COMMITS
September 2, 2026
G
[peft] fix: LoRA adapter init no longer depends on tensor parallel size (#5843)
GavinZhu-GMI committed
O
fix(training): avoid double-counting MTP MoE metric layers (#5795)
Oded Ovadia committed
H
fix: dequantize GLM 5 FP8 weights on import (#5851)
Hersh Godse committed
Y
feat(model): add Muse Glimmer support (#5479)
Yu Yao committed
S
chore(deps): AUT-2186 bump Megatron-Core nightly pins (#5891)
svcnemo-autobot committed
September 1, 2026
Y
[docs] fix: correct Nemotron 3.5 Lightning page (#5936)
Yu Yao committed
Y
[build] chore: bump MCore main to 3134ff9 (#5932)
Yu Yao committed
J
feat(training): support MFSDP V2 expert parallelism (#5521)
Jingyue Wu committed
J
fix(peft): replicate adapters for duplicated linears (#5900)
JackRao123 committed
L
fix(model): handle singleton packed Qwen3-VL batches (#5904)
leooop committed
Y
Y
[misc] fix: Honor async generation sequence limit (#5926)
Yu Yao committed
Y
[training] fix: persist checkpoints on planned exit (#5922)
Yu Yao committed
Y
[peft] fix: Track recompute patches by model lifetime (#5929)
Yu Yao committed
Y
[build] chore: bump MCore main to 23dc642 (#5912)
Yu Yao committed
C
fix(training): restrict HybridEP padding to THD (#5915)
Chen Cui committed
C
fix(peft): support hybrid-stack activation recompute (#5586)
Charles Tang committed
B
perf(deepseek): add DeepSeek V4 Flash Blackwell recipes (#5903)
Bhargav committed
R
feat(perf): add Nemotron 3.5 Lightning B200 and B300 recipes (#5793)
Raghav Hrishikeshan Mukundan committed
M
Revert "feat(recipe): Add NT3 Ultra NVFP4 perf-recipes" (#5920)
malay-nagda committed
August 31, 2026
Y
[peft] fix: Support partial canonical LoRA runtime (#5901)
Yu Yao committed
R
feat(recipe): Add NT3 Ultra NVFP4 perf-recipes (#5798)
Ravi Ghadia committed
K
refactor(data): remove redundant histogram reconstruction in create_hist (#5863)
Kenta Izumi committed
A
fix(data): keep object-store schemes out of memmap index paths (#5643)
Asaf Gardin committed
Y
fix(docs): refresh DeepSeek V4 verification pages (#5910)
Yu Yao committed
C
feat(recipes): modernize DeepSeek V4 library configs (#5554)
Chen Cui committed
M
no overlap_param_gather_with_optimizer_step (#5847)
malay-nagda committed
Y