Repository navigation

feat(dynamo): refit, FP8 and speculative decoding via the vLLM worker extension - #4725

Open

jthomson04 wants to merge 11 commits into

NVIDIA-NeMo:mainfrom

jthomson04:jthomson04/dynamo-extension-refit

Open

feat(dynamo): refit, FP8 and speculative decoding via the vLLM worker extension#4725
jthomson04 wants to merge 11 commits into
NVIDIA-NeMo:mainfrom
jthomson04:jthomson04/dynamo-extension-refit

Commits on Oct 23, 2026
- build(dynamo): run managed Dynamo from a nemo-rl venv on vLLM 0.29
  
  jthomson04
  and
  claude
  committed
  9bdba86
- feat(dynamo): nccl_reshard refit through NeMo RL's worker extension
  
  jthomson04
  and
  claude
  committed
  7bc3c6d
- feat(dynamo): route all refit and FP8 through NeMo RL's worker extension
  
  jthomson04
  and
  claude
  committed
  1df8a08
- feat(dynamo): speculative decoding with drafter refit
  
  jthomson04
  and
  claude
  committed
  724c0fa
- fix(dynamo): keep the SingleController refit deadline for Dynamo
  
  jthomson04
  and
  claude
  committed
  a4d6c3f
- style: satisfy pyrefly and ruff import order
  
  jthomson04
  and
  claude
  committed
  61c83f5
- fix(dynamo): address review on the extension refit path
  
  jthomson04
  and
  claude
  committed
  6cda886
- refactor(generation): share the spec-decode clamp and MTP drafter setup
  
  jthomson04
  and
  claude
  committed
  8c1738e
- perf(dynamo): pause managed Dynamo in place for refit, as vLLM does
  
  jthomson04
  and
  claude
  committed
  ea04798
- fix(dynamo): follow upstream MXFP8 prequantize and draft co-training
  
  jthomson04
  and
  claude
  committed
  bb51bbc
- fix(dynamo): apply the Nemotron-H fp32 LM head in managed engines
  
  jthomson04
  and
  claude
  committed
  e2fb3ab