SceneWorks Inference

Unified source repository for SceneWorks' backend-neutral inference contracts, MLX and Candle engines, model providers, conformance suites, and platform runtime bundles.

Migration status

This repository was assembled from the existing core-llm, mlx-llm, candle-llm, mlx-gen, and candle-gen histories. Phase 1 preserved and verified those histories; Phase 2 moved the imported packages into their ownership paths and normalized them under one Cargo workspace and lockfile. Phase 3 provides dependency-aware CI selection, supply-chain policy, immutable real-weight fixture pins, and deterministic source/SBOM release tooling. Phase 4 adds validated runtime-macos, runtime-cuda, and runtime-cpu composition boundaries for explicit media, LLM, and snapshot-preparation catalogs; the former link-time provider registries have been removed. Phase 5 published immutable release runtime-2026.07.0 and cut the SceneWorks and ChatWorks lockfiles over to its exact commit. Local product validation is complete; hosted SceneWorks validation requires its scoped credential for this private repository.

After the cutover, the legacy MLX generation repository received one further epic (10834, sequential component residency). That post-cutover drift was reconciled into the monorepo and published as runtime-2026.07.2; see docs/migration/RUNTIME_2026_07_2_CHECKPOINT.md. Legacy development is halted pending the durable decision on whether the monorepo is the sole live development line.

The current authoritative migration plan and release-set baseline live in the SceneWorks repository under documents/rearchitecture/.

Layout

crates/contracts/  Backend-neutral contracts and conformance suites
crates/bundles/    Named supported platform compositions and catalog validation
crates/llm/        MLX and Candle LLM engines
crates/media/      MLX and Candle media engines and provider families
docs/              Migration maps, architecture, compatibility, and release records

Documentation

Start in docs/. The essentials:

• Getting Started — how a product depends on a runtime bundle, builds its validated registries, and loads/runs media generators, LLMs, and snapshot preparers.
• Model Catalog Reference — every shipped provider id per platform, with the MLX/Candle deltas.
• Architecture Rationale — why the repositories were consolidated, why provider discovery is explicit, the alternatives considered, and the tradeoffs accepted.

See docs/migration/PHASE_2_CHECKPOINT.md for normalization invariants and docs/migration/PHASE_3_CHECKPOINT.md for the published release-train evidence, and docs/migration/PHASE_4_CHECKPOINT.md for the explicit bundle composition and platform results.

Contribution boundaries and security reporting are documented in CONTRIBUTING.md and SECURITY.md.

Validate the normalized graph with:

./scripts/check-workspace.py