pin v2 payload: engine 2611011701001, cuda 2610311711001, cuda12 2610311711002, sbsa 2610311711003, vulkan 2611011644001, macos 2610311711005, media 2611011351002

.gitattributes

@@ -337,3 +337,7 @@ components/media/2611011351002/oc-espeak-linux-x86_64.so filter=lfs diff=lfs mer
 components/media/2611011351002/oc-espeak-linux-aarch64.so filter=lfs diff=lfs merge=lfs -text
 components/media/2611011351002/oc-espeak-win-x86_64.dll filter=lfs diff=lfs merge=lfs -text
 components/media/2611011351002/oc-espeak-macos-aarch64.dylib filter=lfs diff=lfs merge=lfs -text
+components/engine/2611011701001/opencoti-0.10.5-c14-2611011701001 filter=lfs diff=lfs merge=lfs -text
+components/engine/2611011701001/opencoti-0.10.5-c14-2611011701001.exe filter=lfs diff=lfs merge=lfs -text
+components/vulkan/2611011644001/ggml-vulkan-x86_64.so filter=lfs diff=lfs merge=lfs -text
+components/vulkan/2611011644001/ggml-vulkan-win-x86_64.dll filter=lfs diff=lfs merge=lfs -text

components/cuda/2610311711001/BUILD_INFO.cuda.md

@@ -2,7 +2,7 @@
 
 Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
 
-Built from the engine source of build 2610312243001; gated floor (engine-min) 2610312243001; gated with 2610312243001, 2611010802001, 2611011335001, 2611011211001.
+Built from the engine source of build 2610312243001; gated floor (engine-min) 2610312243001; gated with 2610312243001, 2611010802001, 2611011335001, 2611011619001, 2611011701001.
 
 | file | platform | kind | sha256 | bytes |
 |---|---|---|---|---|

components/cuda12/2610311711002/BUILD_INFO.cuda12.md

@@ -2,7 +2,7 @@
 
 Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
 
-Built from the engine source of build 2610312243001; gated floor (engine-min) 2610312243001; gated with 2611010802001, 2611011335001, 2611011211001.
+Built from the engine source of build 2610312243001; gated floor (engine-min) 2610312243001; gated with 2611010802001, 2611011335001, 2611011619001, 2611011701001.
 
 | file | platform | kind | sha256 | bytes |
 |---|---|---|---|---|

components/engine/2611011701001/BUILD_INFO.engine.md

@@ -0,0 +1,49 @@
+# engine 2611011701001
+
+Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
+
+| file | platform | kind | sha256 | bytes |
+|---|---|---|---|---|
+| `opencoti-0.10.5-c14-2611011701001` | any | bin | `63042e6ef20b9b3819e2275e6d73d023b31eb1ff69e419342f6b670728282be6` | 146932862 |
+| `opencoti-0.10.5-c14-2611011701001.exe` | win-x86_64 | bin | `63042e6ef20b9b3819e2275e6d73d023b31eb1ff69e419342f6b670728282be6` | 146932862 |
+
+ABI `ggml` = `bac5050ddc8c58c51f0b059a7f9753b33af0ed79d355090e7c9a17ffaaaf340a` (exported) — sha256 of this list:
+
+```
+llama.cpp/ggml/include/ggml-alloc.h 94e4cd069b9313b2ceb35dacec901981e0bb478d8bb31035b7126be091998c23
+llama.cpp/ggml/include/ggml-backend.h 764d1640ad62579513c4837539708f89e8603c9dbeae053a13d1cbd84a82c865
+llama.cpp/ggml/include/ggml-cuda.h 674ea064021cf5b9970e78f1209e513dd23350b94a44a705a27eec1423fe4b0a
+llama.cpp/ggml/include/ggml-metal.h 322f36cd30f3e9e7aad7b5b9bc63078012fd0b7706ac4e24f5721b604f3d8980
+llama.cpp/ggml/include/ggml-neo-pipeline.h 190f39e183542d16588a3104d8cb4359bf52fc7a62afe8d8cc4c822f4085086d
+llama.cpp/ggml/include/ggml-oc-kvarn-range.h 76635fdd8978632f670640ddf5125a666441e3d47c680bb8b5c0a099e4861202
+llama.cpp/ggml/include/ggml-oc-ledger.h c1be11d69a57a1e721f1ae8720b87887d239ac0183d4277fbaa07f1be257d20c
+llama.cpp/ggml/include/ggml-vulkan.h f9df4f7161a84ce7600d53c3fac8a62837d5716676c9f38cd340699ef1c070d9
+llama.cpp/ggml/include/ggml.h ecfbf114f9e9151d22e30ef6d5139570c5c41b595843ca4f99125bd0196709aa
+llama.cpp/ggml/include/opencoti-devtype.h 94dea946173449bb76bffd7f733ab822fe658fe7198992de750ac012d29308ec
+llama.cpp/ggml/src/ggml-backend-impl.h f9fdfe6bae61d1387593b40fc0f90ceee490c6a339ec6d7388c51e4fca318d8f
+llamafile/gpu_backend.h 4929c36c034b6589e88653bcb7a5c5a225743a07320f64c0857d92bdaf048208
+```
+
+ABI `media` = `b099409b3029a6e380268e31d1caefa6706569961d32e72b3543fbc73903d95c` (exported) — sha256 of this list:
+
+```
+llamafile/oc-audiocpp/audiocpp.h 6ad0b55d5f28adc9ad60a45bf9a7caa428c4f5441389a6932d6670a8e2a90d73
+llamafile/oc-codec/oc_codec_sidecar.h f2c9001d1414aa2a55533b867b79c52a52c682558e0234c5bebd30bf97b91005
+```
+
+ABI `loader` = `d6ff34022f1d3fb5af56ca0c80e3f4e552bf09f1bebdf0cb5f314a454bf38dbc` (exported) — sha256 of this list:
+
+```
+.cosmocc/4.0.2/bin/ape-m1.c 78af79e20abbb3f99a97355959b59bd58949564064adb754f0302ebabb4bf7a3
+```
+
+**c14 dev snapshot — the development build of the c14 release, for integration testing.** Engine 2611011701001 (`opencoti-0.10.5-c14`, dev channel) is built from the same source as the c14 release engine 2611011619001 (the 12336 source files the two trees share are identical by sha256 (0 differ; clean-room /mnt/sdc/opencoti-cleanroom-c14r4b at 41ab985d3d)). The libraries are the c14 release index's. This engine ran the glibc 2.28 route gate (Vulkan + CUDA) on bs2 with the release libraries (cuda, cuda12, vulkan); the release gates ran on 2611011619001 (see the release notes).
+
+**c14 r4 (2026-11-01) — a re-cut of c14.** Same tag, new engine 2611011619001 (r3: 2611011335001, r2: 2611010802001, r1: 2610312243001; bare `0f22e977cd6c`, r3 `2261d16fef14`): `release/c14-r4` = the c14 r3 source + `0826-c14-sfx-media-only` (bug-4319, roadmap 57.13 / op-320, xollama #2072) + `0827-c14-vk-mmv-one-col-scalar-acc` (roadmap 57.12 / op-319; shaders only — it changes the Vulkan libraries, not the engine's behaviour): `--sfx-model` joins the launcher's media-only list in `llamafile/main.cpp`, so a `--server --sfx-model X` without `-m` boots (on r3 it exited "missing required -m MODEL.gguf"). One line in the launcher; both Vulkan libraries rebuilt (0827), every other library r3's bytes; the feature rows are r3's 78. Gated: the sound-effects gate with the two media-only arms (bs2 GPU1, r3 = the failing control), 40 control / glibc / bonsai ident, 42 / 43 / 44, 55 (12a-g).
+
+**c14 r3 (2026-11-01) — a re-cut of c14.** Same tag, new engine 2611011335001 (r2: 2611010802001, r1: 2610312243001; bare `2261d16fef14`, r2 `9f4bfaf682a2`): `release/c14-r3` = the c14 r2 source + `0824-c14-sfx-route` (roadmap 57.10 / op-317: `--sfx-model`, `POST /v1/audio/generations` mp3 | wav | pcm | json through the `oc-audiocpp` sidecar's `stable_audio` family, feature `audio_generations_v1`, `/props media.sfx`, one clip at a time with 503 + Retry-After, 501 without a model; bug-4316: the sidecar thread's CPU mask widened around load and run — a 2 s clip 170 s → 4–5 s on bs2, speech gains the same) + `0825-c14-vk-mmv-one-col` (roadmap 57.11 / op-318, bug-4315: the host side of the one-column integer-dot mat-vec pipelines — the engine's part is the pipeline selection in `ggml-vulkan.cpp`, shipped inside the Vulkan libraries; the bare carries no Vulkan code). Chain `0001`–`0825`. Built in the clean-room from `release/c14-r3` (85dab0deae) with the id minted before the build (the first clean-room run of the day went without it and said `build unknown` — bug-4318; the lane now refuses). Feature rows: r2's 77 plus `audio_generations_v1` (78). Owner 2026-11-01 12:55: "cut now the c14 re-release with it, publish on hf dev and hf pub". Gates: the scoped re-cut set + the two reproducers (the sound-effects route gate on bs2 GPU1; the Windows gate's `vulkan2-igpu` arm on eleven2go against the c13 r6 library).
+
+**c14 r2 (2026-11-01) — a re-cut of c14.** Same tag, new engine 2611010802001 (r1: 2610312243001; bare `9f4bfaf682a2`, r1 `72d949a3cbdd`): `release/c14-r2` = the c14 tag + `0823-c14-cu12-win-published-name` (bug-4313) — on Windows the CUDA 12 legacy library is now found beside the engine under its published name `ggml-cuda-cu12-win-x86_64.dll` (r1 searched `ggml-cuda-cu12.dll` and the Linux spelling only, so a consumer staging the published names got no CUDA below cc 7.5 and the CUDA 13 library silently above). Every library is the c14 bytes. Scoped gates on the r2 bytes with r1 as the control: the Windows gate's cuda12 arm under the published names (r1 FAIL, r2 PASS), control, old-glibc, library identity, split-form boots (bs2, eleven2go, Mac mini), the staged-bytes gate.
+
+**c14 (2026-10-31) — a new cut.** `--version`: `llamafile v0.10.5 (opencoti-0.10.5-c14) build 2610312243001`; chain 0001–0822 (c13 r6 ended at 0717). Built in the clean-room from `release/c14`. What changes: `GET /polykv/pools` (and `/kv`, `/elastic`) shows a pool op at once — create, fork, release, pin, admission force the read snapshot (0822, bug-4304; the list could stay stale on an idle engine since c13); no admission 429 without PolyKV (0718, bug-4140); the PolyKV gate books what it admits and refuses from a measured law (0721, 0724); the admission law `t = W(total rows) + k D` with a composition-matched k and one accounting for the PolyKV and elastic gates (0773, 0777, 0812, 0818); prefill and decode steps alternate by default so one prompt never stalls the decoders for its whole length (`--pd-step-budget alternate`, 0722, 0815) with the opt-in bounded prompt step `--pd-prefill-step-ratio` (0819); a session's step reads only its own live cells on CUDA, Vulkan and Metal (0725–0731, 0762), 256-cell KV pages by default (0770); PolyKV decode reads each shared block once for all its sessions (0764–0769, −23 % at 16 sessions); the KQ mask built on the device in row chunks (0785–0789); run-time KV commits that never fail (0813) and the Windows commit guard (0814); the lazy / split pinned KVarN host arena (0756–0758); `--moe-strategy` (0739, 0740); the drafter changes (0754, 0772, 0791–0794, 0806–0808: `--auto-mtp-policy adaptive`, opt-in); the CUDA 12 legacy refusal (0792, bug-4247); LFM3 and LFM3-VL (0817); the outer llamafile 0.10.6 commits (0816); the xollama #26 fixes (0738). Features: the c13 r6 set, no new names. Gated on these bytes (bare 72d949a3cbdd): the old-glibc run gate (glibc 2.28 image, Vulkan + CUDA, every read route twice) on the side dir and again on the staged bytes; the composition smoke on the RTX 3090 (14/14, 17/17, 17/17, with the 0822 plain-pool-list check); split-form boots with an empty HOME on bs2 (CUDA, Vulkan, the forced cuda12 arm), on eleven2go (RTX 3090 CUDA and Vulkan, RX 9070 XT Vulkan: 66/66 layers on the device, needle found, 68.9 / 62.8 / 72.0 tok/s) and on the Mac mini (Metal, signed loader); the slot-lazy gate (bug-4061) CUDA 5/5, CPU 6/6, Vulkan 5/6 with the +22 MiB at default slots accepted by the owner (bug-4308); the Bonsai 2 Vulkan gate (needle 3/3 at 112k, speed, KLD 0.000305, parity at knife edges only, old-vs-new library ident); the RADV iGPU rig 4/4. The table is in RELEASE_NOTES.md.
+Feature rows: c13 r6's 73 plus four pin-only rows for xollama's capability gate (mails #2041/#2042, 2026-10-31): `arch_lfm3_v1` (0817), `vulkan_matvec_tune_v1` (0802/0803/0810, opt-in), `auto_mtp_adaptive_v1` (0807/0808, opt-in), `moe_strategy_v1` (0739/0740). /props `features` of this engine does not list them (unchanged from c13 r6); the pin is the capability source for consumers.

components/engine/2611011701001/opencoti-0.10.5-c14-2611011701001

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63042e6ef20b9b3819e2275e6d73d023b31eb1ff69e419342f6b670728282be6
+size 146932862

components/engine/2611011701001/opencoti-0.10.5-c14-2611011701001.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63042e6ef20b9b3819e2275e6d73d023b31eb1ff69e419342f6b670728282be6
+size 146932862

components/macos/2610311711005/BUILD_INFO.macos.md

@@ -2,7 +2,7 @@
 
 Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
 
-Built from the engine source of build 2610312243001; gated floor (engine-min) 2610312243001; gated with 2610312243001, 2611010802001, 2611011335001.
+Built from the engine source of build 2610312243001; gated floor (engine-min) 2610312243001; gated with 2610312243001, 2611010802001, 2611011335001, 2611011619001.
 
 | file | platform | kind | sha256 | bytes |
 |---|---|---|---|---|

components/media/2611011351002/BUILD_INFO.media.md

@@ -2,7 +2,7 @@
 
 Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
 
-Built from the engine source of build 2611011335001; gated floor (engine-min) 2611011211001; gated with 2611011335001, 2611011211001.
+Built from the engine source of build 2611011335001; gated floor (engine-min) 2611011335001; gated with 2611011335001.
 
 | file | platform | kind | sha256 | bytes |
 |---|---|---|---|---|

components/vulkan/2611011644001/BUILD_INFO.vulkan.md

@@ -0,0 +1,33 @@
+# vulkan 2611011644001
+
+Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
+
+Built from the engine source of build 2611011619001; gated floor (engine-min) 2611011619001; gated with 2611011619001, 2611011701001.
+
+| file | platform | kind | sha256 | bytes |
+|---|---|---|---|---|
+| `ggml-vulkan-x86_64.so` | x86_64 | vulkan | `c6792230501f91bab6aae05278d059542c4f953acde3d8eb42e5d83c72d20a56` | 71955824 |
+| `ggml-vulkan-win-x86_64.dll` | win-x86_64 | vulkan | `1affc561f45d03d73c61562057db8350106f57968300de6d12d31421a6864303` | 73009528 |
+
+ABI `ggml` = `bac5050ddc8c58c51f0b059a7f9753b33af0ed79d355090e7c9a17ffaaaf340a` (exported) — sha256 of this list:
+
+```
+llama.cpp/ggml/include/ggml-alloc.h 94e4cd069b9313b2ceb35dacec901981e0bb478d8bb31035b7126be091998c23
+llama.cpp/ggml/include/ggml-backend.h 764d1640ad62579513c4837539708f89e8603c9dbeae053a13d1cbd84a82c865
+llama.cpp/ggml/include/ggml-cuda.h 674ea064021cf5b9970e78f1209e513dd23350b94a44a705a27eec1423fe4b0a
+llama.cpp/ggml/include/ggml-metal.h 322f36cd30f3e9e7aad7b5b9bc63078012fd0b7706ac4e24f5721b604f3d8980
+llama.cpp/ggml/include/ggml-neo-pipeline.h 190f39e183542d16588a3104d8cb4359bf52fc7a62afe8d8cc4c822f4085086d
+llama.cpp/ggml/include/ggml-oc-kvarn-range.h 76635fdd8978632f670640ddf5125a666441e3d47c680bb8b5c0a099e4861202
+llama.cpp/ggml/include/ggml-oc-ledger.h c1be11d69a57a1e721f1ae8720b87887d239ac0183d4277fbaa07f1be257d20c
+llama.cpp/ggml/include/ggml-vulkan.h f9df4f7161a84ce7600d53c3fac8a62837d5716676c9f38cd340699ef1c070d9
+llama.cpp/ggml/include/ggml.h ecfbf114f9e9151d22e30ef6d5139570c5c41b595843ca4f99125bd0196709aa
+llama.cpp/ggml/include/opencoti-devtype.h 94dea946173449bb76bffd7f733ab822fe658fe7198992de750ac012d29308ec
+llama.cpp/ggml/src/ggml-backend-impl.h f9fdfe6bae61d1387593b40fc0f90ceee490c6a339ec6d7388c51e4fca318d8f
+llamafile/gpu_backend.h 4929c36c034b6589e88653bcb7a5c5a225743a07320f64c0857d92bdaf048208
+```
+
+**c14 r4 (2026-11-01) — rebuilt, version `2611011644001`** (r3 `2611011351001`): `0827-c14-vk-mmv-one-col-scalar-acc` (roadmap 57.12 / op-319, the last 2 % of bug-4315): under `MMV_TILE1` every `mmvq_dot_product` takes ONE `inout` scalar accumulator instead of the whole `temp[NUM_COLS][NUM_ROWS]`, and every per-column loop bound / local array size is the literal 1 (`MMV_TILE_N`); `iter()` has no tile loop. The tiled (2..8-column) pipelines are unchanged. Measured library-only on the r3 engine (eleven2go, Qwen3-4B Q4_K_M, interleaved 3 × 256 greedy): Renoir iGPU (Windows AMD driver) 9.92 / 9.93 vs r3 8.77–8.92 vs c13 r6 9.09–9.11 tok/s; RX 9070 XT 181.1 vs 181.8 vs 180.5; RTX 3090 194.2 vs 194.4 vs 194.1; the greedy output sha identical to the r3 library's on all three devices. `ggml-vulkan-x86_64.so` `c6792230501f` (r3 `954c12f03caa`), `ggml-vulkan-win-x86_64.dll` `1affc561f45d` (r3 `1f4703b1bb51`), both from the clean-room `release/c14-r4` on the c14 lane caches (incremental). Gated: 41 RADV rig (`.so` vs the r3 library), 43 Windows arms incl. `vulkan2-igpu` vs the c13 r6 library, 40 bonsai ident (r3 vs r4), 44, 55.
+
+**c14 r3 (2026-11-01) — both Vulkan libraries REBUILT** from the c14 r3 clean-room on bs2 (`ggml-vulkan-x86_64.so` `954c12f03caa`, the glibc 2.28 lane, floor GLIBC_2.27, 72.4 MB; `ggml-vulkan-win-x86_64.dll` `1f4703b1bb51`, the mingw lane, PE32+, imports vulkan-1.dll only, 73.4 MB; r2 / c14: `aabaf8a7e26d` / `957802ab0856`). One patch over c14: `0825-c14-vk-mmv-one-col` (roadmap 57.11 / op-318, bug-4315, xollama #2064) — c14's decode on the Renoir iGPU under the Windows AMD driver (32.0.21045) was 26 % below c13 r6 (6.63 vs 9.10 tok/s on Qwen3-4B Q4_K_M, 3 × 256 greedy; qwen3:8b 4.18 vs 5.65 through xollama) because 0797's column tile made the B cache of `mul_mat_vecq.comp` two spec-constant-sized module arrays (`cache_b_qs[MMV_COL_TILE][..]`, `cache_b_ds[MMV_COL_TILE]`) in EVERY integer-dot mat-vec pipeline, the one-column (decode) pipelines included, and that compiler keeps them out of registers (tile clamped to one column 7.49, a fixed `[8]` array 5.32). The generator now builds every integer-dot mat-vec shader twice — tiled, and a `_1c` variant with `MMV_TILE1` (the pre-0797 one-column shape: a 1-D fixed array and a scalar, `CB_QS` / `CB_DS` macros); the host picks the `_1c` pipelines for one column; the `mul_mat_id` shaders (always one column) get the shape directly; the tile spec constant is clamped to the column count. Measured, library-only cells on the c14 r2 engine: the iGPU 8.90 tok/s (c13 r6 9.10 — the remaining 2 % is c15 row 57.12 / op-319), RX 9070 XT 181 vs 182, RTX 3090 194 vs 194; the greedy 256-token output sha identical across c14 r2 / r3 / c13 r6 on every device; the RADV rig (solidPC 5600G) neutral at 9.5. The generated shader header gained the `_1c` arrays, so the Windows lane recompiled its 169 shader units once (plan 169 of 169, expected). Gated (r3): the Windows gate's `vulkan2-igpu` arm against the sha-pinned c13 r6 DLL (PASS ≥ 0.97), RADV rig, Vulkan identity on bs2, split-form boots on bs2 / eleven2go, old-glibc.
+
+**c14 (2026-10-31):** built from the c14 clean-room on bs2 (the Linux library in the glibc 2.28 lane, floor GLIBC_2.27; the Windows DLL cross-built in the mingw lane). Carries: the KVarN store in its parallel (workspace) form (0741) with the RX 9070 XT's flush tile in device memory (0800); the V pass staging a tile's record words (0742); the scalar flash attention at head dim >= 256 on RDNA (0743: RX 9070 XT gemma-4 A4B f16 16k prefill 746 -> 2424 tok/s); the head-512 coopmat1 FA in two 32-lane subgroups and the Q split (0781, 0799); the native KVarN attention and the plain FA over each row's live cells (0726, 0728, 0730, 0731); the multi-column mat-vec for every weight type (0797), PrismML's PTQ1_0 mat-vec and signed FWHT fusion (0745, 0746), Bonsai 2 coopmat2 / integer-dot mat-mat (0744), IQ4_NL from 3 columns on NVIDIA (0805), split-k for narrow outputs (0784); the opt-in mat-vec pick `--vulkan-matvec-tune` (0802, 0803, 0810); no pipeline compile on the decode path (0809: bs2 cold 16k decode 5.1 -> 99.0 tok/s) and the Windows pipeline cache kept (0811); the row-chunk KQ mask on the device (0786); KV layers larger than the device's maximum buffer held in segments (0759); past-window KVarN record prefetch (0755); the teardown staging cycle (0804); the record-reading coopmat1 KVarN FA (0798, opt-in). Gated with the c14 engine 2610312243001 (the re-cut on 0822): the RADV iGPU rig (solidPC 5600G, RADV RENOIR, image opencoti-vulkan-radv) on the STAGED bytes beside the engine, the c13 r6 library as the other arm: f16 and kvarn3 cells, every needle found, 36/36 layers on Vulkan0, tg 30.8 / 28.1 tok/s (c13 r6: 32.4 / 28.6); the flash-attention variant set from `nm` scalar 10 / cm1 9 / cm2 3 = 22, a superset of c13 r6's 10 / 3 / 3 (the coopmat1 KVarN FA). Gated: the old-glibc Vulkan arm (bs2, glibc 2.28 image, NVIDIA ICD, every read route twice); the Bonsai 2 Vulkan gate on bs2 (RTX PRO 6000): needle 3/3 at 112k, pp/tg on both files, KLD 0.000305 mean vs the reference CUDA base on both files (bar 0.001), greedy parity vs CUDA with every divergence a knife edge on the same swapped pair, old-vs-new library ident (Qwen3-14B Q8_0 bit-identical; Omnimerge v6 Q4_K_M the same 256 greedy tokens with the chosen logprob within 0.052 — the k-quant mat-vec kernels changed on purpose, op-291); the slot-lazy gate on Vulkan 5/6 (default slots +22 MiB over -np 1, the compute buffer 208 vs 198 MiB, accepted by the owner for c14 — bug-4308); the RADV iGPU rig (solidPC 5600G) f16 and kvarn3 needles 4/4 beside c13 r6's library; eleven2go's RTX 3090 (Vulkan0, 62.8 tok/s) and RX 9070 XT (Vulkan1, 72.0 tok/s) with the Windows DLL from the executable directory, 66/66 layers, needle found.

components/vulkan/2611011644001/ggml-vulkan-win-x86_64.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1affc561f45d03d73c61562057db8350106f57968300de6d12d31421a6864303
+size 73009528

components/vulkan/2611011644001/ggml-vulkan-x86_64.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6792230501f91bab6aae05278d059542c4f953acde3d8eb42e5d83c72d20a56
+size 71955824