Release 4.0.0

[maryamziaa]

Proposed change(s)

Package release 4.0.0 changes (for PackageWorks).

Useful links (Github issues, JIRA tickets, ML-Agents forum threads etc.)

Types of change(s)

• Bug fix
• New feature
• Code refactor
• Breaking change
• Documentation update
• Other (please describe):

Checklist

• Added tests that prove my fix is effective or that my feature works
• Updated the changelog (if applicable)
• Updated the documentation (if applicable)
• Updated the migration guide (if applicable)

Other comments

[havokentity]

Does this release work with cuda 12.8 for GeForce 5090 support? I've tried Release 22 & 23, I tried the dev branch as well and they both don't seem to support training on the 5090 (Blackwell ). Works fine on the 4090 though.

[maryamziaa]

Does this release work with cuda 12.8 for GeForce 5090 support? I've tried Release 22 & 23, I tried the dev branch as well and they both don't seem to support training on the 5090 (Blackwell ). Works fine on the 4090 though.

Hi, ML-Agents (upcoming release or previous ones) can work with CUDA 12.8 and the RTX 5090, but there’s an important detail: it depends on your pytorch version.

ML-Agents uses pytorch under the hood for training, and currently, only pytorch 2.7+ (with CUDA 12.8 support) works properly with the RTX 5090. Earlier versions of pytorch don’t yet support the 5090’s architecture (sm_120), so training would fall back to CPU or fail.

[havokentity]

hi, thank you for the reply. I got it upgraded to pytorch 2.7 day before yest and forgot to update you but I've noticed some issues, maybe they are pytorch related. The training will just silently fail and I can't even ctrl+c to end the process if the screen saver kicks in or if I run multiple training simulataneously. It's like the environment loses track of some of the instances. Otherwise if I disable my screensaver I am successfully able to train PPO or SAC 28 out of 30 times. Sometimes it may randomly fail as well. Just a FYI incase you haven't run into the issue yet.