Loading LLaMA ckpt速度特别慢

[lhz-menarchy]

您好，我想问一下为什么在这段加载llama的代码中加载速度特别慢：

we enforce loading to llama checkpoint

    # if load: # this is deprecated


     ckpts = sorted(Path(llama_ckpt_dir).glob("*.pth"))
     for ckpt in tqdm(ckpts, desc="Loading LLaMA ckpt"):
         ckpt = torch.load(ckpt, map_location='cuda:0')
         names = self.llama.state_dict().keys()
         ckpt_names = ckpt.keys()
         for n in ckpt_names:
             if n not in names:
                 print(f"Warning: {n} not in llama model")
         self.llama.load_state_dict(ckpt, strict=False)
     self.llama_keys = ["llama." + i for i in ckpt_names]

[Max-Fu]

Hi,

This usually loads pretty fast on our server (< 1min for the 7b model). How long does it take on your end?