feat(bundle): llama.cpp CPU Qwen3-Embedding-0.6B embedding bundle#111
Merged
Conversation
Adds a CPU-only embedding bundle so hosts without a compatible GPU (or that can't reach grackle-embed) can run semantic search locally. Serves Qwen3-Embedding-0.6B (Q8_0 GGUF, 1024-dim) via llama.cpp with an OpenAI-compatible /v1/embeddings endpoint on 127.0.0.1:8007, registering the llamacpp-cpu-embed provider. Same model / vector space as the GPU embed bundles. Runs on macOS/Windows Docker Desktop; first request auto-downloads the GGUF via -hf and caches it. Regenerated registry/add-ons.json via build-registry. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
manifest declared contextLen 32768 but docker-compose serves --ctx-size 8192, so inputs over 8K tokens would be silently rejected despite the advertised capacity. The CPU bundle intentionally caps ctx at 8192 for RAM; embedding inputs are capped at 8000 chars upstream, so 8192 is ample. Lower the declared contextLen (manifest + regenerated registry entry) to match reality. Vector space is unchanged (1024-dim, same model) — embeddings stay interchangeable with the GPU bundles; only max input length differs.
kh0pper
added a commit
that referenced
this pull request
Jun 29, 2026
The CPU embedding bundle (#111) binds host port 8007, but the row was never added to docs/developers/port-allocation.md, so the Port Allocation Check CI (scripts/check-port-allocation.js) failed on the PR and on the main push. Add the 8007 row; check now passes (43 ports, all documented). Co-authored-by: kh0pper <kevin.hopper@maestro.press>
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.
What
A CPU-only embedding bundle so hosts without a compatible GPU — or that can't reach a shared embedder like
grackle-embed— can run Crow's semantic search locally.Serves Qwen3-Embedding-0.6B (Q8_0 GGUF, 1024-dim) via llama.cpp with an OpenAI-compatible
/v1/embeddingsendpoint on127.0.0.1:8007, registering thellamacpp-cpu-embedprovider. Same model / vector space as the GPUvllm-cuda-embedandllamacpp-vulkan-qwen3-embedbundles, so embeddings are interchangeable.Why
The existing embedding bundles are GPU + Linux-only (
vllm-cuda-embed→ NVIDIA,llamacpp-vulkan-qwen3-embed→ ROCm/gfx1151). There was no option for a Mac/Windows Docker Desktop host or any CPU-only box.Highlights
gpu_arch: ["cpu"], port bound to127.0.0.1.-hfand caches it in a Docker volume (no manual model fetch).dashboard_settings.embed_provider = 'llamacpp-cpu-embed'(orCROW_EMBED_PROVIDER).Changes
bundles/llamacpp-cpu-qwen3-embed/—manifest.json,docker-compose.yml,README.md.registry/add-ons.json— regenerated vianpm run build-registry(single entry added, no churn).Validation
npm run build-registry→ 89 bundles, 0 invalid/draft/untracked.npm run test:bundle-contract→ 25/25 pass.🤖 Generated with Claude Code