facebookresearch/chameleon

Single GPU bug with python -m chameleon.miniviewer

MaureenZOU opened this issue · 6 comments

MaureenZOU commented

One trick for the follow up with one gpu is adding the following two lines in chameleon.py, otherwise it would hang forever.

if world_size > 1:
    dist.broadcast_object_list(to_continue, src=0)

if world_size > 1:
    dist.broadcast_object_list(req, src=0)
neoyxm commented

Hi,
when using the single gpu , I still got hangup even adding these two lines on inferencing by 30b or 7b.

On loading the model, the returning value of world_size == 0 and no worker was started.

Such a mysterious code of this project :(

MaureenZOU commented

Interesting, may check the cuda env, my problem fixed by adding those two lines.

neoyxm commented

Interesting, may check the cuda env, my problem fixed by adding those two lines.

I hard-code the world-size to 1, then got the error:
File "/data/app/ai/chameleon/chameleon/inference/loader.py", line 52, in load_model with open(src_dir / "params.json", "r") as f: FileNotFoundError: [Errno 2] No such file or directory: '/data/app/ai/models/chameleon/models/30b/params.json'

:(

MaureenZOU commented

This seems that you haven't download the checkpoint correctly. Could you please check the folder?

MaureenZOU commented

BTW, I am running the miniviewer.

neoyxm commented

BTW, I am running the miniviewer.

thanks , I missed consolidated.pth . But its size is also huge, oops .