diff --git a/vllm-v0.6.2/vllm/model_executor/layers/vocab_parallel_embedding.py b/vllm-v0.6.2/vllm/model_executor/layers/vocab_parallel_embedding.py
index 52771f5..776ec33 100644
--- a/vllm-v0.6.2/vllm/model_executor/layers/vocab_parallel_embedding.py
+++ b/vllm-v0.6.2/vllm/model_executor/layers/vocab_parallel_embedding.py
@@ -38,6 +38,9 @@ class UnquantizedEmbeddingMethod(QuantizeMethodBase):
               layer: torch.nn.Module,
               x: torch.Tensor,
               bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+        # MLU F.linear requires matching dtypes
+        if x.dtype != layer.weight.dtype:
+            x = x.to(layer.weight.dtype)
         return F.linear(x, layer.weight, bias)
 
     def embedding(self, layer: torch.nn.Module,