drumih/turbo-fieldfare

TurboFieldfare is a local, GPU-accelerated inference engine for running large language models, specifically optimized for Apple Silicon via Metal kernels. It provides a Mac application and a local HTTP server to handle multimodal interactions, including vision processing and LaTeX rendering, while managing model loading, tokenization, and conversation state. The system also includes CLI tools and a format library for packaging, validating, and installing model weights securely.

The app now supports optional Gemma 4 image processing. This change introduces a new vision inference pipeline built on Metal, including GPU-based bicubic image resizing, attention mechanisms, linear layers, and shared-expert MoE kernels, alongside a new secure model directory reader for loading vision assets.

Sources · high confidence

Added three new Ruby scripts to enforce build integrity and repository hygiene: \check\_app\_version.rb\ verifies that the app's internal version string is not more than one release behind the latest GitHub tag, allowing for the standard tag-to-bump window; \check\_swiftui\_macros.rb\ scans source files to ensure no production code uses \@State\ or \\#Preview\ macros, which require Xcode platform plugins and are incompatible with Command Line Tools builds; and \check\_tracked\_symlinks.rb\ fails the build if any symlinks are tracked in Git, preventing data-loss issues where symlinks might overwrite local directories. A corresponding test script, \swift\_macro\_probe\_test.rb\, validates the macro detection logic and ensures the build system correctly handles platform plugin removal.

Scripts · high confidence

Add server test fixtures and HTTP server tests; Added comprehensive test coverage for Mac presentation layer components; Added test coverage for vision image processing and multimodal prefill logic; Added test infrastructure for transcript rendering validation; Added tests for CLI argument parsing and image handling logic; Added tests for Gemma 4 tokenization, chat templates, and decoder configuration; Added tests for Mac app configuration, history, and image handling; Added tests for Metal command buffer diagnostics, conversation lineage validation, and generation loop controls; Added tests for TurboFieldfareDecodeService conversation and load safety; Added tests for TurboFieldfareRepack CLI, format parsing, and remote installation; Added tests for context admission and updated runtime configuration assertions; Added tests for prefill chunking, replay coercion, and cancellation behavior; Added tests for the .gturbo v1 format contract and vision companion support; Expanded test coverage for TurboFieldfare core kernels; Expanded test coverage for model loading, Metal context, and IO security; Updated KVCacheManager tests for rewind semantics and context limits.

The project now includes a new local server executable (TurboFieldfareServer) built on SwiftNIO, allowing users to run the application as a network service. This change updates the SwiftNIO dependency to version 2.101.3 and introduces the SwiftMath library to support mathematical rendering in the Mac presentation layer. Additionally, a new TurboFieldfareFormat target and associated test suites have been added to handle format validation and compatibility.

(dependencies) · high confidence

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

• CAI 70 → 66 (-3.7)
• Rubric changed (rubric-2026.09.11 → rubric-2026.09.18) — scores are not directly comparable.

• Code Health 78 → 79 (+1.6)
• Architecture 98 → 90 (-8.2)
• Maturity 76 → 76 (-0.5)
• Readiness 54 → 54 (-0.4)
• Security 100 → 100 (+0.0)
• Performance 71 (new)

• Documentation: no installation or build instructions
• Documentation: no installation or build instructions (docs/experiments/summaries/01-model-install-and-expert-io.md)
• Documentation: no installation or build instructions (docs/experiments/summaries/02-decode-moe-int4-and-router.md)
• Documentation: no usage examples (README.md)
• Hotspot: Sources/TurboFieldfare/Kernels/Attention/Attention.swift (Sources/TurboFieldfare/Kernels/Attention/Attention.swift)
• Hotspot: Sources/TurboFieldfare/Kernels/TensorCore/MPPPrefillInt4QMM.swift (Sources/TurboFieldfare/Kernels/TensorCore/MPPPrefillInt4QMM.swift)
• Hotspot: Sources/TurboFieldfare/Runtime/Generation/MultimodalConversation.swift (Sources/TurboFieldfare/Runtime/Generation/MultimodalConversation.swift)
• Hotspot: Sources/TurboFieldfare/Runtime/Generation/Sampler.swift (Sources/TurboFieldfare/Runtime/Generation/Sampler.swift)
• Hotspot: Sources/TurboFieldfare/Runtime/Vision/Preprocessing/Gemma4ImageGeometry.swift (Sources/TurboFieldfare/Runtime/Vision/Preprocessing/Gemma4ImageGeometry.swift)
• Hotspot: Sources/TurboFieldfare/Runtime/Vision/VisionRuntime.swift (Sources/TurboFieldfare/Runtime/Vision/VisionRuntime.swift)
• Hotspot: Sources/TurboFieldfare/Tokenization/StructuredAssistantDecoder.swift (Sources/TurboFieldfare/Tokenization/StructuredAssistantDecoder.swift)
• Hotspot: Sources/TurboFieldfareApp/Core/Installation/DownloadETAEstimator.swift (Sources/TurboFieldfareApp/Core/Installation/DownloadETAEstimator.swift)
• Hotspot: Sources/TurboFieldfareApp/MacPresentation/ResponseMarkdownRenderer.swift (Sources/TurboFieldfareApp/MacPresentation/ResponseMarkdownRenderer.swift)
• Hotspot: Sources/TurboFieldfareDecodeService/DecodeServiceOutbox.swift (Sources/TurboFieldfareDecodeService/DecodeServiceOutbox.swift)
• Hotspot: Sources/TurboFieldfareRepack/Core/Format/Safetensors.swift (Sources/TurboFieldfareRepack/Core/Format/Safetensors.swift)
• Hotspot: Sources/TurboFieldfareRepack/Core/Planning/RepackPlanner.swift (Sources/TurboFieldfareRepack/Core/Planning/RepackPlanner.swift)
• Hotspot: Sources/TurboFieldfareRepack/Core/Remote/RemoteStreamingRepacker.swift (Sources/TurboFieldfareRepack/Core/Remote/RemoteStreamingRepacker.swift)
• Hotspot: Sources/TurboFieldfareServer/Core/GemmaToolSchema.swift (Sources/TurboFieldfareServer/Core/GemmaToolSchema.swift)
• Hotspot: Sources/TurboFieldfareServer/Core/HTTPServer.swift (Sources/TurboFieldfareServer/Core/HTTPServer.swift)
• Hotspot: Sources/TurboFieldfareServer/Core/ServerPromptCache.swift (Sources/TurboFieldfareServer/Core/ServerPromptCache.swift)
• …and 1 more

• Ambiguous overloading: Two methods named loadExpert differ only by the presence of a slot parameter. While technically distinct signatures, this can be confusing for API consumers who might expect the slot to be an optional parameter or a separate caching step. It implies two different loading strategies (direct vs. cached) under the same name.
• ClassTooLong: IncrementalTranscriptView (Sources/TurboFieldfareApp/Mac/Generation/OutputPaneView.swift)
• Coverage not measured — Swift suite
• Duplicate intent: ManifestArch and ArchConfig contain identical property names and types for model architecture details (e.g., slidingWindow, hiddenSize, numHeads). ManifestArch is a parsed data structure from a file, while ArchConfig is a runtime configuration object. While they serve different lifecycle stages, having two nearly identical types with significant property overlap increases maintenance burden and risk of drift.
• Duplicated block (10–18 lines × 2) (Sources/TurboFieldfare/Runtime/Generation/RawCompletion.swift)
• Duplicated block (11 lines × 2) (Sources/TurboFieldfare/Infrastructure/ModelIO/ModelTypes.swift)
• Duplicated block (12 lines × 2) (Sources/TurboFieldfare/Infrastructure/ModelIO/ModelTypes.swift)
• Duplicated block (21 lines × 2) (Sources/TurboFieldfareApp/Core/State/AppModel.swift)
• Duplicated block (5 lines × 2) (Sources/TurboFieldfareRepack/Core/Planning/RangeCopyPlanner.swift)
• Duplicated block (7–8 lines × 3) (Sources/TurboFieldfareApp/Core/State/AppModel.swift)
• Duplicated block (9 lines × 2) (Sources/TurboFieldfareApp/Mac/Components/GenerateControl.swift)
• FunctionTooLong: Run.swift.run (Sources/TurboFieldfareCLI/Run.swift)
• Hotspot: Sources/TurboFieldfare/Runtime/Generation/RawCompletion.swift (Sources/TurboFieldfare/Runtime/Generation/RawCompletion.swift)
• Hotspot: Sources/TurboFieldfareApp/Core/Inference/AppInferenceError.swift (Sources/TurboFieldfareApp/Core/Inference/AppInferenceError.swift)
• Hotspot: Sources/TurboFieldfareApp/Mac/Diagnostics/InspectorView.swift (Sources/TurboFieldfareApp/Mac/Diagnostics/InspectorView.swift)
• Hotspot: Sources/TurboFieldfareCLI/Run.swift (Sources/TurboFieldfareCLI/Run.swift)
• Hotspot: Sources/TurboFieldfareRepack/Core/Workflow/Errors.swift (Sources/TurboFieldfareRepack/Core/Workflow/Errors.swift)
• InspectorView.visionSection (cognitive 35) (Sources/TurboFieldfareApp/Mac/Diagnostics/InspectorView.swift)
• InspectorView.visionSection (cyclomatic 29) (Sources/TurboFieldfareApp/Mac/Diagnostics/InspectorView.swift)
• Members sharing a duplicated core (18 members, 50+ identical tokens) (Sources/TurboFieldfare/Infrastructure/ModelIO/ModelTypes.swift)
• …and 26 more

• 3 commits — 2 feature/other, 1 fixes

• Sources/TurboFieldfareApp — 2 commits
• Sources/TurboFieldfare — 1 commit

• fix: Fix SwiftUI macro compatibility for Command Line Tools builds (#186)
• change: Add Mac text-size settings and open an empty chat on launch (#179)
• change: Support longer contexts and speed up full attention (#182)

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

drumih/turbo-fieldfare was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.