[Models][Qwen] Replace pad with cat for better performance (#26486)

Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>

vllm/model_executor/models/dots_ocr.py

@@ -680,7 +680,7 @@ class DotsVisionTransformer(nn.Module):
             dim=0,
             dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
         )
-        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
+        cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens])
 
         max_seqlen, seqlens = self.compute_attn_mask_seqlen(cu_seqlens)
         for blk in self.blocks: