Skip to content

Pull requests: NovaSky-AI/SkyRL

New pull request New
248 Open 1,534 Closed
248 Open 1,534 Closed
Author
Filter by author
Loading
Label
Filter by label
Loading
Use alt + click/return to exclude labels
or ⇧ + click/return for logical OR
Projects
Filter by project
Loading
Milestones
Filter by milestone
Loading
Reviews
Filter by reviews
No reviews Review required Approved review Changes requested
Assignee
Filter by who’s assigned
Assigned to nobody Loading
Sort
Sort by
Newest Oldest Most commented Least commented Recently updated Least recently updated Best match
Most reactions
👍 👎 😄 🎉 😕 ❤️ 🚀 👀

Pull requests list

[eval][1/3] Introduce EvalDispatcher (towards async evals)
#2198 opened Sep 12, 2026 by kyuds Collaborator • 2/2 Loading…
8
Prototype LoRA RDT adapter transport
#2196 opened Sep 10, 2026 by hershg Contributor • Draft
Compact duplicate LoRA adapter tensors
#2195 opened Sep 10, 2026 by hershg Contributor • Draft
Use vLLM 0.28 draft sync endpoints and add MTP tests
#2194 opened Sep 10, 2026 by zanderjiang Contributor Loading…
3
Accept pathlib.Path in checkpoint I/O helpers
#2193 opened Sep 9, 2026 by ethantsliu Loading…
1
[GLM 5.3 Tests] Reproduce trainer batch-size logprob sensitivity
#2192 opened Sep 9, 2026 by hershg Contributor Loading…
8
fix(tinker): finalize failed saves for unloaded models
#2186 opened Sep 9, 2026 by LOGO127 Loading…
8
[GLM 5.3 Tests] Check trainer/inference LoRA logprob agreement
#2182 opened Sep 8, 2026 by hershg Contributor Loading…
14
[GLM 5.3 Tests] B300 profile and profiled Tinker client
#2181 opened Sep 8, 2026 by hershg Contributor Loading…
41
[megatron] GLM-5.3-Flash (glm5_next) support: KDA + NoPE-MLA/DSA + mHC hybrid MoE, 4-layer GPU CI entry
#2179 opened Sep 8, 2026 by erictang000 Collaborator Loading…
5
[eval] Consolidate Eval Functions for Step-Wise and Non-Step Wise
#2174 opened Sep 7, 2026 by kyuds Collaborator • 1/2 Loading…
18
Support final normalization of streamed Megatron gradients
#2173 opened Sep 7, 2026 by oran-goodman • Draft
fix(megatron): snapshot precision-aware optimizer state
#2172 opened Sep 7, 2026 by hershg Contributor Loading…
7
fix(sft): exclude first tokens from pretokenized loss windows
#2171 opened Sep 6, 2026 by feng1201 Loading…
3
[tinker] Load harness for the API server's sampling path at 131k concurrency
#2166 opened Sep 5, 2026 by avigyabb Collaborator • Draft
[tinker] Decode vLLM completion bodies straight into numpy with pysimdjson
#2165 opened Sep 5, 2026 by avigyabb Collaborator Loading…
3
[tinker] Encode forwarded sample results to proto once and serve them as-is
#2164 opened Sep 5, 2026 by avigyabb Collaborator Loading…
8
[tinker] Forward samples with aiohttp instead of httpx
#2161 opened Sep 5, 2026 by avigyabb Collaborator Loading…
8
[megatron] Skip frozen-master backload for adapter-only LoRA weight sync
#2157 opened Sep 4, 2026 by avigyabb Collaborator • Draft
fix(megatron): stage common state for cloud resume
#2154 opened Sep 3, 2026 by bvolpato Contributor • Draft
fix(megatron): isolate FP32 adapter optimizer state
#2153 opened Sep 3, 2026 by bvolpato Contributor • Draft
fix(megatron): refresh masters after weights-only load
#2149 opened Sep 3, 2026 by bvolpato Contributor • Draft
fix(rollout): preserve sampled EOS probabilities
#2148 opened Sep 3, 2026 by bvolpato Contributor • Draft
fix(inference): reset requests when discarding KV cache
#2147 opened Sep 3, 2026 by bvolpato Contributor • Draft
fix(megatron): synchronize complete accumulation windows
#2146 opened Sep 3, 2026 by bvolpato Contributor • Draft
ProTip! Updated in the last three days: updated:>2026-09-08.