Files
club-3090/scripts/tests
noonghunna 458c473908 feat(ik-llama): PRISM-PRO-DQ + APEX-MTP presets, conformed to the <quant>/ layout
Ports VykosX's #223 preset families onto the post-refactor convention
(PR-A #231 layout + PR-B #234 registry-derived launchers).

- 8 ik-llama presets under models/<model>/ik-llama/compose/<topology>/<quant>/:
  - qwen3.6-27b: ex0bit-prism-pro-dq/{mtp,long,two-stage} (single) + {mtp,mtp-vision} (dual)
  - qwen3.6-35b-a3b: mudler-apex-compact/{mtp,long} (single) + mudler-apex-quality/mtp (dual)
    [one repo, two quant files — the canonical two-slug case]
- +1 ../ depth on all relocated mounts (models-cache + APEX chat-template).
- weights variants added to models/*.yml: ex0bit-prism-pro-dq, mudler-apex-compact,
  mudler-apex-quality.
- 8 compose_registry.py entries (weights_variant=slug, kvcalc_key=SKIP); launchable
  via the registry — no launch.sh edits needed (PR-B derives them).
- catalog-size guard 47 -> 55.

Dropped from #223 as obsolete/out-of-convention: the launch.sh hardcoded-array edits
(superseded by PR-B) and the root UPSTREAM_CHANGES.md.

Validation: 7/7 gates PASS (registry-disk incl. quant-slug<->weights, mounts-resolve,
model-weights-registry, switch+launch parity, default-resolver, launch-compat) +
kv-calc 22/22. Structurally validated; live boot is weights-gated (community GGUFs not
on our rig) — ships community-experimental, crediting VykosX's reported numbers.

Co-Authored-By: VykosX <VykosX@users.noreply.github.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-26 19:51:00 +00:00
..