buzz + update ds4

This commit is contained in:
ari committed 2026-08-05 16:42:57 -04:00
1 parent d48198e199
commit 4cc2123eff
13 files changed
+615 -208

No files matched your search

+1
View File
@@ -246,6 +246,7 @@ in
hermes-agent.packages.${pkgs.stdenv.hostPlatform.system}.default
# claude-desktop.packages.x86_64-linux.default
claude-code
buzz-desktop
]);
+1
View File
@@ -76,6 +76,7 @@ in
HSA_ENABLE_SDMA = "0";
};
openFirewall = true;
autoStart = false;
};
};
networkDrives = {
+2
View File
@@ -21,6 +21,8 @@
ds4-cuda = final.callPackage ../../pkgs/ds4 { backend = "cuda"; };
# 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.
buzz-desktop = final.callPackage ../../pkgs/buzz-desktop { };
})
(final: prev: {
pythonPackagesExtensions = prev.pythonPackagesExtensions ++ [
+1
View File
@@ -30,6 +30,7 @@ let
# need their toolchains and are wired in modules/core/default.nix.
ds4 = final.callPackage ../pkgs/ds4 { };
niri-touchscreen-gestures = final.callPackage ../pkgs/niri-touchscreen-gestures { };
buzz-desktop = final.callPackage ../pkgs/buzz-desktop { };
})
];
}
+316
View File
@@ -0,0 +1,316 @@
{ config, lib, pkgs, ... }:
with lib;
let
cfg = config.skyg.nixos.server.services.colibri;
skygUser = config.skyg.user;
# The colibri package variant to use: colibri (cpu), colibri-rocm.
colibriPkg = cfg.package;
# Build the `coli serve` command line from options.
# optionalString returns "" when false; filter those out so we don't get
# spurious leading/trailing whitespace in the final command line.
serverArgs = with cfg;
concatStringsSep " " (
filter (x: x != "") [
"--model ${escapeShellArg model}"
"--host ${escapeShellArg host}"
"--port ${toString port}"
(optionalString (modelId != null) "--model-id ${escapeShellArg modelId}")
(optionalString (ram != null) "--ram ${toString ram}")
(optionalString (ctx != null) "--ctx ${toString ctx}")
(optionalString (cap != null) "--cap ${toString cap}")
"--ngen ${toString ngen}"
(optionalString (topp != null) "--topp ${toString topp}")
(optionalString (topk != null) "--topk ${toString topk}")
(optionalString (temp != null) "--temp ${toString temp}")
(optionalString (policy != null) "--policy ${policy}")
(optionalString (gpu != null) "--gpu ${escapeShellArg gpu}")
(optionalString (vram != null) "--vram ${toString vram}")
"--max-queue ${toString maxQueue}"
"--queue-timeout ${toString queueTimeout}"
"--kv-slots ${toString kvSlots}"
]
++ map (o: "--cors-origin ${escapeShellArg o}") corsOrigins
++ map (h: "--allowed-host ${escapeShellArg h}") allowedHosts
);
in
{
options = {
skyg.nixos.server.services.colibri = {
enable = mkEnableOption
"colibri-serve, the OpenAI-compatible HTTP API for colibrì (GLM-5.2/OLMoE local inference).";
package = mkPackageOption pkgs "colibri" {
default = pkgs.colibri;
};
user = mkOption {
description = "The user to run coli serve as.";
type = types.str;
default = skygUser.name;
};
group = mkOption {
description = "The group to run coli serve as.";
type = types.str;
default = "users";
};
model = mkOption {
description = "Path to the (converted) model directory.";
type = types.str;
example = "/nvme/glm52_i4";
};
modelMirror = mkOption {
description = ''
Second copy of the model on another drive: expert reads are split
across both SSDs. `coli` only reads this via the COLI_MODEL_MIRROR
environment variable, so it's exported rather than passed as a flag.
'';
default = null;
example = "/nvme2/glm52_i4";
type = types.nullOr types.str;
};
diskWeights = mkOption {
description = "Primary,mirror disk bandwidth ratio (e.g. \"9,3\"); default is measured at startup. Exported as COLI_DISK_WEIGHTS.";
default = null;
example = "9,3";
type = types.nullOr types.str;
};
host = mkOption {
description = "Address the coli serve HTTP API binds to.";
default = "127.0.0.1";
example = "0.0.0.0";
type = types.str;
};
port = mkOption {
description = "TCP port the coli serve HTTP API listens on.";
default = 8000;
type = types.port;
};
modelId = mkOption {
description = "Model name reported by the OpenAI-compatible API (default: derived from the model's architecture).";
default = null;
example = "glm-5.2-colibri";
type = types.nullOr types.str;
};
apiKeyFile = mkOption {
description = ''
Path to a file containing the bearer token clients must send. Read at
service start and exported as COLI_API_KEY (never passed as a CLI
argument, so it doesn't show up in `ps`). Leave null to run without
authentication (fine for host-only binding on 127.0.0.1).
'';
default = null;
example = "/run/agenix/colibri-api-key";
type = types.nullOr types.path;
};
ram = mkOption {
description = "RAM budget in GB (0 or null = auto: the engine uses ~88% of available RAM).";
default = null;
example = 64;
type = types.nullOr types.int;
};
ctx = mkOption {
description = "Maximum context length (number of tokens; null = auto).";
default = null;
example = 32768;
type = types.nullOr types.int;
};
cap = mkOption {
description = "Cache slots per layer (null = auto).";
default = null;
example = 4;
type = types.nullOr types.int;
};
ngen = mkOption {
description = "Maximum response tokens (a safety net; stop tokens usually end generation first).";
default = 1024;
type = types.int;
};
topp = mkOption {
description = "Adaptive expert top-p (null = engine default).";
default = null;
example = 0.9;
type = types.nullOr types.float;
};
topk = mkOption {
description = "Fixed expert top-k (null = engine default).";
default = null;
type = types.nullOr types.int;
};
temp = mkOption {
description = "Sampling temperature (0 = greedy; null = engine default, nucleus ~0.95).";
default = null;
example = 0.7;
type = types.nullOr types.float;
};
policy = mkOption {
description = "Resource policy. Explicit topk/topp still override it.";
default = null;
type = types.nullOr (types.enum [ "quality" "balanced" "experimental-fast" ]);
};
gpu = mkOption {
description = "GPU device selection: \"auto\", \"none\", or a device list such as \"0,1\".";
default = null;
example = "auto";
type = types.nullOr types.str;
};
vram = mkOption {
description = "Total VRAM budget in GB (null = auto).";
default = null;
example = 24;
type = types.nullOr types.int;
};
maxQueue = mkOption {
description = "Maximum queued requests before the API rejects new ones.";
default = 8;
type = types.int;
};
queueTimeout = mkOption {
description = "Seconds a request may wait in queue before timing out.";
default = 300;
type = types.int;
};
kvSlots = mkOption {
description = "KV cache slots (non-GLM architectures currently support exactly 1).";
default = 1;
type = types.int;
};
corsOrigins = mkOption {
description = "Origins allowed by CORS (repeatable). Empty = CORS disabled.";
default = [ ];
example = [ "https://chat.example.com" ];
type = types.listOf types.str;
};
allowedHosts = mkOption {
description = "Additional Host headers accepted by the DNS-rebinding guard (repeatable).";
default = [ ];
example = [ "colibri.lan" ];
type = types.listOf types.str;
};
environment = mkOption {
description = ''
Extra environment variables for the coli serve process. Useful for
GPU backends — e.g. ROCm Strix Halo may need HSA_OVERRIDE_GFX_VERSION.
Example for ROCm on Strix Halo:
{ HSA_OVERRIDE_GFX_VERSION = "11.0.2"; }
'';
default = { };
example = { HSA_OVERRIDE_GFX_VERSION = "11.0.2"; };
type = types.attrsOf types.str;
};
openFirewall = mkOption {
description = "Open the coli serve port in the firewall.";
default = false;
type = types.bool;
};
extraArgs = mkOption {
description = "Additional arguments to append to the coli serve command line.";
default = [ ];
example = [ "--auto-tier" ];
type = types.listOf types.str;
};
extraServiceConfig = mkOption {
description = "Extra systemd service unit config (e.g. SupplementaryGroups, TimeoutStopSec).";
default = { };
example = {
SupplementaryGroups = [ "render" "video" ];
TimeoutStopSec = 120;
};
type = types.attrsOf types.str;
};
};
};
config = mkIf cfg.enable {
assertions = [
{
assertion = cfg.model != "";
message = "skyg.nixos.server.services.colibri: `model` must be set to a model directory path.";
}
];
# Ensure the colibri package is available in the system closure (for manual use).
environment.systemPackages = [ colibriPkg ];
systemd.services.colibri-serve = {
description = "coli serve — colibrì local inference HTTP API (GLM-5.2/OLMoE)";
documentation = [
"https://github.com/JustVugg/colibri"
"https://github.com/JustVugg/colibri#readme"
];
after = [ "network-online.target" ];
wants = [ "network-online.target" ];
wantedBy = [ "multi-user.target" ];
serviceConfig = {
Type = "simple";
User = cfg.user;
Group = cfg.group;
# Construct the ExecStart as a shell script so we can export env vars
# that would otherwise be lost with a raw binary + Environment= (e.g.
# COLI_MODEL_MIRROR, COLI_API_KEY read from a file). Use
# writeShellScript for a clean script in the store.
ExecStart = pkgs.writeShellScript "colibri-serve-start" ''
set -eu
${optionalString (cfg.modelMirror != null) "export COLI_MODEL_MIRROR=${escapeShellArg cfg.modelMirror}"}
${optionalString (cfg.diskWeights != null) "export COLI_DISK_WEIGHTS=${escapeShellArg cfg.diskWeights}"}
${optionalString (cfg.apiKeyFile != null) ''export COLI_API_KEY="$(cat ${escapeShellArg cfg.apiKeyFile})"''}
${concatStringsSep "\n" (mapAttrsToList (name: value: "export ${name}=${escapeShellArg value}") cfg.environment)}
exec ${colibriPkg}/bin/coli serve ${serverArgs} ${escapeShellArgs cfg.extraArgs}
'';
# No custom ExecStop: `coli serve` runs in the foreground and systemd's
# default cgroup-wide kill on stop takes down its spawned engine
# subprocess with it (same reasoning as ds4-server).
Restart = "on-failure";
RestartSec = 10;
StartLimitBurst = 3;
# Security hardening
NoNewPrivileges = true;
ProtectHome = true;
ProtectSystem = "full";
PrivateTmp = true;
CapabilityBoundingSet = "";
SystemCallFilter = "@system-service @resources";
} // cfg.extraServiceConfig;
};
networking.firewall =
mkIf cfg.openFirewall { allowedTCPPorts = [ cfg.port ]; };
};
}
@@ -3,6 +3,7 @@
imports = [
./audiobookshelf.nix
./ai-services.nix
./colibri.nix
./comfyui
./ds4.nix
./jellyfin.nix
+10 -1
View File
@@ -143,6 +143,15 @@ in
type = types.bool;
};
autoStart = mkOption {
description = ''
Whether to start ds4-server automatically at boot.
Set to false to only start it on demand via `systemctl start ds4-server`.
'';
default = true;
type = types.bool;
};
extraArgs = mkOption {
description = "Additional arguments to append to the ds4-server command line.";
default = [ ];
@@ -182,7 +191,7 @@ in
after = [ "network-online.target" ];
wants = [ "network-online.target" ];
wantedBy = [ "multi-user.target" ];
wantedBy = lib.optionals cfg.autoStart [ "multi-user.target" ];
serviceConfig = {
Type = "simple";
+1
View File
@@ -10,6 +10,7 @@ Packages here are exposed via an overlay defined in `modules/core/default.nix`.
| `grok-cli/` | xAI Grok CLI — prebuilt static binary, fetched per-platform. Exposed as `pkgs.grok-cli`. |
| `ds4/` | DwarfStar (`antirez/ds4`) DeepSeek V4 local inference engine, built from source. Backend-parameterized (`backend = "cpu" \| "rocm" \| "cuda"`); tracks `main` and builds per-host. Exposed as `pkgs.ds4` (cpu, default), `pkgs.ds4-rocm` (Strix Halo / gfx1151), `pkgs.ds4-cuda`. Installs the `ds4`, `ds4-server`, `ds4-bench`, `ds4-eval`, `ds4-agent` binaries plus a `ds4-download-model` helper for fetching GGUF weights. Model weights are a runtime concern and not packaged. |
| `niri-touchscreen-gestures/` | Python daemon that reads multi-touch events from a touchscreen via `evdev`, detects 2/3/4-finger swipes (up/down/left/right), and dispatches configurable `niri msg action` commands. Configured via TOML; exposed as `pkgs.niri-touchscreen-gestures`. |
| `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). |
## Adding a New Package
+128
View File
@@ -0,0 +1,128 @@
{ lib
, fetchurl
, appimageTools
, runCommand
}:
# Buzz Desktop — self-hostable workspace for humans + AI agents (block/buzz).
#
# Upstream ships a Type-2 AppImage. On NixOS we need:
# 1. FHS libs missing from appimageTools.defaultFhsEnvArgs (elfutils, zstd, …)
# 2. WebKit DMA-BUF disabled — Buzz only auto-enables this when it detects a
# real AppImage (`APPIMAGE` env). wrapAppImage runs the extracted AppDir.
# 3. Explicit GST_PLUGIN_SYSTEM_PATH_1_0 for WebKit media elements.
# 4. Drop the upstream AppRun hook that hardcodes `GDK_BACKEND=x11` (breaks
# Wayland clipboard under niri).
# 5. Drop AppImage-bundled OpenSSL 3.0 — AppRun prepends $APPDIR/usr/lib to
# LD_LIBRARY_PATH, so that old libssl shadows nixpkgs OpenSSL 3.2+ and
# breaks GStreamer plugins / curl (`OPENSSL_3.2.0' not found`).
#
# To bump: set `version`, then:
# nix hash file --type sha256 --sri <(curl -fsSL \
# https://github.com/block/buzz/releases/download/desktop-v${version}/Buzz_${version}_amd64.AppImage)
let
pname = "buzz-desktop";
version = "0.5.5";
src = fetchurl {
url = "https://github.com/block/buzz/releases/download/desktop-v${version}/Buzz_${version}_amd64.AppImage";
hash = "sha256-zFHK2mN9YZcSHpXwgyisGcu/7t0+mSIotVLPQ4k+K90=";
};
extracted = appimageTools.extractType2 {
inherit pname version src;
};
appdir = runCommand "${pname}-appdir" { } ''
cp -a ${extracted}/. "$out"
chmod -R u+w "$out"
# Prefer Wayland (clipboard) over hard-coded X11 from linuxdeploy.
hook="$out/apprun-hooks/linuxdeploy-plugin-gtk.sh"
if [ -f "$hook" ]; then
sed -i \
-e 's/^export GDK_BACKEND=x11/# export GDK_BACKEND=x11 # nixos: allow Wayland for clipboard/' \
"$hook"
fi
# AppRun.wrapped forces LD_LIBRARY_PATH=$APPDIR/usr/lib first. The bundled
# libssl/libcrypto are OpenSSL 3.0.x and lack OPENSSL_3.2.0, which nixpkgs
# curl / gstreamer plugins need. Remove them so the FHS OpenSSL is used
# (libssl.so.3 soname is ABI-compatible for this use).
rm -f \
"$out/usr/lib"/libssl.so \
"$out/usr/lib"/libssl.so.* \
"$out/usr/lib"/libcrypto.so \
"$out/usr/lib"/libcrypto.so.*
'';
in
appimageTools.wrapAppImage {
inherit pname version;
src = appdir;
extraPkgs = pkgs: with pkgs; [
elfutils # libelf.so.1
zstd # libzstd.so.1
openssl # replaces removed AppImage OpenSSL
curl
webkitgtk_4_1
libsoup_3
libayatana-appindicator
# WebKit media: appsink/appsrc (base), autoaudiosink (good)
gst_all_1.gstreamer
gst_all_1.gst-plugins-base
gst_all_1.gst-plugins-good
gst_all_1.gst-plugins-bad
gst_all_1.gst-plugins-ugly
gst_all_1.gst-libav
# Clipboard / desktop integration
wl-clipboard
xdg-utils
];
profile = ''
# WebKit blank-window workaround (Buzz sets this itself only for real AppImages).
export WEBKIT_DISABLE_DMABUF_RENDERER=1
export WEBKIT_DISABLE_COMPOSITING_MODE=1
# Do not force X11 — Wayland clipboard with niri.
unset GDK_BACKEND
export GTK_USE_PORTAL=1
# GStreamer plugins from the FHS root (after AppDir OpenSSL removal they
# can load nixpkgs plugins that need OpenSSL 3.2+).
export GST_PLUGIN_SYSTEM_PATH_1_0="/usr/lib/gstreamer-1.0''${GST_PLUGIN_SYSTEM_PATH_1_0:+:$GST_PLUGIN_SYSTEM_PATH_1_0}"
export GST_PLUGIN_PATH="/usr/lib/gstreamer-1.0''${GST_PLUGIN_PATH:+:$GST_PLUGIN_PATH}"
'';
extraInstallCommands = ''
install -Dm444 ${appdir}/usr/share/applications/Buzz.desktop \
$out/share/applications/buzz-desktop.desktop
substituteInPlace $out/share/applications/buzz-desktop.desktop \
--replace-fail 'Exec=buzz-desktop' 'Exec=${pname}' \
--replace-fail 'Icon=buzz-desktop' 'Icon=${pname}'
for size in 32x32 128x128 '256x256@2'; do
if [ -f ${appdir}/usr/share/icons/hicolor/$size/apps/buzz-desktop.png ]; then
install -Dm644 ${appdir}/usr/share/icons/hicolor/$size/apps/buzz-desktop.png \
$out/share/icons/hicolor/$size/apps/${pname}.png
fi
done
if [ ! -e $out/share/icons/hicolor/128x128/apps/${pname}.png ] \
&& [ -f ${appdir}/buzz-desktop.png ]; then
install -Dm644 ${appdir}/buzz-desktop.png \
$out/share/icons/hicolor/128x128/apps/${pname}.png
fi
'';
meta = {
description = "Buzz desktop — self-hostable workspace for humans and AI agents";
homepage = "https://github.com/block/buzz";
changelog = "https://github.com/block/buzz/releases/tag/desktop-v${version}";
license = lib.licenses.asl20;
sourceProvenance = [ lib.sourceTypes.binaryNativeCode ];
platforms = [ "x86_64-linux" ];
mainProgram = "buzz-desktop";
};
}
+115
View File
@@ -0,0 +1,115 @@
{ lib
, stdenv
, fetchFromGitHub
, makeWrapper
, python3
, rocmPackages
, backend ? "cpu" # one of "cpu" | "rocm"
, rocmArch ? "gfx1151" # AMD Strix Halo (Framework Desktop iGPU)
}:
# colibrì (JustVugg/colibri) — run GLM-5.2 (744B MoE) and other frontier MoE
# models locally from a pure-C engine with experts streamed from disk.
# Upstream ships its own `make install` target (installs `coli` to bin/ and
# the engine binaries + support .py modules to libexec/colibri/), and `coli`
# already knows how to find libexec/colibri/ relative to its own install
# path — so we lean on that instead of re-deriving the file layout by hand.
#
# This derivation is backend-parameterized like pkgs/ds4: pick the GPU
# backend with `backend` ("cpu" | "rocm"). The overlay in
# modules/core/default.nix wires up `colibri` (cpu) and `colibri-rocm`.
#
# To bump: set `rev` to the new tag's commit and refresh `hash` (start from
# lib.fakeHash and copy the SRI hash from the build error).
assert lib.elem backend [ "cpu" "rocm" ];
let
# Only `coli convert` / `coli bench` need these; chat/serve just exec the
# compiled engine binary.
pythonEnv = python3.withPackages (
ps: with ps; [
torch
safetensors
huggingface-hub
numpy
tokenizers
datasets
]
);
hipMakeFlags = lib.optionals (backend == "rocm") [
"HIP=1"
"HIP_ARCH=${rocmArch}"
"ROCM_HOME=${rocmPackages.clr}"
"HIPCC=${rocmPackages.clr}/bin/hipcc"
];
in
stdenv.mkDerivation {
pname = "colibri" + lib.optionalString (backend == "rocm") "-rocm";
version = "1.5.0";
src = fetchFromGitHub {
owner = "JustVugg";
repo = "colibri";
rev = "8f512fc8c2f48ffa18cd624cd4a5bcaae4a4abfc"; # tag v1.5.0
hash = lib.fakeHash; # replace with the real hash on first build attempt
};
nativeBuildInputs = [ makeWrapper ];
buildInputs = lib.optionals (backend == "rocm") [
rocmPackages.clr
rocmPackages.rocm-runtime
rocmPackages.hipblas
];
# python3 is needed by checkPhase: `make test-c` shells out to
# `python3 tools/run_tests.py`.
nativeCheckInputs = [ python3 ];
# Portable AVX2 binary by default; ARCH=native for single-host builds.
ARCH = if stdenv.hostPlatform.isx86_64 then "x86-64-v3" else "native";
makeFlags = [ "-C" "c" ] ++ hipMakeFlags;
buildFlags = [ "colibri" "olmoe" ];
# Use upstream's own install target instead of hand-copying files — it
# installs `coli` to bin/ and the engine + support modules to
# libexec/colibri/, which `coli` already knows how to find on its own.
installPhase = ''
runHook preInstall
make -C c install PREFIX=$out ARCH="$ARCH" ${lib.escapeShellArgs hipMakeFlags}
runHook postInstall
'';
# coli ships `#!/usr/bin/env python3`; point env at pythonEnv so the
# converter/bench subcommands resolve torch etc. No COLI_ENGINE/PYTHONPATH
# override needed — coli auto-detects bin/ + libexec/colibri/ relative to
# its own installed path.
postFixup = ''
wrapProgram $out/bin/coli --prefix PATH : ${lib.makeBinPath [ pythonEnv ]}
'';
# rocm skips: test-c would relink the engine without HIP=1, clobbering the
# GPU build, and needs real hardware anyway.
doCheck = backend == "cpu";
checkPhase = ''
runHook preCheck
cd c
make test-c
cd ..
runHook postCheck
'';
meta = {
description = "Run GLM-5.2 (744B MoE) and other frontier MoE models on hardware you already own";
homepage = "https://github.com/JustVugg/colibri";
license = lib.licenses.asl20;
sourceProvenance = [ lib.sourceTypes.fromSource ];
platforms =
if backend == "rocm"
then lib.platforms.linux
else lib.platforms.linux ++ lib.platforms.darwin;
mainProgram = "coli";
};
}
-194
View File
@@ -1,194 +0,0 @@
{
description = "colibrì — run GLM-5.2 (744B MoE) on a consumer machine with ~25 GB RAM";
# Reproducibility: these inputs track a branch, so torch/numpy/etc. float across
# rebuilds. For deterministic builds run `nix flake lock` once and COMMIT the
# generated flake.lock (it pins each input to a commit SHA). (#D2)
inputs = {
nixpkgs.url = "github:NixOS/nixpkgs/nixos-26.05";
flake-utils.url = "github:numtide/flake-utils";
};
outputs =
{ self
, nixpkgs
, flake-utils
,
}:
flake-utils.lib.eachDefaultSystem (
system:
let
pkgs = import nixpkgs { inherit system; };
# Python with the packages needed by the offline converter tools
pythonEnv = pkgs.python3.withPackages (
ps:
with ps; [
torch
safetensors
huggingface-hub
numpy
tokenizers
datasets
]
);
colibri = pkgs.stdenv.mkDerivation {
pname = "colibri";
version = "1.0";
src = ./.;
nativeBuildInputs = with pkgs; [ makeWrapper ];
buildInputs = with pkgs; [
gcc
gmp
];
# python3 is needed by checkPhase: `make test-c` shells out to
# `python3 tools/run_tests.py` (see c/Makefile, PYTHON ?= python3).
nativeCheckInputs = with pkgs; [ python3 ];
# Use x86-64-v3 (AVX2) for a portable binary; override with ARCH=native for local builds
ARCH =
if pkgs.stdenv.hostPlatform.isx86_64
then "x86-64-v3"
else "native";
buildPhase = ''
runHook preBuild
make -C c glm ARCH="$ARCH"
runHook postBuild
'';
installPhase = ''
runHook preInstall
# Self-contained layout under $out/lib/colibri that mirrors the
# source tree `coli` runs in (see the path-resolution logic at the
# top of c/coli): the engine, the coli CLI script, the support
# modules it imports (openai_server.py, resource_plan.py,
# doctor.py), and tools/ all sit next to each other.
mkdir -p $out/lib/colibri/tools $out/bin
cp c/glm $out/lib/colibri/glm
cp c/coli $out/lib/colibri/coli
chmod +x $out/lib/colibri/coli
cp c/openai_server.py c/resource_plan.py c/doctor.py c/version.py \
$out/lib/colibri/
cp -r c/tools/* $out/lib/colibri/tools/
# $out/bin holds the user-facing entry points.
ln -s ../lib/colibri/glm $out/bin/glm
# Wrap coli: point it at the bundled engine (COLI_ENGINE) so it is
# found by default, and at the module dir (PYTHONPATH) so
# `import openai_server` / `resource_plan` / `doctor` resolve.
makeWrapper ${pythonEnv}/bin/python $out/bin/coli \
--add-flags "$out/lib/colibri/coli" \
--set-default COLI_ENGINE "$out/lib/colibri/glm" \
--set PYTHONPATH "$out/lib/colibri:${pythonEnv}/${pkgs.python3.sitePackages}"
runHook postInstall
'';
checkPhase = ''
runHook preCheck
cd c
make test-c
cd ..
runHook postCheck
'';
doCheck = true;
meta = with pkgs.lib; {
description = "Run GLM-5.2 (744B MoE) on a consumer machine with ~25 GB RAM";
homepage = "https://github.com/JustVugg/colibri";
license = licenses.asl20;
platforms = with platforms; linux ++ darwin;
mainProgram = "coli";
};
};
# ROCm/HIP variant: same package as `colibri`, plus the optional GPU
# expert tier built for AMD via the engine's HIP=1 path (see c/Makefile).
# Tuned for the Framework Desktop's Radeon 8060S/8050S iGPU == gfx1151.
# Linux-only (HIP=1 errors out on non-Linux in the Makefile).
#
# nix build .#colibri-rocm
# COLI_MODEL=/nvme/glm52_i4 CUDA_EXPERT_GB=auto nix run .#colibri-rocm -- chat
#
# If a gfx1151 kernel misbehaves on your ROCm version, run with
# HSA_OVERRIDE_GFX_VERSION=11.0.0
colibri-rocm = colibri.overrideAttrs (old: {
pname = "colibri-rocm";
buildInputs =
old.buildInputs
++ (with pkgs.rocmPackages; [ clr rocm-runtime hipblas ]);
buildPhase = ''
runHook preBuild
make -C c glm \
ARCH="$ARCH" \
HIP=1 \
HIP_ARCH="gfx1151" \
ROCM_HOME="${pkgs.rocmPackages.clr}" \
HIPCC="${pkgs.rocmPackages.clr}/bin/hipcc"
runHook postBuild
'';
# test-c would relink the engine WITHOUT HIP=1 (clobbering the GPU
# build), and the GPU path needs real hardware — so skip the suite.
doCheck = false;
meta = old.meta // { platforms = pkgs.lib.platforms.linux; };
});
in
{
packages =
{
default = colibri;
inherit colibri;
}
# Only expose the ROCm package on Linux (HIP build is Linux-only).
// pkgs.lib.optionalAttrs pkgs.stdenv.hostPlatform.isLinux {
inherit colibri-rocm;
};
apps = {
default = {
type = "app";
program = pkgs.lib.getExe colibri;
};
glm = {
type = "app";
program = "${colibri}/share/colibri/glm";
};
};
formatter = pkgs.alejandra;
devShells.default = pkgs.mkShell {
inputsFrom = [ colibri ];
packages = with pkgs; [
pythonEnv
gcc
gnumake
clang-tools # clangd / clang-tidy for IDE support
pkg-config
];
shellHook = ''
echo "🐦 colibrì dev shell"
echo " gcc: $(gcc --version | head -1)"
echo " python: $(python3 --version)"
echo ""
echo "Build the engine: make -C c glm"
echo "Run the converter: python c/coli convert --model /path/to/glm52_i4"
echo "Chat: COLI_MODEL=/path/to/glm52_i4 ./c/glm ..."
'';
};
}
);
}
+3 -3
View File
@@ -111,13 +111,13 @@ let
in
stdenv.mkDerivation (finalAttrs: {
pname = "ds4" + lib.optionalString (backend != "cpu") "-${backend}";
version = "0-unstable-2026-06-17";
version = "0-unstable-2026-08-05";
src = fetchFromGitHub {
owner = "antirez";
repo = "ds4";
rev = "80ebbc396aee40eedc1d829222f3362d10fa4c6c";
hash = "sha256-Ieuc72GHZs20ModQfnvI5Me31n4Pj+WFYtsuqaKJceo=";
rev = "b0309611041655f4e45671cfd9c9886aff161406";
hash = "sha256-yBPQqX8oI9fElGiXfz72iWwIS8ZQpK4Bl9TWLZSa6JU=";
};
enableParallelBuilding = true;
+36 -10
View File
@@ -24,23 +24,42 @@ OpenAI-compatible client at.
## 1. Get the weights
Weights are **not** packaged. They live at the project's Hugging Face repo
[`antirez/deepseek-v4-gguf`](https://huggingface.co/antirez/deepseek-v4-gguf)
(public, MIT, no token needed) and only these custom GGUFs work — not arbitrary
DeepSeek GGUFs.
Weights are **not** packaged. They live at the project's Hugging Face repos
([`antirez/deepseek-v4-gguf`](https://huggingface.co/antirez/deepseek-v4-gguf)
for DeepSeek V4, [`antirez/GLM-5.2-GGUF`](https://huggingface.co/antirez/GLM-5.2-GGUF)
and [`unsloth/GLM-5.2-GGUF`](https://huggingface.co/unsloth/GLM-5.2-GGUF) for
GLM 5.2) — public, MIT, no token needed — and only these custom GGUFs work, not
arbitrary DeepSeek/GLM GGUFs.
`ds4-download-model` writes into `./gguf` and links `./ds4flash.gguf` in the
current directory (override the root with `DS4_HOME`, or the gguf dir with
`DS4_GGUF_DIR`):
```sh
ds4-download-model q2-imatrix # ~81 GB, the pick for 96/128 GB machines
ds4-download-model --help # all targets (q4-imatrix, pro-*, mtp, ...)
# DeepSeek V4 Flash (0731 checkpoints)
ds4-download-model ds4f-q2 # ~81 GB, the pick for 96/128 GB machines
ds4-download-model ds4f-q2-q4 # ~98 GB, last 6 expert layers at q4
ds4-download-model ds4f-q4 # ~153 GB, for >=256 GB machines
ds4-download-model ds4f-mxfp4 # ~156 GB, native MXFP4 (Metal/CUDA)
ds4-download-model ds4f-dspark # ~6 GB, optional DSpark speculative draft
# DeepSeek V4 PRO (512 GB / distributed)
ds4-download-model pro-q2-imatrix # single-file PRO q2
ds4-download-model pro-q4-split # both halves of PRO Q4 split
# GLM 5.2
ds4-download-model glm-unsloth-q4 # Unsloth UD-Q4_K_XL, 11 shards
ds4-download-model glm-antirez-iq2xxs # routed IQ2_XXS, reduced-memory
ds4-download-model glm-antirez-q2 # routed Q2_K, ~262 GB
ds4-download-model glm-antirez-q4 # routed Q4_K, ~434 GB
ds4-download-model --help # full target list
```
Pick by machine memory: `q2-imatrix` (128 GB), `q4-imatrix` (≥256 GB),
`pro-*` (512 GB / distributed). The small/Flash quants download via `curl`; the
huge **PRO** files need the `hf` CLI (not bundled — `nix shell nixpkgs#huggingface-hub`).
Pick by machine memory: `ds4f-q2` (96/128 GB), `ds4f-q4` (>=256 GB),
`pro-*` (512 GB / distributed), `glm-*` (per-target sizes above). The small
Flash quants download via `curl`; the MXFP4, PRO, and GLM files need the `hf`
CLI (not bundled — `nix shell nixpkgs#huggingface-hub`).
## 2. Run it
@@ -49,7 +68,14 @@ huge **PRO** files need the `hf` CLI (not bundled — `nix shell nixpkgs#hugging
ds4
# explicit model path
ds4 -m /path/to/gguf/DeepSeek-V4-Flash-IQ2XXS-...-imatrix.gguf
ds4 -m /path/to/gguf/DeepSeek-V4-Flash-IQ2XXS-...-imatrix-0731.gguf
# DSpark speculative decoding (greedy only; needs the ds4f-dspark support GGUF)
ds4 -m ./ds4flash.gguf \
--dspark --mtp ./gguf/DeepSeek-V4-Flash-DSpark-support-0731.gguf --temp 0
# GLM 5.2 (experimental greedy MTP via --glm-mtp)
ds4 -m gguf/GLM-5.2-UD-IQ2_XXS_RoutedIQ2XXS_blk78Q2K.gguf --glm-mtp-timing --temp 0
# HTTP server on the LAN
ds4-server -m ./ds4flash.gguf \