FastDeploy2.0:Error reading file: SafeTensorError::MetadataIncompleteBuffer

一、现象

GPU:4张英伟达A6000
python3.10
大模型：DeepSeek-R1-Distill-Qwen-32B（aisudio上下载得）

执行如下命令：

python -m fastdeploy.entrypoints.openai.api_server        --model /workspace/models/DeepSeek-R1-Distill-Qwen-32B_baidu        --port 8180        --metrics-port 8181        --engine-worker-queue-port 8182        --max-model-len 8192        --max-num-seqs 1  --tensor-parallel-size 4   --reasoning-parser qwen3

报如下错误：

Traceback (most recent call last):
File "/usr/local/lib/python3.10/dist-packages/fastdeploy/engine/../worker/worker_process.py", line 730, in
run_worker_proc()
File "/usr/local/lib/python3.10/dist-packages/fastdeploy/engine/../worker/worker_process.py", line 711, in run_worker_proc
worker_proc.load_model()
File "/usr/local/lib/python3.10/dist-packages/fastdeploy/engine/../worker/worker_process.py", line 409, in load_model
self.worker.load_model()
File "/usr/local/lib/python3.10/dist-packages/fastdeploy/worker/gpu_worker.py", line 160, in load_model
self.model_runner.load_model()
File "/usr/local/lib/python3.10/dist-packages/fastdeploy/worker/gpu_model_runner.py", line 725, in load_model
self.model = get_model_from_loader(fd_config=self.fd_config)
File "/usr/local/lib/python3.10/dist-packages/fastdeploy/model_executor/model_loader.py", line 54, in get_model_from_loader
model = model_loader.load_model(fd_config)
File "/usr/local/lib/python3.10/dist-packages/fastdeploy/model_executor/model_loader.py", line 115, in load_model
state_dict = load_composite_checkpoint(
File "/usr/local/lib/python3.10/dist-packages/fastdeploy/model_executor/load_weight_utils.py", line 311, in load_composite_checkpoint
state_dict = load_tp_checkpoint(model_path,
File "/usr/local/lib/python3.10/dist-packages/paddleformers/transformers/model_utils.py", line 3260, in load_tp_checkpoint
return load_sharded_checkpoint_as_one(folder, return_numpy=return_numpy)
File "/usr/local/lib/python3.10/dist-packages/paddleformers/transformers/model_utils.py", line 3237, in load_sharded_checkpoint_as_one
state_dict = loader(os.path.join(folder, shard_file))
File "/usr/local/lib/python3.10/dist-packages/paddleformers/utils/safetensors.py", line 317, in fast_load_file
with fast_safe_open(filename, framework="np") as f:
File "/usr/local/lib/python3.10/dist-packages/paddleformers/utils/safetensors.py", line 289, in init
self.base, self.tensors_decs, self.metadata = read_metadata(self.file)
File "/usr/local/lib/python3.10/dist-packages/paddleformers/utils/safetensors.py", line 111, in read_metadata
raise ValueError("SafeTensorError::MetadataIncompleteBuffer")
ValueError: SafeTensorError::MetadataIncompleteBuffer

二、原因

模型文件下载不全。

   你说模型文件下载不全，可翻来覆去瞅了半天，愣是没看出哪个文件“缺胳膊少腿”——这感觉，就像在满屋子双胞胎里找谁多长了一根头发。模型文件一多，数着数着就眼冒金星，脑瓜子嗡嗡的，真不想重头再下一遍，光是想想，头发都快掉光了！别急，兄弟，咱打工人也有智慧闪光的时刻——难道非得靠“人肉点数大法”和“玄学肉眼扫描术”？当然不！其实早有妙招：校验和（checksum）才是你的数字侦探，轻轻一算，哪个文件“发育不良”立马现形。省时省力，还能顺带拯救一下岌岌可危的发际线，这波不亏！

检验代码如下：在检验model-00003-of-000008.safetensors时报错，再认证核对，发现是8.13G，实际应该是8.17G。

from paddleformers.utils.safetensors import fast_safe_opentry:with fast_safe_open("/workspace/models/DeepSeek-R1-Distill-Qwen-32B_baidu/model-00001-of-000008.safetensors", framework="np") as f:print("00001:", list(f.keys()))with fast_safe_open("/workspace/models/DeepSeek-R1-Distill-Qwen-32B_baidu/model-00002-of-000008.safetensors", framework="np") as f:print("00002:", list(f.keys()))with fast_safe_open("/workspace/models/DeepSeek-R1-Distill-Qwen-32B_baidu/model-00003-of-000008.safetensors", framework="np") as f:print("00003:", list(f.keys()))with fast_safe_open("/workspace/models/DeepSeek-R1-Distill-Qwen-32B_baidu/model-00004-of-000008.safetensors", framework="np") as f:print("00004:", list(f.keys()))with fast_safe_open("/workspace/models/DeepSeek-R1-Distill-Qwen-32B_baidu/model-00005-of-000008.safetensors", framework="np") as f:print("00005:", list(f.keys()))with fast_safe_open("/workspace/models/DeepSeek-R1-Distill-Qwen-32B_baidu/model-00006-of-000008.safetensors", framework="np") as f:print("00006:", list(f.keys()))with fast_safe_open("/workspace/models/DeepSeek-R1-Distill-Qwen-32B_baidu/model-00007-of-000008.safetensors", framework="np") as f:print("00007:", list(f.keys()))with fast_safe_open("/workspace/models/DeepSeek-R1-Distill-Qwen-32B_baidu/model-00008-of-000008.safetensors", framework="np") as f:print("00008:", list(f.keys()))
except Exception as e:print("Error reading file:", e)

三、解决方案

单独下载第三个模型文件就行。

aistudio download model-00003-of-000008.safetensors --model PaddleNLP/DeepSeek-R1-Distill-Qwen-32B --local_dir /mnt/data/models/DeepSeek-R1-Distill-Qwen-32B_baidu/