		216
217	if __name__ == '__main__':	217	if __name__ == '__main__':
218	- option = {}
219	- option["NPU_FUZZY_COMPILE_BLACKLIST"] = "BatchMultiClassNonMaxSuppression"
220	- torch.npu.set_option(option)
221	main()	218	main()

MPyTorch/contrib/cv/detection/SSD/tools/train.py+0-3

	@@ -197,7 +197,4 @@ def main():
197		197
198	if __name__ == '__main__':	198	if __name__ == '__main__':
199	torch_npu.npu.set_compile_mode(jit_compile=False)	199	torch_npu.npu.set_compile_mode(jit_compile=False)
200	- option = {}
201	- option["NPU_FUZZY_COMPILE_BLACKLIST"] = "BatchMultiClassNonMaxSuppression"
202	- torch.npu.set_option(option)
203	main()	200	main()

MPyTorch/dev/cv/image_classification/ATS_ID2682_for_PyTorch/examples/cats_and_dogs.py+0-3

	@@ -82,9 +82,6 @@ import os
82	import apex	82	import apex
83	#模糊编译	83	#模糊编译
84	torch.npu.set_start_fuzz_compile_step(3)	84	torch.npu.set_start_fuzz_compile_step(3)
85	-option = {}
86	-option["NPU_FUZZY_COMPILE_BLACKLIST"] = "LayerNormGrad"
87	-torch.npu.set_option(option)
88	try:	85	try:
89	from apex import amp	86	from apex import amp
90	except:	87	except:

MPyTorch/dev/cv/image_classification/FasterRCNN_ID0100_for_PyTorch/tools/train_net.py+0-4

	@@ -54,10 +54,6 @@ from detectron2.evaluation import (
54	verify_results,	54	verify_results,
55	)	55	)
56	from detectron2.modeling import GeneralizedRCNNWithTTA	56	from detectron2.modeling import GeneralizedRCNNWithTTA
57	-#算子黑名单
58	-option = {}
59	-option["NPU_FUZZY_COMPILE_BLACKLIST"] = "NMSWithMask,BatchMultiClassNonMaxSuppression,AxpyV2"
60	-torch.npu.set_option(option)
61	class Trainer(DefaultTrainer):	57	class Trainer(DefaultTrainer):
62	"""	58	"""
63	We use the "DefaultTrainer" which contains pre-defined default logic for	59	We use the "DefaultTrainer" which contains pre-defined default logic for