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 14, 05:30 UTC
[Docs] Fix warnings in vllm/profiler and vllm/transformers_utils (#25220)
Signed-off-by: windsonsea <haifeng.yao@daocloud.io>
Michael Yao committed
Sep 21, 2025 at 07:39 UTC
367a480bd3534edf27a8dac3c6f7ea8af9d1ed45
3 files changed
+4
-4
mkdocs.yaml
+1
@@ -102,6 +102,7 @@ plugins:
102
- https://numpy.org/doc/stable/objects.inv
103
- https://pytorch.org/docs/stable/objects.inv
104
- https://psutil.readthedocs.io/en/stable/objects.inv
105
+ - https://huggingface.co/docs/transformers/main/en/objects.inv
106
107
markdown_extensions:
108
- attr_list
vllm/profiler/layerwise_profile.py
+2
-2
@@ -353,8 +353,8 @@ class layerwise_profile(profile):
353
354
Args:
355
num_running_seqs (Optional[int], optional): When given,
356
- num_running_seqs will be passed to LayerProfileResults for metadata
357
- update. Defaults to None.
356
+ num_running_seqs will be passed to LayerProfileResults
357
+ for metadata update. Defaults to None.
358
"""
359
super().__init__(
360
activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
vllm/transformers_utils/configs/jais.py
+1
-2
@@ -74,8 +74,7 @@ class JAISConfig(PretrainedConfig):
74
use_cache (`bool`, *optional*, defaults to `True`):
75
Whether or not the model should return the last key/values
76
attentions (not used by all models).
77
- scale_attn_by_inverse_layer_idx
78
- (`bool`, *optional*, defaults to `False`):
77
+ scale_attn_by_inverse_layer_idx (`bool`, *optional*, default `True`):
78
Whether to additionally scale attention weights
79
by `1 / layer_idx + 1`.
80
reorder_and_upcast_attn (`bool`, *optional*, defaults to `False`):