Skip to content

Release v0.6.9 - #320

Merged
Andyyyy64 merged 6 commits into
mainfrom
v0.6.9
Nov 19, 2026
Merged

Andyyyy64 merged 6 commits into
mainfrom
v0.6.9

Conversation

@Andyyyy64

Copy link
Copy Markdown
Owner

This brings v0.6.9 into main for the 0.6.9 release. The branch holds #308, #309, #314, #316, #317 and the release prep #319, each squash-merged with CI passing.

#309 reads Artificial Analysis scores live again by converting AA's index v5 onto the v4.3.2 scale, which moves the top of the ranking. #308 adds an engine card for llama-modchip. #314 stops a row on the GPU from listing an engine that runs only on the CPU there. #316 detects a Jetson AGX Thor as a GPU, and #317 re-reads eight engine cards.

I'm merging with a merge commit, not a squash, so main keeps each PR as its own commit. main has one commit v0.6.9 doesn't, the README badge from #318, and the two touch different files, so there is nothing to resolve.

#309 changes how scripts/build_index.py reads AA scores, so the index picks it up on its next build from main.

devangpratap and others added 6 commits November 14, 2026 21:54
* Convert AA index v5 scores onto the v4.3.2 scale

AA's leaderboard moved to Intelligence Index v5, so the fetcher has
served only the snapshot. Fit v4.3.2 = slope * v5 + intercept on four
releases read on both versions, stored in aa_index.json, and convert
5.x values before normalizing. Other versions still fall back to the
snapshot. Map the 31 open v5 rows that had no entry.

* Keep uploads scored and clamped scores out of the anchor

Lists the five uploads that AA's per-benchmark results used to fill next
to their originals in aa_names.json, so they keep a score once the index
maps the release. A general score at 0 or 100 no longer joins the anchor
that puts those results on the general scale. Adds gpt-oss-120b as a
fifth v5 pair, now that AA gives it a v5 value. CHANGELOG entry under
Unreleased.

Refs #244
llama-modchip is a llama.cpp fork for large MoE models on the CPU. Its own
quant types do not load in mainline: LUT-VQ3-64 is VQ3_0 at 3.00 bpw, and
Fractal-64 reuses type 42 with a 30-byte block, which mainline sizes at 18
bytes and refuses. Standard GGUFs still load, so the card is a runner and
lists VQ3_0 under `gguf.extra` the way ik_llama.cpp lists its own types.

The platforms are the builds the fork actually ships: x86-64 AVX2 and
AVX-512 prebuilt binaries plus a source build for ARM64 on Linux, a MinGW-w64
UCRT64 build on Windows, and the patched Metal engine on Apple silicon. CUDA
is Windows only and experimental, the -Cuda switch of the build script, since
install.sh builds no GPU engine and the CUDA and Vulkan binaries the Windows
hybrid pack downloads are mainline llama.cpp's.

There is no `probe`: the patched server is llama-server, so its /props reply
is mainline's and would claim the port for this engine on any llama.cpp
server. `detect_binaries` names the two files install.sh leaves in the
clone, and the note says the installer does not put them on PATH.

Refs #294
…#314)

* fix(engines): keep CPU-only builds off rows placed on the GPU

engines_for checked that an engine runs on the machine, not that it runs
on the GPU. ik_llama.cpp, which has no ROCm or SYCL build and needs CUDA
compute capability 7.5, was listed for GPU rows on AMD, Intel and older
NVIDIA GPUs, where it runs only on the CPU. A row that puts weights on the
GPU now counts only engines with a GPU build here.

* test(engines): llama-modchip runs only on the CPU on Linux with an NVIDIA GPU
Thor's GPU is named "NVIDIA Thor" by NVML and nvidia-smi, without the
"(nvgpu)" suffix of Orin and Xavier, and both report its memory as
unsupported. The name matched no unified-memory marker, so the GPU was
dropped. Recognise the name, map the T5000 module (p3834-0008) to
"Jetson AGX Thor", and add its bandwidth and compute capability.

Co-authored-by: Will Schenk <wschenk@thor.monkey-cardassian.ts.net>
Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
llama-modchip 3.0 gives its own quant types their own IDs (FRACTAL_64,
VQ3_64), Strata 0.2.16 no longer runs DeepSeek V4.1 Flash, ik_llama.cpp
loads mainline FP8 GGUFs, vLLM 0.34.0 ships ROCm 10.0 wheels only and
Lemonade's CUDA build checks the driver. llama.cpp 1.0.0, TurboFieldfare
0.10.0 and WARP 0.8.2 move only their versions.
@Andyyyy64
Andyyyy64 merged commit 3f51a8f into main Nov 19, 2026
12 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants