[feat] 2/n Adding NVFP4 Humans& Quantization - #1967

Draft

kailash109 wants to merge 21 commits into

NovaSky-AI:mainfrom

kailash109:nvfp4

Commits on Jul 21, 2026

Add FP8 training and rollout weight synchronization
jinghanyao1-hub
committed
b2f59b6
Guard FP8 runtime setup for FSDP configs
jinghanyao1-hub
committed
ea2e6c1
Batch MoE tensors during serialized FP8 weight sync
jinghanyao1-hub
committed
4a5cfa5

Commits on Jul 22, 2026

expert mxfp8 added config + modal script for perf testing
kailash109
committed
78956b9
increae timeout, benchmark w larger batch size
kailash109
committed
28bfa2c

Commits on Jul 23, 2026

Commits on Jul 27, 2026

Commits on Jul 28, 2026

rename normalize moe target
kailash109
committed
41a0394
remove vllm weight config
kailash109
committed
3fe13b7
remove unnecessary target dataclasses
kailash109
committed
323007f
simplify registry (no need for dyamic registration)
kailash109
committed
f2a44fc
remove more weight-naming bs
kailash109
committed
3dd2078
refactor -- unified quantization policy + weight sync strategy now define entire customizable quantization configurations
kailash109
committed
94ec50c
refactoring to quantization strategy + model quantization layout -- add runtime env var setting for NVfp4 prep
kailash109
committed
711e3a5
delete outdated concurrency benchmark
kailash109
committed
acc2ab9
gemini code review fixes
kailash109
committed
bc3b51a

Commits on Jul 31, 2026

e2e humans& nvfp4 rl init
kailash109
committed
1112b3f
remove non TE quantization path
kailash109
committed
e3176c2