Grad norm issue

[LukeBailey181]

Hi! Im training gpt-oss-120b with CISPO loss. I get the following grad norm returned on a regular batch of on policy data:

OptimStepResponse(metrics={'unclipped_grad_l2:mean': 13468.94921875})

How exactly is unclipped_grad_l2:mean computed, as the reported grad norm is dramatically too large.

Thanks in advance for any help with this!

[GnarlyMshtep]

Are you using the tinker API? I am under the impression that OptimStepResponse returns empty metrics. (I know the SkyRL backend returns grad norm and a few other metrics)