alibaba/ROLL

View on GitHub

An Efficient and User-Friendly Scaling Library for Reinforcement Learning with Large Language Models

Stars: 3.4k
Forks: 312
Open beginner issues: 0
Indexed issues: 101
Avg merge: 1h 2m
Merged PRs (30d): 2

Dominant language: Python
License: Apache-2.0
Last GitHub push: Aug 27, 2026
Latest indexed: Sep 14, 2026
Contributing guide: No contributing guide
Code of conduct: No code of conduct
Beginner labels: good first issue

Open All

101 open issues indexed

How do you compare ROLL with verl Open

alibaba/ROLL#6 · 6 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Why use DeepSpeed's implementation instead of torch's FSDP? Open

alibaba/ROLL#20 · 3 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
ROLL now supports the RLVR pipeline for Qwen2.5 VL Open

alibaba/ROLL#34 · 5 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Question about "_create_workers" Open

alibaba/ROLL#75 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
QwenVL Training Error When Using Specific GPUs on a Single Server Open

alibaba/ROLL#78 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
🚀 Recent Updates Summary for ROLL Project Open

enhancement

alibaba/ROLL#90 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 14, 2026
Benchmark Comparation with Verl Open

alibaba/ROLL#101 · 3 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Request for TOPR Training Examples or Configs on GSM8k/MATH Open

alibaba/ROLL#102 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Does ROLL support DAPO training for LLMs or VLMs? Open

alibaba/ROLL#103 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 14, 2026
LLM_judge question Open

alibaba/ROLL#106 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Issue with metrics of difficulty_mask_ratio and entropy Open

alibaba/ROLL#108 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Request to provide the XMind source file for a clearer version of the roll_map_high.png image. Open

alibaba/ROLL#109 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Does ROLL support Google TPU? Open

alibaba/ROLL#110 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
🚀 [2025/07/31] Recent Updates Summary for ROLL Project Open

enhancement

alibaba/ROLL#112 · 5 comments · 5 reactions · 1 assignee · Indexed Sep 14, 2026
Lack of Practical Documentation & Examples for New Users in ROLL RL Framework Open

documentation

alibaba/ROLL#115 · 2 comments · 1 reaction · 1 assignee · Indexed Sep 14, 2026
terminated和done Open

alibaba/ROLL#123 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
ifeval_rule_reward_worker bug Open

alibaba/ROLL#125 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Enhanced RL Environment: A Proposed Design for LLM/VLM Scenarios Open

enhancement good first issue

alibaba/ROLL#140 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 14, 2026
previous Ray instance has since crashed Open

alibaba/ROLL#147 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
deepspeed zero3 model_update error Open

alibaba/ROLL#150 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
support for 'ray job submit --working-dir . -- python start_xxx_pipeline.py ' start Open

alibaba/ROLL#154 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Error occurs when saving the model during megtron training Open

alibaba/ROLL#167 · 5 comments · 0 reactions · 1 assignee · Indexed Sep 14, 2026
框架设置中权重offload和reload交换过程中显存超出ray的最大显存设置，导致ray主动杀进程的情况。 Open

alibaba/ROLL#171 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
🚀 [2025/09/25] Recent Updates Summary for ROLL Project Open

alibaba/ROLL#173 · 9 comments · 2 reactions · 1 assignee · Indexed Sep 14, 2026
Cannot run quick start example Open

alibaba/ROLL#183 · 3 comments · 0 reactions · 1 assignee · Indexed Sep 14, 2026
Unexpected result nan？ Open

alibaba/ROLL#193 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Video text RL training? Open

alibaba/ROLL#203 · 4 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
间断显卡编号的设备映射问题（English translation: device_mapping cannot handle discontinuous GPU indexes） Open

alibaba/ROLL#205 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
🚀 [2025/10/31] Recent Updates Summary for ROLL Project——ROLL Flash Open

alibaba/ROLL#207 · 2 comments · 1 reaction · 0 assignees · Indexed Sep 14, 2026
Will support Qwen-Image/Qwen-Image-Edit RL pipeline? Open

alibaba/ROLL#210 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Inquires about custom reward and tool use Open

alibaba/ROLL#213 · 6 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
oom Open

alibaba/ROLL#219 · 7 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
IPRO的评测集会开源吗 Open

alibaba/ROLL#229 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
ROLL 是否可以加速reward模型（非LLM）计算与LLM GRPO的高效协同训练 Open

alibaba/ROLL#230 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Ray OOM prevention kills vLLM actors during engine initialization (collective_rpc/offload_states) unless GRPO_GROUP=1 Open

alibaba/ROLL#234 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 14, 2026
是否支持Wan2.2 5B的pipeline? Open

alibaba/ROLL#238 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
How to use a dataset in chat format Open

alibaba/ROLL#241 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
有关训推差异和old-log-pro Open

alibaba/ROLL#242 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
IPRO的权重会开源吗？自己train需要什么配置呢 Open

alibaba/ROLL#252 · 0 comments · 1 reaction · 0 assignees · Indexed Sep 14, 2026
Failed to register worker to Raylet: IOError Open

alibaba/ROLL#263 · 6 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
可以添加一些recipe吗，比如retool或者search-r1的复现。 Open

alibaba/ROLL#264 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 14, 2026
weight update progress过于缓慢 Open

alibaba/ROLL#279 · 3 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
请教：如何训练react类型多轮人机对话 Open

alibaba/ROLL#280 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
华为硬件的 feature 支持 Open

alibaba/ROLL#282 · 4 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
on-policy下，log-prob和old-logprob数值不一致，有一些崩溃现象 Open

alibaba/ROLL#284 · 4 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
ROLL目前是否有考虑一些VLA的模型的整合？我看verl已经弄了一下openvla-oft模型了 Open

alibaba/ROLL#285 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
🚀 [2025/12/8] Recent Updates Summary for ROLL Project Open

alibaba/ROLL#286 · 2 comments · 2 reactions · 0 assignees · Indexed Sep 14, 2026
rope_factor args will be ignored if do rope_scaling during the initialization of McaGPTModel Open

alibaba/ROLL#287 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Missing ETP Support in mca_adapter Under Expert Parallelism Open

alibaba/ROLL#290 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Agentic pipeline data prepare Open

alibaba/ROLL#291 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026

alibaba/ROLL

Get new issues in your inbox