no_split_module_classes not needed

This commit is contained in:
81300 2023-02-01 19:47:26 +02:00
parent 0a0d289537
commit c515282f5c
No known key found for this signature in database

View file

@ -178,7 +178,7 @@ def load_model(model_name):
# DeepSpeed ZeRO-3
elif args.deepspeed:
model = AutoModelForCausalLM.from_pretrained(Path(f"models/{model_name}", no_split_module_classes=["GPTJBlock"]))
model = AutoModelForCausalLM.from_pretrained(Path(f"models/{model_name}"))
model = deepspeed.initialize(model=model,
config_params=ds_config,
model_parameters=None,