update code for NPU qwen2 (#12094)

* update code

* fix

python/llm/src/ipex_llm/transformers/npu_models/lm_head.py

@@ -85,7 +85,6 @@ class LMHeadLinear(NNFactory):
         Returns:
             np.ndarray: result
         """
-        self.prefetchWeights(1, verify_size=False)
         self.set_input_tensor(X, 0)
         self.elapsed = backend_lib.run(self._mm)
         if len(self.out) == 1:

python/llm/src/ipex_llm/transformers/npu_models/qwen2_mp.py

@@ -990,6 +990,7 @@ def gen_qwen2_fused_model_forward(prefill_runner, decode_runner):
         if inputs_embeds is None:
             inputs_embeds = self.embed_tokens(input_ids)
 
+        if seq_length > 1:
             past_key_values_length = 0
 
             from ipex_llm.transformers.npu_models.kv import DynamicFusedNormalCache
@@ -1019,6 +1020,9 @@ def gen_qwen2_fused_model_forward(prefill_runner, decode_runner):
                 past_key_values_length,
                 sliding_window=self.config.sliding_window,
             )
+        else:
+            attention_mask = None
+            position_ids = None
 
         # embed positions
         hidden_states = inputs_embeds