[ROADMAP] TeleFuser Q3 2026

[lzx1413]

Overall Goal

1. Improve open-source project infrastructure
2. Fix existing bugs and stabilize service/API
3. Add quantifiable regression tests
4. Support new feature integration, such as KV cache, streaming inference, and distributed inference

🧠 New Model Integrations

• SwiftVR feasibility study / initial integration

New Feature Support

• Streaming Inference with KV Management
• Disaggregated Inference
• Multi-instance server support

⚡ Performance & Kernels

• NVF4 for world model
• SVDQuant support

🔗 Distributed Inference

• Ulysses SP for LingBot-World
  • Current status: not supported yet

🌐 API & Service

• Refactor service module and fix bugs
• Stabilize task lifecycle and OpenAI-compatible API

🧪 Testing

• Add quantifiable regression test suite
• Add pipeline smoke tests
• Add service/API regression tests
• Add packaging/install validation tests

📚 Documentation & Tooling

• Support online document website
• Improve contributor docs for testing, service, and examples

Deployment

• Make pip install package available

---

Progress Tracking

Status: 🔜 Not Started

Area	Progress	Key PRs/Issues
New Models	⬜ 0/1	SwiftVR
New Features	⬜ 0/3	KV streaming, disaggregated inference, multi-instance server
Performance	⬜ 0/2	NVF4, SVDQuant
Distributed	⬜ 0/1	LingBot-World Ulysses SP
API & Service	⬜ 0/2	service refactor, API/task lifecycle stability
Testing	⬜ 0/4	regression suite, pipeline smoke, service/API, install validation
Docs & Tooling	🟩 1/2	online docs done; contributor docs pending
Deployment	⬜ 0/1	pip install package

[ActivePeter]

Is there any difference between SGLang’s KV cache implementation and TeleFuser’s? Fluxon has been working on SGLang KV cache integration recently.