Merged
Conversation
…code project - Regenerate project.pbxproj with: - MLXInferenceCore .swift files as direct compile sources - XCLocalSwiftPackageReference for mlx-swift and mlx-swift-lm - XCSwiftPackageProductDependency for MLX, MLXLLM, MLXLMCommon - Add ModelManagementView (download manager, disk usage, delete) - Add ModelDownloadManager to MLXInferenceCore - Wire progress callbacks in InferenceEngine.load()
ModelStorage (new):
- macOS: Library/Caches/huggingface/hub/ (matches defaultHubApi)
- iOS: Library/Application Support/SwiftLMChat/Models/ + isExcludedFromBackup
- Platform-agnostic scan, sizeOnDisk, delete primitives
ModelDownloader (new, iOS only):
- URLSession background session (survives app suspension)
- HuggingFace API file enumeration (GET /api/models/{id})
- Per-file download with progress streaming
- macOS: LLMModelFactory handles download directly (no change)
ModelDownloadManager refactor:
- Built on ModelStorage + ModelDownloader layers
- NWPathMonitor for WiFi/cellular/offline detection
- iOS RAM budget: 40% (vs 75% macOS) via modelsForDevice()
- Cellular threshold: warn before >200MB downloads on cellular
- updateProgress() / clearProgress() for InferenceEngine bridge
InferenceEngine:
- UIApplication.didReceiveMemoryWarningNotification → auto-unload (iOS)
- ProcessInfo.thermalStateDidChangeNotification → ThermalLevel @published
- Critical thermal → stop generation immediately
- HubApi.downloadBase redirected to ModelStorage.cacheRoot
ModelPickerView:
- Network status banner (offline / cellular warning)
- Thermal warning banner
- Cellular confirmation dialog before large downloads
- handleModelTap() blocks download when offline
SwiftLMChat.entitlements (new):
- com.apple.developer.kernel.increased-memory-limit
- UIBackgroundModes: fetch, processing
Package.swift: add Hub product to MLXInferenceCore dependencies
…load on foreground InferenceEngine: - willResignActiveNotification → stopGeneration() + unload() + save backgroundedModelId - didBecomeActiveNotification → reload backgroundedModelId (or lastLoadedModelId) - autoOffloadOnBackground: Bool (default true on iOS, false on macOS) - Observers consolidated into [NSObjectProtocol] for clean deinit - Reactive memory warning still kept as safety fallback - Thermal observer migrated to same consolidated array - Background unload sets .idle (not .error) — clean UX on return
ExpertStreamingConfig (new, MLXLMCommon): - Replaces EXPERIMENTAL_SSD_STREAM env var with a proper Swift API - .mmapPageCache mode: APFS page-cache (iOS + macOS without directIO) - .directNVMe mode: pread() at 5GB/s NVMe (macOS default for MoE) - activate(modelDirectory:useDirectIO:) + deactivate() - legacyEnvPath shim for any remaining C-level consumers SwitchLayers.swift: - ExpertStreamingConfig.shared.isEnabled replaces env var gate - #if os(macOS) / #else: directNVMe path locked to macOS only - iOS always routes to mmap prefault fallback (was dead code before) Load.swift / LayerPartitioning.swift: - Both env var gates replaced with ExpertStreamingConfig.shared.isEnabled InferenceEngine.load(): - MoE models get config.lazyLoad = true + ExpertStreamingConfig.activate() - macOS: useDirectIO=true (5GB/s NVMe pread) - iOS: useDirectIO=false (APFS mmap, ~2-3GB/s, fits in sandbox) - Deactivated on error or unload() ModelCatalog: - ramRequiredGB for MoE = peak-resident (active experts only) - Qwen3 30B MoE: ramRequired=4.5GB (targets iPad Pro M4 8GB+) - DeepSeek R1 0528: ramRequired=8GB (targets iPad Pro M4 16GB+) - Qwen3.5 122B: ramRequired=12GB (macOS / iPad Pro M4 Max 32GB) This enables 30B-class MoE reasoning models on iPad Pro M4 without any system swap — purely via OS page-cache eviction.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.