Hi,

I am using clustered GPU(driver -1GPU and Worker-3GPU), and caching model data into unity catalog but while loading model checkpoint shards its always use driver memory and failed due insufficient memory.

How to use complete cluster GPU while loading HF models.

Thanks