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, 01:15 UTC
[Bugfix] Fix mm budget setting for Qwen Omni models (#33634)
Signed-off-by: Roger Wang <hey@rogerw.io>
Roger Wang committed
Feb 2, 2026 at 20:56 UTC
10546f925aef5d805e41f0f40c6610d08ff1a037
1 file changed
+5
vllm/multimodal/budget.py
+5
@@ -72,9 +72,14 @@ class MultiModalBudget:
72
mm_counts=dict.fromkeys(active_modalities, 1),
73
)
74
75
+ # Some models (e.g., Qwen3Omni with use_audio_in_video=True) share
76
+ # placeholders between modalities, so not all active modalities will
77
+ # have their own entry in the returned dict. We filter to only include
78
+ # modalities that have independent placeholder tokens.
79
mm_max_toks_per_item = {
80
modality: all_mm_max_toks_per_item[modality]
81
for modality in active_modalities
82
+ if modality in all_mm_max_toks_per_item
83
}
84
85
encoder_compute_budget, encoder_cache_size = compute_mm_encoder_budget(