COMMITS
August 10, 2026
Z
K
feat(speculative): add Kimi K3 DSpark training (#3263)
khazzz1c committed
K
Z
fix(training): all-reduce grad-norm scalars on the mesh device (#3461)
Zeyu Zhou committed
A
fix(peft): align frozen LoRA tensors with compute dtype (#3470)
Alexandros Koumparoulis committed
A
docs: announce MuseGlimmer support (#3478)
Alexandros Koumparoulis committed
H
docs(models): add MuseGlimmer recipes and coverage (#3475)
Huiying committed
A
fix(docs): generate model coverage tables at build time (#3472)
Alexandros Koumparoulis committed
August 9, 2026
Y
fix(checkpoint): save ranked RNG state per global rank (#3437)
Yuhe Zhang committed
A
docs: improve accessibility for agents (#3424)
Alexandros Koumparoulis committed
Y
fix(vlm): correct CP gradients and vision sharding (#3414)
Yuhe Zhang committed
Z
fix(bagel): type SFT backend before AutoModel init (#3449)
Zeyu Zhou committed
Y
fix(peft): enable v5 expert adapters for Nemotron and MiniMax (#3439)
Yuhe Zhang committed
K
feat(streaming): add EAGLE-3 producer and loader over the local store (#3085)
Kashif Rasul committed
August 8, 2026
A
fix(gemma4): exact router scalar and fp32 reference routing for MoE parity (#3456)
Abhishree Thittenamane committed
A
feat: nemotron 3.5 cfg (#3442)
Adil committed
August 7, 2026
O
fix(retrieval): export Sentence Transformers metadata (#3021)
Oliver Holworthy committed
Y
test(checkpoint): stabilize MoE checkpoint parity gates (#3440)
Yuhe Zhang committed
H
fix: normalize MoE auxiliary loss during gradient accumulation (#3359)
Huiying committed
Y
test(checkpoint): redesign resume robustness around a shared trajectory (#3427)
Yuhe Zhang committed
Y
fix(retrieval): repair Ministral3 training recipe (#3441)
Yuhe Zhang committed
A
ci(vlm): raise Slurm wall time for MiniMax-M3 and Gemma4 recipes (#3448)
Abhishree Thittenamane committed
A
fix: Gemma4-31B CoderForge CP8/64K data processing and recipe (#3446)
Abhishree Thittenamane committed
K
feat(kimi_linear): context parallelism and packed-sequence support (#3243)
khazzz1c committed
S
fix(kimi_k25_vl): keep PEFT expert LoRA keys under the base_model.model. prefix (#3431)
stanley1208 committed
H
fix(#1755): Fix PEFT adapter outputs for v5 support (#3284)
hari committed
August 6, 2026
Y
fix(distributed): avoid duplicate FSDP2 prefetch all-gathers (#3411)
Yuhe Zhang committed
August 5, 2026
O
fix(retrieval): use stock Ministral embedding backbone (#3103)
Oliver Holworthy committed
A
ci(convergence): re-baseline Qwen3-MoE IFEval to 0.5989 (#3416)
Abhishree Thittenamane committed
S
feat(model_init): add FP8 pre-flight check for the force_hf path (#2350)
stanley1208 committed