A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Models][Qwen] Replace `pad` with `cat` for better performance (#26486)
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>
Lukas Geiger committed
Oct 9, 2025 at 15:51 UTC
2c1c7dfb350eb7aadf7dba05f4dd883f84833a83
4 files changed
+6
-5
vllm/model_executor/models/dots_ocr.py
+1
-1
@@ -680,7 +680,7 @@ class DotsVisionTransformer(nn.Module):
680
dim=0,
681
dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
682
)
683
- cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
683
+ cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens])
684
685
max_seqlen, seqlens = self.compute_attn_mask_seqlen(cu_seqlens)
686
for blk in self.blocks:
vllm/model_executor/models/ernie45_vl.py
+3
-2
@@ -574,11 +574,12 @@ class Ernie4_5_VisionTransformer(nn.Module):
574
grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
575
).cumsum(dim=0, dtype=torch.int32)
576
577
+ zeros = cu_seqlens.new_zeros(1)
578
if num_pad > 0:
578
- cu_seqlens = F.pad(cu_seqlens, (1, 1), value=0)
579
+ cu_seqlens = torch.cat([zeros, cu_seqlens, zeros])
580
cu_seqlens[-1] = cu_seqlens[-2] + num_pad
581
else:
581
- cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
582
+ cu_seqlens = torch.cat([zeros, cu_seqlens])
583
584
# add batch size
585
if hidden_states.ndim == 2:
vllm/model_executor/models/qwen3_vl.py
+1
-1
@@ -539,7 +539,7 @@ class Qwen3_VisionTransformer(nn.Module):
539
dim=0,
540
dtype=grid_thw_tensor.dtype if torch.jit.is_tracing() else torch.int32,
541
)
542
- cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
542
+ cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens])
543
544
hidden_states = hidden_states.unsqueeze(1)
545
rotary_pos_emb = rotary_pos_emb.to(hidden_states.device)
vllm/model_executor/models/siglip2navit.py
+1
-1
@@ -592,7 +592,7 @@ class Siglip2Encoder(nn.Module):
592
# for more information
593
dtype=grid_thws.dtype if torch.jit.is_tracing() else torch.int32,
594
)
595
- cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
595
+ cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens])
596
597
reverse_indices = torch.argsort(window_index)
598