AnyLanguageModel (Techopolis Fork)

Fork Relationship

Origin: https://github.com/Techopolis/AnyLanguageModel.git
Upstream: https://github.com/mattt/AnyLanguageModel.git
Consumer: Perspective Intelligence app (/Packages/PerspectiveLLM/ references this via local path)

These files exist only in our fork and should never conflict with upstream:

Sources/AnyLanguageModel/DownloadableLanguageModel.swift — DownloadableLanguageModel protocol + MLXModelDownloadManager
Sources/AnyLanguageModel/Models/MLXLanguageModel.swift — contains the download manager extension at the end of file (after upstream's JSONValue.toSendable())
scripts/validate-merge.sh — post-merge validation
scripts/check-upstream.sh — upstream divergence tracking

Check divergence: ./scripts/check-upstream.sh
Prefer rebase over merge when possible: git rebase upstream/main
- Rebase resolves conflicts per-commit (smaller, easier to reason about)
- Use merge only when rebase would rewrite published/shared history
Validate after sync: ./scripts/validate-merge.sh
Rebuild Perspective Intelligence to verify consumer compatibility

After merging upstream, watch for these residual fork artifacts:

Duplicate type declarations — Our fork previously defined GPUMemoryConfiguration, GPUMemoryManager, and SessionCacheEntry at the module level. Upstream defines them inside MLXLanguageModel. After merging, only the nested (upstream) versions should exist.
Old cache API — Our fork used hashTokenPrefix, prefillTokenHash, resolveCache(for:...). Upstream uses prefixTokens, cacheConfigSignature, resolveCache(session:...). The fork versions are dead code after merge.
Scope-less GPU calls — Our fork's markActive() / markIdle() take no args. Upstream's take GPUMemoryConfiguration / scope: UUID. If you see parameterless calls, they're from the old fork.

The validate-merge.sh script checks for all of these automatically.

swift build                              # SPM build (macOS)
swift build --traits MLX                 # with MLX trait