diff --git a/modules/llm/Miqu-1-70b/Miqu-1-70b_fine_tune.py b/modules/llm/Miqu-1-70b/Miqu-1-70b_fine_tune.py index 56f86c983236dd9e8f81235987520b0c3d90385d..9c3d283f49c72a92cd46a6416af30a3a5050e9a7 100644 --- a/modules/llm/Miqu-1-70b/Miqu-1-70b_fine_tune.py +++ b/modules/llm/Miqu-1-70b/Miqu-1-70b_fine_tune.py @@ -99,7 +99,14 @@ def generate(base_model, new_model): load_in_4bit=True, attn_implementation="flash_attention_2" ) - model = PeftModel.from_pretrained(base_model_reload, new_model) + model = PeftModel.from_pretrained( + base_model_reload, + new_model, + device_map="auto", + load_in_8bit=False, + load_in_4bit=True, + attn_implementation="flash_attention_2" + ) model = model.merge_and_unload() tokenizer = transformers.AutoTokenizer.from_pretrained(base_model) @@ -151,7 +158,7 @@ def generate(base_model, new_model): base_model = "../../../models/Miqu-1-70b" new_model = "../../../models/Fine-tuned_Miqu-1-70b" -fine_tune(base_model, new_model) +#fine_tune(base_model, new_model) generate(base_model, new_model) print("========== Program finished ==========")