feat(refit): support BF16 to MXFP8 NCCL reshard - #3477

Open

seonjinn wants to merge 16 commits into

NVIDIA-NeMo:mainfrom

seonjinn:sna/nccl-reshard-bf16-mxfp8

Commits on Aug 3, 2026

feat(refit): support BF16 to MXFP8 NCCL reshard
seonjinn
committed
f3dcd2a
test(refit): cover matching FP8 reshard routing
seonjinn
committed
4d1522b

Commits on Aug 5, 2026

fix(vllm): restore grouped MXFP8 quantization shape
seonjinn
committed
d63ed16
Merge remote-tracking branch 'origin/main' into sna/nccl-reshard-bf16-mxfp8
seonjinn
committed
6f57c1b

Commits on Aug 7, 2026

fix(vllm): initialize MXFP8 MoE kernel after refit load
seonjinn
committed
f7964a6
Merge upstream main into sna/nccl-reshard-bf16-mxfp8
seonjinn
committed
d420d1c

Commits on Aug 11, 2026

Merge remote-tracking branch 'upstream/main' into sna/nccl-reshard-bf16-mxfp8
seonjinn
committed
83738f9

Commits on Aug 12, 2026

Commits on Aug 13, 2026

Merge remote-tracking branch 'upstream/main' into sna/nccl-reshard-bf16-mxfp8
seonjinn
committed
ffb6d47
test(refit): harden MXFP8 receiver metadata checks
seonjinn
committed
98d00ab
fix(refit): clarify canonical trainer precision
seonjinn
committed
778c218
fix(refit): address BF16 to MXFP8 review feedback
seonjinn
committed
9d2a1d0
fix(vllm): reject MXFP8 without FP8 precision
seonjinn
committed
68bf4c9

Commits on Aug 17, 2026

fix(refit): support runtime MXFP8 scale parameters
seonjinn
committed
6f2b4fb
refactor(refit): name BF16 to MXFP8 quantization spec
seonjinn
committed
31ee1c7