Skip to content

Conversation

balancap
Copy link
Contributor

@balancap balancap commented Oct 2, 2024

The main trick is to use dqt_dtype = jnp.bfloat16 instead of jnp.float32 for "dequantization". As presented in bug ticket jax-ml/jax#24051, there is still some issues with the latest JAX version when no relu or abs-max is used.

The main trick is to use `dqt_dtype = jnp.bfloat16` instead of `jnp.float32` for "dequantization".
As presented in bug ticket jax-ml/jax#24051, there is still some issues with the latest JAX version
when no `relu` or `abs-max` is used.
@balancap balancap self-assigned this Oct 2, 2024
@balancap balancap added the bug Something isn't working label Oct 2, 2024
@balancap balancap merged commit 365e5d9 into main Oct 2, 2024
2 checks passed
@balancap balancap deleted the fix-fp8-matmul-notebook-latest-jax branch October 2, 2024 08:49
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

bug Something isn't working

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant