llama-cpp: llama-server with DFlash2 speculative decoding for gfx1151

v0.8.0 is the first upstream release with DFlash2 drafters; nixpkgs has 0.6.0 (unstable) and an older tree (stable), so the engine is built here. HIP build for Strix Halo. The NixOS module exposes llama-server behind skyg.nixos.server.services.llama-cpp; fwdesk wires it to Qwen3.8-27B + its DFlash2 drafter, disabled until the weights are on disk.
This commit is contained in:
hermenix committed 2026-11-11 16:01:32 +00:00
1 parent 62c41a6f8e
commit ce92096042
9 files changed
+392

No files matched your search

+3
View File
@@ -22,6 +22,9 @@
# colibrì (JustVugg/colibri) — GLM-5.2/OLMoE local inference, see pkgs/colibri.
colibri = final.callPackage ../../pkgs/colibri { };
colibri-rocm = final.callPackage ../../pkgs/colibri { backend = "rocm"; };
# llama.cpp v0.8.0 for gfx1151 — DFlash2 speculative decoding, see pkgs/llama-cpp.
llama-cpp-halo = final.callPackage ../../pkgs/llama-cpp { };
llama-cpp-halo-cpu = final.callPackage ../../pkgs/llama-cpp { backend = "cpu"; };
# Touchscreen gesture bridge for niri, see pkgs/niri-touchscreen-gestures.
niri-touchscreen-gestures = final.callPackage ../../pkgs/niri-touchscreen-gestures { };
# Buzz Desktop AppImage wrapper, see pkgs/buzz-desktop.