chore: pin llamacpp-upstream backend to b10063 (and fix macOS backend download URLs) - #187
Closed
cashcon57 wants to merge 2 commits into
Closed
chore: pin llamacpp-upstream backend to b10063 (and fix macOS backend download URLs)#187cashcon57 wants to merge 2 commits into
cashcon57 wants to merge 2 commits into
Conversation
b9937 predates the ternary Q2_0 Metal kernels that landed upstream in ggml-org/llama.cpp#25419 (merged Jul 14, first shipped in b9994); the CPU kernels landed earlier in #24448. Bumping to b10063 lets the upstream backend load ternary models like Ternary-Bonsai in the mainline Q2_0_g64 format on macOS and CPU. Verified locally on an M3 Max with the b10063 macos-arm64 release binaries and Ternary-Bonsai-8B-Q2_0_g64: Metal backend initializes, the model loads (254 q2_0 tensors, file type Q2_0), and generation is coherent. All eight asset names referenced by the fallback manifest exist on the b10063 release. Mirrors: PINNED_BACKEND_TAG (extension), the offline fallback manifest (backend.ts), and LLAMACPP_UPSTREAM_TAG (Makefile). The remote manifest in atomic-chat-conf needs the matching one-line bump.
ggml-org publishes macOS release assets only as .tar.gz (and always has), but getBackendDownloadUrl and getBackendArchiveName fell through to the Windows .zip naming for macos-* backends, so every runtime backend download on macOS 404ed and enforcePinnedBackendVersion silently kept the previous backend. The bug was masked because macOS normally receives the backend bundled into the app at build time; it surfaces as soon as the pinned tag is newer than the bundled backend. Route macos-* through the .tar.gz naming (the Tauri decompress command already handles both formats) and add archive-name tests. Verified in a dev build: the pin now downloads and installs llama-b10063-bin-macos-arm64.tar.gz successfully, and a ternary Q2_0_g64 model loads and generates on the installed backend.
cashcon57
changed the title
Contributor
Author
|
Closing as superseded: main now pins the upstream backend via LLAMACPP_UPSTREAM_PINNED_TAG (currently b10205), which covers what this PR was for. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Describe Your Changes
Bumps the pinned upstream llama.cpp backend from
b9937tob10063so thedesktop app can run ternary models (e.g. the new Ternary-Bonsai family) on the
upstream backend — plus a fix this bump surfaced: runtime backend downloads
on macOS were silently 404ing because the extension built
.zipasset URLswhile ggml-org publishes macOS builds only as
.tar.gz. The bug was maskeduntil now because macOS normally gets the backend bundled into the app at
build time; the moment the pin outruns the bundled backend,
enforcePinnedBackendVersionhits the 404 and quietly keeps the old backend.Second commit routes
macos-*through.tar.gznaming (the Tauridecompresscommand already handles both) with archive-name tests.Why: mainline llama.cpp ternary
Q2_0support landed after our pin. TheCPU kernels merged Jul 7 (ggml-org/llama.cpp#24448) and the Metal kernels
merged Jul 14 (ggml-org/llama.cpp#25419, first shipped in
b9994);b9937was cut Jul 9, so it has CPU but not Metal. Today no backend available in the
desktop app can load these models, while the mainline
Q2_0_g64GGUFs runfine on current upstream builds. (Vulkan and CUDA ternary kernels are still in
review upstream: ggml-org/llama.cpp#25430 and #25707, so those platforms gain
nothing yet but lose nothing either.)
What's bumped (the three mirrored pins):
PINNED_BACKEND_TAGinextensions/llamacpp-upstream-extension/src/index.tsextensions/llamacpp-upstream-extension/src/backend.tsLLAMACPP_UPSTREAM_TAGinMakefileThe remote manifest in
atomic-chat-conf/backends/manifest.jsonneeds thematching bump; opening that separately since it goes live for all installs the
moment it merges.
Verification
b10063macos-arm64release binaries,prism-ml/Ternary-Bonsai-8B-ggufQ2_0_g64(the mainline-format file)loads on an M3 Max with
254 q2_0 tensors/file type = Q2_0, Metalinitializes, and generation is coherent.
enforcePinnedBackendVersionfailed withHTTP status 404 Not Found; withit, the pin downloaded
llama-b10063-bin-macos-arm64.tar.gz, loggedBackend b10063/macos-arm64 installed successfully, and the installedbinary reports
version: 10063. The ternary 8B then loads and generatesthrough the app on that backend (~704 t/s prompt / ~85 t/s gen at 5k
context, clean stops, tool calls working).
(win cpu/cuda-12.4/cuda-13.3/vulkan, ubuntu x64/vulkan, both cudart zips)
exist verbatim on the
b10063release.failures are identical on
main(verified by stash-compare).Fixes Issues
Self Checklist