pin v2 payload: engine 2610041714001, cuda 2610040656001, cuda12 2610040730001, sbsa 2610040704001, vulkan 2610041656001, macos 2610041000001, media 2610040945001

.gitattributes

@@ -234,3 +234,5 @@ components/engine/2610041643001/opencoti-0.10.5-c7-2610041643001 filter=lfs diff
 components/engine/2610041643001/opencoti-0.10.5-c7-2610041643001.exe filter=lfs diff=lfs merge=lfs -text
 components/vulkan/2610041656001/ggml-vulkan-x86_64.so filter=lfs diff=lfs merge=lfs -text
 components/vulkan/2610041656001/ggml-vulkan-win-x86_64.dll filter=lfs diff=lfs merge=lfs -text
+components/engine/2610041714001/opencoti-0.10.5-c7-2610041714001 filter=lfs diff=lfs merge=lfs -text
+components/engine/2610041714001/opencoti-0.10.5-c7-2610041714001.exe filter=lfs diff=lfs merge=lfs -text

components/engine/2610041714001/BUILD_INFO.engine.md

@@ -0,0 +1,59 @@
+# engine 2610041714001
+
+Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
+
+| file | platform | kind | sha256 | bytes |
+|---|---|---|---|---|
+| `opencoti-0.10.5-c7-2610041714001` | any | bin | `3118324fe3339cbaf1f1f91bb27b8af63e6379aa8f6c85e9115bc9e8a8b2d05e` | 134677211 |
+| `opencoti-0.10.5-c7-2610041714001.exe` | win-x86_64 | bin | `3118324fe3339cbaf1f1f91bb27b8af63e6379aa8f6c85e9115bc9e8a8b2d05e` | 134677211 |
+
+ABI `ggml` = `41ba479c24d7e3a624c7b13aa3cbe357e41fac04a504c54ec2ee170de0bff9a1` (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 af6dc423ee7c6c3f67a6a187d6dfbfeb94e52ddc1db27c45ee7440289bc44282
+llama.cpp/ggml/include/ggml-vulkan.h 7eae5dad2cc7bb4d3eca828539f441816d5fd59fdc7b224d49aa229fc7b7248c
+llama.cpp/ggml/include/ggml.h f2bb8b127e3e601a78fe32242ffab7235e6a4db3bab5370b72fb6bacba2b2433
+llama.cpp/ggml/include/opencoti-devtype.h 94dea946173449bb76bffd7f733ab822fe658fe7198992de750ac012d29308ec
+llama.cpp/ggml/src/ggml-backend-impl.h f9fdfe6bae61d1387593b40fc0f90ceee490c6a339ec6d7388c51e4fca318d8f
+llamafile/gpu_backend.h d17c504fbab50ff9376105a1f2c79d99fcde4b431886b17b90718a2a359278de
+```
+
+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
+```
+
+The APE: one file for Linux x86_64 / aarch64, Windows and macOS arm64; the `.exe` row is the same bytes under the name Windows needs.
+
+**Replaces engine 2610041643001 — fix for a crash (bug-3921, patch 0557 ring-stats-decode-thread).** `GET /props` and `GET /kv` killed the engine (exit 139) when the Vulkan library was loaded on a host with glibc older than 2.34 (measured: Debian 11, glibc 2.31): the report ran on an HTTP thread and called into the Vulkan library, whose `pthread_mutex_lock` faults on any thread but the main one there. Every engine of this index since 2610040837001 has it; CUDA was not affected; glibc 2.34+ hosts and Windows did not show it. The figures under `rolling_kv.live.ring` are now taken on the decode thread and are one graph older. Host side only: the Vulkan library of this index is unchanged.
+
+Chain through 0557, on 0556 shutdown-gpu-policy (bug-3919; docs/features/server_shutdown.md), on top of 0555 graceful-shutdown and 0554 abi-export:
+- `POST /shutdown` — loopback only (403 otherwise), 503 while the model is still loading, behind the API key when one is set. The same path as SIGTERM. Feature `graceful_shutdown_v1`.
+- The stop path frees the media engines, the models and the contexts (`shutdown: models and contexts freed`), then for every Vulkan device waits for it to be idle and destroys it (`ggml_vulkan: shutdown: Vulkan0 idle and destroyed`), then the instance (`… N device(s) destroyed, 0 kept, instance destroyed`), exit 0.
+- `OPENCOTI_SHUTDOWN_GPU=auto|off|free|device|full` selects how far that goes; `auto` = `full` on every platform. `off` = nothing released (the exit as it was before 0555): `shutdown: GPU state left to the process exit (…)`. `free` = models only; `device` = devices destroyed, instance kept. Feature `shutdown_gpu_policy_v1`.
+- Deadline `OPENCOTI_SHUTDOWN_DEADLINE_MS` (default 10000, 0 = none): `the teardown did not finish within the deadline, exiting now`, exit 0.
+- Parent watch: `OPENCOTI_PARENT_PID=<pid>` (on Windows a Windows process id) — `parent watch: armed on pid N`; when that process is gone, `shutdown: parent pid N is gone` and the same stop path. `NOT armed` when the pid cannot be watched. Feature `parent_watch_v1`.
+- `--list-devices` destroys the Vulkan instance before it exits; `OPENCOTI_LIST_DEVICE_IDS=1` appends ` id=<PCI address | uuid:<32 hex>>` to each line (the stock line is unchanged without it). Feature `list_device_ids_v1`.
+- The Vulkan teardown needs the `vulkan` component of this index; with an older Vulkan library the engine says the library predates the device teardown and leaves the devices.
+No interface header changed: the `ggml`, `media` and `loader` digests are those of engine 2610040950001, so every other component fits unchanged.
+
+AMD Radeon on Windows: AMD Software 26.9.2 (display driver 32.0.32015.2008) or later is the minimum for Vulkan. On 26.8.1 (32.0.31041.1004) a model left idle on an RX 9070 XT times out in the driver (0x141, amdkmdag.sys) with any Vulkan application, this engine and stock llama.cpp alike; the engine does not work around it.
+
+Gated on this exact file: the old-glibc run gate, on bs2 inside the glibc 2.28 image (`scripts/bs2-oldglibc-gate.sh`, RTX PRO 6000, with the cuda and vulkan libraries of this index) and on solidPC (glibc 2.31, RTX 3090), `scripts/http-thread-dso-gate.sh` — Vulkan and CUDA: one generation, every read route twice (hot and after 4 s idle), POST /shutdown 200, exit 0 (the previous engine fails this at /props on Vulkan). Linux x86_64 (bs2), `.opencoti/graceful-shutdown-gate.sh` all arms — POST /shutdown and SIGTERM on Vulkan (exit 0 in 2.0-3.0 s, device and instance destroyed), non-loopback 403, stop during a stream, deadline, parent watch armed / not armed, the `off` / `free` / `device` modes, CUDA POST /shutdown, list-devices teardown and ids; Vulkan rolling window, 80k-token prompt, 5 host windows: ring figures equal to the previous engine (48 tiles, copy 47.9 ms, 54.7 GB/s). NOT run on this file: Windows and macOS, Linux aarch64, the bs2 2M-context swarm.
+Run by xollama on an earlier engine (2610041105001, same stop path with the default `full`), Windows 11 + RX 9070 XT on driver 32.0.32015.2008: POST /shutdown + reopen 10 of 10 clean, 103.3-103.6 tok/s; route, teardown lines and parent watch as specified (mails #768, #778).
+
+Image / video features are listed for the platforms they were gated on.

components/engine/2610041714001/opencoti-0.10.5-c7-2610041714001

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3118324fe3339cbaf1f1f91bb27b8af63e6379aa8f6c85e9115bc9e8a8b2d05e
+size 134677211

components/engine/2610041714001/opencoti-0.10.5-c7-2610041714001.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3118324fe3339cbaf1f1f91bb27b8af63e6379aa8f6c85e9115bc9e8a8b2d05e
+size 134677211