5fa1a76
1
The dtype of the online weights is mostly irrelevant, unless you are using torch_dtype="auto" when initializing a model using model = AutoModelForCausalLM.from_pretrained("path", torch_dtype = "auto").
