Skip to content

Conversation

@daniellepintz
Copy link
Contributor

The current example config does not work because it runs on slurm. Fixing to a command that does work and runs locally.

The current example config does not work because it runs on slurm. Fixing to a command that does work and runs locally.
@meta-cla meta-cla bot added the CLA Signed This label is managed by the Meta Open Source bot. label Dec 5, 2025
@felipemello1
Copy link
Contributor

hey, happy to approve, but can you help me understand why 1.7b would run but 30b wouldnt?

@halflearned
Copy link

hey I saw this issue today. imo this isn't the right fix. the problem is that the qwen3_32b.yaml file contains the attributes hosts:1, which assumes a scheduler/remote host. just removing those lines from that file fixed the issue for me when running locally.

Copy link
Contributor

@JenniferWang JenniferWang left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

See the comments above

@daniellepintz
Copy link
Contributor Author

yes indeed the problem is qwen3_32b.yaml is configured to use slurm. do we want to change this?

@JenniferWang
Copy link
Contributor

@felipemello1 @daniellepintz this is, generally speaking a UX problem -- what's our default examples to be based on? Local DevGPU with 8 card?

On the other hand, for now, we can certainly augment the example run command with different mode override.

Local DevGPU: 
python -m tests.sandbox.rl_trainer.main --config apps/grpo/qwen3_1_7b.yaml

Slurm: 
python -m tests.sandbox.rl_trainer.main --config apps/grpo/qwen3_1_7b.yaml provisioner.launcher='slurm' 

MAST:
...

Given that the comment is accurate. I suggest closing the PR and include the improvement in the UX work.

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

Labels

CLA Signed This label is managed by the Meta Open Source bot.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

5 participants