Skip to content

Mixed precision training support #491

Closed as duplicate
Closed as duplicate

Description

@ogencoglu

Does compute_gramian() support mixed precision training? I couldn't make it work at least.

Activity

  1. PierreQuinton commented on Dec 19, 2025

    @PierreQuinton
    Contributor

    Thanks for the issue!

    It is not clear to me what the problem/goal is. Could you provide us with a minimal example of what you want to do?

  2. ValerianRey commented on Apr 24, 2026

    @ValerianRey
    Member

    Had a bug a few days ago with autojac.jac related to bfloat16. It wasn't autogram but it still may be related. I'll try to post a simple reproducer.

  3. ValerianRey commented on May 22, 2026

    @ValerianRey
    Member

    The bug I had with AMP is documented here: #699
    It's very possible that this bug also affects autogram.Engine.compute_gramian, and not just autojac (we use vmap quite similarly internally). So it's possible that the two issues are related. I'll try to fix that soon.

    @ogencoglu do you remember what went wrong when you tried compute_gramian on automatic mixed precision? Could it be the same error?

  4. ogencoglu commented on May 22, 2026

    @ogencoglu
    Author

    Unfortunately, I don't remember.

  5. ValerianRey commented on Jul 30, 2026

    @ValerianRey
    Member

    Closing in favor of #699

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions