fix(benchmark): restrict remote model code execution-pytorch-AtomGit

Mbenchmarks/llm/baichuan2-7B-Chat/README.md+4-0

	@@ -34,6 +34,10 @@ pip install -r ../utils/requirements.txt
34	python ../utils/download_hf.py --model baichuan-inc/Baichuan2-7B-Chat --save_path ./Baichuan2-7B-Chat	34	python ../utils/download_hf.py --model baichuan-inc/Baichuan2-7B-Chat --save_path ./Baichuan2-7B-Chat
35	```	35	```
36		36
		37	+> [!WARNING]
		38	+> Baichuan2 依赖模型仓中的自定义 Python 代码。脚本默认不执行远程代码；仅在确认模型来源可信后使用 `--trust-remote-code`。
		39	+> 使用远程模型仓时必须同时通过 `--revision` 指定完整的 40 位 commit hash，避免已审核代码被后续更新替换。优先使用本地已校验的模型目录。
		40	+
37	## 2. 数据获取	41	## 2. 数据获取
38		42
39	训练数据集来自 LlamaFactory 仓库示例数据：	43	训练数据集来自 LlamaFactory 仓库示例数据：

Mbenchmarks/llm/baichuan2-7B-Chat/run_baichuan2.sh+2-1

	@@ -20,4 +20,5 @@ python train_baichuan2_7B.py \
20	--use_lora \	20	--use_lora \
21	--use_bf16 \	21	--use_bf16 \
22	--gradient_checkpointing \	22	--gradient_checkpointing \
23	- > logs/train_baichuan.log 2>&1	23	+ "$@" \
		24	+ > logs/train_baichuan.log 2>&1

Mbenchmarks/llm/baichuan2-7B-Chat/train_baichuan2_7B.py+20-4

	@@ -46,7 +46,8 @@ class Baichuan2Trainer:
46		46
47	self.tokenizer = AutoTokenizer.from_pretrained(	47	self.tokenizer = AutoTokenizer.from_pretrained(
48	self.args.model_path,	48	self.args.model_path,
49	- trust_remote_code=True,	49	+ revision=self.args.revision,
		50	+ trust_remote_code=self.args.trust_remote_code,
50	padding_side="right",	51	padding_side="right",
51	model_max_length=self.args.max_length,	52	model_max_length=self.args.max_length,
52	)	53	)
	@@ -56,7 +57,8 @@ class Baichuan2Trainer:
56		57
57	self.model = AutoModelForCausalLM.from_pretrained(	58	self.model = AutoModelForCausalLM.from_pretrained(
58	self.args.model_path,	59	self.args.model_path,
59	- trust_remote_code=True,	60	+ revision=self.args.revision,
		61	+ trust_remote_code=self.args.trust_remote_code,
60	torch_dtype=torch.bfloat16 if self.args.use_bf16 else torch.float16,	62	torch_dtype=torch.bfloat16 if self.args.use_bf16 else torch.float16,
61	use_cache=not self.args.gradient_checkpointing,	63	use_cache=not self.args.gradient_checkpointing,
62	)	64	)
	@@ -238,6 +240,10 @@ def build_argparser():
238	parser.add_argument("--mfusion", action="store_true",	240	parser.add_argument("--mfusion", action="store_true",
239	help="Enable MFusion for graph fusion optimization")	241	help="Enable MFusion for graph fusion optimization")
240	parser.add_argument("--model_path", type=str, required=True)	242	parser.add_argument("--model_path", type=str, required=True)
		243	+ parser.add_argument("--trust-remote-code", action="store_true",
		244	+ help="Allow execution of custom code from the model repository")
		245	+ parser.add_argument("--revision", type=str, default=None,
		246	+ help="Full commit hash for a remote model repository")
241	parser.add_argument("--data_path", type=str, required=True)	247	parser.add_argument("--data_path", type=str, required=True)
242	parser.add_argument("--output_dir", type=str, default="./baichuan2-finetuned")	248	parser.add_argument("--output_dir", type=str, default="./baichuan2-finetuned")
243	parser.add_argument("--num_epochs", type=int, default=3)	249	parser.add_argument("--num_epochs", type=int, default=3)
	@@ -278,7 +284,17 @@ def build_argparser():
278		284
279		285
280	def main():	286	def main():
281	- args = build_argparser().parse_args()	287	+ parser = build_argparser()
		288	+ args = parser.parse_args()
		289	+ if args.trust_remote_code and not os.path.isdir(args.model_path):
		290	+ revision = args.revision or ""
		291	+ is_commit_hash = len(revision) == 40 and all(
		292	+ char in "0123456789abcdefABCDEF" for char in revision
		293	+ )
		294	+ if not is_commit_hash:
		295	+ parser.error(
		296	+ "--revision must be a full commit hash when enabling remote code"
		297	+ )
282	detect_device_type()	298	detect_device_type()
283	os.environ['TORCHINDUCTOR_NPU_BACKEND']=args.npu_backend	299	os.environ['TORCHINDUCTOR_NPU_BACKEND']=args.npu_backend
284	if args.npu_backend == "akg":	300	if args.npu_backend == "akg":
	@@ -305,4 +321,4 @@ def main():
305	print(f"op_compile_time:{op_compile_time * 1e3} ms", )	321	print(f"op_compile_time:{op_compile_time * 1e3} ms", )
306		322
307	if __name__ == "__main__":	323	if __name__ == "__main__":
308	- main()	324	+ main()

Mbenchmarks/llm/glm4-9B-chat/README.md+3-0

	@@ -23,6 +23,9 @@
23		23
24	可在链接页面中 `Files and versions` 一栏直接下载。	24	可在链接页面中 `Files and versions` 一栏直接下载。
25		25
		26	+> [!NOTE]
		27	+> 本脚本使用 Transformers 内置的 GLM4 实现，不执行模型仓自定义 Python 代码。请使用与 `transformers==4.57.1` 兼容的原生 GLM4 模型。
		28	+
26	### 环境提示	29	### 环境提示
27		30
28	本项目依赖已整理到 `requirements.txt`，可直接安装：	31	本项目依赖已整理到 `requirements.txt`，可直接安装：

Mbenchmarks/llm/glm4-9B-chat/train_glm4_9B.py+0-2

	@@ -72,7 +72,6 @@ class GLM4Trainer:
72	self.args.model_path,	72	self.args.model_path,
73	quantization_config=bnb_config if self.args.use_4bit else None,	73	quantization_config=bnb_config if self.args.use_4bit else None,
74	torch_dtype=torch.bfloat16 if self.args.use_bf16 else torch.float16,	74	torch_dtype=torch.bfloat16 if self.args.use_bf16 else torch.float16,
75	- trust_remote_code=True,
76	)	75	)
77		76
78	logger.info(f"Moving model to {self.args.device_type}...")	77	logger.info(f"Moving model to {self.args.device_type}...")
	@@ -80,7 +79,6 @@ class GLM4Trainer:
80		79
81	self.tokenizer = AutoTokenizer.from_pretrained(	80	self.tokenizer = AutoTokenizer.from_pretrained(
82	self.args.model_path,	81	self.args.model_path,
83	- trust_remote_code=True
84	)	82	)
85		83
86	if self.tokenizer.pad_token is None:	84	if self.tokenizer.pad_token is None:

Mbenchmarks/llm/gpt-oss-20B/train_gpt_oss_20B.py+1-3

	@@ -45,7 +45,6 @@ class GPT_OSS_20BTrainer:
45	self.model = AutoModelForCausalLM.from_pretrained(	45	self.model = AutoModelForCausalLM.from_pretrained(
46	self.args.model_path,	46	self.args.model_path,
47	torch_dtype=torch.bfloat16 if self.args.use_bf16 else torch.float16,	47	torch_dtype=torch.bfloat16 if self.args.use_bf16 else torch.float16,
48	- trust_remote_code=True
49	)	48	)
50		49
51	logger.info(f"Moving model to {self.args.device_type}...")	50	logger.info(f"Moving model to {self.args.device_type}...")
	@@ -53,7 +52,6 @@ class GPT_OSS_20BTrainer:
53		52
54	self.tokenizer = AutoTokenizer.from_pretrained(	53	self.tokenizer = AutoTokenizer.from_pretrained(
55	self.args.model_path,	54	self.args.model_path,
56	- trust_remote_code=True
57	)	55	)
58		56
59	if self.tokenizer.pad_token is None:	57	if self.tokenizer.pad_token is None:
	@@ -390,4 +388,4 @@ def main():
390	print("="*50)	388	print("="*50)
391		389
392	if __name__ == "__main__":	390	if __name__ == "__main__":
393	- main()	391	+ main()

Mbenchmarks/llm/llama3&llama3.2/train_llama3.py+0-2

	@@ -56,14 +56,12 @@ class LLama3Trainer:
56	self.args.model_path,	56	self.args.model_path,
57	quantization_config=bnb_config if self.args.use_4bit else None,	57	quantization_config=bnb_config if self.args.use_4bit else None,
58	torch_dtype=torch.bfloat16 if self.args.use_bf16 else torch.float32,	58	torch_dtype=torch.bfloat16 if self.args.use_bf16 else torch.float32,
59	- trust_remote_code=True
60	)	59	)
61		60
62	logger.info(f"Moving model to {self.args.device_type}...")	61	logger.info(f"Moving model to {self.args.device_type}...")
63	self.model = self.model.to(self.args.device_type)	62	self.model = self.model.to(self.args.device_type)
64	self.tokenizer = AutoTokenizer.from_pretrained(	63	self.tokenizer = AutoTokenizer.from_pretrained(
65	self.args.model_path,	64	self.args.model_path,
66	- trust_remote_code=True
67	)	65	)
68		66
69	if self.tokenizer.pad_token is None:	67	if self.tokenizer.pad_token is None:

Mbenchmarks/llm/mamba-codestral-7B/train_mamba2_7B.py+2-4

	@@ -58,7 +58,7 @@ def setup_environment():
58	def parse_args():	58	def parse_args():
59	parser = argparse.ArgumentParser(description="Mamba Codestral LoRA finetune")	59	parser = argparse.ArgumentParser(description="Mamba Codestral LoRA finetune")
60		60
61	- parser.add_argument("--model_path", type=str, default="/home/zhangyican/workspace/q4_data/Mamba-Codestral-7B-v0.1", help="Base model path")	61	+ parser.add_argument("--model_path", type=str, required=True, help="Base model path")
62	parser.add_argument("--data_file", type=str, default="./c4_demo.jsonl", help="Training data file path")	62	parser.add_argument("--data_file", type=str, default="./c4_demo.jsonl", help="Training data file path")
63	parser.add_argument("--output_dir", type=str, default="./mamba_codestral_lora_no_quant", help="Output directory")	63	parser.add_argument("--output_dir", type=str, default="./mamba_codestral_lora_no_quant", help="Output directory")
64	parser.add_argument("--epochs", type=int, default=3, help="Number of training epochs")	64	parser.add_argument("--epochs", type=int, default=3, help="Number of training epochs")
	@@ -103,7 +103,6 @@ def main():
103	print(f"Loading tokenizer from: {args.model_path}")	103	print(f"Loading tokenizer from: {args.model_path}")
104	tokenizer = AutoTokenizer.from_pretrained(	104	tokenizer = AutoTokenizer.from_pretrained(
105	args.model_path,	105	args.model_path,
106	- trust_remote_code=True,
107	use_fast=True	106	use_fast=True
108	)	107	)
109	if tokenizer.pad_token is None:	108	if tokenizer.pad_token is None:
	@@ -113,7 +112,6 @@ def main():
113	base_model = AutoModelForCausalLM.from_pretrained(	112	base_model = AutoModelForCausalLM.from_pretrained(
114	args.model_path,	113	args.model_path,
115	torch_dtype=torch.bfloat16, # Use bfloat16 precision	114	torch_dtype=torch.bfloat16, # Use bfloat16 precision
116	- trust_remote_code=True,
117	use_cache=False, # Disable cache to save memory	115	use_cache=False, # Disable cache to save memory
118	)	116	)
119		117
	@@ -221,4 +219,4 @@ def main():
221	print("Training complete!")	219	print("Training complete!")
222		220
223	if __name__ == "__main__":	221	if __name__ == "__main__":
224	- main()	222	+ main()

Mbenchmarks/llm/qwen2vl-2B-Instruct/train_qwen2vl.py+2-5

	@@ -69,8 +69,7 @@ def patch_remove_ops_from_generate_list(op_names=None):
69		69
70	def parse_args():	70	def parse_args():
71	p = argparse.ArgumentParser(description="Qwen2-VL 2B Single GPU Training")	71	p = argparse.ArgumentParser(description="Qwen2-VL 2B Single GPU Training")
72	- p.add_argument("--model_name_or_path", type=str,	72	+ p.add_argument("--model_name_or_path", type=str, required=True)
73	- default="/data/zyc/Qwen2-VL-2B-Instruct")
74	p.add_argument("--use_lora", type=str2bool, default=False)	73	p.add_argument("--use_lora", type=str2bool, default=False)
75	p.add_argument("--lora_r", type=int, default=64)	74	p.add_argument("--lora_r", type=int, default=64)
76	p.add_argument("--lora_alpha", type=int, default=16)	75	p.add_argument("--lora_alpha", type=int, default=16)
	@@ -247,7 +246,6 @@ def main():
247	args.model_name_or_path,	246	args.model_name_or_path,
248	min_pixels=args.min_pixels,	247	min_pixels=args.min_pixels,
249	max_pixels=args.max_pixels,	248	max_pixels=args.max_pixels,
250	- trust_remote_code=True,
251	)	249	)
252	print(f"Processor loading time: {time.time() - t0:.2f}s")	250	print(f"Processor loading time: {time.time() - t0:.2f}s")
253	print(f" min_pixels = {args.min_pixels}, max_pixels = {args.max_pixels}")	251	print(f" min_pixels = {args.min_pixels}, max_pixels = {args.max_pixels}")
	@@ -260,7 +258,6 @@ def main():
260	model = Qwen2VLForConditionalGeneration.from_pretrained(	258	model = Qwen2VLForConditionalGeneration.from_pretrained(
261	args.model_name_or_path,	259	args.model_name_or_path,
262	torch_dtype=torch.bfloat16,	260	torch_dtype=torch.bfloat16,
263	- trust_remote_code=True,
264	)	261	)
265	print(f"Model loading time: {time.time() - t0:.2f}s")	262	print(f"Model loading time: {time.time() - t0:.2f}s")
266		263
	@@ -368,4 +365,4 @@ def main():
368	processor.save_pretrained(final_output_dir)	365	processor.save_pretrained(final_output_dir)
369		366
370	if __name__ == "__main__":	367	if __name__ == "__main__":
371	- main()	368	+ main()

Mbenchmarks/llm/qwen3-4B/train_qwen3_4B.py+1-3

	@@ -57,7 +57,6 @@ class Qwen3Trainer:
57	self.args.model_path,	57	self.args.model_path,
58	quantization_config=bnb_config if self.args.use_4bit else None,	58	quantization_config=bnb_config if self.args.use_4bit else None,
59	torch_dtype=torch.bfloat16 if self.args.use_bf16 else torch.float32,	59	torch_dtype=torch.bfloat16 if self.args.use_bf16 else torch.float32,
60	- trust_remote_code=True
61	)	60	)
62		61
63	logger.info(f"Moving model to {self.args.device_type}...")	62	logger.info(f"Moving model to {self.args.device_type}...")
	@@ -65,7 +64,6 @@ class Qwen3Trainer:
65		64
66	self.tokenizer = AutoTokenizer.from_pretrained(	65	self.tokenizer = AutoTokenizer.from_pretrained(
67	self.args.model_path,	66	self.args.model_path,
68	- trust_remote_code=True
69	)	67	)
70		68
71	if self.tokenizer.pad_token is None:	69	if self.tokenizer.pad_token is None:
	@@ -380,4 +378,4 @@ def main():
380	print("="*50)	378	print("="*50)
381		379
382	if __name__ == "__main__":	380	if __name__ == "__main__":
383	- main()	381	+ main()