copy fused rms norm's reuslt to avoid <unk> (#9909)

python/llm/src/bigdl/llm/transformers/models/baichuan2.py

@@ -54,6 +54,8 @@ def baichuan_13b_rms_norm_forward(self, hidden_states):
                                             self.epsilon)
         # if nelement == 0, means fused norm failed, go back to python implement.
         if result.nelement != 0:
+            # We should copy this result to avoid <unk> by unknown reason on Arc GPUs.
+            result = result.clone()
             return result
     input_dtype = hidden_states.dtype
     hidden_states = hidden_states.to(torch.float32)

python/llm/src/bigdl/llm/transformers/models/chatglm2.py

@@ -88,6 +88,8 @@ def chatglm_rms_norm_forward(self, hidden_states):
                                             self.eps)
         # if nelement == 0, means fused norm failed, go back to python implement.
         if result.nelement != 0:
+            # We should copy this result to avoid <unk> by unknown reason on Arc GPUs.
+            result = result.clone()
             return result
     input_dtype = hidden_states.dtype
     hidden_states = hidden_states.to(torch.float32)

python/llm/src/bigdl/llm/transformers/models/llama.py

@@ -91,6 +91,8 @@ def llama_rms_norm_forward(self, hidden_states):
                                             self.variance_epsilon)
         # if nelement == 0, means fused norm failed, go back to python implement.
         if result.nelement != 0:
+            # We should copy this result to avoid <unk> by unknown reason on Arc GPUs.
+            result = result.clone()
             return result
     input_dtype = hidden_states.dtype
     hidden_states = hidden_states.to(torch.float32)