vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

flake.nix: unclutter darwin support

Browse source
This commit is contained in:
Someone Serge 2023-12-24 19:35:32 +00:00 • committed by Philip Taron
parent b28426a952
commit d73272f18b
2 changed files with 26 additions and 40 deletions
Download patch file Download diff file Expand all files Collapse all files

14
.devops/nix/overlay.nix
View file

@ -1,17 +1,5 @@
final: prev: final: prev:
let
inherit (final.stdenv) isAarch64 isDarwin;
darwinSpecific =
if isAarch64 then
{ inherit (final.darwin.apple_sdk_11_0.frameworks) Accelerate MetalKit; }
else
{ inherit (final.darwin.apple_sdk.frameworks) Accelerate CoreGraphics CoreVideo; };
osSpecific = if isDarwin then darwinSpecific else { };
in
{ {
llama-cpp = final.callPackage ./package.nix osSpecific; llama-cpp = final.callPackage ./package.nix { };
} }

52
.devops/nix/package.nix
View file

@ -11,14 +11,18 @@
mpi, mpi,
openblas, # TODO: Use the generic `blas` so users could switch betwen alternative implementations openblas, # TODO: Use the generic `blas` so users could switch betwen alternative implementations
cudaPackages, cudaPackages,
darwin,
rocmPackages, rocmPackages,
clblast, clblast,
Accelerate ? null, useBlas ? builtins.all (x: !x) [
MetalKit ? null, useCuda
CoreVideo ? null, useMetalKit
CoreGraphics ? null, useOpenCL
useOpenCL ? false, useRocm
],
useCuda ? config.cudaSupport, useCuda ? config.cudaSupport,
useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL,
useOpenCL ? false,
useRocm ? config.rocmSupport, useRocm ? config.rocmSupport,
}@inputs: }@inputs:
@ -29,7 +33,6 @@ let
optionals optionals
versionOlder versionOlder
; ;
isDefault = !useOpenCL && !useCuda && !useRocm;
# It's necessary to consistently use backendStdenv when building with CUDA support, # It's necessary to consistently use backendStdenv when building with CUDA support,
# otherwise we get libstdc++ errors downstream. # otherwise we get libstdc++ errors downstream.
@ -44,7 +47,7 @@ let
" (CUDA accelerated)" " (CUDA accelerated)"
else if useRocm then else if useRocm then
" (ROCm accelerated)" " (ROCm accelerated)"
else if (MetalKit != null) then else if useMetalKit then
" (MetalKit accelerated)" " (MetalKit accelerated)"
else else
""; "";
@ -70,13 +73,16 @@ let
] ]
); );
# See ./overlay.nix for where these dependencies are passed in. # apple_sdk is supposed to choose sane defaults, no need to handle isAarch64
defaultBuildInputs = builtins.filter (p: p != null) [ # separately
Accelerate darwinBuildInputs =
MetalKit with darwin.apple_sdk.frameworks;
CoreVideo [ Accelerate ]
CoreGraphics ++ optionals useMetalKit [ MetalKit ]
]; ++ optionals (!useMetalKit) [
CoreVideo
CoreGraphics
];
cudaBuildInputs = with cudaPackages; [ cudaBuildInputs = with cudaPackages; [
cuda_cccl.dev # <nv/target> cuda_cccl.dev # <nv/target>
@ -121,7 +127,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
++ optionals useOpenCL [ clblast ] ++ optionals useOpenCL [ clblast ]
++ optionals useCuda cudaBuildInputs ++ optionals useCuda cudaBuildInputs
++ optionals useRocm rocmBuildInputs ++ optionals useRocm rocmBuildInputs
++ optionals isDefault defaultBuildInputs; ++ optionals effectiveStdenv.isDarwin darwinBuildInputs;
cmakeFlags = cmakeFlags =
[ [
@ -129,6 +135,8 @@ effectiveStdenv.mkDerivation (finalAttrs: {
(cmakeBool "LLAMA_BUILD_SERVER" true) (cmakeBool "LLAMA_BUILD_SERVER" true)
(cmakeBool "BUILD_SHARED_LIBS" true) (cmakeBool "BUILD_SHARED_LIBS" true)
(cmakeBool "CMAKE_SKIP_BUILD_RPATH" true) (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
(cmakeBool "LLAMA_METAL" useMetalKit)
(cmakeBool "LLAMA_BLAS" useBlas)
] ]
++ optionals useOpenCL [ (cmakeBool "LLAMA_CLBLAST" true) ] ++ optionals useOpenCL [ (cmakeBool "LLAMA_CLBLAST" true) ]
++ optionals useCuda [ (cmakeBool "LLAMA_CUBLAS" true) ] ++ optionals useCuda [ (cmakeBool "LLAMA_CUBLAS" true) ]
@ -143,18 +151,8 @@ effectiveStdenv.mkDerivation (finalAttrs: {
# Should likely use `rocmPackages.clr.gpuTargets`. # Should likely use `rocmPackages.clr.gpuTargets`.
"-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102" "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
] ]
++ optionals isDefault ( ++ optionals useMetalKit [ (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1") ]
if (MetalKit != null) then ++ optionals useBlas [ (lib.cmakeFeature "LLAMA_BLAS_VENDOR" "OpenBLAS") ];
[
"-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
"-DLLAMA_METAL=ON"
]
else
[
"-DLLAMA_BLAS=ON"
"-DLLAMA_BLAS_VENDOR=OpenBLAS"
]
);
# TODO(SomeoneSerge): It's better to add proper install targets at the CMake level, # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
# if they haven't been added yet. # if they haven't been added yet.