https://dujh22.github.io/LLMDailyDigestWeb/tech/ Recent content in 技术 on LLM-DailyDigest · 大模型研究日报 Hugo zh-CN Thu, 27 Aug 2026 23:41:07 +0800 https://dujh22.github.io/LLMDailyDigestWeb/tech/common-open-source-frameworks/ Thu, 27 Aug 2026 00:00:00 +0800 https://dujh22.github.io/LLMDailyDigestWeb/tech/common-open-source-frameworks/ <h1 id="常用开源框架">常用开源框架</h1> <p>围绕大模型研发全流程（<strong>后训练 → 推理服务 → 评测</strong>）的常用开源框架，以下仓库本站均已 fork，便于跟读源码与二次开发。</p> <table> <thead> <tr> <th>框架</th> <th>定位</th> <th>上游仓库</th> <th>本站 fork</th> </tr> </thead> <tbody> <tr> <td>verl</td> <td>LLM 强化学习后训练</td> <td><a href="https://github.com/verl-project/verl"target="_blank" rel="external nofollow noopener noreferrer">verl-project/verl<i class="fa-solid fa-external-link-alt fa-fw fa-xs ms-1 text-secondary" aria-hidden="true"></i></a></td> <td><a href="https://github.com/dujh22/verl"target="_blank" rel="external nofollow noopener noreferrer">dujh22/verl<i class="fa-solid fa-external-link-alt fa-fw fa-xs ms-1 text-secondary" aria-hidden="true"></i></a></td> </tr> <tr> <td>SGLang</td> <td>LLM / 多模态推理服务</td> <td><a href="https://github.com/sgl-project/sglang"target="_blank" rel="external nofollow noopener noreferrer">sgl-project/sglang<i class="fa-solid fa-external-link-alt fa-fw fa-xs ms-1 text-secondary" aria-hidden="true"></i></a></td> <td><a href="https://github.com/dujh22/sglang"target="_blank" rel="external nofollow noopener noreferrer">dujh22/sglang<i class="fa-solid fa-external-link-alt fa-fw fa-xs ms-1 text-secondary" aria-hidden="true"></i></a></td> </tr> <tr> <td>SkyThought</td> <td>推理模型训练全流程</td> <td><a href="https://github.com/NovaSky-AI/SkyThought"target="_blank" rel="external nofollow noopener noreferrer">NovaSky-AI/SkyThought<i class="fa-solid fa-external-link-alt fa-fw fa-xs ms-1 text-secondary" aria-hidden="true"></i></a></td> <td><a href="https://github.com/dujh22/SkyThought"target="_blank" rel="external nofollow noopener noreferrer">dujh22/SkyThought<i class="fa-solid fa-external-link-alt fa-fw fa-xs ms-1 text-secondary" aria-hidden="true"></i></a></td> </tr> <tr> <td>OpenJudge</td> <td>AI 应用评测与质量奖励</td> <td><a href="https://github.com/agentscope-ai/OpenJudge"target="_blank" rel="external nofollow noopener noreferrer">agentscope-ai/OpenJudge<i class="fa-solid fa-external-link-alt fa-fw fa-xs ms-1 text-secondary" aria-hidden="true"></i></a></td> <td><a href="https://github.com/dujh22/OpenJudge"target="_blank" rel="external nofollow noopener noreferrer">dujh22/OpenJudge<i class="fa-solid fa-external-link-alt fa-fw fa-xs ms-1 text-secondary" aria-hidden="true"></i></a></td> </tr> </tbody> </table> <h2 id="verl--llm-强化学习后训练框架">verl — LLM 强化学习后训练框架</h2> <p><a href="https://github.com/verl-project/verl"target="_blank" rel="external nofollow noopener noreferrer">verl<i class="fa-solid fa-external-link-alt fa-fw fa-xs ms-1 text-secondary" aria-hidden="true"></i></a>（Volcano Engine Reinforcement Learning for LLMs）由字节跳动 Seed 团队发起、verl 社区维护，是 <a href="https://arxiv.org/abs/2409.19256"target="_blank" rel="external nofollow noopener noreferrer">HybridFlow<i class="fa-solid fa-external-link-alt fa-fw fa-xs ms-1 text-secondary" aria-hidden="true"></i></a>（EuroSys 2025）论文的开源实现，主打灵活、高效、可落地的 LLM RL 后训练。</p>