Reported in #39 by @clort81 — the `luce-spec/llama-cpp-dflash` repo
returns 404. The DFlash work consolidated into Luce-Org/lucebox-hub
(verified: github.com/Luce-Org/lucebox-hub returns 200, contains
dflash/ + pflash/ subdirs and dflash/deps/llama.cpp submodule).
Affected files:
- docs/engines/README.md (2 link sites in comparison table)
- docs/engines/LLAMA_CPP.md (4 sites: intro, "Pros" table, build clone
command, "See also" links)
- models/qwen3.6-27b/llama-cpp/README.md (2 link sites)
Plus collateral updates:
- Build clone path /opt/llama-cpp-dflash → /opt/lucebox-hub (matches
the new repo name; was a 3-replace via path globbing)
- HF model path luce-spec/dflash-qwen3.6-27b-N5 (401 gated) →
z-lab/Qwen3.6-27B-DFlash (200 public, the actually-shipping draft)
+ local-dir adjusted to /mnt/models/huggingface/z-lab/... matching
the canonical HF model path convention
- `git clone --recurse-submodules` flag added since lucebox-hub uses
submodules for its bundled llama.cpp fork (in dflash/deps/llama.cpp)
Updates URL framing in user-facing prose to acknowledge that
lucebox-hub is a separate harness containing a llama.cpp fork rather
than just being a llama.cpp fork. The recipe section build commands
should be re-verified against the lucebox-hub README before treating
them as canonical — this commit only updates the URL/path; the multi-
step build instructions in docs/engines/LLAMA_CPP.md may need a
follow-up walkthrough.
CHANGELOG references to luce-spec preserved as historical context (the
links were valid at the time the CHANGELOG entries were written).
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
