News for hackers. Stuff that matters.

A comment on Ollama Turbo. Parent comment

Re:Ollama Turbo by reissbaker on Tuesday August 05, 2025 @07:30PM (#44805804)
It was natively trained in FP4. Probably both to reduce VRAM usage at inference time (fits on a single H100), and to allow better utilization of B200s (which are especially fast for FP4).
[ Parent | on HN ]
Re:Ollama Turbo by irthomasthomas on Tuesday August 05, 2025 @07:53PM (#44805976)
Interesting, thanks. I didn't know you could even train at FP4 on H100s
[ Parent | on HN ]
Re:Ollama Turbo by reissbaker on Thursday August 07, 2025 @04:50PM (#44830186)
It's impressive they got it to work — the lowest I'd heard of this far was native FP8 training.
[ Parent | on HN ]