Fix tensor device and dtype placement in Qwen2VL model (#26219)

Signed-off-by: Cyrus Leung <[email protected]>
Co-authored-by: Yuanfeng Li <[email protected]>
Co-authored-by: Cyrus Leung <[email protected]>

Author: 3 people

vllm/model_executor/models/qwen2_vl.py

@@ -720,7 +720,7 @@ def forward(
         rotary_pos_emb = self.rot_pos_emb(grid_thw)
 
         # compute cu_seqlens
-        grid_thw_ = torch.tensor(grid_thw)
+        grid_thw_ = torch.tensor(grid_thw, device=x.device, dtype=torch.long)
         cu_seqlens = torch.repeat_interleave(grid_thw_[:, 1] * grid_thw_[:, 2],
                                              grid_thw_[:, 0]).cumsum(
                                                  dim=0, dtype=torch.int32)