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 13, 13:15 UTC
[Model][Qwen3VL] Slighly speedup `fast_pos_embed_interpolate` (#28434)
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>
Lukas Geiger committed
Nov 11, 2025 at 10:35 UTC
9973e6e04ad3e4a6c74c51a2dc87b2d3ddc4837f
1 file changed
+2
-2
vllm/model_executor/models/qwen3_vl.py
+2
-2
@@ -491,8 +491,8 @@ class Qwen3_VisionTransformer(nn.Module):
491
weights = weights.to(dtype=self.dtype)
492
493
embeds = self.pos_embed(indices)
494
- weighted_embeds = embeds * weights
495
- combined = weighted_embeds.sum(dim=0)
494
+ embeds *= weights
495
+ combined = embeds.sum(dim=0)
496
497
combined = combined.reshape(
498
h // m_size, m_size, w // m_size, m_size, hidden_dim