Skip to content

GRPO training: policy_loss / policy_entropy / raw_grad_norm are NaN on every batch (rollout & reward are fine) #116

Description

@lzt013

Rollout completes normally and both reward and advantage look correct, but as soon as the policy update step begins, policy_loss, policy_entropy, and raw_grad_norm are all NaN on every single batch — multiple consecutive steps, without exception.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions