Refactor kernel

cookbook/transformers/ep_fsdp2_lora_qwen3_5_moe.py

@@ -17,7 +17,7 @@
 from twinkle.model import TransformersModel
 from twinkle.preprocessor import SelfCognitionProcessor
 from twinkle.utils.framework import Torch
-from twinkle.kernel import kernelize_model
+from twinkle.kernel import kernelize, npu_builtin
 
 logger = get_logger()
 args = CLI.from_args()
@@ -95,7 +95,7 @@ def train():
     )
     # npu patch
     if Torch.is_npu_available():
-        model = kernelize_model(model, mode='train', device='npu')
+        model = kernelize(model, npu_builtin(model))
     lora_cfg = _build_lora_config(ENABLE_EP)
     model.add_adapter_to_model(args.lora.adapter_name, lora_cfg,
                                gradient_accumulation_steps=args.training.gradient_accumulation_steps)

cookbook/transformers/fsdp2.py

@@ -12,7 +12,7 @@
 from twinkle.model import TransformersModel
 from twinkle.preprocessor import SelfCognitionProcessor
 from twinkle.utils.framework import Torch
-from twinkle.kernel import kernelize_model
+from twinkle.kernel import kernelize, npu_builtin
 
 logger = get_logger()
 args = CLI.from_args()
@@ -59,7 +59,7 @@ def train():
     model.model._no_split_modules = {'Qwen3_5DecoderLayer'}
     # npu patch
     if Torch.is_npu_available():
-        model = kernelize_model(model, mode='train', device='npu')
+        model = kernelize(model, npu_builtin(model))
 
     lora_config = LoraConfig(**args.get_lora_args())
     model.add_adapter_to_model(

cookbook/transformers/sp_fsdp_dense.py

@@ -9,7 +9,7 @@
 from twinkle.model import TransformersModel
 from twinkle.preprocessor import SelfCognitionProcessor
 from twinkle.utils.framework import Torch
-from twinkle.kernel import kernelize_model
+from twinkle.kernel import kernelize, npu_builtin
 
 logger = get_logger()
 args = CLI.from_args()
@@ -68,7 +68,7 @@ def train():
     )
     # npu patch
     if Torch.is_npu_available():
-        model = kernelize_model(model, mode='train', device='npu')
+        model = kernelize(model, npu_builtin(model))
     lora_config = LoraConfig(**args.get_lora_args())
     model.add_adapter_to_model(args.lora.adapter_name, lora_config,
                                gradient_accumulation_steps=args.training.gradient_accumulation_steps)