インフラ
SGLang
Open-source serving engine that streams realtime/causal video over WebSocket — incremental frames with reusable KV-cache state.
docs.sglang.io · 掲載≠推奨 · 2026-09-03 再確認
- 運営
- Open source · SGLang project
- 形態
- WebSocket realtime-video endpoint
- モデル
- Realtime pipelines: SANA-WM · LingBot World
- 利用状況
- Self-host (open source)
- 役割
- Serving layer for your own live-gen endpoint
SGLang とは?
SGLang — best known as an LLM serving framework — has been extended to diffusion with a realtime path: a /v1/realtime_video/generate WebSocket endpoint streams video frames incrementally while keeping a live session whose causal KV-cache is reused across chunks until disconnect. It ships realtime pipelines for world models (LingBot World, NVIDIA SANA-WM) and is the self-hosted serving layer if you want to run your own live-gen endpoint rather than rent one.
SGLang のできること
- A /v1/realtime_video/generate WebSocket endpoint that streams frames incrementally
- Reuses a session’s causal KV-cache state across chunks until disconnect — no regeneration from scratch
- Ships realtime pipelines for world models (LingBot World, NVIDIA SANA-WM)
- MessagePack framing plus delta-gzip frame compression for low-bandwidth streaming
- Open source — self-host the serving layer rather than renting a live-gen endpoint
SGLang は誰向け?
- Serving your own realtime/causal video models behind a WebSocket API
- Building world-model streaming (camera-controlled exploration) pipelines
- Self-hosting live generation to control cost and data flow
SGLang の長所と短所
長所
- Open source with real engineering — an LLM serving framework extended to diffusion
- Reusable KV-cache state is the right primitive for continuous streaming
- Avoids per-second vendor markup on always-on live generation
短所
- A serving engine — you still supply and operate the model
- Realtime diffusion support is newer and narrower than its LLM serving core
- World-model pipelines are research/niche as of Sep 2026
世間の声
公開報道からの引用 — 収集日 2026-09-03。
“Realtime and causal video pipelines generate video incrementally and reuse state across chunks.”
よくある質問
What is SGLang’s realtime video support?
A WebSocket endpoint (/v1/realtime_video/generate) that generates video incrementally and reuses a live session’s KV-cache state across chunks.
Is SGLang open source?
Yes — SGLang is an open-source serving framework, and the diffusion/realtime extensions are part of the same project.
What can I serve with it?
It ships realtime pipelines for world models like LingBot World and NVIDIA SANA-WM, plus your own causal diffusion models.
