COMMITS
July 17, 2026
Y
J
Avoid extra MFSDP v2 model-weight sync memcpy (#5834)
Jingyue Wu committed
Y
fix bug where Gemma4 is not working with recompute_granularity = "full" (#5324)
Yury Parfenov committed
J
J
Allow parameterless FSDP root modules (#5711)
Jingyue Wu committed
O
Fix for sequence-level aux MoE loss being dependent on batch size (#5798)
OlegSudakov committed
July 16, 2026
S
Inference: Extend default cuda-graph coverage to 512 tokens (#5797)
Siddharth Singh committed
S
test(mfsdp): AUT-881 mark test_overlaps_communication_and_compute flaky (#5848)
svcnemo-autobot committed
M
Fix averaging for MoE z-loss metric tracking (#3199)
Marks101 committed
J
Pass device IDs to cleanup barrier (#5702)
Jingyue Wu committed
K
Refactor data parallel coordinator to enable modular handlers (#5550)
Keshav Santhanam committed
S
return prefix cache hits data from the chat completions api (#5609)
Siddharth Singh committed
O
chore(skills): remove Open Plugin manifest (superseded) (#5842)
oliver könig committed
O
chore(skills): add Regent Open Plugin manifest (#5840)
oliver könig committed
J
Overlap FSDP communication with compute (#5719)
Jingyue Wu committed
J
Clarify NVIDIA email signing guidance (#5699)
Jingyue Wu committed
J
Pair frozen FSDP backward hooks (#5710)
Jingyue Wu committed
P
Add GPTModel to HybridModel migration guide (#5698)
Philip Petrakian committed
July 15, 2026
S
IMA fix by making the copy of book keeping buffer to GPU blocking (#5715)
Shanmugam Ramasamy committed
A
ci: Enhance nightly/mr/weekly error reporting (#5831)
Ajay committed
K
Set num_splits to 0 for FA4 inference (#5804)
Keshav Santhanam committed
S
chore(tests): AUT-851 move NCCL defaults from run_ci_test.sh to conftest (#5826)
svcnemo-autobot committed
Y
S
Fix configured norm epsilon in MambaLayer (#5750)
Shanhao committed
J
Avoid FSDP unit terminology in MFSDP v2 (#5793)
Jingyue Wu committed
J
Missing moe_router_dtype causes unexpected downcast in ModelOpt example (#5810)
Jinhang Choi committed
S
Inference: Add the nemotron_v3 reasoning parser (#5634)
Siddharth Singh committed
G
chore: rotate oncall schedule
github-actions[bot] committed
A
Pin cudnn-fe and cuTeDSL version (#5812)
Ajay committed