ROLL是一个高效且用户友好的强化学习库，专为大型语言模型（LLMs）设计，能够充分利用大规模GPU资源。它提升了LLM在与人类偏好对齐、复杂推理和多轮交互等关键领域的表现。ROLL采用多角色分布式架构，集成了先进技术，支持灵活的资源分配和任务调度，适合技术先锋、产品开发者和算法研究者。该库提供了丰富的奖励和环境选项、易于配置的设备映射以及多种强化学习算�
active 2025-05-28 → 2026-08-08 (UTC)
Activity over time
Daily event counts in the loaded window
Line chart, 438 days from 2025-05-28 to 2026-08-08. Pushes: 279 total, peak 10 in a day. Pull requests: 162 total, peak 7 in a day. Issues: 137 total, peak 6 in a day. Comments: 306 total, peak 9 in a day. Stars: 1,556 total, peak 172 in a day.
- Pushes
- Pull requests
- Issues
- Comments
- Stars
Stars, PRs, issues and forks are under-captured in the later part of this window. GH Archive progressively stopped capturing non-push events during 2026 — −95% or worse by the end of the window. Every series here except Pushes fades for that reason, so a decline above reflects the archive, not this repository. Pushes stay reliable throughout, so read them, and the contributor counts derived from them, as the real signal. Data health has the measurements.
Top contributors
Pushes, PRs, issues, reviews and comments — stars and forks excluded, so this is contribution rather than popularity
| Contributor | Contributions | Pushes | PRs | Comments |
|---|---|---|---|---|
| PanAndy | 220 | 62 | 56 | 76 |
| github-actions[bot] | 163 | 163 | 0 | 0 |
| kkkky123 | 54 | 25 | 29 | 0 |
| StephenRi | 53 | 15 | 15 | 11 |
| HuangJoJo | 49 | 10 | 1 | 24 |
| CLAassistant | 29 | 0 | 0 | 29 |
| FightingZhen | 22 | 0 | 0 | 16 |
| jingyushen | 18 | 3 | 12 | 2 |
| noemotiovon | 14 | 0 | 1 | 10 |
| wwxFromTju | 14 | 0 | 0 | 11 |
| millioniron | 10 | 0 | 1 | 7 |
| chocoded | 9 | 0 | 2 | 7 |
| hydrozhao | 8 | 0 | 2 | 6 |
| guoshengCS | 7 | 0 | 0 | 7 |
| tt0718 | 7 | 0 | 0 | 5 |
| lowdy1 | 7 | 0 | 0 | 6 |
| taoluo | 6 | 0 | 4 | 2 |
| liu-zichen | 6 | 1 | 1 | 4 |
| Schnabel-8 | 6 | 0 | 0 | 6 |
| lyz22233 | 5 | 0 | 0 | 1 |
Recent activity
Latest issues, pull requests and releases
- Issue comment#411chocoded2026-05-09 10:24Error: Qwen3.5-35B-A3B lora sft with mcore-adapter, error occurs when saving a ckpt.
- Issue comment#411Unofish2026-05-09 10:14Error: Qwen3.5-35B-A3B lora sft with mcore-adapter, error occurs when saving a ckpt.
- Issue comment#437CLAassistant2026-05-08 04:23feat: support mm_token_type_ids and 3D RoPE alignment for Qwen2/3-VL …
- Pull request#437Wangxiaoxiaoa2026-05-08 04:22
- Issue comment#372chocoded2026-04-28 06:11Does LORA works with Qwen3.5, when using the Megatron backend?
- Issue comment#300zfnice2026-04-28 06:06ipro的训练代码中奖励函数是不可微的，训练中梯度不会更新，和论文中描述不一致，请问是哪里有问题呢？
- Issue comment#425CLAassistant2026-04-20 06:27feat: add pyproject.toml for pip install and fix unit tests
- Issue comment#405HuangJoJo2026-04-17 06:25Add OpenReward Intergration
- Issue comment#219HuangJoJo2026-04-16 01:57oom
- Issue comment#419YisuZhou2026-04-13 04:27mcore-adapter从VLM的hfconfig中获取参数异常
- Issue comment#419guoshengCS2026-04-13 03:58mcore-adapter从VLM的hfconfig中获取参数异常
- Issue#416Unofish2026-04-08 09:09Error: Qwen3.5-35B-A3B lora sft with mcore-adapter, error occurs under batch_size > 1.
- Issue comment#416Unofish2026-04-08 09:09Error: Qwen3.5-35B-A3B lora sft with mcore-adapter, error occurs under batch_size > 1.
- Issue comment#416chocoded2026-04-08 06:40Error: Qwen3.5-35B-A3B lora sft with mcore-adapter, error occurs under batch_size > 1.
- Pull request#404ParagEkbote2026-04-07 09:21
- Issue comment#411Unofish2026-03-31 03:10Error: Qwen3.5-35B-A3B lora sft with mcore-adapter, error occurs when saving a ckpt.
- Issue comment#408HuangJoJo2026-03-30 05:37expandable_segments:True hardcoded in CUDA platform files causes error on older kernel; PYTORCH_CUDA_ALLOC_CONF ignored
- Issue#408cchen-hhz2026-03-28 11:23expandable_segments:True hardcoded in CUDA platform files causes error on older kernel; PYTORCH_CUDA_ALLOC_CONF ignored
- Issue comment#401PanAndy2026-03-27 03:37ROLL with OpenReward Environments
- Issue#405shamanez2026-03-27 03:35Add OpenReward Intergration
- Issue comment#398SF-evil2026-03-26 07:58Torch 2.8.0 + vLLM 0.10.2 开启 use_kl_loss 后训练qwen2.5-vl-3B-agentic崩溃，Torch 2.6.0 + vLLM 0.8.4 正常
- Issue comment#398SF-evil2026-03-26 07:56Torch 2.8.0 + vLLM 0.10.2 开启 use_kl_loss 后训练qwen2.5-vl-3B-agentic崩溃，Torch 2.6.0 + vLLM 0.8.4 正常
- Issue comment#398SF-evil2026-03-26 07:56Torch 2.8.0 + vLLM 0.10.2 开启 use_kl_loss 后训练qwen2.5-vl-3B-agentic崩溃，Torch 2.6.0 + vLLM 0.8.4 正常
- Issue comment#402PanAndy2026-03-26 07:26Add Trackio as a New Backend for Experiment Visualization
- Pull request#399joeyzyz2026-03-24 09:44
Totals cover only the window loaded into ClickHouse and count events, not GitHub's lifetime totals — 1,556 stars here means stars gained during the window, not the repo's star count.