Repository navigation
Qualify post-v0.6.0 upstream correctness fixes and consumer compatibility #112
Description
Triage
P2 / enhancement / remaining owner qualification. The v0.8.0 Apple P1 #154 is repaired by #153's published v0.8.0-1 and consumed by core 0.17.1/0.17.2. Remaining device/model gates are not closed by that repair. No instruction to adopt raw main blindly.
Current status — 2026-11-07
| Layer | Revision |
|---|---|
| Raw stable | v0.9.0, cd959bae99ff4bed22bef2604f45241c09bbb8c8. |
| Raw nightly / main | b12184, c0ab6e5c0bab3fb21e6cb9d2371f3c41d85ea775; inspected main cf659f240b232e15700b1721bbb43c1d973c52e9,227 beyond bundled source. |
| Published owner | v0.8.0-3, owner 703c0f51fe9f8ebfa941c70504ad0b13c7ca79cb, upstream 8642541b255bfb5dc39cd351a178be09013a55d6, carries0001/0002/0004. No new family/performance feature. |
| Main/local | c6010608 pins v0.8.0-2; core#1746, d14513ae, prepares -3 bindings/checksums/SwiftPM and Apple0.18 floor. Draft/unpublished; refresh against current main before final acceptance. |
| Web owner | Immutable assets v0.1.82, bridge e57b9f68, same8642541b; main82/public79. |
| Public core | 0.17.5, native v0.8.0-1/upstream8642541b; Apple companions0.17.3. |
Current exact evidence and next gates. Draft#175,7f15b5d1, now passes89 hosted checks after ISA/dispatch containment review, superseding the unprepared fingerprint blocker. Independent exact-head audit remains pending;0.9 publication is held until core0.18. Carries0001/0002 apply;0003/0004 are detected fixed only for0.9.
ARM FP16#174 is a confirmed P2 numerical defect in shipped v0.8.0-3 Mac/optimized Linux ARM64 modules, not merely suspected exposure. Draft repair#177 still needs current-head independent acceptance, patch-policy and corrected-artifact/platform gates. Public-model impact and regression introduction remain unmeasured.
Raw-only new Kimi-K3.1-Pro text derivative, earlier Limite/Violetto, Mistral Large4, Clover AR, Metis1.6, xing4_0 and Ripple need converter/metadata/template/tool/reference/model/backend qualification; no distinct new public Dart API is demonstrated. Speech preparation draft#176 remains source-only; retain Pocket#163, Chatterbox#164 and Drama3#165 acceptance separately.
Family status at v0.8.0 (acceptance item 4):
| Family | Status |
|---|---|
| LFM3 text, VL, Omni, Audio, d1-2B, d1-Omni-2B | Qualified, #130 closed |
| Soprano | Qualified, #123 closed |
| EmbeddingGemma2 BF16 | Native CPU and Metal reference vectors; Web worker and direct match at cosine ≥ 0.99997 (#1618) |
| Decision 2/3, pplx-decider, LFM2 d1-3B and d1-omni | Not run. Consumer semantics: llamadart#1125 |
| Qwen3.5 embedding variants, K2 Horizon, Cohere2 vision, MiniCPM-V 4.7, Prism Bonsai 2, maion-coder, Nemotron Super 3.5 VL/Bolt | Not run |
| Muse Spark Open, Reflection Beam, Kolibri-1, SenseNova U1, Flint, DeepSeek-V4.1 | Not run. Beam: llamadart#1165 |
| LiliCorr | Not run: llamadart#1071. Voxtral stateful audio: llamadart#327 |
Fix controls at v0.8.0 (item 3): compared with v0.7.0-3 on #1611: combining-mark tokenizer #30845, Qwen-VL resize #30736, set_embeddings re-reserve #31086 with patch 0003 retired, CPU/Metal greedy agreement on Qwen3.5, Qwen3-TTS state API. Not compared: the Metal alias race #30795, CUDA and Vulkan fix controls beyond #79/#116, DFlash and MTP (#30688, #30698, #31077), Laguna #30896, and CUDA gpt-oss #30950/#30951.
Upstream limits kept (item 6): patches 0001 and 0002 still apply. Mali #30106 is merged and in v0.8.0; LFM3 text and Soprano pass on Mali-G715 Vulkan, but llamadart#948 stays open on its own gates. ON_DEVICE #30318 remains open. Quantized-KV #30541 merged e56f006e, first b12121 after stable0.9; absent owner/public, with its independent acceptance retained.
Earlier status and candidate notes — 2026-10-21
Published llamadart 0.15.5 consumes native v0.7.0-1: owner daffc7fc8bdb62bbc09f723a74e761f6f32d549f, upstream stable v0.7.0 620d2225cc436ab0b001e3200e590e3b286ef9ab, including #117's guarded mixed-batch processing. Published core and current main both use Web v0.1.66 at the same upstream revision. Platform/model qualification remains core #1189, not automatically complete.
Latest raw nightly b11786 is 479a52779ff4c4674b81a226f8511816e15dc8dd; raw main eb182d85c0e304a69ecc0be0c5e12884691c116a is 123 commits beyond the owner source (39 since the previous scan; six after b11786). Raw stable v0.7.0 and owner artifact/pins are unchanged. New Web owner assets v0.1.67 retain that upstream source; consumer repin core #1260 is pending in #1282. Earlier candidate gates remain below:
| Change | First containing raw release / required evidence |
|---|---|
Combining-mark tokenizer repair #30845, d8d92973c8c21bb4dd5d592f0fec26fe2cd1e654 |
b11753. Fixes #30787. Pinned TEKKEN/GPT4O/MINIMAX_M2 source still has the affected letter-only lookaheads; compare Hindi/Thai combining marks, exact token IDs and unaffected ASCII controls through native and Web. This is upstream defect/source exposure, not a newly proven Dart PR regression or locally reproduced binary failure. |
Qwen-VL resize parity #30736, f711c8f3038020d71f69b8ced09162b014fd7335 |
b11747. Removes unintended letterboxing and matches half-even resizing for affected Qwen-derived projectors. Upstream says existing GGUFs need no reconversion. Qualify native/direct/worker image token counts and OCR/embedding controls; name untested projector families. |
SWA checkpoint invalidation #30666, f174553e1621c6a22a64e7bfcbb58d3a5c49e6fb |
b11738. The fix is in llama-server: do not label Dart state/session paths repaired or affected without a call-path reproduction. |
50/50 SLRU expert cache #30637, 33d6135cbe20e90e67d9db1c874fbb4a9dc64b19 |
Present in b11738 (tag b11736). Core #1123 still owns the opt-in budget API/qualification. Preserve disabled defaults, target/draft isolation, eviction/memory/output controls; upstream benchmark observations are not a Dart speedup. |
| Hybrid-DNA tokenization #30825, SYCL reordered get_rows #30430 | b11742 / b11748. Existing-model/backend corrections, not new families. Test only applicable formats/backends; shipped platform coverage must not be inferred from raw SYCL support. |
Jinja unique #30819, 0ce8cd11416c8840498e6701bcbe39ad4b858c2b |
b11739; upstream latest GLM-5.3-Flash tool-template compatibility. Audit the actual Dart renderer/template separately; do not assume missing support in Dart merely from upstream's former missing filter. |
Earlier candidate gates remain below; their fixes are now included in b11753 but remain absent from the owner artifact.
- CUDA Flash Attention shared-memory WAR repair #30573,
99db719bdf07f7d537431ca556a0512bf3d995b6, first b11724: upstream sanitizer hazard reduction is evidence for candidate testing, not a demonstrated current Dart output defect or measured consumer speedup. - Metal MOE_REDUCE alias-race repair #30795,
098ce526b546b5a190327c1da82929cee051b9ac, was after b11724 and is now contained in b11753. Audit alias/fusion controls on the exact candidate; do not assert that it fixes owner Qualify Intel Mac AMD Metal output correctness and fail-safe selection #116's Intel/AMD report without reproduction. - Qwen3 bidirectional attention / dense_2 projection #30607,
412a363c655b479279d3050d0d1c67a5733697b4, adds the Voyage-4-nano embedding variant, not a separate general Qwen family. Test conversion, projected vector dimensions/pooling and reference retrieval before public support. It was raw-main-only after b11724 and is now contained in b11753. - Raw-only families since the pinned source: Muse Spark Open #30616 (text/vision/audio), Reflection Beam #30670 (core #1165) and Kolibri-1 #30673 (first b11709). Qualify converter/tokenizer, metadata, templates/tools and representative modalities separately. Granite Switch Decision variants (core #1125), Mellum2.1 MTP and Voxtral realtime streaming are variant/capability work, not additional newly added families.
- Noncontiguous SUM #28851, Qwen4exp CUDA fusion #30602, Hexagon long-context/multistream FA #30731 / #30218, MMQ Stream-K #30553 and CPU vectorization #30791 remain raw-only. Preserve output/memory controls and measure on real applicable hardware.
- MTP multi-slot issue #27572 was closed by the stale bot as not planned, not repaired. Retain its concurrency/acceptance watch until causal fix evidence exists.
Before adoption: audit wrapper exports and public headers, including the new mtmd audio-stream boundary; regenerate bindings only as needed; test state/cancellation/version-skew and affected models; qualify matching Web artifacts, provenance/checksums and consumer docs. This is not a pin-only upgrade.
Evidence and benefit
Historical comparison: v0.6.0-3 bundles llama.cpp d81235049384534c167caea52b85a694f6103d14 plus its two manifest-recorded patches. On 2026-10-15 raw main was 08698c81ed1053b7be1ade00d1e127daf1ae9ba3. Changes assessed in this historical v0.6.0 comparison include:
- CPU AVX2/F16C regression repair: ggml-cpu : keep the FP16->FP32 lookup table on x86 with F16C (#30342) ggml-org/llama.cpp#30343 (685f70894b66; first nightly b11635). Our pin predates both the regression and repair.
- Metal/WebGPU duplicate expert IDs: metal : fix mul_mat_id with duplicate ids ggml-org/llama.cpp#30373 and webgpu: (Fix CI) fix mul_mat_id with duplicate expert ids ggml-org/llama.cpp#30451
- common-chat parts and ggml symbol visibility: chat: replace output string with parts ggml-org/llama.cpp#30389 and ggml: Change default visibility of symbols to hidden ggml-org/llama.cpp#30014
- LiliCorr draft checkpoints: spec : add lilicorr support ggml-org/llama.cpp#30506
Current provenance: https://github.com/leehack/llamadart-native/releases/download/v0.7.0/assets.json
Post-v0.7 watch / remaining qualification
Historical 2026-10-17 raw main/b11707 was 1268edac579110aeb1407f7928f53193ed4e28d4. Beam (#30670, first b11701; consumer llamadart#1165) and Granite Switch decisions (#30676/#30693; llamadart#1125) are newer than the published artifact. Do not conflate them with the v0.7 families now source-consumable. Native #116 owns Intel/AMD Metal correctness qualification.
Upstream #30688 reports a v0.6 -> v0.7 Vulkan MTP slowdown with exact single-commit revert evidence for #30020; published source contains that change, but the custom-model/near-full-VRAM server setup has not been reproduced through our wrapper. Separate #30695 was closed with measured RADV parity after #30468; its fixing commit is after v0.7.0, so the current artifact does not contain it. Do not combine these two mechanisms. #30698 is a Gemma4 MTP+projector server startup report; verify our production call path before asserting exposure. Mellum CUDA #30704's introducing #30584 commit is also after v0.7.0: raw-candidate risk, not established current-pin regression.
Additional acceptance: compare production MTP with/without draft state, VRAM headroom and visible-device counts on available real Vulkan hardware; preserve output/acceptance parity and distinguish first-request reallocation from steady-state performance. Recheck ON_DEVICE/quantized-KV and Nemotron Metal limitations; never disable safety precision checks merely to restore throughput.
Acceptance
- Select an exact upstream candidate and release identity under existing policy; do not call a different source a v0.6.0 wrapper rebuild. Record raw/artifact/consumer revisions separately.
- Audit all wrapper exports, C/C++ ABI/header changes, common-chat call sites, ownership/cancellation/state paths and retained patch necessity.
- Run issue-specific positive/negative controls for applicable fixes; compare affected CPU quantized operations and representative Metal/Vulkan/CUDA/Web paths only where actual hardware exists.
- Qualify family metadata, outputs and conversion needs: Decision 2/3, pplx-decider, LiquidAI d1-3B/d1-omni, EmbeddingGemma2, Qwen3.5 embedding variants, K2 Horizon, Cohere2 vision, MiniCPM-V 4.7, Prism Bonsai 2, maion-coder, Nemotron Super 3.5 VL/Bolt, Muse Spark Open and LiliCorr. Assess Voxtral's new stateful audio boundary with llamadart#327; Mellum2.1 MTP is a capability addition, not a new model family. Do not promote absent models to supported.
- Coordinate matching bridge source/artifact qualification and generated Dart bindings, capability probes, templates/grammars, support matrix and consumer smokes before any downstream adoption.
- Preserve open Mali #30106, ON_DEVICE #30318/quantized-KV #30541 and other remaining upstream limitations; source update alone does not close them.
Boundaries
Tracking/qualification only. This issue does not authorize publication, workflow dispatch, paid hardware runs, downstream pin changes, or weakening supported-channel policy. No PR-caused regression is alleged without a causal reproduction. Preserve immutable historical artifacts and record unavailable models/hardware explicitly; CPU/software fallback is not hardware GPU qualification.
Activity
leehack commentedon Oct 15, 2026
Consumer follow-ups now tracked separately: llamadart#1071 (LiliCorr/DFlash qualification) and llamadart#1072 (multimodal embeddings). Neither is marked public-supported until owner publication and its consumer acceptance gates pass.
leehack commentedon Oct 16, 2026
Native implementation is prepared locally on
feature/v070-native-contracts(commitsdd3634d,c1a51fe; PR pending):Final local gates: 213 Python tests (one platform skip), 48/48 v0.6.0 native contracts, 49/49 v0.7.0 contracts including the real-weight embedding smoke; 86 wrapper exports pass on macOS arm64. Hosted cross-platform and artifact gates remain pending; this is not publication or downstream adoption.
Missing work now has focused trackers: native #113/#114; leehack/llamadart#1123 (expert budget), #1124 (draft microbatch), #1125 (decision families), #1126 (embedding prefixes/dimensions). Existing leehack/llamadart#1071/#1072 retain LiliCorr and typed multimodal embedding scope. Native guarded mixed-batch processing/capability handling, regenerated Dart bindings and model-backed consumers remain under this issue/#1072.
leehack commentedon Oct 17, 2026
Material delivery/qualification update (2026-10-17): native v0.7.0 and matching Web v0.1.63 are published, consumed by core #1141, and shipped in 0.14.0. Refreshed the status block without checking off unproven family/device acceptance. Added post-v0.7 Beam/Granite tracking and bounded GPU risks: #30688 vs separately repaired #30695, server-only exposure checks, and raw-only Mellum CUDA #30704. Intel/AMD Metal now has focused owner #116. No new PR-caused regression is asserted from upstream reports alone.
leehack commentedon Oct 18, 2026
Material update 2026-10-18: v0.7.0-1 is published with #117's guarded processing; core main adopts it and Web v0.1.65, but pub.dev is still 0.14.0. Added raw-only Kolibri-1 (first b11709) to this owner umbrella instead of creating a speculative per-family API issue. Remaining platform/model checks are core #1189, including focused Windows discovery #1199 and bridge #162. No new GPU performance or universal family-support claim.
leehack commentedon Oct 19, 2026
2026-10-19 material update: public core 0.15.1 now consumes v0.7.0-1; owner source remains
620d2225. Refreshed the current block with raw b11724's CUDA WAR repair, later main-only Metal alias-race fix and Voyage-4-nano/Qwen3 projection. Kept family vs variant distinctions and bounded ABI/model/device acceptance. #27572's stale-bot closure is not a fix. No raw pin bump or current-pin exposure is asserted.leehack commentedon Oct 20, 2026
2026-10-20: b11753 /
d8d92973is 84 commits beyond unchanged owner v0.7.0-1. Added combining-mark tokenization #30845, Qwen-VL resize #30736, server-only SWA invalidation, SLRU cache and bounded parity controls to the current block. Core 0.15.4 and Web v0.1.66 still use upstream620d2225. These are candidate correctness/performance gains, not public delivery or automatic pin approval; existing family/device acceptance stays open.leehack commentedon Oct 21, 2026
2026-10-21 candidate delta; P2 qualification, not a delivered runtime upgrade.
8baabd53, first b11785) changes MISTRAL4/MELLUM down-projection handling. Retest Mellum CUDA #30704; its condition excludes LAGUNA and does not fix #30896. Preserve the reported ≤31-token ubatch control.Owner v0.7.0-1 / upstream
620d2225cc436ab0b001e3200e590e3b286ef9abremains consumed by published core 0.15.5. Discovery 39571215746 succeeded with dispatch skipped under stable-only policy; no new artifact or automatic pin bump occurred. Existing acceptance stays open.leehack commentedon Oct 22, 2026
2026-10-22 layer/candidate delta — retain P2 owner qualification.
298314bd0ff4650363e0e63b81399d5618869a63; adopted by core 0.15.6 and llama.cpp companion 0.15.2620d2225cc436ab0b001e3200e590e3b286ef9ab, unchangede197cef4554d466e0becac20e453a355daba8e5d; master7170cad4d02f7907517613931840cbb2367399a3, 146 commits beyond bundled sourceThe rebuild first delivers #118 app-local VC runtimes, #121 ARM64 OpenMP removal and #122 TTS codebook reset; #119 strengthens regressions, not patch retirement. Release 39576675443 passed all 25 active jobs. Its manifest hash is
92b9f1ea6d3ec3dd826122ffde3cf21b2b22036c368e826d86e927284c6908f4; #79/core#782/#114 carries remain. Local manager checkout still has older artifact pins; it was not moved.Before raw adoption:
0a1ef309cdbb9f2024f58a38fb4c42352c5db569, first b11796, removes input.seed/state bytes and changes audio_process to take an explicit state object. Current owner Seed every Qwen3-TTS codebook from the request on each synthesis #122 uses the old fields/three-argument call. Migrate state/RNG/output ownership and reset/cancel/free semantics, regenerate affected bindings, and preserve original TTS regressions; this is not proof the new TTS repair can be retired.7775ffb7, first b11801) and Flint #30928 (a37cc461, first b11803) remain raw-only. U1 supports understanding/text/image-to-text, rejecting generation branches. Flint needs converter/tokenizer/expert-attention and actual Dart-template controls; upstream reports a potential adaptive-scaling slowdown. Audit hex escapes #30929 and namespace shadowing #30941 against our renderer before alleging a Dart gap. Existing Muse Spark/Reflection Beam/Kolibri and decision/embedding variants retain their earlier gates.Watch reconciliation: #30894 is fixed by #30960, first b11813, including load-time pixtral→Tekken handling of existing GGUFs; qualify Hindi/Thai/ASCII/official-Pixtral token IDs. #30905 is fixed by #30940,
d88040c7, master-only after b11813; current native Dart tokenToPiece-based detokenization is not established exposure. Both fixes remain absent from the artifact.New open CUDA gpt-oss reports #30950/#30951 have implicated expressions in pinned source, but our binary was not reproduced. Add exact published CUDA/affected-quant/batch≤8 versus >8, finite-output and CPU/stock-format controls; no PR-caused P1 or universal model claim. CPU #30906, draft-only DeepSeek #30944 and server-slot #30961 are separate watches. Laguna #30896, MTP/Metal and candidate PR watches remain unresolved; EAGLE4 #30925 is a request, not support.
Discovery 39749950877 succeeds with stable-only dispatch skipped; not a stalled release. No raw pin, new-family public API, patch retirement or new hardware qualification is approved.
leehack commentedon Oct 23, 2026
2026-10-23 candidate delta — retain P2 qualification; no pin-only upgrade.
Published native v0.7.0-2, owner
298314bd0ff4650363e0e63b81399d5618869a63, and public core 0.15.6/current main still bundle stable v0.7.0620d2225cc436ab0b001e3200e590e3b286ef9ab. Raw nightly/master b11836,f5027fb21307eb264aed483242603d68c87b0fe3, is 167 commits beyond bundled source, 21 since yesterday.#30940 detokenize repair first appears in b11814; native Dart exposure remains unestablished. Raw-only base families since pin remain Muse Spark Open, Reflection Beam, Kolibri-1, SenseNova U1 and Flint; none newly added today. Their conversion/templates/tools/modalities need representative validation, not an unrelated-model smoke.
Watch changes: CUDA #30950 has open #31019; cache #30961 has open #30997, neither delivered. EAGLE-4 #30930 closed unmerged over license concerns, not implemented support. New Flint #31000 lacks original request/SSE evidence; no Dart parser exposure established.
Raw b11836 is not wholly CI-green: UBSan server job failed test_cancel_request (1 failed/346 passed/4 skipped), cause/owner exposure unknown. Owner discovery succeeds with stable-only dispatch skipped, not a stalled release. Existing #79/#114/#116, Laguna/MTP/state and actual-hardware gates remain; no new carry retirement or raw adoption is authorized.
leehack commentedon Oct 24, 2026
2026-10-24 material candidate delta — retain P2 qualification, not a raw repin.
Owner/latest artifact remains v0.7.0-2,
298314bd0ff4650363e0e63b81399d5618869a63, upstream stable620d2225cc436ab0b001e3200e590e3b286ef9ab, now consumed by public core0.16.0. Latest raw nightly b11852 is7894e8b92481a8bb8ffe94f73b656ffe0dd01e10(+183); raw master6b9ba4f2e8c1436511956f0b3c309e2c9e642071is +186/+19 since yesterday.Watch corrections: #31019 closed unmerged, not a #30950 CUDA fix; #30951 remains open. New Gemma4 MTP #31077 needs checkpoint/full-KV-clear/reprocessing, MTP-off and target/draft reuse controls. MiniMax-M3.1 scales #31063 matches omitted scale arguments in pinned source, not a reproduced owner binary failure. Keep conversion/indexer checks distinct. DSpark log-SNR #30657 and server-cache #30997 remain open; DFlash2 probability still needs candidate distributions/verifier/seed wiring, not source adoption alone. EAGLE4 remains closed unmerged.
New sanitizer run40050711577 passes; old cancel-test failure39918657001 is not causally explained. Some b11852 GPU checks remain queued, so no blanket all-green claim. Owner discovery40077278600 succeeds with stable-only dispatch skipped. Existing hardware/state/carry gates remain; no patch retirement or workflow action.
leehack commentedon Oct 25, 2026
2026-10-25 material raw delta — retain P2 qualification; current pins unchanged.
Public core0.16.2/main081690b and latest owner v0.7.0-2, owner298314bd0ff4650363e0e63b81399d5618869a63, still consume stable upstream620d2225cc436ab0b001e3200e590e3b286ef9ab. Latest raw b11871 is b3fa69955ea81c3b60e77e3c1e2224c4a826ffb9 (+202); masterb083c384f735ee0c4b5b187edcf75a22a675370d is +204/+18 since yesterday (one late-scan commit beyond344be0a6).
Muse Spark Open, Reflection Beam, Kolibri-1, SenseNova U1 and Flint, plus Soprano TTS #123, retain their raw-only qualification lanes. No new separate family ticket is justified without another concrete downstream gap.
Native#95 now has stock Windows Server2022 x64/no-machine-CRT CPU15/15 evidence and is closed; ARM64 execution remains unrun. Intel Iris BF16#31091 is not established current-artifact exposure: Apple presets disable GGML_METAL_USE_BF16; #31090 closed unmerged, not a repair. Keep AMD#116 separate.
Raw b11871 is not wholly green: Vulkan CM and undefined CTest show test-llama-context timeouts; cause/owner exposure unproved, other GPU jobs queued. Owner discovery40177038216 succeeds with stable-only dispatch skipped. Public JSON-depth hardening#30242 first ships b11871; no new vulnerability/exposure assertion or security ticket is made. CUDA/Laguna/MTP/DSpark/cache/EAGLE watches otherwise unchanged.
leehack commentedon Oct 26, 2026
2026-10-26 raw candidate delta — retain P2; no owner/raw repin.
Latest owner remains v0.7.0-2, owner298314bd0ff4650363e0e63b81399d5618869a63, upstreamstable620d2225cc436ab0b001e3200e590e3b286ef9ab. Published core0.16.2 and main6305b34c retain it; release prep#1429 does not change pins. Raw b11890 is a074215352c382e66a108597bcf7f4f31e00d111 (+221); master9125a2ebe8e5a38f0f4232170c134a29ed730b66 is +227/+23 since yesterday (one late-scan commit beyond9b8446b9).
Watch reconciliation: Laguna candidate#31147 is open, not a fix. SparkX2.6 DSpark#31167 closed unmerged; replacement#31168 open (existing architecture/draft capability, not a new family). MTP#26425 closed not_planned/stale, not repaired. #30950 contributor stopped pursuing after #31019's unmerged closure; defect evidence is not thereby retracted. Other CUDA/MTP/state/device/Soprano watches unchanged.
Raw b11890 is not green: WebGPU NVIDIA fails Grok seq-1 state equality after fragmented-cache restoration; macOSx64 and serverMetal time out. Worker/timeout tuning or dropping a lane is not causal repair proof. Late-scan main-only #31162, 9125a2eb, adds the existing mutex around shared memset_params_buf setup/use. Its author traces this race to parallel Grok fragmented-cache CI; the merged code is a concrete repair candidate, but b11890 does not contain it and complete fixed-head CI/owner exposure was not established here. #31085 reports upstream Windows CUDA tag/binary skew, not independently downloaded or established in this owner artifact. Public hardening#29755/#31153 first b11882/b11888 is noted high-level only, without a new vulnerability/exposure claim. Owner publication is unchanged/healthy; no new carry retirement, test run or workflow action.
leehack commentedon Oct 27, 2026
2026-10-27 candidate reconciliation — retain P2 and the separate 0.17 preparation gate.
Published owner v0.7.0-2, owner298314bd0ff4650363e0e63b81399d5618869a63/upstream620d2225cc436ab0b001e3200e590e3b286ef9ab, is unchanged and already consumed by public core0.16.4/main4509171d. Latest raw b11923 is de5122d879d00d2895c36be7e0cee94105aeeabd (+254); inspected masterfae9ac561f7c20b320d89103be67372381485653 is +256/+29 since yesterday.
Draft#125, head1dd46c9e3784697b0e39b7f2284850805951a348, selects b11912/961c376a2daa42220ec09663b9bde32c5fba1f70 (+243). It records independent acceptance, wrapper/export/provenance and seeded CPU TTS evidence; non-publishing artifact preparation was still running. This is not a published artifact or consumer adoption. Keep it draft until0.16.5 ships. New SVE execution coverage and affected-family/device qualification remain explicit; carry0003 was retained, so #114 retirement is unproved.
Material raw changes:
Post-b11912 OpenCL Q3_K/skipped-ID improvements and SCALE repair are not in the selected candidate; b11922 state-test corrections are not runtime repairs. Latestb11923 Vulkan/llvmpipe context tests abort and server cancellation fails; #31239 is an open repair proposal, not clearance.
DeepSeek olddraft#30817 closed unmerged/superseded by#31086; Laya#29363 closed unmerged. No new merged base family today. Prior raw-only families, CUDA/DSpark/EAGLE/state/security watches retain their existing gates. Next: finish candidate artifact/model/platform evidence and wrapper/binding/template/public capability review; only a qualified owner release can precede consumer pins/checksums/SwiftPM/docs.
5 remaining items
leehack commentedon Oct 31, 2026
2026-10-31: the owner/source adoption milestone is reached, but Apple release qualification has a confirmed P1. Retain this umbrella at P2; #154 owns the P1, not a duplicate.
Published nativev0.8.0 bundles upstream8642541b255bfb5dc39cd351a178be09013a55d6; core main77166b047a620755c3de9995ca00ce13130420b6 adopts it through #1611, regenerated bindings and the large-schema grammar repair, with Webv0.1.79/#1618. Public core0.17.0 remains v0.7.0-3/620d2225. Main is not public0.17.1; release#1645 remains held.
#154's exact-device evidence supersedes treating the old Metal compile report as only a raw watch: upstream#31146's constexpr BF16 kernels fail on affected pre-26 Apple OS compilers. Proposed #153 carries upstream#31360 as0004 for v0.8.0-1; the repair first appears raw b11983, not a stable release. It is not yet merged/published. Require immutable rebuilt artifacts, consumer/SwiftPM/checksum repin, exact-head re-audit and affected-OS device controls before release. Public0.17.0 is not upgraded or declared affected by this finding.
#149 fixes Linux LFM3 speech in source; #146 still needs a rebuilt artifact. #123/#130 now have bounded real-model evidence and successor tickets; core#1615/#1616 retain Dart-facing scope. No new family landed in today's twelve raw commits.
Raw latest b12020/master0ec4027a adds LFM3 decoder quantization#31386, reasoning-none grammar#31380 and SYCL copy#31120, absent stable0.8. New #31421 reports tiny SLRU expert-cache performance loss; core's opt-in moeCacheBytes makes it a candidate control, not a proven consumer regression. #31457's longstanding iSWA rollback report is not established core exposure: inspected620d/864 defaults use swa_full=true and core preserves that default. #31456 is an unverified longstanding GPTNeoX activation report.
Raw WebGPU context reallocation failures persist; server sanitizer now passes. DFlash/getter/Laguna and remaining hardware gates stay unchanged. No new carry retirement, model/device run, release or pin mutation occurred in this scan.
leehack commentedon Nov 1, 2026
2026-11-01: the Apple recovery is published and consumed; remaining raw qualification stays P2.
Nativev0.8.0-1, ownerf950319714bbc5e81e73b39c455dcdc91382f2d1, retains upstream8642541b255bfb5dc39cd351a178be09013a55d6 and carries0001/0002/0004. Release41189395508 passed required smoke; #153/#154's Metal repair and #149/#146's Linux LFM3 mask fix are in this artifact. Public core0.17.2, Apple companion0.17.1 and main26dcbcf5 adopt it; Web79 stays on the same upstream. The old release hold is superseded, not an instruction to broaden untested-OS claims. No further pin bump is needed for the latest owner artifact. #156 merged afterward and is not the provenance workflow used to publish this build.
Raw stable remainsv0.8.0; b12032/masterdde879bc4f954b809892fd0d660515f42536394a is72 commits ahead. New candidate fixes include Gemma4 resize#30922, shared-KV device rotation#31406, InternVL CLS order#30886, LFM short-conv batching#30831 and LFM3 tensor split#31431. They require affected-model/projector and native/direct/worker or multi-device controls; batching measurements are not single-request Dart speedups.
The only new architectures after8642541b are Limite/Violetto#29433 and Mistral Large4#31376, both raw-only. Qualify conversion/metadata/templates/reference output; Limite's fixed mathematical system/think template is not arbitrary system/tool-role support. FP8 E4M3 is a format addition. No new family or public-header change landed in today's twelve commits; earlier raw batched-copy APIs/F8 types still need adoption review.
New watches: #31472 speculative temp-zero drift has no complete repair; #31480 reports Vulkan/RPC flash-attention gibberish; #31495 custom-HIP MTP prefill loss; #31482 MTP/unified-KV probe abort. Wrapper call paths make focused controls relevant, but none is established as a new Dart regression. #31479's BACKEND_DL=OFF trigger does not match the owner's ON preset.
Fresh raw WebGPU and Vulkan tests fail; redirected logs were unavailable, so causes are not newly verified. Existing #79/#116/#150/#151 and other qualification gates remain. No new model/device execution or upstream write occurred.
leehack commentedon Nov 2, 2026
2026-11-02 candidate update — retain P2 qualification; no native repin.
Published owner remains v0.8.0-1 / f950319 on upstream 8642541b255bfb5dc39cd351a178be09013a55d6, now also selected by public core 0.17.5. Raw b12034 is 1c5ed702c9e540ac14cc8c78100ea3984bc0e227; master 4e910a86e8e8818e3619738ae372563a4e0cb42b is 75 commits beyond the pin. New merged changes are WebGPU scratch allocation #31422, explicit test thread linkage #31414, and dependencies. No new family or public-header delta landed today.
Watch changes:
Prior WebGPU failure is superseded by actual passing NVIDIA/Apple Test steps in 41414049204. Vulkan41414049136 has passing Apple/Intel rows but NVIDIA rows were queued; no blanket raw-CI clearance. #146 is now closed on published Linux payload/mask evidence; public speech wiring stays core #1615.
Limite/Violetto and Mistral Large4 remain the only new architectures after the pin, raw-only. Existing export/binding, template/conversion, model/device and carry gates remain; no new model run, artifact or pin mutation occurred.
leehack commentedon Nov 3, 2026
2026-11-03: newer owner artifact is published; raw model/ABI changes remain separate. Retain P2 qualification.
Native v0.8.0-2, owner 8ae5e3f, retains upstream 8642541b and carries 0001/0002/0004. Ownership/creation fixes and legacy-export retirement are in this qualified publication. Public core 0.17.5 and current main still pin -1; adoption is draft core#1735, not an automatic pin bump. Preserve export/old-consumer, settlement/ownership, Apple companion, model lifecycle and exact-head gates; the release remains a mutable prerelease (#151).
Raw b12066 is 713319e2; master740251a920abf92b7744d9c0518ca157db0db579 is 120 commits beyond stable. Clover#31306, first b12060, adds AR-only support, not self-speculation. Limite/Violetto, Mistral Large4 and Clover are the new architectures after this pin; conversion, metadata/templates and representative reference controls precede support claims.
PocketTTS#31429, first b12061, is an existing speech pipeline's predefined-voice/no-encoder variant, not another text architecture. Distinct owner capability/voice/lifecycle work is now #163. Its return-by-value audio-info layout grows 16→24 bytes on the inspected LP64 host; input voice storage and caller-owned voice-bitmap helpers change. Require platform ABI/wrapper/generated-binding review before future raw adoption.
Other raw corrections: Vulkan MoE cache#31114, Metal copy#31581, CUDA graph eviction#30662 and unified-KV#29510/#31570. No consumer speedup or local fix qualification is inferred. DFlash copies#31506 merged at b12063, but #31586 reports CUDA failure with DFlash/tensor split on two 3090s; this is post-pin candidate risk. Arc decision-output #31560 and throughput #31557 are reporter-only hardware watches. #31504/#31505/#31513 remain partial/open.
Retain 0001/0002. Stable still needs 0004; a future raw candidate containing #31360 should skip it through existing fix detection, not delete it blindly. 0003 is already absent from current artifacts; #145 owns old-row/patch cleanup. New Metal copy work does not establish #116 repair. No new model/device job or upstream write occurred.
leehack commentedon Nov 4, 2026
2026-11-04: latest owner adoption is merged; newer raw models/ABI/fixes remain separate P2 qualification work.
Core#1737 merged a2c429be13a0a0ec9af75e1ebd36ac249edc4926: current main29c19ec8 pins native v0.8.0-2 / upstream8642541b255bfb5dc39cd351a178be09013a55d6. Public core0.17.5 still selects -1 at the same upstream. The latest artifact is already source-adopted; another main pin bump is unnecessary. #1403's final 40-cycle Metal acceptance supports its source-delivered closure, not published-core or broad hardware credit. Release/Apple delivery stays core#1703.
Raw stable remains v0.8.0. Latest published nightly is b12102, bbb02c93e9b83255b70ac1a755bb02b0a251f9a7; current master13ec8f614cfef964fe97bbf872bc436dbc484a1f is149 commits beyond the bundled source. New general architecture Metis1.6#31134, firstb12089, joins post-pin Limite/Violetto, Mistral Large4 and Clover. All remain raw-only; metadata/conversion, state, templates/tools and representative validation precede public support.
New Nemotron3.5 Transcribe#31471, firstb12087, is ASR on the existing Nemotron family, not a new general LLM architecture. It needs offline/40-second-window/speaker-cache and thinking-off prompt validation; no public MTMD-header delta proves a new owner ABI requirement. Existing custom Dart speech prompt adapters can express the basic prompt/text path, but no built-in Nemotron adapter or qualification exists. Chatterbox#30993, firstb12097, has distinct voice/paired-sequence/one-shot-vocoder owner work now tracked by #164. Keep Pocket#163 separate. Neither speech family is in -2/public Dart.
Raw-only gains include Gemma4 unified-KV prefill#31356, firstb12096; MTP checkpoint#30141, firstb12091, whose final implementation restores pending_h only, not verify_h or demonstrated target-output parity; Q8_1 CUDA#31247 firstb12081; strided-view bound#31584 firstb12102. Master-only shared-KV device/FA#31635, e853ec9a, fixes #31614 with local CUDA/RPC controls, not actual two-GPU qualification. No Dart speed/session-fix claim follows.
Watch correction: DFlash#31586 was explained as extra VRAM demand/OOM, not fixed by a code commit. #31560 stays open despite a second tester's non-reproduction; no Arc repair established. Existing backend/state/quantized-KV/temporary-zero drift and ABI/immutability tickets remain. Retain carries0001/0002; stable still needs0004, while future fixed raw auto-skips it via existing detection;0003 already absent, cleanup stays#145.
New upstream release41807089651 failed pushing b12107 because GitHub returned Internal Server Error: one failed tag job,19 downstream builds skipped, not a compile or zero-job failure. Owner CI/provenance are green; its stable autodetector correctly skips an already-built stable release. Future raw adoption still requires wrapper/header/export audit, generated bindings where changed, matching Web/artifacts/Apple metadata and model/device/lifecycle validation. No scanner source, workflow, device/model or pin change.
leehack commentedon Nov 5, 2026
2026-11-05: new raw stable v0.9.0 is published; latest owner/main remain on qualified0.8. Retain P2 owner qualification, without adding0.9 automatically to core0.18 scope.
Stable v0.9.0, cd959bae99ff4bed22bef2604f45241c09bbb8c8 / nightly-tag b12110, is150 commits beyond bundled8642541b. It contains Limite/Violetto, Mistral Large4, Clover AR, Metis1.6, Chatterbox, Nemotron3.5 ASR and predefined Pocket voices plus Gemma4 prefill/shared-context, MTP pending-state and DFlash device-copy changes. These remain absent owner v0.8.0-2; main1f341905 pins -2, public core0.17.5 pins -1 at the same8642541b. No additional -2 pin implementation is needed.
A0.9 candidate is not pin-only: audit F8 types, batched tensor/FlashAttention rows/communicator APIs, layer-input/embedding-row extensions and MTMD return-by-value audio-info growth, voice/Chatterbox/state-reference APIs and helper-input layout; regenerate affected bindings and qualify wrapper/old-runtime/model/state/lifecycle paths plus matching Web/Apple/artifact metadata. Existing0004 fix detection matches0.9 and automatically skips that carry for the candidate; keep it for0.8. Retain0001/0002 and#145's0003 cleanup.
New post-stable Drama3#31613, firstb12116, has distinct sampling/speaker/codebook/stream lifetime work now #165. Chatterbox#164 now includes raw#31686 Unicode preprocessing; Pocket#163 remains separate. No new public speech support is inferred.
Tracked quantized ON_DEVICE#30541 merged e56f006ed6b416ae8b7ccb5243b204743f09df1e, firstb12121, after stable0.9 and absent owner/public. Main's inspected checkpoints use PARTIAL_ONLY, so no ordinary Dart session exposure/repair claim follows. ON_DEVICE#30318 remains open. Other post-stable candidates include WebGPU#31348, Metal fusion boundaries#31688 and recurrent-state copies#31660; upstream measurements are not Dart speed gains.
Latest rawb12125 is7d516e25ead4653169236b3f90d13f7bd8629724; main9c9074e14e029a16fc3298e6bc6a3ef00a0b9c11 is179 beyond pin. DFlash2 multi-slot#31656 and existing Arc/temp-zero/replay controls remain open; candidate reports on unmerged#31499 do not fix#30698.
Automation recovered: later b12109 first publishes the previously master-only shared-context fix; b12107 never published. The first stable release attempt41826623079 lacked a successful nightly; 41832787898 subsequently published0.9 successfully. Owner detector41818939735 ran before0.9 and correctly skipped already-built0.8; that is not a failed0.9 update. No workflow/pin/publication/model/device execution by this scan.
leehack commentedon Nov 6, 2026
2026-11-06: latest qualified wrapper is -3, but main/public and raw0.9 delivery remain separate.
Published v0.8.0-3, source703c0f51fe9f8ebfa941c70504ad0b13c7ca79cb, still bundles8642541b255bfb5dc39cd351a178be09013a55d6 with0001/0002/0004 and hook contract2. It removes the final constant-false need_embd query while retaining need_embd_nextn; export gate92required/15retired. No new family/performance feature. Main/local2e784dac still pin -2; public core0.17.5 pins -1. Draft core#1746,0f9e2c75, already prepares -3 with authentic bindings/checksums/SwiftPM and completed scoped ordinary JIT/AOT/Apple/Metal acceptance. Do not use -3 as an unreviewed old-consumer override. Core/Apple publication remains core#1703; no further -3 implementation is identified.
Correct yesterday's automation snapshot: autodetector42016623730 successfully dispatched0.9. Build42016701448 completed compilation but refused Linuxarm64 Vulkan and Androidarm64 Vulkan/OpenCL ISA audits for the unaudited ggml/KleidiAI fingerprint pair. Smoke/package/publication skipped; no0.9 owner artifact. Next: audit exact0.9 dispatch/ISA changes and real-artifact containment before any allowlist update—never refresh fingerprints blindly. This is fail-closed qualification, not a compiler/network/zero-job failure. Separately, mirror-timeout#169 is repaired by#173; do not keep reporting that old stall.
Raw stable remainsv0.9.0/cd959bae. Latestb12167 is240d68fcfe0c2a1f298c8e18ca181368b1914259; main18ab62a8ec5144ae5d2f65e78374ba69652d75dd is214 beyond bundled source. New raw-only families xing4_0#31589,45b6604b, and Ripple#31692,fecdc81e with immediate build repair#31749, first appear in b12167. Existing ops/converter/metadata/templates/reference-output and MTP/state/backend qualification fit this umbrella; no distinct new Dart API requirement demonstrated. Earlier four post-pin architectures and speech#163/#164/#165 retain their own limits.
New focused ARM FP16#174 is P2 artifact/ISA numerical qualification prompted by raw#31734, not proven owner exposure or a new0.18 blocker. Chatterbox#164 now records merged classifier-free-guidance metadata/query; Unicode#31686 remains open.
Other raw gates: YaRN#30206, firstb12140 adds ggml_rope_set_truncate and changes public/internal RoPE helper declarations; include in future ABI/binding audit. KV shifting#31683 and server checkpoints#30137 need separate model/state versus server-only scope. Vulkan FA scratch#30487 fixes upstream#30048 firstb12142; ReBAR#31658 fixes upstream#27097 only afterb12167. No Dart session/speed or actual hardware claim follows. Existing long Vulkan/RPC#31505, Arc, temp-zero and ON_DEVICE watches remain unresolved;#30541 remains raw-only fixed. Public auth hardening is not evidence of Dart downloader exposure; no security disclosure made.
Retain owner-before-consumer, exact ABI/export/model/lifecycle, matching Web/assets, Apple and device gates before0.9/raw adoption. No workflow, source, pin, model/device or publication change by this scan.
leehack commentedon Nov 7, 2026
2026-11-07: v0.9 qualification has progressed; it is still source-only and held, not a published runtime upgrade.
Draft #175, head 7f15b5d, now passes 89 hosted checks after auditing the exact ggml/KleidiAI pair and ISA/dispatch containment. This supersedes yesterday's unprepared fingerprint blocker; the original refusal remains valid fail-closed evidence. Independent exact-head adversarial review remains pending, and the candidate is explicitly held until core0.18 publication. Carries0001/0002 still apply;0003/0004 are detected fixed for0.9, retained for older rows.
Correct yesterday's exposure wording: #174 now confirms synthetic FP16 overflow in shipped v0.8.0-3 Mac ARM64 and optimized Linux ARM64 modules. Draft #177 prepares widening accumulation; independent review, patch-policy exception and corrected-artifact qualification remain. Public-model impact and introduction causality are unmeasured, not a proven new core regression.
Raw stable remains v0.9.0/cd959bae; latest b12184 is c0ab6e5c0bab3fb21e6cb9d2371f3c41d85ea775; inspected main cf659f240b232e15700b1721bbb43c1d973c52e9 is227 commits beyond bundled8642541b. New Kimi-K3.1-Pro text derivative #31779, bcb0152876e8f0b684ba70d3defcab6cdf35adb5, firstb12180, adds kimi_k31 loader/conversion and sparse-indexer/full-MLA hybrid; vision is skipped and upstream tests an8-layer slice, not full-model/device qualification. No new Dart API demonstrated: retain converter/metadata/template/tool/reference/model/backend controls here rather than create an empty family ticket.
DeepSeek-V4.1 tensor-split #31447, fd27b2034905f8dadd8bd6d2cc44595f4929d391, firstb12182, increases meta views32→256 and repairs an upstream first-decode assertion. It remains absent owner/public; upstream multi-GPU evidence does not qualify our wrapper. Daily delta has no new public-header requirement. Existing MTP/DFlash/EAGLE, Arc, Vulkan/RPC, temp-zero and ON_DEVICE watches remain separate.
Published owner remains v0.8.0-3/703c0f51/upstream8642541b; local/main c6010608 pins -2, public0.17.5 pins -1. Core#1746, d14513ae, already prepares -3; no additional -3 pin implementation is identified. New speech preparation is #176, with distinct #163/#164 gates. No workflow/source/pin/model/device/publication changes by this scan.
leehack commentedon Nov 8, 2026
2026-11-08: owner0.9 is published and adopted on main; yesterday's source-only/hold snapshot is superseded.
Publication42433633399 produced v0.9.0, ownera604639131105c42a25a3e7384ca10314d385f80/upstreamcd959bae99ff4bed22bef2604f45241c09bbb8c8,14 artifacts and required/passed smoke. Carries0001/0002/0005/0006 remain;0003/0004 are retired for this baseline, retained on older rows. #174's accepted publication includes corrected dot/GEMM, strict tests/ISA controls and explicitly accepted bounded throughput costs; do not repeat it as an unresolved unpublished repair.
Core#1784 merged185b56f6947a98940783db0c103c5760ee83deb5 with regenerated bindings, checksums/SwiftPM and independent exact-head acceptance. Public0.18.0 still pins0.8.0-3/8642541b; local clean5305fd81 also retains0.8.0-3. No further0.9 pin implementation is needed on main, but core/all-companion0.19 publication and its focused regression gates remain core#1786. Web assets82 still bundle8642541b: intentional checked skew, not synchronized0.9 support.
Raw stable remains0.9/cd959bae. Latest b12202 isd6be065572cdb7bf64014c8c5771e210c9570e44; inspected maina691b2345844ce1b430201fc5fd404ce062e05f2 is95 beyond owner. New raw-only MageVL-Cowork#31819,9f0e5098/firstb12192, needs converter/projector/RoPE/image/reference/precision controls; no distinct new Dart API demonstrated. Kimi, xing4_0, Ripple and Drama3 remain post-stable; stable0.9 architecture additions are artifact-consumable, not automatically model-qualified/public0.18 support.
Future raw is NOT pin-only: #30966,e95426f1/firstb12191, inserts moe_cache_cpu_share into llama_context_params, shifting later fields. Audit owner ABI, regenerate/version-gate bindings, preserve default-zero and target/draft isolation (#1123). CPU F32-src1#31854,12ab5979/main-only, changes dot/GEMM traits/source fingerprints: reassess0005/0006 with original numerical/ISA/performance controls, not blind deletion. BF16 policy#31851,firstb12202, needs affected-family/backend precision checks. Log-mel#31808,firstb12193, changes CPU preprocessing/concurrency. PEG UTF8#31861 firstb12199 is upstream common-parser work, not established exposure of Dart's own parser.
Draft-cache#31497,firstb12201, fixes the example; our published wrapper already clears draft cache (#113), so no new Dart fix credit. EAGLE#29313 and MTP/RPC#28484 closed stale/not_planned, not fixed. Arc#31560 now has reporter faulty-card suspicion, not general regression proof. Existing ON_DEVICE/MTP/DFlash/Unicode and device watches retain their acceptance. No source, workflow/model/device or publication execution by this scan.
leehack commentedon Nov 9, 2026
2026-11-09: native v0.9.0 is now public through core 0.19.0, including guarded LFM3/Soprano and Chatterbox/Pocket adapters, voices and version-skew refusal. Local checkout ec4bf4f89dc224296ec0c8e6201ff13a3339e51e is clean and uses the same artifact; GitHub main moved to1fdc222f through the concurrent0.19.1 batch merge, without changing the llama.cpp pin. Latest owner remains v0.9.0, a604639 / upstream cd959bae99ff4bed22bef2604f45241c09bbb8c8, carries 0001/0002/0005/0006, 94 required wrapper symbols. No additional stable pin implementation is needed. Published speech evidence remains a bounded native subset, not all-platform, broad-language, Web-family or Pocket encoder-cloning qualification.
Raw stable is unchanged. Latest b12212 is d78f4255d4d27ebef8dfd27f58f9ee3c8e4b7afb; inspected main a0ad695cb212db3850f163121ddc687513448d42 is 105 commits beyond the owner (+10 since yesterday).
Material raw-only gates:
Owner-main #185, b998594, retires source patch0003 and the v0.7 CI row only. Published bytes are unchanged; patch0004/v0.8 row remain. #186 already owns release immutability; v0.9 remains mutable with consumer digest checks, not enforced immutable.
No new family today: MageVL, Kimi, xing4_0, Ripple and Drama3 remain post-stable raw gates. Ripple DFlash2 #31904 is open, not delivered. Precision #31903 closed unmerged as duplicate of open #31896; qwen4exp MTP #31914 duplicates open #30860, not a fix. Other ON_DEVICE/MTP/Unicode and greedy-parity watches are unchanged. No source, pin, setting, workflow or model/device execution by this scan.
leehack commentedon Nov 10, 2026
2026-11-10: stable/owner/public pins are unchanged. Public core0.19.3 and clean main/local77b01da1f6db7cde751b618700574b63f3eb74a9 consume nativev0.9.0, a604639 / upstreamcd959bae99ff4bed22bef2604f45241c09bbb8c8, 14 artifacts,94 wrapper symbols,carries0001/0002/0005/0006. No additional stable pin implementation. Owner mainb9985944 still has only the recorded source-row retirement, not new published bytes.
Latest rawb12244 is d5bbcd1f31dba754f716a26c3a4784502a6b9ede; inspected main0d9eacadf2e19ff5d49df29800afc88a8e844123 is148 beyond owner (+43 today). New candidate gates:
Fix reconciliation: tools+response-format#31956,3b4d6ab3/firstb12242, fixes#31342; resampler#31964 fixes#31696; MetalFA#31662 firstb12221 and MoltenVK#31890 main-only (fixes#31892) remain absent owner/public. Raw allOf#31834 and tool-string#31937 do not automatically repair Dart's independent converter/parser: current non-object/non-enum allOf is explicitly refused. #29410 was reporter-retracted after proper A/B, and#31189 was conversion-dependency compatibility, not new runtime fixes.
No newly merged architecture today. DSpine#31941, DSpark#31372 and RippleDFlash2#31904 remain unmerged; preserve prior raw-family qualification and existing CPU/Unicode/ON_DEVICE/MTP watches. No carry retirement, source/pin/workflow/model/device or publication changes by this scan.
leehack commentedon Nov 11, 2026
2026-11-11: public 0.19.5 and clean main/local
4ecd129aretain native v0.9.0: ownera604639131105c42a25a3e7384ca10314d385f80, upstreamcd959bae99ff4bed22bef2604f45241c09bbb8c8, 14 artifacts/94 required exports, carries 0001/0002/0005/0006. Raw stable is unchanged; no additional stable repin is needed.Latest b12283 is
f3b407ca1bfeeecebf9e9ef19fcd6eed565bd70a; inspected main eeb21047 is 180 commits beyond owner (+32 daily).Material raw-only gates:
Watch corrections:
Prior context/helper ABI, audio-temperature default, stateful audio, source-fingerprint and old-runtime gates remain. Select and qualify an owner candidate, regenerate affected bindings, then checksums/SwiftPM/bridge and public mapping/docs; this is not an automatic raw pin bump. Existing P2/enhancement unchanged. No source/pin/workflow/model/device execution by this scan.
Full current-pin inventory (cd959bae → eeb21047): new base text architectures are xing4_0 #31589 and Ripple/ricursive #31692, both first verified published nightly b12167 (release inventory gap, do not invent intermediate publications). Kimi K3.1 text #31779, first b12180, is a K3 derivative/new loader identifier; vision extends it in b12259. MageVL-Cowork #31819, b12192, adds a LlavaOnevision2 projector over existing Qwen text. Drama3/FishOmni #31613, b12116, adds speech/codebook/FishDAC over Qwen3.5-MoE, NOT a new base text architecture; guarded owner/public adapter work stays #165. All remain absent owner0.9/public0.19.5. Generic text/image APIs may suffice for the first groups after conversion/template/reference/device qualification; no distinct new API is demonstrated. Pocket/LFM3/Chatterbox/Gemma audio, PLE/sharded components, NVFP4 scales and FP8 KV are existing-family component/configuration changes, not additional families. Ripple DFlash2, DSpine and Spark DSpark remain unmerged.
leehack commentedon Nov 12, 2026
2026-11-12: public0.19.6 and clean main/local
fb7ca7d5retain native v0.9.0, ownera604639131105c42a25a3e7384ca10314d385f80/ upstreamcd959bae99ff4bed22bef2604f45241c09bbb8c8, 14 artifacts, 94 required exports and carries0001/0002/0005/0006. Owner main remainsb9985944, no open PR/new failed or stalled owner run. Manifest digest0d2119cdcf34d347a96a996423094498de3393026efb6ad1a96ba286b621c39e; no additional stable repin is ready.Raw stable is unchanged. Latest b12317 is
e0949d09b130707b3fa93d2cff616024231cf939; inspected main0d69255c is219 commits over owner (+39 daily).New focused owner: #187 tracks CPU Flash Attention F32-V repair#29810, first verified published b12302, closing upstream29774. Exact bundled ops.cpp retains the affected F16 accumulator; existing closed ARM dot/GEMM#174 and carries0005/0006 do NOT repair it. Source exposure is confirmed; published-module/Dart reproduction is unrun. P2/bug acceptance requires exact provider/ISA, long-context positive and FA-off/F32-V/short-context controls, fail-before/pass-after, numerical/performance and corrected-artifact/consumer gates. Do not reopen174 or infer PR causality.
Other raw-only adoption gates:
No new merged family today: post-pin xing4_0/Ripple, KimiK3.1 text+vision derivative, MageVL projector and Drama3 speech remain absent owner0.9/public0.19.6. Zaya1#32071, DSpine, Spark DSpark and Ripple DFlash2 remain unmerged; old closed predecessors are not delivery. Embedding32036/32044 remain open; ARM31734/MTP31077 dispositions are unchanged. #1071 source is consumable but model-pair qualification remains; #1072 closed API delivery delegates remaining platform/media gates to#1189.
No daily public-header delta; accumulated context/audio/helper ABI, audio-default, carry, binding/version-skew, bridge/checksum/SwiftPM and actual-model/device gates remain. Keep P2/enhancement; this scan performed no source/pin/workflow/model/device execution.
leehack commentedon Nov 13, 2026
2026-11-13: current public0.19.8/local
768feb72/GitHub maina2a5cf29retain native v0.9.0: ownera604639131105c42a25a3e7384ca10314d385f80, upstreamcd959bae99ff4bed22bef2604f45241c09bbb8c8, 14 artifacts/94 required exports/carries0001/0002/0005/0006. Owner mainb9985944has no open PR/new failed or stalled release automation. The active sibling checkout differs; these facts use GitHub/exact release objects, not its modified submodule.Raw latest/main b12360,
2d8a32502909ca065c41c4cf5f2fb1c427c047fc, is250 commits over owner (+31 daily). Stable release remains0.9.0; #32190's source1.0.0 bump is NOT a published stable owner/1.0 release.New migration gate, not an automatic pin bump:
Raw-only gains: ARM NEON FA4x4#32128 and Intel Vulkan FA prefill#29357, first b12336; CUDA few-row MMQ#31533, b12342, relevant to speculative verification; Qwen3.5 conversion merge#32102, b12343; Hexagon Q5#32094, b12345. Require actual ISA/shape/checkpoint/output, warmed target/draft acceptance and throughput controls—upstream measurements are not Dart speed claims.
Watch changes: K2 Horizon closer#30335, first b12360, causally closes30246; unclosed tool-tag cases remain outside scope and Dart parsing is independent. Public MTMD hardening32129–32132 is now merged/first b12336, not yesterday's open state; no owner artifact contains it. Keep downstream security-impact/disclosure review separately approval-gated; no new vulnerability details are recorded here. Server29689 is fixed by32133/b12340, not an in-process Dart repair.
CPU-FA#187 and embedding32036/32044 remain open without new published-owner reproduction/candidate. Raw sampler32141, CUDA32139/unmerged32177 and HIP32178 are new qualification watches, not established owner/Dart regressions. Backend sampling is used by Dart speculation, but affected configurations must be demonstrated. Lavapipe30563 closes not_planned/Mesa attribution; multimodal DFlash27408 closes stale—not fixes.
No new merged family today; xing4_0/Ripple, KimiK3.1 derivative, MageVL projector and Drama3 speech remain the post-pin raw-only inventory. Zaya1/Inkling/DSpine/Spark DSpark/Ripple DFlash2 remain unmerged. #1071 still needs pair qualification despite consumable source; #1072 delegates remaining media/platform evidence to#1189. Existing P2/enhancement/acceptance unchanged; no source/pin/workflow/device execution by this scan.
leehack commentedon Nov 14, 2026
2026-11-14: raw stable v1.0.0 is now published (Nov13 10:20 UTC), source
2d8a32502909ca065c41c4cf5f2fb1c427c047fc. Latest nightly/main b12377,32bdc3a89e6528d3e419024c833c5dabf32c7541, is267 commits over owner/+17 daily.Latest owner0.9.0, a604639/upstreamcd959bae, and publiccore0.19.9/main-local6b18ff51 remain unchanged:14 artifacts/94 exports/carries0001/0002/0005/0006, manifest0d2119cdcf34d347a96a996423094498de3393026efb6ad1a96ba286b621c39e. Owner mainb9985944 has no open PR/new failed or stalled run. Its auto_native_release workflow is intentionally workflow_dispatch-only via#178; no automatic1.0 build is expected.
New raw-only migration gate: DFlash refactor#32016, firstb12377 (AFTER stable1.0), adds draft-dflash2/draft-lilicorr plus llama_dflash_variant/query. Generic draft-dflash retains compatibility, but owner draft-context/count/admission helpers hardcode old types. New enum members shift n-gram ordinals: audit legacy type_mask; current Dart uses type_names/mask0, so no current-pin ordinal regression is established. Update helpers/validation, optional exports and enum/query bindings/probes before alias admission. Core#1071 retains LiliCorr routing/model-pair controls; raw now rejects draft-dspark+LiliCorr. No new public strategy is inherently needed to retain generic routing.
Other first-containing raw gains: grammar left factoring#29557,b12362 (recursive/large grammar, immutable-rule ownership, tools/schema/reasoning controls); default-off Vulkan keepalive#29267,b12363 (not a proven eviction fix); draft-p-min#31294,b12364 (WARNING ONLY, behavior unchanged); Vulkan TOP_K_MASK#31544,b12365; BF16 RMS_NORM#32135,b12366; Metal offset-overflow#32221,b12374 causally closes30237; Mistral workaround removal#32230,b12375. Require applicable backend/ISA/shape/checkpoint/output/state/performance controls; independent Dart parsing does not inherit raw template repairs.
Stable1.0 now contains CPU-FA#29810 and public MTMD hardening, but owner/public do not; #187 records first stable containment without new binary reproduction or closure. Preserve private security-impact/disclosure review. Prior1.0 header/API/ggml pruning, RPC matching, carry/export/layout/binding/version-skew gates remain; no automatic pin-only update.
Open raw watches include V10032245 (1.0-vs0.8 report, not0.9 causal proof), Vulkan32216/unmerged32244, Pascal32229, AMX32197 and Hexagon32198. No owner/Dart exposure or PR causality is established by those reports; prior embedding/sampler/CUDA/HIP gates remain.
No new merged family today: post-pin xing/Ripple/KimiK3.1/MageVL/Drama3 remain raw-only, while new draft aliases/Hexagon formats are capabilities, not families. Zaya/Inkling/DSpine/Spark/Ripple DFlash2 proposals remain unmerged. P2/enhancement unchanged; no source/pin/workflow/model/device execution by this scan.