149 lines
5.1 KiB
Nix
149 lines
5.1 KiB
Nix
{ 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
|
|
]
|
|
);
|
|
|
|
# ROCm inputs for the HIP build. hipcc is the raw ROCm compiler, not the
|
|
# nix cc-wrapper, so it needs explicit -I include paths to find the HIP
|
|
# runtime + rocWMMA headers. rocWMMA defines COLI_GPU_HAS_WMMA in
|
|
# backend_gpu_compat.h; without its include dir that macro is undeclared and
|
|
# backend_cuda.cu fails to compile.
|
|
rocmInputs = lib.optionals (backend == "rocm") [
|
|
rocmPackages.clr # provides hipcc + HIP runtime
|
|
rocmPackages.rocm-runtime
|
|
rocmPackages.hipblas
|
|
rocmPackages.hipblas-common
|
|
rocmPackages.hipblaslt
|
|
rocmPackages.rocblas
|
|
rocmPackages.rocwmma # gfx1151 backend uses rocWMMA headers
|
|
rocmPackages.hipcub
|
|
rocmPackages.rocprim
|
|
rocmPackages.rocthrust
|
|
];
|
|
|
|
rocmIncludeFlags =
|
|
lib.concatStringsSep " "
|
|
(map (p: "-I${lib.getDev p}/include") rocmInputs);
|
|
|
|
hipMakeFlags = lib.optionals (backend == "rocm") [
|
|
"HIP=1"
|
|
"HIP_ARCH=${rocmArch}"
|
|
"ROCM_HOME=${rocmPackages.clr}"
|
|
"HIPCC=${rocmPackages.clr}/bin/hipcc"
|
|
];
|
|
|
|
# HIPCCFLAGS is the Makefile's GPU compile flags (GPUFLAGS). It contains
|
|
# spaces, so it must go through makeFlagsArray (a quoted bash array) rather
|
|
# than makeFlags, which the generic builder word-splits. This reproduces the
|
|
# upstream default flags and appends the rocm include paths.
|
|
hipMakeFlagsArray = lib.optionals (backend == "rocm") [
|
|
"HIPCCFLAGS=-O3 -std=c++17 -x hip --offload-arch=${rocmArch} -Wall -Wextra -fPIE ${rocmIncludeFlags}"
|
|
];
|
|
in
|
|
stdenv.mkDerivation {
|
|
pname = "colibri" + lib.optionalString (backend == "rocm") "-rocm";
|
|
version = "2.1.0";
|
|
|
|
src = fetchFromGitHub {
|
|
owner = "JustVugg";
|
|
repo = "colibri";
|
|
rev = "8579957378ca922ad4340a6392513485cbc50c2c"; # tag v1.5.0
|
|
hash = "sha256-wuSLM50nAzcFWDC7fJb3dfAvuq1sc69quT1y0ANq9Pk=";
|
|
};
|
|
|
|
nativeBuildInputs = [ makeWrapper ];
|
|
buildInputs = rocmInputs;
|
|
|
|
# python3 is needed by checkPhase: `make test-c` shells out to
|
|
# `python3 tools/run_tests.py`.
|
|
nativeCheckInputs = [ python3 ];
|
|
|
|
# Portable AVX2 binary by default; ARCH=native for single-host builds.
|
|
ARCH = if stdenv.hostPlatform.isx86_64 then "x86-64-v3" else "native";
|
|
|
|
makeFlags = [ "-C" "c" ] ++ hipMakeFlags;
|
|
|
|
# Pass space-containing make variables as a quoted array so they aren't
|
|
# word-split (see hipMakeFlagsArray note above).
|
|
preBuild = lib.optionalString (hipMakeFlagsArray != [ ]) ''
|
|
makeFlagsArray+=(${lib.escapeShellArgs hipMakeFlagsArray})
|
|
'';
|
|
|
|
buildFlags = [ "colibri" "olmoe" "deepseek-v4" ];
|
|
|
|
# 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";
|
|
};
|
|
}
|