COMMITS
August 25, 2026
Y
Q
Let gradient checkpointing skip layers with every_n_layers (#48200)
Quentin Gallouédec committed
R
Disable daily nightly CI (#48292)
Rémi Ouazan committed
F
Pipeline parallel naive inference (#47289)
Ferdinand Mom committed
E
gs (#48288)
eustlb committed
T
CI: fix muse OOMs (#48284)
Tarek Ziade committed
W
Fix `BayesianDetectorModel.from_pretrained()` by calling `post_init()` (#48254)
Woojin Paik committed
R
[Fix] Avoid duplicating tests in CI (#48287)
Rémi Ouazan committed
A
[`GDN`] Fix recurrent FLA fallback (#48266)
Anton Vlasjuk committed
Q
Fix `tie_word_embeddings` not lifted from `text_config` for some VLM configs (BC regression) (#45857)
Quentin Gallouédec committed
K
replace xpu-smi subprocess call in benchmark_v2 (#48083)
kaixuanliu committed
T
ignore mlinter ci file (#48267)
Tarek Ziade committed
A
Fix dtype mismatch in grouped_mm_fallback for LoRA training on Mamba+… (#47933)
Aakriti Adhikari committed
August 24, 2026
Q
Compute MoE load-balancing loss per layer to avoid giant one-hot materialization −99.7% @ 128k (#48131)
Quentin Gallouédec committed
M
deterministic layer_types buffer registration in multiple models (#48162)
Moritz Wörmann committed
Y
Fix nemotron_h save_pretrained emitting singular backbone.embedding.weight (#48075)
Yuekai Zhang committed
M
`force_accelerate_hooks` should not hide the signature it wraps (#48156)
Marc Sun committed
X
A
S
[docs] Cache crop (#47950)
Steven Liu committed
M
Fix a typo in a use of a local variable field_ in a test (#48184)
Matej Aleksandrov committed
V
doc: Fix typo in SigLIP2 Flash Attention code example (#48197)
Vimal Sahani committed
S
[serge] Fix 2 integration tests regressed by commit 16780c86b20a (PR #47622) (#48134)
sergereview[bot] committed
J
[xcodec2] Fix flex attention and flash dispatch tests (#48244)
jiqing-feng committed
R
[Fix] Export crashes on kernel-decorated function (#47808)
Rémi Ouazan committed
August 23, 2026
Y
[Gemma4] Fix stale expected values in integration tests (#48233)
Yih-Dar committed
August 21, 2026
Y
T
Post two CI badges on a PR: CPU PR CI and GPU run-slow (#48190)
Tarek Ziade committed
Y
T
Assign a reviewer even when a codeowner has left, and route models by modality (#48085)
Tarek Ziade committed