HN
Hacker News
Top
New
Best
Ask
Show
Jobs
Comment by phh | Hacker News Reader
Parent
Full thread
phh
·
You're running fp32 models, fp16 or quantized?
View on HN
tome
·
FP16 for calculating all activations. Some data is stored as FP8 at rest.
Reply on news.ycombinator.com