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 6, 15:45 UTC
[Bugfix] Add safety check and fallback for null scaling factor (#36106)
Signed-off-by: Yuanheng Zhao <jonathan.zhaoyh@gmail.com> Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Yuanheng Zhao committed
Mar 16, 2026 at 22:27 UTC
8d8855fdae00830221025e4a8ba8267596372056
1 file changed
+9
vllm/config/model.py
+9
@@ -2021,6 +2021,15 @@ def _get_and_verify_max_len(
2021
2022
if rope_type == "yarn":
2023
derived_max_model_len = rp["original_max_position_embeddings"]
2024
+ if scaling_factor is None:
2025
+ # Fallback the factor to 1.0 if a user assigned `null`
2026
+ logger.warning_once(
2027
+ "The model's RoPE configuration has a null scaling "
2028
+ "factor which is unexpected. This likely indicates a bug "
2029
+ "in the model's HuggingFace config.json. Please notify the "
2030
+ "model vendor. Falling back the value to 1.0. "
2031
+ )
2032
+ scaling_factor = 1.0
2033
# Do this outside loop since all layer types should have the same scaling
2034
derived_max_model_len *= scaling_factor
2035