Skip to content

registry: ornith-1.5:fast aliases the MQ4R speed SKU - #680

Open
Kaden-Schutt wants to merge 2 commits into
masterfrom
registry/ornith-fast-alias
Open

registry: ornith-1.5:fast aliases the MQ4R speed SKU#680
Kaden-Schutt wants to merge 2 commits into
masterfrom
registry/ornith-fast-alias

Conversation

@Kaden-Schutt

Copy link
Copy Markdown
Collaborator

Summary

ornith-1.5:fast and ornith-1.5:35b-a3b-fastornith-1.5:35b-a3b-mq4r, matching how qwen3.8:fast and muse-glimmer:fast name their speed SKUs. The bare ornith-1.5:35b-a3b tag is unchanged.

Why

hipfire bench --spec off --backend noslots --workload stateless, gfx1201, 3 runs, stddev < 1.4:

artifact decode tok/s
ornith-1.5-35b-a3b.mq4 (bare tag) 139.7
ornith-1.5-35b-a3b.mq4r (#664 router fix) 205.7

The 206 matches docs/perf-checkpoints/2026-07-13-redline-mq4r-110-to-204.md.

Which surface(s) does this touch?

  • policy filesregistry/ (curated models.json + generated v1.json, kept in sync so the scheduled regeneration is a no-op)

Test plan

  • cargo test -p hipfire-registry --lib — 18 passed (new alias assertions included)
  • python3 -m pytest scripts/test_registry_gen_ornith.py — new test_fast_aliases_point_at_the_mq4r_speed_sku passes; test_curated_entry_uses_the_canonical_tag_and_hyphenated_files fails on master already (expects recommended_settings without reasoning_effort: xhigh) and is untouched here.

No hardware route: registry aliases only.

The bare ornith-1.5:35b-a3b tag stays on the plain MQ4 trunk. The MQ4V2
router-fix artifact from PR #664 (ornith-1.5:35b-a3b-mq4r) benches 205.7
tok/s AR on gfx1201 vs 139.7 for the trunk, so it gets the :fast alias the
same way qwen3.8:fast and muse-glimmer:fast name their speed SKUs. Both the
curated models.json and the generated v1.json carry the aliases so the
scheduled regeneration is a no-op.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant