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, 00:00 UTC
[Bugfix] Fix startup hang for Granite Speech (#33699)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Feb 3, 2026 at 23:57 UTC
18e7cbbb158a86bdc76585e64ada795bf1c0d435
1 file changed
+8
-8
vllm/multimodal/budget.py
+8
-8
@@ -54,17 +54,17 @@ class MultiModalBudget:
54
self.max_model_len = model_config.max_model_len
55
self.max_num_reqs = scheduler_config.max_num_seqs
56
57
- cache = mm_registry.processor_only_cache_from_config(vllm_config)
58
- processor = mm_registry.create_processor(model_config, cache=cache)
57
+ with set_default_torch_num_threads(): # Avoid hang during startup
58
+ cache = mm_registry.processor_only_cache_from_config(vllm_config)
59
+ processor = mm_registry.create_processor(model_config, cache=cache)
60
60
- self.cache = cache
61
- self.mm_limits = mm_limits = processor.info.allowed_mm_limits
61
+ self.cache = cache
62
+ self.mm_limits = mm_limits = processor.info.allowed_mm_limits
63
63
- active_modalities = {
64
- modality for modality, limit in mm_limits.items() if limit > 0
65
- }
64
+ active_modalities = {
65
+ modality for modality, limit in mm_limits.items() if limit > 0
66
+ }
67
67
- with set_default_torch_num_threads(): # Avoid hang during startup
68
all_mm_max_toks_per_item = get_mm_max_toks_per_item(
69
model_config,
70
mm_registry,