DeepSeek-V2 icon indicating copy to clipboard operation
DeepSeek-V2 copied to clipboard

8 * A100 启动巨慢,有启动成功的勇士不

Open CarryChang opened this issue 1 year ago • 2 comments
trafficstars

CarryChang avatar May 08 '24 08:05 CarryChang

建议使用vllm启动https://github.com/vllm-project/vllm/pull/4650

zwd003 avatar May 08 '24 11:05 zwd003

HuggingFace代码中accelerate库对模型的显存分配计算有问题,目前示例代码已修改,预计大幅缩短模型加载速度。

加载模型的代码修改为:

model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, device_map="sequential", torch_dtype=torch.bfloat16, max_memory=max_memory, attn_implementation="eager")

stack-heap-overflow avatar May 09 '24 06:05 stack-heap-overflow