feat(sglang): add end-to-end training and refit integration - #3330
#3330Draft
Kh4L wants to merge 26 commits into
NVIDIA-NeMo:mainNVIDIA-NeMo/RL:mainfrom
Kh4L:integration/sglang-training-shareableKh4L/NemoRL:integration/sglang-training-shareableCopy head branch name to clipboard
Draft
feat(sglang): add end-to-end training and refit integration#3330Kh4L wants to merge 26 commits intoNVIDIA-NeMo:mainNVIDIA-NeMo/RL:mainfrom Kh4L:integration/sglang-training-shareableKh4L/NemoRL:integration/sglang-training-shareableCopy head branch name to clipboard
Kh4L wants to merge 26 commits into
NVIDIA-NeMo:mainNVIDIA-NeMo/RL:mainfrom
Kh4L:integration/sglang-training-shareableKh4L/NemoRL:integration/sglang-training-shareableCopy head branch name to clipboard
Commits
Commits on Jul 24, 2026
feat(sglang): engine fault tolerance for rollouts
Show description for beff788authored andcommittedfeat(sglang): megatron backend weight refit for sglang rollouts
Show description for 214f028authored andcommittedfeat(sglang): mxfp8 quantized rollout and refit
Show description for 0365f85authored andcommittedfix(sglang): avoid ncclCommSplit hang in the refit weight-update group
Show description for 9235213andcommittedfeat(sglang): enable async-GRPO NeMo-Gym rollouts on the SGLang backend
Show description for 78e4982committedfix(sglang): dispatch NeMo-Gym context limit on backend, not vllm_cfg presence
Show description for 4f5cea3committedfix(sglang): harden async Gym config handling
Show description for 58a6c41committedfeat(nemo-gym): opt-in truncation of non-contiguous rollout episodes
Show description for 3667506committedfix(nemo-gym): drop call-site default for truncate_noncontiguous_episodes
Show description for 099a16dcommittedfix(nemo-gym): centralize rollout contiguity defaults
Show description for e2e84d4committedfix(sglang): preserve runtime ownership across serialization
Show description for 186152bcommittedfix(sglang): require cache invalidation before refit
Show description for e3396aecommittedfeat(sglang): emit refit lifecycle markers
Show description for 26342afcommittedtest(sglang): cover async backend selection and refit loop
Show description for 5850729committedfix(sglang): validate runtime configuration
Show description for 8f9ff6fcommitteddocs(sglang): add reproducible training and refit recipes
Show description for 25e32bfcommittedfix(sglang): bound and recover distributed refits
Show description for c7c1611committedfix(sglang): harden async backend lifecycle
Show description for 6c07857committedfix(sglang): fail closed across startup and refit
Show description for bd2461fcommitteddocs(sglang): add strict public refit reproducers
Show description for dbec358committedfix(sglang): make public evidence revision-exact
Show description for dd5c30ccommittedchore(gym): pin NeMo-Gym submodule to the token-exact SGLang adapter
Show description for 7833c77committedfix(sglang): dispatch the SGLang refit without importing a policy worker
Show description for dfbf3decommittedfix(sglang): make the refit reproducers runnable and their checks collectable
Show description for da23bedcommitted
Commits on Jul 25, 2026
test(sglang): add a four-GPU-per-node async Gym recipe variant
Show description for d322b52committed
Commits on Jul 26, 2026
fix(sglang): declare uv_venv_dir so the async Gym recipes compose
Show description for 76d68b3committed