Skip to content

student pre-training diverge #11

Description

@DennisMenn

Thanks for sharing this great work! While reproducing the code, I ran into two issues:

1. Missing .prev_sample
In trainer_far_wan_anyflow_onpolicy.py:120, .prev_sample appears to be missing from the scheduler step output.

2. Training instability
Training the student with train_farwan1b_student_shift5_81f_480p_lr5e-5_6k_b32.yml, grad_norm is large (~150) on the first iteration and losses diverge after ~100 steps.

Could you take a look when you have a chance? Happy to share more details from my runs. Thanks!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions