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

Signed-off-by: Lukas Geiger <[email protected]>

Author: lgeiger

vllm/model_executor/models/dots_ocr.py

@@ -680,7 +680,7 @@ def forward(
             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:

vllm/model_executor/models/ernie45_vl.py

@@ -574,11 +574,12 @@ def forward(
             grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
         ).cumsum(dim=0, dtype=torch.int32)
 
+        zeros = cu_seqlens.new_zeros(1)
         if num_pad > 0:
-            cu_seqlens = F.pad(cu_seqlens, (1, 1), value=0)
+            cu_seqlens = torch.cat([zeros, cu_seqlens, zeros])
             cu_seqlens[-1] = cu_seqlens[-2] + num_pad
         else:
-            cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
+            cu_seqlens = torch.cat([zeros, cu_seqlens])
 
         # add batch size
         if hidden_states.ndim == 2:

vllm/model_executor/models/qwen3_vl.py

@@ -539,7 +539,7 @@ def forward(
             dim=0,
             dtype=grid_thw_tensor.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])
 
         hidden_states = hidden_states.unsqueeze(1)
         rotary_pos_emb = rotary_pos_emb.to(hidden_states.device)

vllm/model_executor/models/siglip2navit.py

@@ -592,7 +592,7 @@ def forward(
             # for more information
             dtype=grid_thws.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])
 
         reverse_indices = torch.argsort(window_index)