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 3, 19:00 UTC
[Bugfix][Qwen3-VL] Fix pipeline-parallel deepstack initialization (#42394)
Signed-off-by: MrZ20 <2609716663@qq.com>
SILONG ZENG committed
May 13, 2026 at 18:58 UTC
cee6751e548357478a9943cae5786062b7b95127
2 files changed
+2
-2
vllm/model_executor/models/qwen3_vl.py
+1
-1
@@ -1697,7 +1697,7 @@ class Qwen3VLForConditionalGeneration(
1697
if not get_pp_group().is_first_rank and hasattr(
1698
config.vision_config, "deepstack_visual_indexes"
1699
):
1700
- assert self.language_model.start_layer >= len(
1700
+ assert self.language_model.model.start_layer >= len(
1701
config.vision_config.deepstack_visual_indexes
1702
), (
1703
"start_layer should be greater than or equal to "
vllm/model_executor/models/qwen3_vl_moe.py
+1
-1
@@ -454,7 +454,7 @@ class Qwen3VLMoeForConditionalGeneration(
454
if not get_pp_group().is_first_rank and hasattr(
455
config.vision_config, "deepstack_visual_indexes"
456
):
457
- assert self.language_model.start_layer >= len(
457
+ assert self.language_model.model.start_layer >= len(
458
config.vision_config.deepstack_visual_indexes
459
), (
460
"start_layer should be greater than or equal to "