Skip to content

Pull requests: NovaSky-AI/SkyRL

New pull request New
125 Open 1,055 Closed
125 Open 1,055 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

fix(docker): optimize Dockerfile.megatron to reduce image size by 1.21 GB run_train_megatron_gpu_ci
#1499 opened Apr 11, 2026 by dinhxuanvu Loading…
7
[train][multimodal][3/3] Trainer changes to extract multi-modal outputs from GeneratorOutput
#1498 opened Apr 11, 2026 by nithinvc Contributor Loading…
9
[skyrl][tinker] Use VLLMRenderer in SkyRL train backend
#1496 opened Apr 10, 2026 by nithinvc Contributor Loading…
1
10
[train][multimodal][1/3] Add vision support to generate() in new inference stack
#1494 opened Apr 10, 2026 by nithinvc Contributor Loading…
3 tasks done
12
[tinker] Fix single request batching in TinkerEngine
#1489 opened Apr 10, 2026 by pcmoritz Collaborator Loading…
3
[multimodal] add language_model_only flag for models like qwen3.5
#1487 opened Apr 9, 2026 by erictang000 Collaborator Loading…
8
[train][multimodal][2/3] Add multi-turn VLM generator
#1486 opened Apr 9, 2026 by nithinvc Contributor Loading…
2 tasks done
20
[skyrl][tinker] Multi-modal Tinker Sampling
#1484 opened Apr 9, 2026 by nithinvc Contributor Loading…
3 tasks done
4
[fix][train] Prompt-based mini-batching for step-wise training
#1483 opened Apr 9, 2026 by CharlieFRuan Member Loading…
3 tasks done
5
Add ppo as alias for dual_clip policy loss type
#1481 opened Apr 8, 2026 by j316chuck • Draft
[feat][train] Enable new inference codepath by default
#1480 opened Apr 8, 2026 by SumanthRH Member • Draft
Add prefix-aware merging for step-wise training
#1479 opened Apr 8, 2026 by CharlieFRuan Member Loading…
3 tasks done
4
feat: add max_tokens_per_microbatch config for token-based micro-batching
#1477 opened Apr 8, 2026 by erictang000 Collaborator Loading…
5
[CI] Migrate GPU CI to run on new inference codepath run_train_gpu_ci
#1476 opened Apr 8, 2026 by SumanthRH Member • Draft
feat: native Atropos-SHM integration and modular ingestion layer
#1473 opened Apr 7, 2026 by RUFFY-369 Loading…
10
[train] Enable expandable_segments to reduce GPU memory fragmentation run_train_gpu_ci
#1470 opened Apr 7, 2026 by CharlieFRuan Member • Draft
5 tasks done
1
7
[tinker] Support prompt_logprobs in SkyRLTrainBackend sample() path
#1461 opened Apr 6, 2026 by pbokc Contributor Loading…
18
[tinker] Support KL loss in SkyRLTrainBackend
#1460 opened Apr 5, 2026 by pbokc Contributor Loading…
8
feat: LLM-synthesized hints for failed trajectories
#1456 opened Apr 4, 2026 by dzorlu Loading…
4 tasks
6
[skyrl-train] feat: add native GMPO policy loss with validation and tests
#1449 opened Apr 2, 2026 by taivu1998 Loading…
3
Fix event-loop blocking in one-step-off async save/export paths
#1446 opened Apr 2, 2026 by taivu1998 Loading…
1
3
Change default KL estimator from k3 to k2 for loss-based KL
#1445 opened Apr 2, 2026 by taivu1998 Loading…
2
[skyrl-train] Flip grpo_norm_by_std default to false
#1443 opened Apr 2, 2026 by taivu1998 Loading…
1
[skyrl-train] Add trainer-side max_response_length for Dr. GRPO normalization and DAPO overlong handling
#1440 opened Apr 2, 2026 by taivu1998 Loading…
2
[WIP][tx] Add initial implementation of RayJaxBackend
#1418 opened Mar 31, 2026 by andrewsykim Contributor • Draft
1
1
ProTip! Exclude everything labeled bug with -label:bug.