[NPU Feature] Qwen3.5 NPU FLA and fused-operator patches

cookbook/transformers/ep_fsdp2_lora_qwen3_5_moe.py

@@ -16,6 +16,8 @@
 from twinkle.dataset import Dataset, DatasetMeta
 from twinkle.model import TransformersModel
 from twinkle.preprocessor import SelfCognitionProcessor
+from twinkle.utils.framework import Torch
+from twinkle.kernel import kernelize_model
 
 logger = get_logger()
 
@@ -102,6 +104,9 @@ def train():
             }
         },
     )
+    # npu patch
+    if Torch.is_npu_available():
+        model = kernelize_model(model, mode='train', device='npu')
     lora_cfg = _build_lora_config(ENABLE_EP)
     model.add_adapter_to_model(ADAPTER_NAME, lora_cfg, gradient_accumulation_steps=GRAD_ACCUM_STEPS)
     model.set_optimizer('AdamW', lr=LR, foreach=False)

cookbook/transformers/fsdp2.py

@@ -9,6 +9,8 @@
 from twinkle.dataset import Dataset, DatasetMeta
 from twinkle.model import TransformersModel
 from twinkle.preprocessor import SelfCognitionProcessor
+from twinkle.utils.framework import Torch
+from twinkle.kernel import kernelize_model
 
 logger = get_logger()
 
@@ -72,7 +74,9 @@ def train():
     # Use a TransformersModel
     model = TransformersModel(model_id=MODEL_ID)
     model.model._no_split_modules = {'Qwen3_5DecoderLayer'}
-
+    # npu patch
+    if Torch.is_npu_available():
+        model = kernelize_model(model, mode='train', device='npu')
     lora_config = LoraConfig(r=8, lora_alpha=32, target_modules='all-linear')
 
     # Add a lora to model, with name `default`

src/twinkle/kernel/__init__.py

@@ -49,7 +49,7 @@ def kernelize_model(
     # so that patched module classes are used when new instances are created.
     if device == 'npu' or (device is None and _is_npu_device(model)):
         try:
-            apply_npu_patch()
+            apply_npu_patch(model)
         except Exception:
             logger.warning('NPU patch failed. Continuing without fused ops.', exc_info=True)