diff --git a/vllm_ascend/models/qwen2_vl.py b/vllm_ascend/models/qwen2_vl.py
index d9b3e03..b601b16 100644
--- a/vllm_ascend/models/qwen2_vl.py
+++ b/vllm_ascend/models/qwen2_vl.py
@@ -314,6 +314,7 @@ class AscendQwen2VisionTransformer(Qwen2VisionTransformer):
         x: torch.Tensor,
         grid_thw: torch.Tensor,
     ) -> torch.Tensor:
+        grid_thw = torch.tensor(grid_thw, dtype=torch.int32)
         # compute cu_seqlens and avoid cumsum to fit operator unpadFA
         cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2],
                                              grid_thw[:,
diff --git a/vllm_ascend/quantization/w8a8.py b/vllm_ascend/quantization/w8a8.py
index 5c7d986..fec542c 100644
--- a/vllm_ascend/quantization/w8a8.py
+++ b/vllm_ascend/quantization/w8a8.py
@@ -99,8 +99,11 @@ class AscendW8A8LinearMethod:
     ) -> torch.Tensor:
         if x.dtype != torch.int8:
             layer_cls_name = layer.__class__.__name__
-            weight_prefetch_method = get_forward_context(
-            ).weight_prefetch_method
+            try:
+                weight_prefetch_method = get_forward_context(
+                ).weight_prefetch_method
+            except AssertionError:
+                weight_prefetch_method = None
 
             # prefetch qkvo_proj.weight preprocess
             if weight_prefetch_method: