huggingface / optimum-nvidia

Apache License 2.0
867 stars 86 forks source link

Use FP8 by default when on a supported device #50

Open laikhtewari opened 8 months ago

laikhtewari commented 8 months ago

If the user does not specify the use_fp8 flag as true or false, the default behavior should be FP8 when running on a supported device and FP16 otherwise. If the user does specify, use their option and error if running on an unsupported device