COMMITS
August 20, 2026
Y
refactor(checkpoint): gate direct loads by adapter capability (#3574)
Yuhe Zhang committed
August 19, 2026
S
fix(kimi_k3): apply the HF renames to expert LoRA keys on save (#3435)
Stanley Mei committed
K
feat(datasets): support SynthTraces agent SFT (#3371)
khazzz1c committed
Y
fix(perf): honor configured peak in MFU logging (#3569)
Yuhe Zhang committed
A
test: add functional coverage for under-tested recipes (#3584)
Alexandros Koumparoulis committed
A
fix(minimax): declare packed-sequence and CP attention ownership (#3547)
Abhishree Thittenamane committed
K
docs(dllm): align DiffusionGemma guides with the SFT recipe (#3550)
Kashif Rasul committed
August 18, 2026
A
perf(glm): skip top-k pipeline carry without IndexShare (#3573)
Alexandros Koumparoulis committed
H
fix(kimi_linear): make long-context recipe CP8 (#3571)
Huiying committed
P
feat(nemotron): allow skipping logits materialization (#3555)
Piotr Żelasko committed
H
fix(checkpoint): restore single-GPU custom-model DCP loading (#3533)
Huiying committed
Y
fix(model_init): stream Devstral FP8 checkpoints (#3413)
Yuhe Zhang committed
Y
test(checkpoint): improve resume correctness diagnostics (#3562)
Yuhe Zhang committed
S
fix(recipes): support MTP with context parallelism (#3544)
Slyne Deng committed
P
fix(moe): stabilize async DeepEP checkpointing (#3522)
Piotr Żelasko committed
August 17, 2026
S
fix(mtp): support context-parallel caller inputs (#3543)
Slyne Deng committed
D
fix(deps): resolve msgpack, wandb, and mistune container CVEs (#3563)
Dong Hyuk Chang committed
O
fix(docker): harmonize non-root runtime contract (#3484)
oliver könig committed
August 16, 2026
F
fix(training): correct dataloader annotation and modernize Optional syntax (#3534)
Faisal Alsrheed committed
S
fix(metric_logger): expose the logger buffer size in the factory builder (#3541)
Sahel Sharifymoghaddam committed
K
fix(checkpoint): restore Gemma4 Unified HF export keys (#3260)
khazzz1c committed
August 15, 2026
H
fix(vlm): use FP32 master weights for Nemotron Omni (#3444)
Huiying committed
H
A
fix(fsdp): uniform reduce dtype and EP-local expert gradients (#3540)
Alexandros Koumparoulis committed
August 14, 2026
A
feat(qwen): add Qwen3.8-27B fine-tuning recipes (#3560)
Alexandros Koumparoulis committed
A
fix(docker): build bitsandbytes for SM121 (#3553)
Alexandros Koumparoulis committed
A
ci: route DGX Spark recipes to GB10 (#3538)
Alexandros Koumparoulis committed
C
fix(minimax_m3_vl): declare and honour logits_to_keep so fused losses are kept (#3511)
coderaBruce committed
H
H
feat(gemma4): add E-series tensor parallelism (#3512)
Huiying committed