Skip to content

Native 5c self-play saves random weights as the initial checkpoint #618

Description

@MischaU8

In 5c, run_train() creates a randomly initialized model and immediately saves it as the initial self-play checkpoint. base.load_model_path is never applied.

Because of this, the self-play pool is bootstrapped with random weights rather than the requested checkpoint and training also starts from those random weights without warning.

Problematic code: src/pufferl.cu:3240

Expected: load base.load_model_path and synchronize async actor weights before saving or registering the initial checkpoint.

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