[model] temporarily support npu fused options on v0, powered by v1 kernels (#9520)

Co-authored-by: frozenleaves <frozen@Mac.local>

src/llamafactory/hparams/model_args.py

@@ -174,6 +174,10 @@ class BaseModelArguments:
         default=True,
         metadata={"help": "Whether or not to use KV cache in generation."},
     )
+    use_v1_kernels: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to use high-performance kernels in training."},
+    )
     infer_dtype: Literal["auto", "float16", "bfloat16", "float32"] = field(
         default="auto",
         metadata={"help": "Data type for model weights and activations at inference."},

src/llamafactory/model/loader.py

@@ -213,6 +213,17 @@ def load_model(
     else:
         model.train()
 
+    # Borrowing the kernel plugins ability of v1 to temporarily apply the NPU fusion operator to v0,
+    # it is turned off by default, and can be discarded after the transition period ends.
+    if model_args.use_v1_kernels and is_trainable:
+        logger.warning_rank0(
+            "You are try to using future feature about kernels, please note that this feature "
+            "is not supported for all models. If get any error, please disable this feature, or report the issue."
+        )
+        from ..v1.plugins.model_plugins.kernels.registry import apply_available_kernels
+
+        model = apply_available_kernels(model)
+
     trainable_params, all_param = count_parameters(model)
     if is_trainable:
         param_stats = (