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

+1
View File
@@ -339,6 +339,7 @@
buzz-desktop = lib.pkgs.${system}.buzz-desktop;
delta-editor = lib.pkgs.${system}.delta-editor;
colibri = lib.pkgs.${system}.colibri;
llama-cpp-halo = lib.pkgs.${system}.llama-cpp-halo;
herdr = herdr.packages.${system}.default;
}
);
+1
View File
@@ -1,6 +1,7 @@
{
imports = [
./ds4.nix
./llama-cpp.nix
./ollama.nix
./libretranslate.nix
./hermenix.nix
+32
View File
@@ -0,0 +1,32 @@
{ ... }:
{
# llama.cpp with DFlash2 speculative decoding.
#
# Off by default: it needs ~19 GB of weights on disk and a first build of the
# ROCm package. Fetch the files, then flip enable:
#
# sudo install -d -o ari -g users /var/lib/llama-cpp
# cd /var/lib/llama-cpp
# curl -L -O https://huggingface.co/ggml-org/Qwen3.8-27B-GGUF/resolve/main/Qwen3.8-27B-Q4_K_M.gguf
# curl -L -O https://huggingface.co/z-lab/Qwen3.8-27B-DFlash2-GGUF/resolve/main/Qwen3.8-27B-DFlash2-Q8_0.gguf
#
# Measured expectation for this box: ~25-40 tok/s, not the 380 tok/s in the
# post that prompted this — that figure is a 3090 copying a document.
skyg.nixos.server.services.llama-cpp = {
enable = false;
model = "/var/lib/llama-cpp/Qwen3.8-27B-Q4_K_M.gguf";
draftModel = "/var/lib/llama-cpp/Qwen3.8-27B-DFlash2-Q8_0.gguf";
specType = "draft-dflash";
specDraftNMax = 7;
ctx = 32768;
kvCacheType = "q8_0";
cacheReuse = 256;
host = "0.0.0.0";
port = 8081;
openFirewall = true;
};
}
+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.
+7
View File
@@ -16,6 +16,7 @@ services/
├── nbot-agent.nix # skyg.nixos.server.services.nbot-agent — nbot's agent system + gh, for nbot's `system`
├── signal-cli.nix # skyg.nixos.server.services.signal-cli — signal-cli HTTP daemon (Hermes Signal bridge)
├── colibri.nix # skyg.nixos.server.services.colibri — coli serve HTTP API (GLM-5.2/OLMoE local inference)
├── llama-cpp.nix # skyg.nixos.server.services.llama-cpp — llama-server (GGUF, DFlash2 speculative decoding)
└── comfyui/ # skyg.nixos.server.services.comfyui — ComfyUI with ROCm/CUDA GPU support
```
@@ -31,6 +32,11 @@ services/
service. `package` picks the backend (`pkgs.colibri`
cpu / `pkgs.colibri-rocm`); `environmentFile` supplies `COLI_API_KEY` via agenix instead of
a CLI flag.
- **`llama-cpp.nix`** runs `llama-server` as a native systemd service. `draftModel` plus
`specType = "draft-dflash"` turns on DFlash2 speculative decoding; `cacheReuse` turns on
prefix reuse for repeated prompts (agents). Model files are read from disk, never packaged —
the unit refuses to start with a clear message when one is missing. Enabled on fwdesk
(off by default: ~19 GB of weights, first build is the ROCm package).
## Option Namespace
@@ -41,6 +47,7 @@ skyg.nixos.server.services.jellyfin.enable
skyg.nixos.server.services.nbot-agent.system
skyg.nixos.server.services.signal-cli.{enable,host,port,configDir,account,environmentFile,openFirewall}
skyg.nixos.server.services.colibri.{enable,package,model,host,port,environmentFile,openFirewall,...}
skyg.nixos.server.services.llama-cpp.{enable,package,model,draftModel,specType,specDraftNMax,cacheReuse,ctx,host,port,openFirewall,...}
skyg.nixos.server.services.comfyui.{enable,mode,port,openFirewall}
```
@@ -4,6 +4,7 @@
./audiobookshelf.nix
./ai-services.nix
./colibri.nix
./llama-cpp.nix
./comfyui
./jellyfin.nix
./nbot-agent.nix
+252
View File
@@ -0,0 +1,252 @@
{ config, lib, pkgs, ... }:
with lib;
let
cfg = config.skyg.nixos.server.services.llama-cpp;
skygUser = config.skyg.user;
# Target model plus optional drafter. The drafter is only useful for a
# strategy that matches its architecture: DFlash2 sidecars want
# `--spec-type draft-dflash`.
modelFiles = [ cfg.model ] ++ optional (cfg.draftModel != null) cfg.draftModel;
serverArgs =
[
"--model"
cfg.model
"--host"
cfg.host
"--port"
(toString cfg.port)
"--ctx-size"
(toString cfg.ctx)
"--n-gpu-layers"
(toString cfg.gpuLayers)
"--parallel"
(toString cfg.parallel)
"--flash-attn"
cfg.flashAttention
"--cache-type-k"
cfg.kvCacheType
"--cache-type-v"
cfg.kvCacheType
# Qwen-style tool calls arrive as Jinja templates.
"--jinja"
]
++ optionals (cfg.draftModel != null) [
"--model-draft"
cfg.draftModel
]
++ optionals (cfg.specType != null) [
"--spec-type"
cfg.specType
]
++ optionals (cfg.specDraftNMax != null) [
"--spec-draft-n-max"
(toString cfg.specDraftNMax)
]
# Reuse the longest matching prefix of the previous prompt instead of
# reprocessing it. This is the "prefix caching" half of the speedup, and it
# works with any model.
++ optionals (cfg.cacheReuse != null) [
"--cache-reuse"
(toString cfg.cacheReuse)
]
++ cfg.extraArgs;
checkModels = pkgs.writeShellScript "llama-cpp-check-models" ''
set -eu
${concatMapStringsSep "\n" (
f: ''
if [ ! -f ${escapeShellArg f} ]; then
echo "llama-cpp: missing model file: ${f}" >&2
exit 1
fi
''
) modelFiles}
'';
in
{
options.skyg.nixos.server.services.llama-cpp = {
enable = mkEnableOption "the llama.cpp HTTP API (llama-server)";
package = mkOption {
type = types.package;
default = pkgs.llama-cpp-halo;
defaultText = literalExpression "pkgs.llama-cpp-halo";
description = ''
Server package. `pkgs.llama-cpp-halo` is the ROCm build for gfx1151;
`pkgs.llama-cpp-halo-cpu` is the same tree without the GPU backend.
'';
};
model = mkOption {
type = types.str;
example = "/var/lib/llama-cpp/Qwen3.8-27B-Q4_K_M.gguf";
description = "Path to the target model (GGUF).";
};
draftModel = mkOption {
type = types.nullOr types.str;
default = null;
example = "/var/lib/llama-cpp/Qwen3.8-27B-DFlash2-Q8_0.gguf";
description = ''
Optional draft model (GGUF). A drafter must be trained for the exact
target model — a mismatch loads and then drafts nothing useful.
'';
};
specType = mkOption {
type = types.nullOr types.str;
default = null;
example = "draft-dflash";
description = ''
Speculative decoding strategy passed to `--spec-type`. Use
`draft-dflash` for a DFlash/DFlash2 drafter, `draft-mtp` for a model's
own MTP heads, or `ngram-simple` for prompt-copying workloads (code
edits, document rewriting) where no drafter is needed at all.
'';
};
specDraftNMax = mkOption {
type = types.nullOr types.int;
default = null;
example = 7;
description = ''
Tokens drafted per verification step. Clamped by the drafter's trained
block size; 7 is the value upstream uses for DFlash2.
'';
};
cacheReuse = mkOption {
type = types.nullOr types.int;
default = null;
example = 256;
description = ''
`--cache-reuse`: minimum run of tokens to reuse from a matching prompt
prefix. 256 is a reasonable default for coding agents.
'';
};
ctx = mkOption {
type = types.int;
default = 32768;
description = "Context size in tokens.";
};
kvCacheType = mkOption {
type = types.enum [
"f16"
"bf16"
"q8_0"
"q4_0"
];
default = "q8_0";
description = "KV cache type for both K and V (q8_0 halves KV memory).";
};
flashAttention = mkOption {
type = types.enum [
"on"
"off"
"auto"
];
default = "auto";
description = "`--flash-attn` value.";
};
gpuLayers = mkOption {
type = types.int;
default = 99;
description = ''
Layers offloaded to the GPU. On Strix Halo the GPU shares system RAM,
so all layers fit and 99 is the right value.
'';
};
parallel = mkOption {
type = types.int;
default = 1;
description = "Parallel request slots.";
};
host = mkOption {
type = types.str;
default = "127.0.0.1";
example = "0.0.0.0";
description = "Address llama-server binds to.";
};
port = mkOption {
type = types.port;
default = 8081;
description = "TCP port for the HTTP API.";
};
user = mkOption {
type = types.str;
default = skygUser.name;
description = "User to run the server as. Needs the render/video groups for GPU access.";
};
group = mkOption {
type = types.str;
default = "users";
description = "Group to run the server as.";
};
environment = mkOption {
type = types.attrsOf types.str;
default = {
# Same workaround the box's Ollama and ds4 units use.
HSA_ENABLE_SDMA = "0";
};
description = "Environment variables for the service.";
};
extraArgs = mkOption {
type = types.listOf types.str;
default = [ ];
example = [
"--metrics"
"--no-mmap"
];
description = "Extra arguments appended to the llama-server command line.";
};
openFirewall = mkEnableOption "the llama-server port in the firewall";
};
config = mkIf cfg.enable {
assertions = [
{
assertion = cfg.specType == null || cfg.draftModel != null || cfg.specType != "draft-dflash";
message = "services.llama-cpp: specType = \"draft-dflash\" needs a draftModel.";
}
];
systemd.services.llama-cpp = {
description = "llama.cpp server (local LLM inference)";
wantedBy = [ "multi-user.target" ];
after = [ "network.target" ];
environment = cfg.environment;
serviceConfig = {
User = cfg.user;
Group = cfg.group;
ExecStartPre = checkModels;
ExecStart = "${lib.getExe' cfg.package "llama-server"} ${escapeShellArgs serverArgs}";
Restart = "on-failure";
RestartSec = 5;
# Model files live here; the unit may create it, weights are downloaded
# by hand (they are tens of GB and not part of the system closure).
StateDirectory = "llama-cpp";
WorkingDirectory = "/var/lib/llama-cpp";
LimitNOFILE = 65536;
TimeoutStopSec = 120;
};
};
networking.firewall.allowedTCPPorts = mkIf cfg.openFirewall [ cfg.port ];
};
}
+1
View File
@@ -12,6 +12,7 @@ Packages here are exposed via an overlay defined in `modules/core/default.nix`.
| `buzz-desktop/` | Block Buzz desktop (Type-2 AppImage from `block/buzz`). Wrapped with `appimageTools.wrapType2` plus extra FHS libs (`elfutils`, `zstd`, WebKit/tray, GStreamer good/bad) so it runs on NixOS. Exposed as `pkgs.buzz-desktop` (`buzz-desktop` binary + desktop entry / icons). |
| `buzz-cli/` | Buzz CLI (`buzz`, `crates/buzz-cli` of `block/buzz`), built with `rustPlatform.buildRustPackage` from the same tag as `buzz-desktop`. Hermes's Buzz platform shells out to it. Bump with `update.sh` (sets the source hash, resets `cargoHash` to `lib.fakeHash`; build once and paste the `got:` hash). Robots call it with `pkgs.callPackage` (their nixpkgs has no host overlay). |
| `delta-editor/` | Delta Editor (Zed Industries' AI-native editor, `delta.dev`) prebuilt tarball. No stable URL exists — the release API returns a short-lived signed R2 link — so `src` is a fixed-output derivation that resolves the API at build time. Interpreter/RPATH patched (`--force-rpath`, `dontPatchELF`), dlopen'd graphics libs (`libglvnd`, `vulkan-loader`, `wayland`) added via wrapper, `XKB_CONFIG_ROOT` set for the bundled libxkbcommon. Exposed as `pkgs.delta-editor` (`delta-editor` binary + desktop entry / icons); named `-editor` to avoid clashing with the `delta` git diff pager. |
| `llama-cpp/` | llama.cpp pinned to v0.8.0, the first release with DFlash2 drafters (`--spec-type draft-dflash`). nixpkgs only carries v0.6.0 in the unstable input and an older tree in the stable one, so the server is built here. HIP build for Strix Halo (`CMAKE_HIP_ARCHITECTURES=gfx1151`, ROCm clang++, `autoPatchelfHook`). Backend-parameterized: `pkgs.llama-cpp-halo` (rocm, default) and `pkgs.llama-cpp-halo-cpu` (same tree, no GPU backend). Web UI is off — it downloads prebuilt assets at build time, which the sandbox cannot do. Model weights are not packaged. |
## Adding a New Package
+94
View File
@@ -0,0 +1,94 @@
{
lib,
stdenv,
fetchFromGitHub,
cmake,
ninja,
pkg-config,
autoPatchelfHook,
rocmPackages,
# "rocm" builds the gfx1151 engine. "cpu" is the same tree without the GPU
# backend, so the recipe can be checked on a machine with no AMD card.
backend ? "rocm",
rocmArch ? "gfx1151",
}:
# llama.cpp, pinned to the first release that carries DFlash2 drafters
# (v0.8.0, 2026-10-28). nixpkgs only has v0.6.0 in the unstable input and an
# older tree in the stable one, and DFlash2 landed upstream after both — hence
# this package. Built for the Framework Desktop iGPU (Strix Halo, gfx1151).
#
# To bump: set `version` and `rev`, refresh `hash` with
# `nix-prefetch-url --unpack https://github.com/ggml-org/llama.cpp/archive/refs/tags/vX.Y.Z.tar.gz`
# then `nix hash convert --to sri <hash>`. Check the release notes for a new
# `--spec-type` before assuming a drafter still loads.
assert lib.elem backend [
"cpu"
"rocm"
];
let
rocmInputs = [
rocmPackages.clr
rocmPackages.rocm-runtime
rocmPackages.hipblas
rocmPackages.rocblas
];
in
stdenv.mkDerivation {
pname = "llama-cpp-halo" + lib.optionalString (backend == "cpu") "-cpu";
version = "0.8.0";
src = fetchFromGitHub {
owner = "ggml-org";
repo = "llama.cpp";
tag = "v0.8.0";
hash = "sha256-WQAtD8Hos6fEtgPQId7nZiLGyP4asAMKJYEXZid1nqc=";
};
nativeBuildInputs = [
cmake
ninja
pkg-config
]
++ lib.optionals (backend == "rocm") [
autoPatchelfHook
# Provides clang++ and the HIP headers; llama.cpp wants ROCm's own clang.
rocmPackages.clr
];
buildInputs = lib.optionals (backend == "rocm") rocmInputs;
# The web UI is a prebuilt asset downloaded at build time, which the sandbox
# cannot do. The API does not need it.
cmakeFlags = [
(lib.cmakeBool "GGML_NATIVE" false)
(lib.cmakeBool "BUILD_SHARED_LIBS" false)
(lib.cmakeBool "LLAMA_BUILD_SERVER" true)
(lib.cmakeBool "LLAMA_BUILD_TESTS" false)
(lib.cmakeBool "LLAMA_BUILD_EXAMPLES" false)
(lib.cmakeBool "LLAMA_BUILD_APP" false)
(lib.cmakeBool "LLAMA_BUILD_UI" false)
(lib.cmakeBool "LLAMA_USE_PREBUILT_UI" false)
(lib.cmakeBool "LLAMA_OPENSSL" false)
# Upstream reads these from git, which the release tarball does not carry.
(lib.cmakeFeature "LLAMA_BUILD_NUMBER" "31309")
(lib.cmakeFeature "LLAMA_BUILD_COMMIT" "8642541b2")
]
++ lib.optionals (backend == "rocm") [
(lib.cmakeBool "GGML_HIP" true)
(lib.cmakeFeature "CMAKE_HIP_COMPILER" "${rocmPackages.clr.hipClangPath}/clang++")
(lib.cmakeFeature "CMAKE_HIP_ARCHITECTURES" rocmArch)
];
doCheck = false;
meta = {
description = "llama.cpp server (DFlash2 speculative decoding) for AMD Strix Halo (gfx1151)";
homepage = "https://github.com/ggml-org/llama.cpp";
license = lib.licenses.mit;
platforms = lib.platforms.linux;
mainProgram = "llama-server";
};
}