support ndhwc of qwen2-vl and qwen2.5-vl

vllm/model_executor/models/qwen2_5_vl.py

@@ -436,6 +436,7 @@ class Qwen2_5_VisionPatchEmbed(nn.Module):
         L, C = x.shape
         x = x.view(L, -1, self.temporal_patch_size, self.patch_size,
                    self.patch_size)
+        x=x.to(memory_format=torch.channels_last_3d)
         x = self.proj(x).view(L, self.hidden_size)
         return x
 

vllm/model_executor/models/qwen2_vl.py

@@ -461,6 +461,7 @@ class Qwen2VisionPatchEmbed(nn.Module):
         L, C = x.shape
         x = x.view(L, -1, self.temporal_patch_size, self.patch_size,
                    self.patch_size)
+        # x=x.to(memory_format=torch.channels_last_3d)
         x = self.proj(x).view(L, self.embed_dim)
         return x