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 11, 20:00 UTC
[Bugfix] Skip generation config fallback for GGUF to prevent multi-process hang (#30209)
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Christina Norman committed
Dec 8, 2025 at 19:52 UTC
e41312a2f5086cc9199e024c5d451e65a303a4d1
1 file changed
+7
vllm/transformers_utils/config.py
+7
@@ -954,6 +954,13 @@ def try_get_generation_config(
954
revision: str | None = None,
955
config_format: str | ConfigFormat = "auto",
956
) -> GenerationConfig | None:
957
+ # GGUF files don't have generation_config.json - their config is embedded
958
+ # in the file header. Skip all filesystem lookups to avoid re-reading the
959
+ # memory-mapped file, which can hang in multi-process scenarios when the
960
+ # EngineCore process already has the file mapped.
961
+ if is_gguf(model):
962
+ return None
963
+
964
try:
965
return GenerationConfig.from_pretrained(
966
model,