Skip to content

perf(runtime): llama prewarm-on-focus, mid-prefill abort, and KV-reuse visibility #681

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
FuJacob merged 2 commits into from
Jun 12, 2026
Merged

perf(runtime): llama prewarm-on-focus, mid-prefill abort, and KV-reuse visibility #681

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
711bca6
perf(runtime): llama prewarm-on-focus, mid-prefill abort, and KV-reus…
FuJacob Jun 12, 2026
00c4968
review: stop prewarm from double-decoding on models that reject parti…
FuJacob Jun 12, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions Cotabby.xcodeproj/project.pbxproj
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -568,6 +568,7 @@
E51FA12B690428CA431328FC /* WritingPaneView.swift in Sources */ = {isa = PBXBuildFile; fileRef = D48B95B6665109B6C6A63B42 /* WritingPaneView.swift */; };
E54F5F03E16859D5A1E3437A /* MacroController.swift in Sources */ = {isa = PBXBuildFile; fileRef = 4638C74239D1DE2DC4D87975 /* MacroController.swift */; };
E5CB34ED76BAE87E8A858112 /* WebContentFieldDetectorTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = 210F9AD332273FE2EB3A9A01 /* WebContentFieldDetectorTests.swift */; };
E64AE96DF2A80A368FDE522D /* LlamaSuggestionEnginePrewarmTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = 26EF16C7439BEB156BD9FB03 /* LlamaSuggestionEnginePrewarmTests.swift */; };
E6EE3C13FA31F261CD734C69 /* DownloadOutcomeClassifier.swift in Sources */ = {isa = PBXBuildFile; fileRef = 3DE1975F3B5F4A70478DBF41 /* DownloadOutcomeClassifier.swift */; };
E853B9C7AF93FA595DC417B2 /* EmojiVariantResolver.swift in Sources */ = {isa = PBXBuildFile; fileRef = 1A8414BEB7E34F57607E37FE /* EmojiVariantResolver.swift */; };
E912D4617AE1376061DF1F00 /* LanguageSupportTests.swift in Sources */ = {isa = PBXBuildFile; fileRef = 4793D4EA5D36D7E5CC216C27 /* LanguageSupportTests.swift */; };
Expand Down Expand Up @@ -696,6 +697,7 @@
24F613F0E2F7046E6532A09C /* OnboardingTemplateFeatureList.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = OnboardingTemplateFeatureList.swift; sourceTree = "<group>"; };
262BE2F1E97389FE8D7A5FB9 /* Cotabby.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = Cotabby.app; sourceTree = BUILT_PRODUCTS_DIR; };
264CA64B2AB1611F82E5B760 /* WelcomeView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = WelcomeView.swift; sourceTree = "<group>"; };
26EF16C7439BEB156BD9FB03 /* LlamaSuggestionEnginePrewarmTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LlamaSuggestionEnginePrewarmTests.swift; sourceTree = "<group>"; };
273B4DC844F79B4BE2C8910F /* FocusPollBackoffTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = FocusPollBackoffTests.swift; sourceTree = "<group>"; };
27A5D63F390E9B7A7FE343FE /* SystemResourceSampler.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = SystemResourceSampler.swift; sourceTree = "<group>"; };
28B7EB84781C0ED57844585E /* OnboardingTemplateTests.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = OnboardingTemplateTests.swift; sourceTree = "<group>"; };
Expand Down Expand Up @@ -1363,6 +1365,7 @@
4793D4EA5D36D7E5CC216C27 /* LanguageSupportTests.swift */,
0CA88BB29BC8727878C99E95 /* LlamaPromptCacheHintTrackerTests.swift */,
AABCC3FD99B1824A81E665F3 /* LlamaSuggestionEngineCancellationTests.swift */,
26EF16C7439BEB156BD9FB03 /* LlamaSuggestionEnginePrewarmTests.swift */,
9030FAAB468119A0236284A6 /* LLMIOFileHandlerTests.swift */,
D8083D44ABCDCFA68A4CD497 /* MacroEngineTests.swift */,
22BE47D1DBF6C23151458836 /* MacroTriggerStateMachineTests.swift */,
Expand Down Expand Up @@ -2334,6 +2337,7 @@
E912D4617AE1376061DF1F00 /* LanguageSupportTests.swift in Sources */,
E38801433B99E65BD7E45A0E /* LlamaPromptCacheHintTrackerTests.swift in Sources */,
BE3CB85508055D159C35020A /* LlamaSuggestionEngineCancellationTests.swift in Sources */,
E64AE96DF2A80A368FDE522D /* LlamaSuggestionEnginePrewarmTests.swift in Sources */,
8429B116328C392DCA018D95 /* MacroEngineTests.swift in Sources */,
3F8CBCBCC45E377DF9ADB216 /* MacroTriggerStateMachineTests.swift in Sources */,
87806DE08881D11F2608A13D /* MarkerSelectionSynthesizerTests.swift in Sources */,
Expand Down
18 changes: 14 additions & 4 deletions Cotabby/Models/SuggestionSubsystemContracts.swift
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -93,10 +93,10 @@ protocol SuggestionGenerating: AnyObject {
/// continuous. Stateless engines may implement this as a no-op.
func resetCachedGenerationContext() async
/// Best-effort warmup hook the coordinator calls after focus arrives on an editable surface.
/// Engines that benefit from prefix caching or weight loading (Apple Foundation Models) use it
/// to prime the next request; engines that do not (llama already keeps its KV cache hot) can
/// rely on the default no-op extension. Failures are intentionally swallowed by implementations
/// because prewarming is opportunistic.
/// Apple Foundation Models primes its session here, and the llama engine prefills the new
/// field's prompt KV (a focus change destroys the previous field's native sequence, so without
/// this the first suggestion in every field pays the full cold prompt decode). Failures are
/// intentionally swallowed by implementations because prewarming is opportunistic.
func prewarm(for request: SuggestionRequest) async
}

Expand All @@ -113,6 +113,16 @@ extension SuggestionGenerating {
protocol LlamaRuntimeGenerating: AnyObject {
func generate(prompt: String, cachedPrefixBytes: Int?, options: LlamaGenerationOptions) async throws -> String
func resetPromptCache()
/// Decodes `prompt` into the native prompt cache without sampling any tokens, so the next
/// `generate` whose prompt extends this one only decodes the typed delta. Best-effort warmup:
/// callers treat failures as "no cache primed", never as a user-facing error.
func prefill(prompt: String, cachedPrefixBytes: Int?, options: LlamaGenerationOptions) async throws
}

extension LlamaRuntimeGenerating {
/// Default no-op so test fakes that only exercise the generate/cancel contract keep compiling;
/// the production manager overrides this with a real prompt prefill.
func prefill(prompt: String, cachedPrefixBytes: Int?, options: LlamaGenerationOptions) async throws {}
}

@MainActor
Expand Down
Loading