Skip to content

Standardize MatFormer tier metadata in exported checkpoints #5

@plugyawn

Description

@plugyawn

Problem

Tiered checkpoint metadata is inconsistent: exported tiered checkpoints can lack matformer_base_intermediate_size and may set matformer_tier=0, which makes auto-detection unreliable for remote repos.

Refs:

  • scripts/export_matformer_tiers.py (tier export)

Expected

Tiered checkpoints should always include:

  • intermediate_size (sliced)
  • matformer_tier (actual tier)
  • matformer_base_intermediate_size (base width before slicing)

Possible Approach

  • Update export/packaging tooling to write these fields consistently.
  • Add validation or a small test to ensure metadata correctness.

Acceptance Criteria

  • Exported tiered checkpoints always include correct tier metadata.
  • Auto-detection works for both local and remote checkpoints.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions