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 9, 23:00 UTC
[Bugfix] fix the ima issue of qwen-vit (#32687)
Signed-off-by: JJJYmmm <92386084+JJJYmmm@users.noreply.github.com>
JJJYmmm committed
Jan 21, 2026 at 01:21 UTC
04a9e064db4dcf57519f1333796ba7face46248b
2 files changed
+3
-2
vllm/model_executor/models/qwen2_5_vl.py
+1
@@ -384,6 +384,7 @@ class Qwen2_5_VisionAttention(nn.Module):
384
qk_reshaped = einops.rearrange(
385
qk, "b s two head head_dim -> (two b) s head head_dim", two=2
386
)
387
+ qk_reshaped = qk_reshaped.contiguous()
388
qk_rotated = self.apply_rotary_emb(
389
qk_reshaped,
390
rotary_pos_emb_cos,
vllm/model_executor/models/qwen3_vl.py
+2
-2
@@ -128,8 +128,8 @@ from .vision import (
128
129
logger = init_logger(__name__)
130
131
-# Official recommended max pixels is 24576 * 32 * 32
132
-_MAX_FRAMES_PER_VIDEO = 24576
131
+# Official recommended max frames is 2048
132
+_MAX_FRAMES_PER_VIDEO = 2048
133
134
135
class Qwen3_VisionPatchEmbed(nn.Module):