Merge pull request #1953 from ShaneTian/model-load-bf16

Fix slow model initialization in bfloat16 dtype.

src/llmtuner/model/loader.py

@@ -67,10 +67,10 @@ def load_model_and_tokenizer(
     model = AutoModelForCausalLM.from_pretrained(
         model_args.model_name_or_path,
         config=config,
-        torch_dtype=model_args.compute_dtype,
         low_cpu_mem_usage=(not is_deepspeed_zero3_enabled()),
         **config_kwargs
     )
+    model = model.to(model_args.compute_dtype) if not getattr(model, "quantization_method", None) else model
     patcher.patch_model(model)
     register_autoclass(config, model, tokenizer)
     if not is_deepspeed_zero3_enabled():
@@ -95,7 +95,6 @@ def load_model_and_tokenizer(
 
     if not is_trainable:
         model.requires_grad_(False) # fix all model params
-        model = model.to(model_args.compute_dtype) if not getattr(model, "quantization_method", None) else model
         model.eval()
     else:
         model.train()