Spaces:

---

osolmaz

/

local-frontier

like 0

Running

Opened on behalf of Onur Solmaz (osolmaz). Work is complete locally and ready for review.

Summary

This finishes the Local Frontier popular-model profile audit. Every over-100k Hugging Face model row now has either an audited profile or an explicit unsupported profile. The compare page now uses profile-backed bounds for model/hardware throughput instead of generated metadata estimates.

What Changed

The main change is replacing rough generated model bounds with audited, self-contained model profiles. The branch also adds schema and engine support needed to evaluate those profiles consistently.

• Added model profile JSON files under profiles/models/.
• Rebuilt assets/local-frontier-profile-data.js.
• Updated assets/local-frontier-model-data.js with audited catalog metadata and corrected weight/KV fields.
• Added and updated Bounds Engine profile handling for dense, MoE, hybrid KV/state, unsupported fail-closed rows, status traces, and comparison metrics.
• Added regression coverage in scripts/test-bounds-engine.mjs.
• Updated docs and progress notes for profile methodology and audit status.

Testing

The local validation gates pass for the committed branch. The SimpleDoc check still reports the pre-existing docs filename/frontmatter migration backlog, which is unrelated to this profile audit.

• node scripts/test-bounds-engine.mjs
• node scripts/validate-model-data.mjs
• node scripts/validate-profiles.mjs
• git diff --check
• npm run slophammer
• npm run check

Known non-blocking check:

• npx -y @simpledoc/simpledoc check fails only on legacy docs migration recommendations.

Risks

The main risk is review size. This branch intentionally touches hundreds of profile files because the requested work was a full manual audit of all popular rows.

• Hugging Face upload may squash the local 503-commit history into one PR commit.
• Some model rows are unsupported by design; they fail closed instead of producing misleading throughput.
• Bounds remain theoretical memory-side decode ceilings, not benchmark results.