Are the weights in float32 of float16 ?

#18
by benjamin-breton-loreal - opened

Are the weights in float32 of float16 ? Saving them in float16 could reduce load time.

benjamin-breton-loreal changed discussion title from Are the weights in float32 of float16 ? Saving them in float16 could reduce load time. to Are the weights in float32 of float16 ?
Databricks org

They're saved in 16-bit (12B params, 24GB on disk, you can see). You can load as fp16 or bf16.

srowen changed discussion status to closed

Sign up or log in to comment