Hacktoberfest 2026: le issue che i maintainer hanno segnato per ottobre, aperte e adatte ai principianti. Sfoglia le issue Hacktoberfest

Complete MLX LightX2V H3 integration and Apple Silicon validation

Aperta
  • Qualcosa non va?
#13,682 1 commento 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
5/5
Tempo stimato
Più di una settimana
Idoneità per principianti
20/100
Tipo di issue
Funzionalità
Chiarezza
Specificata chiaramente
Stato di attività
Attiva
Stack tecnologico
javascript
Ambito
ai, machine-learning

Direzione di ricerca

Start with docs/MINIMAX_H3_LIGHTX2V.md and the MLX request path in server/services/videoGen/prepareParams.js, renderArgs.js, and ensureWeights.js; inspect the MLX sampler and quantized adapter helpers in scripts/generate_minimax_h3.py and minimax_h3_lora.py. Implement the separately typed MLX contract while preserving the CUDA contract and stock arguments, then add deterministic request-to-helper coverage. Completion also requires Apple Silicon text and image stock/profile render pairs with runtime fingerprints, measurements, outputs, and documented quality inspection; full-model validation is currently blocked on hardware, runtime, terms acceptance, and pinned weights.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

apple-silicon area:media blocked effort:high enhancement help wanted macos model:heavy plan

Split from #13637 at the user's request. This successor owns MLX integration and its Apple Silicon validation only. Preserve existing CUDA regression coverage, but CUDA full-model evidence belongs to the separate successor; it does not block this issue's completion. Original provenance is retained in #13637 and #13612.

The Windows/CUDA-first implementation of #13612 provides the original LightX2V v1.3 split adapter, the pinned four-update schedule and dual shifts, explicit terms-gated download, cache/digest admission, a quantization-preserving residual applicator, and existing selector/history integration. The owner requested shipping the CUDA part first. The original MLX integration and full-model paired rendering/quality evidence remain undelivered.

Chosen next step: implement a separately typed MLX sibling of the CUDA contract, retaining the existing MLX sampler as default. Use the original author-hosted, revision/SHA-256-pinned adapter, the pinned MLX runtime's actual point/update convention and video/audio shift support, and PortOS's own quantized LoRA adapter. Do not copy the reference implementation, use repacked weights, extrapolate a CUDA result to MLX, or publish a numerical gain without measured PortOS results.

Affected areas: server/lib/minimaxH3Lightx2v.json and videoSpeedProfiles.js (add a distinct MLX contract; do not mutate the shipped CUDA v1.3 contract), server/lib/mediaModels.js, server/services/videoGen/prepareParams.js, renderArgs.js and ensureWeights.js, scripts/generate_minimax_h3.py and minimax_h3_lora.py, model migration and existing selector/history flow. Validation protocol: docs/MINIMAX_H3_LIGHTX2V.md.

Acceptance criteria:

  • MLX text/first-image profile owns its complete adapter/sampler contract, and the stock path retains its prior arguments. Unsupported runtime pins, modes/canvases, conflicting overrides, missing/corrupt weights, and duplicate adapters are refused before queue admission.
  • Preserve user-edited/custom model rows and all existing settings during the additive migration. Keep downloads explicit and model-terms-gated; render execution stays cache-only.
  • Add deterministic request-to-helper tests for MLX sigma semantics, shifts, alpha/rank scaling and speed + character/style stacking exactly once. Preserve CUDA regression coverage.
  • Run same-machine stock/profile pairs for text and image requests with identical prompt, seed, canvas, frame count and user LoRAs on MLX. Record cold load, inference timing, peak memory, exact runtime fingerprint, actual outputs and genuine inspection of motion/detail, prompt adherence, soundtrack and character/style fidelity. Keep performance claims absent until supported by that evidence.

Hardware blocker / help wanted: these paired full-H3 renders need an Apple Silicon machine with enough unified memory for the existing H3 model . The available validation environment cannot load the full H3 checkpoint; no full-model timing or quality comparison has been performed. The model terms must be accepted and pinned weights explicitly provisioned before rendering. Unblock when those hardware/runtime/weight prerequisites are available; this is not an undecided design choice.

Dispatch: model:heavy for MLX timestep/shift semantics and quantized adapter composition; effort:high for MLX boundary coverage and paired fidelity evidence. Hardware validation requires an actual Apple Silicon operator, so help wanted is intentional.

Related to #13612.

Lingua principale
JavaScript
Stelle
38
Fork
32
Merge medio
44m
PR unite (30g)
990

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di atomantic/PortOS

Tutte le issue di atomantic/PortOS

Issue simili

Altre issue su JavaScript

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.