FP8 Optimization for LingBot-Video 1.3B
ALX-OPEN-SOURCE · reddit · 2026-07-11
The author created a selective FP8 version of LingBot-Video 1.3B for ComfyUI: keeping quality-sensitive layers in BF16 and only converting parts that genuinely benefit from FP8.
On an RTX 5080, sampling time dropped from roughly 4.651s to 3.651s, showing a decent speedup. The author notes that the current workflows and nodes are still quite experimental, and Comfy will likely offer more official support later. His repository also contains some experimental workflows, like first-frame/last-frame video generation, though stability is currently average.
More from Infra
- Engram's random reads don't suit SSDs; CPU-memory over NVLink could serve all 72 GPUs — bookwormengr · 2026-09-11
- 80% of the DIY LLM inference hype posters have already quit — it's brutally hard systems work — abhijithneil · 2026-09-11
- Hugging Face's Ultra Scale Playbook: a free book on training LLMs on GPU clusters — mdancho84 · 2026-09-11
- Is inference latency becoming the biggest bottleneck for production AI agents? — Euphoric_Sea632 · 2026-09-11
- LLM Serving Metrics Thread: Why TPOT and Uptime Make or Break User Experience — abhijithneil · 2026-09-11
- PlanetScale launches sharded Postgres: 768 servers acting as one, 1PB scale — dhruv2038 · 2026-09-11