sgl-project
Development Roadmap (2025 H2)
SpecForge aims to make speculative decoding more accessible by providing ready-to-use Eagle-3 draft models for the most popular open-source language models. It is also committed to supporting future SOTA models as they are open-sourced, ensuring developers can continuously benefit from cutting-edge performance with minimal integration effort.
Dataset
We wish to create a standard, multi-domain, balanced dataset commonly used for our future Eagle3 model release. In this way, we ensure that our released model's performance is actually satisfying on different benchmarks other than MTBench.
- [x] General-Purpose Corpus (1M high-quality examples) @FrankLeeeee
- [ ] Domain-Specific Data (Math & Coding) @FrankLeeeee
- [ ] multi-modal dataset
Model support
We aim to support the current mainstream models.
- [ ] Kimi K2 @FlamingoPg @sleepcoo
- [x] Qwen 3 MoE @yubofredwang @sleepcoo @ZhengHSI
- [x] Qwen 3 Coder @yizhang2077 @zyksir https://github.com/sgl-project/SpecForge/pull/73
- [ ] DeepSeek V3/R1
MultiModal
Support vision-langauge models in collaboration with Nvidia
- [x] VLM @shuaills @yyihuang
Standardization
Standardize the SOP for open-source community.
- [x] Align SpecForge's dependency library versions with sglang's. @yizhang2077 @sleepcoo https://github.com/sgl-project/SpecForge/pull/90
- [x] PyPI @FrankLeeeee
- [x] CI @FrankLeeeee
- [x] Documentation on Github pages @FlamingoPg
- [x] Docker @sleepcoo @FlamingoPg
- [ ] model card @FlamingoPg
New algorithm
- [ ] Eagle 4 (collaborate with EAGLE team)
Training Enhancement
- [x] FlexAttention (for eagle3 attention mask)
- [ ] Sequence Parallel (for long ctx training)
- [ ] Expert parallel
- [ ] FSDP performance tuning @ChenQiaoling00
GPT-OSS Support
- [x] Dataset generation
- [x] Code Adaptation https://github.com/sgl-project/SpecForge/tree/feature/gpt-oss-support
- [x] Aux layers configuration
- [x] hf tp plan
- [x] 20B training
- [x] 120B training
Everyone interested in joining our development, welcome join our slack channel https://sgl-fru7574.slack.com/archives/C09784E3EN6
Greate job! If I want to participate in VLM, what can I do?
You can contact me on Slack. We have an eagle-vlm team. My slack is Chao Wang. @Lzhang-hub
Interested in supporting DS V3/R1, who should I reach out to?
You can search for specforge in the sgl project of slack.
Everyone interested in joining our development, welcome join our slack channel https://sgl-fru7574.slack.com/archives/C09784E3EN6
