Skip to content

Navigation Menu

Appearance settings

View all features
- BY COMPANY SIZE
  Enterprises
  Small and medium teams
  Startups
  Nonprofits
- BY USE CASE
  App Modernization
  DevSecOps
  DevOps
  CI/CD
  View all use cases
- BY INDUSTRY
  Healthcare
  Financial services
  Manufacturing
  Government
  View all industries
View all solutions
- EXPLORE BY TOPIC
  AI
  Software Development
  DevOps
  Security
  View all topics
- EXPLORE BY TYPE
  Customer stories
  Events & webinars
  Ebooks & reports
  Business insights
  GitHub Skills
- SUPPORT & SERVICES
  Documentation
  Customer support
  Community forum
  Trust center
  Partners
View all resources
- COMMUNITY
  GitHub SponsorsFund open source developers
- PROGRAMS
  Security Lab
  Maintainer Community
  Accelerator
  GitHub Stars
  Archive Program
- REPOSITORIES
  Topics
  Trending
  Collections
- ENTERPRISE SOLUTIONS
  Enterprise platformAI-powered developer platform
- AVAILABLE ADD-ONS
  GitHub Advanced SecurityEnterprise-grade security features
  Copilot for BusinessEnterprise-grade AI features
  Premium SupportEnterprise-grade 24/7 support
Pricing

Search code, repositories, users, issues, pull requests...

Search

Clear

Search syntax tips

Provide feedback

We read every piece of feedback, and take your input very seriously.

Include my email address so I can be contacted

Saved searches

Use saved searches to filter your results more quickly

Name

Query

To see all available qualifiers, see our documentation.

Appearance settings

You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session. You switched accounts on another tab or window. Reload to refresh your session.

Dismiss alert

dAppCore / go-mlx Public

Notifications You must be signed in to change notification settings
Fork 0
Star 1

Code
Issues
Pull requests 1
Actions
Projects
Security and quality
Insights

Additional navigation options

Code
Issues
Pull requests
Actions
Projects
Security and quality
Insights

Virgil Lemma foundations#8

Open

Snider wants to merge 1297 commits into

maindAppCore/go-mlx:mainfrom

devdAppCore/go-mlx:devCopy head branch name to clipboard

Conversation Commits1297 (1297)Checks Files changed

Open

Virgil Lemma foundations#8
Snider wants to merge 1297 commits into
mainfrom
dev

Commits

This pull request is big! We're only showing the most recent 250 commits

Commits on May 22, 2026

perf(kv): remove dead normalizeKVSnapshotNativeTensor helper

Snider
and
Virgil
committed
perf(metal): gemma4 cache PerLayerProjectionScale + lift 1/sqrt(2) constant

Snider
and
Virgil
committed
perf(kv): remove dead analysis helpers — HeadVectors / MeanVector / PositionVector

Snider
and
Virgil
committed
perf(metal): gemma4 hoist Squeeze axes out of splitPerLayerInputTensor loop

Snider
and
Virgil
committed
test(metal): document inline-slice-literal cost at attention forward sites

Snider
and
Virgil
committed
perf(safetensors): F32 decode via unsafe-slice memcpy (-34% on 2048-elem path)

Snider
and
Virgil
committed
perf(kv): stream writer direct unsafe-slice — drop scratch memcpy

Snider
and
Virgil
committed
perf(safetensors): F64 decode via unsafe-slice + FCVT (-25% on 2048-elem path)

Snider
and
Virgil
committed
perf(metal): cache AppendState — prefill 26-cache fan-out 27→1 alloc, 3× faster

Snider
and
Virgil
committed
perf(safetensors): hand-rolled JSON encoder for subset-write headers (-38% allocs)

Snider
and
Virgil
committed
merge: cladius-lane-Wave10-W10L (kv residual — Snapshot writer 1217→705 ns -42%, stream-writer direct-unsafe-slice eliminates staging memcpy)
Snider
committed
chore(external): bump go-inference → 22e9e0d (W10-M + W10-N — 4-46× state/scheduler/openai/ollama/parser + 3.5-4.6× jang dequant model-LOAD)
Snider
committed
merge: cladius-lane-Wave10-W10K (gemma4 PerLayerProjectionScale cache + Squeeze axes hoist — -96.2% allocs/forward on 26-layer Gemma 4)
Snider
committed
perf(safetensors): lift sentinel errors in safetensors.go + write.go

Snider
and
Virgil
committed
perf(metal): Slice4 / SliceUpdateInplace4 scalar-pass form — KVCache 14→6 allocs

Snider
and
Virgil
committed
perf(safetensors): F16 + BF16 decode via unsafe-uint16-slice (-13%/-33%)

Snider
and
Virgil
committed
perf(gguf): fold trailer read into shared scratch arena — drop per-tensor [12]byte heap escape
Snider
committed
perf(gguf): unify header into shared scratch buffer — one less per-call alloc
Snider
committed
perf(metal): Slice4 sweep cache.go RotatingKV/FixedKV/PagedKV — broad alloc reduction

Snider
and
Virgil
committed
merge: cladius-lane-Wave10-W10R (safetensors decode unsafe-cast + hand-rolled JSON write — F32 -50% / F64 -44% / BF16 -31% / WriteSubset -38% allocs)
Snider
committed
perf(metal): Transpose4 scalar-pass — SDPAPaged 16-page 17→1 alloc, 1.0× ns

Snider
and
Virgil
committed
perf(gguf): lift sentinel errors in info.go to package vars
Snider
committed
perf(gguf): tensor-name arena — bump-allocate name bytes, zero-copy view per tensor
Snider
committed
perf(eval): pool per-batch int32/float32 scratch slices for FromValues
Snider
committed
perf(hf): folded compare in normalizeKnownArchitecture (-100% allocs / -33% time)

Snider
and
Virgil
committed
perf(hf): pre-count notes slice in estimateTrainingFit (-16% time)

Snider
and
Virgil
committed
perf(metal): pinned_array sync.Pool — KVShape 6→3 allocs across L1-L16384

Snider
and
Virgil
committed
perf(gguf): metadata key arena + unified bulk-string helper
Snider
committed
perf(eval): pool per-batch attention-mask scratch slice
Snider
committed
perf(gguf): metadata value-string arena — bump-allocate string-typed values
Snider
committed
perf(fast_eval): collapse toModelGenerateOptions slice header
Snider
committed
perf(hf): one ReadDir beats five PathGlob in localModelFiles (-29% allocs / -36% time)

Snider
and
Virgil
committed
perf(grpo): asciiHasPrefixFold in cleanGRPOAnswerLine (-1 alloc/call, 3-17× faster)

Snider
and
Virgil
committed
perf(hf): ReadDir + Stat instead of Glob in resolveLocalMetadataRoot (-13% allocs / -36% time)

Snider
and
Virgil
committed
perf(metal): Array.ShapeInto + KVCache.Update — 6→4 allocs / -60% on long runs

Snider
and
Virgil
committed
perf(hf): drop per-candidate stat in resolveLocalMetadataRoot (-4% allocs)

Snider
and
Virgil
committed
merge: cladius-lane-Wave10-W10O (internal/metal substrate primitives — Slice4/SliceUpdateInplace4/Transpose4/ShapeInto/AppendState — KVCache -85% allocs, SDPAPaged -94%, PromptCache -96%)
Snider
committed
perf(grpo): containsFoldASCII + fragment-split in GRPORewardContainsAnswer (14× faster, 2 → 0 allocs)

Snider
and
Virgil
committed
merge: cladius-lane-Wave10-W10V (eval/fast_eval sync.Pool of *[]T scratch — -56%/-83%/-100% allocs across token/loss-mask/attn-mask paths, 64 MiB per ragged batch saved)
Snider
committed
perf(hf): drop redundant slices.SortFunc in localModelFiles

Snider
and
Virgil
committed
perf(grpo): backward LastIndex walk replaces core.Split in ExtractGRPOExpectedAnswer (-1 alloc, 4 → 2 allocs)

Snider
and
Virgil
committed
merge: cladius-lane-Wave10-W10T (gguf info bulk-string arena — TypicalLayers 484→44 allocs -90.9%, VocabHeavy 724→224 -69.1%)
Snider
committed
chore(external): bump go-inference → 8241328 (W10-Y — Open 1.9× faster, PutOpts 0-alloc warm + -17%, decode -3-7%)
Snider
committed
perf(sft): inputs+targets share one int backing in sftStreamingPacker first-add (3 → 2 allocs, 4× faster)

Snider
and
Virgil
committed
perf(hf): first-character dispatch in architectureFromTransformersName (-61% BertRerank / -76% Unknown)

Snider
and
Virgil
committed
perf(sft): unify [][]int + [][]float32 backing in sftBatchFromExamples (3 → 2 allocs, 5× faster)

Snider
and
Virgil
committed
perf(sft): combine inputs+targets+mask backing in buildSFTExample (5 → 4 allocs, 2.8× faster)

Snider
and
Virgil
committed
perf(sft): pre-size out to cap=4 in newSFTBatchBuilder (12 → 10 allocs, -20% B)

Snider
and
Virgil
committed
perf(hf): single-pass JANG/JANGTQ scanner skips haystack for JANGTQ packs (-14% bytes / -7% allocs)

Snider
and
Virgil
committed
perf(lora): adapterConfigPath direct concat — Dir 167→22ns (-87%) / Safetensors 234→60ns (-74%)

Snider
and
Virgil
committed
merge: cladius-lane-Wave10-W10Z (hf perf — PlanFits_LocalCache 118685→54050 ns -54%, Architecture name lookups -61% to -76%, ReadDir-over-PathGlob structural win)
Snider
committed
perf(lora): joinDirChildPattern direct concat — copyModelPackMetadata + ensure + fuseAdapterWeightFiles -10% ns

Snider
and
Virgil
committed
perf(metal): gemma4.go use ShapeInto/Transpose4 — eliminate hot-path []int32 heap allocs

Snider
and
Virgil
committed
merge: cladius-lane-Wave10-W10AB (grpo+sft hot-path — GRPO_CleanAnswerLine_NoMatchAlphaMixedCase 121→7.1 ns 17×, SFT_BatchBuilderFinish 1627→322 ns 5× -6 allocs; ASCII case-fold + unsafe-cast [][]i…
Snider
committed
perf(lora): hoist joinDirChildPattern to adapter.go + route hashAdapter through it — drop the last filepath.Clean trip from Inspect

Snider
and
Virgil
committed
perf(metal): gemma4_vision.go use ShapeInto/Slice4/Transpose4 — vision-tower hot-path allocs

Snider
and
Virgil
committed
perf(lora): isModelWeightMetadataCopySkip case-fold-in-place — uppercase 43→12ns + alloc shed; lowercase -23%

Snider
and
Virgil
committed
perf(metal): gemma3.go use ShapeInto/Transpose4 — model forward + attention output

Snider
and
Virgil
committed
perf(metal): qwen3.go use ShapeInto/Transpose4 — model forward + attention output

Snider
and
Virgil
committed
merge: cladius-lane-Wave10-W10AC (lora pkg adapter+fuse — AdapterConfigPath -87% ns, CopyModelPackMetadata -18%, joinDirChildPattern helper across 5 sites)
Snider
committed
perf(profile): add LookupArchitectureProfileRef for read-only callers

Snider
and
Virgil
committed
chore(external): bump go-inference → W11-B parse-side jsonenc + UnmarshalJSON walkers (anthropic/openai/ollama -20-52%)
Snider
committed
perf(safetensors): NEON FCVT Float16ToFloat32 inline-C wrapper — DecodeFloatData_F16_2048 3445→990 ns -71%

Snider
and
Virgil
committed
perf(metal): gemma4_assistant_decode.go use ShapeInto/Slice4 — assistant decode paths

Snider
and
Virgil
committed
perf(hf): use LookupArchitectureProfileRef on planFit hot path

Snider
and
Virgil
committed
perf(local_tuning): use LookupArchitectureProfileRef in tuningRuntimeForArchitecture

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11D (NEON FCVT Float16ToFloat32 inline-C — F16 decode 3445→990 ns -71% 3.5×, full uint16 bit-parity, darwin+arm64 only with scalar fallback)
Snider
committed
perf(memory): use LookupArchitectureProfileRef in NewPlan resolver

Snider
and
Virgil
committed
perf(agent): linear-scan dupe-URI check for ≤32-entry Validate
Snider
committed
perf(agent): stack-rooted 8-key buffer for indexEntryHash meta sort
Snider
committed
perf(metal): lora.go use Size() instead of Shape()→product loop in regularisation

Snider
and
Virgil
committed
perf(model): use LookupArchitectureProfileRef in pack inspectors

Snider
and
Virgil
committed
test(profile): bench LookupArchitectureProfileRef variants

Snider
and
Virgil
committed
perf(metal): gemma4 + gemma4_assistant_decode follow-up Transpose4 — attention out projection

Snider
and
Virgil
committed
perf(agent): merge empty + trim guards in firstNonEmpty loop body
Snider
committed
merge: cladius-lane-Wave11-W11C (profile defensive clone eliminated — LookupArchitectureProfileRef 0 allocs, PlanFits_SingleRemote 4330→901 ns -79%, PlanFits_LocalCache 103→50 μs -52%, tuningRuntim…
Snider
committed
perf(agent): route indexModel through pooled bytes.Buffer + sha256.Sum256
Snider
committed
merge: cladius-lane-Wave11-W11A (model files sweep — 32 sites across gemma3/4/vision/qwen3/lora/assistant_decode using ShapeInto+Slice4+Transpose4 — 8-30 saved []int32 allocs per layer per token)
Snider
committed
perf(agent): zero-alloc entry-hash equality check via stack [32]byte
Snider
committed
perf(agent): zero-alloc index-hash equality via stack [32]byte
Snider
committed
merge: cladius-lane-Wave10-W10AD (agent package perf — Validate_1000Entries -55% time -99.5% allocs, CheckStateIndexCompatibility 0-alloc via stack [32]byte hash equality, NewStateIndex -63%)
Snider
committed
perf(profile): zero-alloc compactArchitectureName via byte-walk into stack buffer

Snider
and
Virgil
committed

Commits on May 23, 2026

perf(memory): pre-size NewPlan Notes slice when Pack carries architecture

Snider
and
Virgil
committed
perf(bundle): fold SAMI layer arrays into single backing alloc
Snider
committed
merge: cladius-lane-Wave11-W11E (compactArchitectureNameInto byte-walk into [64]byte stack buffer — 0-alloc on Ref Native/TransformersName/Alias paths, -45%/-46% time; bonus -72% on _Normalised via…
Snider
committed
perf(metal): MulScalar/AddScalar inline-C bridge — 1 cgo crossing

Snider
and
Virgil
committed
perf(model): cache HasErrorIssue result in finalizeModelPack
Snider
committed
perf(speculative): per-field assignment avoids struct-literal zero writes
Snider
committed
perf(model): extend dir-index to tokenizer.json — skip syscall on missing
Snider
committed
merge: cladius-lane-Wave11-W11F (MulScalar/AddScalar inline-C bridge — 3 cgo crossings → 1, -16%/-17% per-token, AdamW step cascade, 30+ callers transparently)
Snider
committed
perf(model): drop redundant LookupArchitectureProfileRef in TaskProfiles
Snider
committed
perf(session): byte-scan '<' early-exit on per-token control marker probe

Snider
and
Virgil
committed
perf(model): drop redundant sort.Strings — PathGlob already sorted
Snider
committed
merge: cladius-lane-Wave10-W10AH (speculative.go per-field assignment eliminates struct-literal zero writes — _256Tokens ~10-15%, STEXT -128B)
Snider
committed
perf(bundle): skip redundant TargetKeys clone when adapter came from info
Snider
committed
perf(session): shared backing buffer for key=value label concat — 3 allocs to 2

Snider
and
Virgil
committed
perf(memory): probe profile index directly when input is canonical

Snider
and
Virgil
committed
perf(minimax): findSidecarRef inlines candidate fan-out — -82% time on hit
Snider
committed
perf(minimax): findPackedWeightRef inlines candidate fan-out — 0-alloc on hit
Snider
committed
perf(memory): skip ModelArchitectureProfile struct copy in MoE branch

Snider
and
Virgil
committed
perf(minimax): findProjectionBiasRef inlines candidate fan-out
Snider
committed
perf(session): single-label fast path in agentMemoryLabelsFromInference

Snider
and
Virgil
committed
perf(bundle): stack-buffer fast-path for buildModel + buildAdapter hash payloads
Snider
committed
perf(session): direct map writes when no caller metadata — Typical -34% Full -26% time

Snider
and
Virgil
committed
perf(memory): fill base Plan in place instead of returning by value

Snider
and
Virgil
committed
perf(memory): replace per-class fill with precomputed Plan baselines

Snider
and
Virgil
committed
perf(session): restore Trim safety in inline metadata writes

Snider
and
Virgil
committed
perf(session): pre-grow Generate Builder to 1024 bytes — skip 5-step growth ladder

Snider
and
Virgil
committed
perf(minimax): 4-way unroll ProjectRouterScores dot-product — -73% time
Snider
committed
perf(memory): short-circuit empty architecture before profile/normalize calls

Snider
and
Virgil
committed
perf(fast_eval): modelDecodeGenerate captures *GenerateConfig — 1 alloc, 24 B (was 2/184)

Snider
and
Virgil
committed
merge: cladius-lane-Wave10-W10AG (bundle perf — SAMI layer arrays fold + adapter clone dedup + stack-buffer hash payloads; SAMIFromKV_PrecomputedAnalysis_2048 -50% allocs, New_Typical -3 allocs)
Snider
committed
merge: cladius-lane-Wave10-W10AE (model+m2 perf — findSidecarRef -82%/-93% allocs, findPackedWeightRef -88%/100% allocs gone, ProjectRouterScores -73%, dir-index tokenizer.json + HasErrorIssue cache)
Snider
committed
merge: cladius-lane-Wave10-W10AI (session root files perf — sessionParserControlToken Miss 52.4→3.3 ns -94%, agentMemoryMetadataFromInference Typical 238→171 ns -28%, agentMemoryLabelsFromInference…
Snider
committed
perf(fast_eval): hoist bench-runner base GenerateConfig to runner scope

Snider
and
Virgil
committed
merge: cladius-lane-Wave10-W10AF (memory NewPlan perf — MiniMaxM2Pack -80% allocs -83% time, Qwen3MoEPack -75% -67%, NoPack -39%; precomputed Plan baselines + canonical-index fast path)
Snider
committed
merge: cladius-lane-Wave11-W11H (modelDecodeGenerate base *GenerateConfig — GenerateSpeculative 4→3 allocs -168 B, 2.9× faster on Construct path; coordination note for W11-G to push to 1 alloc via …
Snider
committed
perf(chat): normaliseRole captures Lower(Trim) once — Unknown 14.6→8.4 ns (-42%), Format ~22-30% faster (every message in every Format call hit the doubled-work path)
Snider
committed
test(distill_bench): cover EmitProbe + FormatStepDir + NewCheckpointMetadata hot paths
Snider
committed
perf(chat): drop no-op WriteString("") in formatPlain — 263→248 ns (~-15ns) by removing the per-call branch + length check that produced zero bytes
Snider
committed
perf(distill): pool emitDistillProbe meta + Training, cache temp string — -80% allocs -51% time
Snider
committed
bench(backend): add toRootProbeLogits / toRootToken / toRootAdapterInfo / toRootMetrics dispatch benches

Snider
and
Virgil
committed
perf(medium): mediumModelRoot drop redundant cleanMediumPath of PathDir output — 102→82 ns -20% on .gguf/.safetensors hot path; PathDir result is slash-clean + whitespace-free by construction
Snider
committed
perf(dataset): LoadJSONL per-field record reset + samples pre-size

Snider
and
Virgil
committed
perf(chat): writesRole flag in chatFormatCapacity — Plain back to 2048B (was 2304B); reserve maxNormalisedRoleLen for templates that emit role so aliased roles (gpt/bot/model→assistant) cannot unde…
Snider
committed
chore(external): bump go-inference → W11-G decode.Speculative inner loop (Edge -18.6%, TokensText -14.1%)
Snider
committed
perf(backend): W8-A2 unsafe-cast toRootProbeLogits — Typical -29%, Large -18%

Snider
and
Virgil
committed
perf(dataset): MessagesToSample chat.Config copy-tweak over rebuild

Snider
and
Virgil
committed
perf(medium): walkMedium inline POSIX path concat — 1506→1180 ns -22% on small-tree walk; medium.List contract is POSIX-slash single-segment so filepath.Join's Clean step is dead work, alloc count …
Snider
committed
perf(medium): copyMediumTree inline OS-native destination concat — destinationRoot (MkdirTemp output, no trailing sep) and POSIX-clean relative concat to a clean OS-native path; mirrors walkMedium'…
Snider
committed
perf(chat): split normaliseRole into inline-able canonical fast path + slow alias path — NormaliseRole_Canonical 6.0→0.62 ns (-90%, ~10× faster), Format paths inline the common-case switch (no fn c…
Snider
committed
perf(backend): W8-A2 reinterpret inference.Message → metal.ChatMessage — Chat/ChatStream/ChatChunksStream zero-alloc

Snider
and
Virgil
committed
perf(distill): pool DistillationBatchLoss vocab-sized scratch — -100% allocs -99% bytes
Snider
committed
perf(chat): templateName canonical fast path before Trim/Lower — Architecture 11.0→2.17 ns (-81%), Empty 6.2→2.19 ns (-65%); fires once per Format so the savings compound across every chat-completi…
Snider
committed
perf(dataset): firstNonEmpty variadic → fixed two-arg

Snider
and
Virgil
committed
merge: cladius-lane-Wave10-W10AL (medium.go inline-concat — WalkMedium -10-22%, ModelRoot -20%; options/helpers at floor)
Snider
committed
merge: cladius-lane-Wave10-W10AM (distill EmitProbe pool + temp string cache + BatchLoss scratch pool — EmitProbe 5→1 alloc -80% time, BatchLoss 3→0 allocs -100% B)
Snider
committed
merge: cladius-lane-Wave10-W10AJ (chat normaliseRole/templateName inline-via-fast-path — NormaliseRole_Canonical 6.04→0.62 ns -90% ~10×, TemplateName_Architecture -80%, Format paths -23-32%; fixed …
Snider
committed
perf(backend): drop redundant clone in toRootAdapterInfo — Typical -84% / -64 B / 1 alloc gone

Snider
and
Virgil
committed
bench(backend): add ToRootMetrics_LoRA case demonstrating 0-alloc after redundant-clone drop

Snider
and
Virgil
committed
perf(dataset): MessagesToSample assistantIdx raw-compare fast path

Snider
and
Virgil
committed
perf(dataset): jsonRecord.toSample pointer receiver — eliminate ~256 B per-row stack copy

Snider
and
Virgil
committed
merge: cladius-lane-Wave10-W10AN (backend.go — toRootAdapterInfo -84%/-1 alloc (double-clone bug fixed), toRootProbeLogits -43% W8-A2 unsafe cast, chatMessagesAsMetal reinterpret 0-alloc 0-cost)
Snider
committed
merge: cladius-lane-Wave10-W10AK (dataset/jsonl.go massive wins — LoadJSONL TextOnly -63%, NewJSONL_1000 -73%; per-field record zero + samples pre-size + pointer receiver + firstNonEmpty two-arg + …
Snider
committed
test(gguf): add tokeniser vocab array bench

Snider
and
Virgil
committed
perf(chaptersmoke): slug — single buffer prefix+body, in-place dash trim

Snider
and
Virgil
committed
perf(gguf): []string fast path for tokeniser-embedded vocab arrays

Snider
and
Virgil
committed
perf(chaptersmoke): bundleURI single-alloc — append slug body in place

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11I (gguf []string array specialisation — TokeniserVocab_200k 8.72→5.70 ms -34.7%, -200k allocs per model load)
Snider
committed
perf(chaptersmoke): W9-Y sentinel errors — lift 9 constant-message *Errs

Snider
and
Virgil
committed
perf(chaptersmoke): pre-size countingStore.unique to bundle block count

Snider
and
Virgil
committed
perf(blockcache): pre-compose hash header buffer — collapse 4 length-prefix escapes into 1

Snider
and
Virgil
committed
perf(chaptersmoke): lift Capture labels literal to package var

Snider
and
Virgil
committed
bench(chaptersmoke): add benchmarks for slug / bundleURI / countingStore

Snider
and
Virgil
committed
perf(dataset_stream): ownership-flip flush + lazy first-add + shared int backing

Snider
and
Virgil
committed
merge: cladius-lane-Wave10-W10AO (chaptersmoke — Slug_Empty -58%, BundleURI -40%/-50% allocs, CountingStore hinted -72%; sentinel lift + appendSlugBody helper + newCountingStoreHint)
Snider
committed
perf(tokenizer): DecodeOne single-id fast path — IDToken_PlainToken 9.17 ns/4 B → 2.51 ns/0 B (-73% time, -1 alloc/op per emitted generation token)

Snider
and
Virgil
committed
perf(probe): batch payload-pointer scratch across Recorder.Events — 1409→770 allocs (-45%)

Snider
and
Virgil
committed
chore(external): bump go-inference → W11-L Generator interface (enables -2 allocs Speculative / -1 alloc PromptLookup via pooled-struct caller)
Snider
committed
merge: cladius-lane-Wave11-W11J (dataset_stream datasetPacker sibling-divergence audit — Packed_TightSeq_1000Rows -72.5%, geomean -46.6%; ownership-flip flush + lazy first-add + shared int backing …
Snider
committed
merge: cladius-lane-Wave11-W11K (tokenizer DecodeOne(int32) — IDToken 9.166→2.486 ns -73% / -1 alloc / 0 B; per-token decode 670 ns + 100 allocs saved per second streaming)
Snider
committed
perf(pack): fast-path empty ApplyOptions + direct byte-buffer IssueSummary

Snider
and
Virgil
committed
docs(probe): document CloneEvent per-payload allocation shape vs Events batching

Snider
and
Virgil
committed
perf(inference_contract): toInferenceMemoryPlan cached quantization label — 127→37 ns -71% / -1 alloc per call

Snider
and
Virgil
committed
test(model): add parseConfigProbe parse-only benches

Snider
and
Virgil
committed
feat(bundle): add SaveCompact for cold-storage newlineless JSON

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11O (Bundle SaveCompact — Typical -91.4% disk / -57% time / -47% allocs, Large -92.1% disk; Save kept indented for human-debug, Load auto-detects both forms)
Snider
committed
perf(bundle): FileHash streaming for large files

Snider
and
Virgil
committed
perf(fast_eval): modelDecodeGenerate via pooled struct (W11-M)

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11M (modelDecodeGenerate pooled struct — GenerateSpeculative 3-6→2 allocs/call lands W11-L payoff; Construct 1→0 allocs -50% ns; SpeculativePairConstruct 2→0 allocs -73%…
Snider
committed
merge: cladius-lane-Wave11-W11P (Bundle FileHash size-conditional streaming — 10MB -99.7% B/op +16% faster, 1MB -97% B/op +22% faster, 32KB threshold, bit-exact across block boundaries)
Snider
committed
perf(model): modelConfigProbe hand-rolled UnmarshalJSON walker

Snider
and
Virgil
committed
perf(blockcache): pre-render aligned prefix_tokens labels in New — -4 Itoa allocs per WarmCache

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11N (modelConfigProbe hand-rolled UnmarshalJSON walker via jsonenc — ParseConfigProbe -22-27% time, -3 allocs; ReadModelConfig full path -8-10%)
Snider
committed
perf(memvid/cli): expand bench coverage for golden-path

Snider
and
Virgil
committed
merge: cladius-lane-Wave10-W10AQ (inference_contract toInferenceMemoryPlan -71% time -50% allocs via cached quantization label; shape.go + adapter.go A5 ship-zero at floor)
Snider
committed
test(metal): tokeniser bench coverage for W11-S re-pass

Snider
and
Virgil
committed
bench(metal): add sampler coverage — MinP/Temperature/SuppressedGreedy/RepeatPenalty/HostFallback
Snider
committed
perf(metal): tokeniser indexIn via core.Index (stdlib Rabin-Karp)

Snider
and
Virgil
committed
perf(metal): applyRepeatPenalty pooled scratch + sort-dedup (W11-R)
Snider
committed
perf(metal): suppressTokenLogits pooled scratch + sort-dedup (W11-R)
Snider
committed
perf(memvid/cli): drop dead reassignment in Resolve

Snider
and
Virgil
committed
merge: cladius-lane-Wave10-W10AP (blockcache+probe+pack — ApplyOptions_Defaults -90% (empty-loop escape leak fix), Recorder_Events_128 -45% allocs, BlockCacheID -37% allocs, WarmCache -25% allocs)
Snider
committed
perf(memvid/cli): stack-buffer Put tag keys for ≤16-tag fast path

Snider
and
Virgil
committed
perf(metal): hostUnsuppressedGreedyToken slices.BinarySearch + pooled scratch (W11-R)
Snider
committed
perf(memvid/cli): exact-size Put args slice (drop worst-case over-alloc)

Snider
and
Virgil
committed
perf(metal): cache_quantized reuse shape buffer across Update

Snider
and
Virgil
committed
perf(metal): cache_paged_metal swap W10-G pool for W11-A stack-buffer Slice4

Snider
and
Virgil
committed
perf(memvid/cli): drop duplicate ready() call in view()

Snider
and
Virgil
committed
perf(memvid/filestore): bit-exact binary round-trip parity + Resolve/RoundTrip benches

Snider
and
Virgil
committed
perf(metal): hostUnsuppressedGreedyToken zero-copy float32 view (W11-R)
Snider
committed
perf(metal): cache_quantized Size/Dim over Shape() in hot paths

Snider
and
Virgil
committed
perf(metal): tokeniser Decode single-pass Builder with inline ▁→space

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11Q (pkg/memvid cli shim — Put -128-256 B / -1 alloc, stack-buffer tags, exact-arg-slice; finding: pkg/memvid is shim, real State perf is go-inference state/filestore al…
Snider
committed
perf(metal): decodeGPT2Bytes pre-sized buf + AsString + inline UTF-8

Snider
and
Virgil
committed
perf(metal): scalar-inline TopP/MinP/MinPSampler bridge ops (W11-R)
Snider
committed
perf(metal): normalizeSentencePieceSegment single-pass + zero-alloc passthrough

Snider
and
Virgil
committed
perf(metal): cache_paged_metal pool scorePages slice in SDPAPaged

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11R (sample.go golden-path — hostUnsuppressedGreedyToken 1MB→35B (-99.997%) zero-copy MLX view; applyRepeatPenalty + suppressTokenLogits -84% B / -67% allocs via sync.Po…
Snider
committed
perf(metal): cache PagedKVCache storageKVPair slice-free conversion

Snider
and
Virgil
committed
perf(metal): cache_paged_metal swap Zeros for Zeros4 in page-grow

Snider
and
Virgil
committed
perf(metal): cache cacheTail NumDims fast-path skips Shape() allocs

Snider
and
Virgil
committed
perf(metal): cache_quantized NumDims gate in Update entry

Snider
and
Virgil
committed
perf(metal): metal_kernel Apply inline-C inputVec collapse — N+2→1 cgo crossings

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11U (QuantizedKVCache_Q8Q8 -69% allocs / -71.7% B; shape-buffer reuse + NumDims/Dim/Size accessors + slice-free PagedKV pair + cacheTail fast-path; A5 revert on cacheFlo…
Snider
committed
perf(metal): cache_paged_metal embed shape scratch as [4]int32 fields

Snider
and
Virgil
committed
perf(metal): prompt_cache Slice4 + ShapeInto on golden-path restore

Snider
and
Virgil
committed
perf(metal): bpeMerge direct min-heap bypasses container/heap interface boxing

Snider
and
Virgil
committed
perf(metal): generate.go use zero-copy view in inspectAttentionCache (W11-X)

Snider
and
Virgil
committed
perf(metal): prompt_cache drop Shape() heap allocs in validators

Snider
and
Virgil
committed
perf(metal): encodeSentencePieceSegment / encodeGPT2Segment via splitRunes zero-alloc

Snider
and
Virgil
committed
perf(metal): kv_snapshot.go use zero-copy views in inspectKVCacheRangeWithOptions (W11-X)

Snider
and
Virgil
committed
perf(metal): cache_paged_metal short-circuit empty GO_MLX_PAGED_KV_PAGE_SIZE parse

Snider
and
Virgil
committed
perf(metal): cachedBPETokens defer-free RLock unwind

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11T (PagedKV W10-G→W11-A substrate evolution + cache_paged_metal.go deleted; SDPAPaged_16Pages 1→0 alloc -98% B; Zeros4 scalar-pass + scorePages pool + embedded [4]int32…
Snider
committed
perf(metal): Zeros4 rank-4 scalar-pass + fixed-cache restore swap

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11S (tokenizer full pass — indexIn -93%, Encode_Typical -49%/-1 alloc, bpeMerge_Long -17 allocs via typed heap, normalizeSP -48%; Decode/Encode cascade)
Snider
committed
perf(metal): metal_kernel add ApplyOne fast path — single cgo crossing for 1-output kernels

Snider
and
Virgil
committed
perf(metal): migrate single-output kernel callers to ApplyOne — 1 alloc/op savings

Snider
and
Virgil
committed
test(metal): bench_test.go realistic benches for cache snapshot fan-outs (W11-X)

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11X (zero-copy view sweep — inspectAttentionCache 33.5→16.8 MB -50% / -64% time, inspectKVCacheRange -25% B -41% time; A5-honest revert on probe.go (cgo overhead > saved…
Snider
committed
perf(metal): prompt_cache appendRestore — skip per-cache slice literal

Snider
and
Virgil
committed
perf(metal): prompt_cache lazy snapshotOffsets on failure path only

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11W (prompt_cache.go residual — copyCachePrefix 3→0 allocs, 26-cache Gemma 4 restore -156 allocs, RoundTrip -17% ns + -33B; Zeros4 convergent with W11-T)
Snider
committed
perf(metal): metal_kernel add DispatchOne — collapse cfg+apply into single cgo crossing

Snider
and
Virgil
committed
perf(metal): migrate single-output kernel callers to DispatchOne — drop cfg ceremony

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11V (metal_kernel.go 3 new substrate primitives — apply_inline + ApplyOne + DispatchOne; MoE_RouterProjection_H2048_E32 -21.6%, geomean -37% B/op -29.3% allocs; 23 calle…
Snider
committed
feat(metal): add Reshape1 + Reshape2 scalar-pass primitives (W11-AC)

Snider
and
Virgil
committed
feat(metal): add Slice1 / Slice2 / SliceUpdateInplace2 scalar-pass primitives (W11-AC)

Snider
and
Virgil
committed
feat(metal): Slice4WithStream / SliceUpdateInplace4WithStream

Snider
and
Virgil
committed
test(metal): fast bench gaps — nativePagedSingleToken + singleTokenCacheUpdate (W11-Y)

Snider
and
Virgil
committed
perf(metal): fast nativePagedSingleTokenAttention page-handle pool (W11-Y)

Snider
and
Virgil
committed
perf(metal): packQ4Cached use Reshape1 / Reshape2 / Slice2 (W11-AC)

Snider
and
Virgil
committed
fix(metal): pinned_array uintptr_t payload — drop unsafe.Pointer(id) cast (go vet clean)

Snider
and
Virgil
committed
feat(metal): materialiseFloat32ViewFast skips Materialize for contiguous float32

Snider
and
Virgil
committed
perf(metal): cache use stream-passing Slice4 variant

Snider
and
Virgil
committed
perf(metal): unpackQ4 use Reshape1 + Slice1 (W11-AC)

Snider
and
Virgil
committed
perf(metal): fast SDPA mode-string cache (W11-Y)

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11AD (Slice4WithStream + SliceUpdateInplace4WithStream substrate consistency; A5 diagnosis: 1028 alloc floor is newArray SetFinalizer + arrayPool type-assertion, not Def…
Snider
committed
perf(metal): probe.go use materialiseFloat32ViewFast in summarizeProbeLogitsCompact

Snider
and
Virgil
committed
perf(metal): maxAll use Reshape1 (W11-AC)

Snider
and
Virgil
committed
perf(metal): pinned_array buffer pool — KVShape 3->2 allocs, 120->56 B/op (-53%)

Snider
and
Virgil
committed
test(metal): fast singleTokenCausalMask bench coverage (W11-Y)

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11AC (Reshape1/Reshape2 + Slice1/Slice2/SliceUpdateInplace2 scalar-pass primitives — QuantizedKVCache_Q8Q4 -82% allocs / -53% B, Q8Q8 -50% allocs; scalar-pass family com…
Snider
committed
merge: cladius-lane-Wave11-W11AF (pinned_array.go vet cleared via runtime/cgo.Handle pattern + pinnedRawArrayBuffer sync.Pool — all KV shapes 120 B/3 allocs → 56 B/2 allocs -53%/-33%)
Snider
committed
merge: cladius-lane-Wave11-W11Y (fast.go nativePagedSingleToken sync.Pool — 2Pages -53%, 4-16Pages -43-46%; SDPA mode-string cache; A5 reverts on ShapeInto + scalar-cache showing cgo-stack-array + …
Snider
committed
docs(metal): materialiseFloat32ViewFast contract — caller must pre-Eval arr

Snider
and
Virgil
committed
perf(metal): generate.go use materialiseFloat32ViewFast in inspectAttentionCache

Snider
and
Virgil
committed
perf(metal): kv_snapshot.go use materialiseFloat32ViewFast in inspectKVCacheRangeWithOptions

Snider
and
Virgil
committed
merge: cladius-lane-Wave11-W11AE (materialiseFloat32ViewFast skips Materialize for contiguous-float32 — InspectAttentionCache -83%, InspectKVCacheRange -74%; threshold inverts vs W11-X — fast-path …
Snider
committed

Footer

© 2026 GitHub, Inc.

Footer navigation

Terms
Privacy
Security
Status
Community
Docs
Contact

You can’t perform that action at this time.