You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Goal: Low precision weights allow testing a 405B on one H100 node
Currently, the weights are stored in bf16 (see #100 for a fix on that front). However, jax supports other datatypes, and further libraries such as https://github.com/neuralmagic/compressed-tensors support yet more.
Adding support for alternative dtypes (and, ideally, lower-precision datatypes) would significantly simplify testing at scale.
The text was updated successfully, but these errors were encountered:
Goal: Low precision weights allow testing a 405B on one H100 node
Currently, the weights are stored in bf16 (see #100 for a fix on that front). However, jax supports other datatypes, and further libraries such as https://github.com/neuralmagic/compressed-tensors support yet more.
Adding support for alternative dtypes (and, ideally, lower-precision datatypes) would significantly simplify testing at scale.
The text was updated successfully, but these errors were encountered: