github/gpt4all
1
0
Fork 0
mirror of https://github.com/nomic-ai/gpt4all.git synced 2025-09-05 02:20:28 +00:00
Code Issues Packages Projects Releases Wiki Activity

support the llama.cpp CUDA backend (#2310)

Browse Source 
* rebase onto llama.cpp commit ggerganov/llama.cpp@d46dbc76f
* support for CUDA backend (enabled by default)
* partial support for Occam's Vulkan backend (disabled by default)
* partial support for HIP/ROCm backend (disabled by default)
* sync llama.cpp.cmake with upstream llama.cpp CMakeLists.txt
* changes to GPT4All backend, bindings, and chat UI to handle choice of llama.cpp backend (Kompute or CUDA)
* ship CUDA runtime with installed version
* make device selection in the UI on macOS actually do something
* model whitelist: remove dbrx, mamba, persimmon, plamo; add internlm and starcoder2

Signed-off-by: Jared Van Bortel <jared@nomic.ai>
This commit is contained in:
Jared Van Bortel
2024-05-15 15:27:50 -04:00
committed by GitHub
parent a618ca5699
commit d2a99d9bc6
22 changed files with 1360 additions and 773 deletions
Download Patch File Download Diff File Expand all files Collapse all files

80
gpt4all-backend/CMakeLists.txt
View File

@@ -2,15 +2,23 @@ cmake_minimum_required(VERSION 3.16)
set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
if(APPLE)
option(BUILD_UNIVERSAL "Build a Universal binary on macOS" ON)
if(BUILD_UNIVERSAL)
if (APPLE)
option(BUILD_UNIVERSAL "Build a Universal binary on macOS" ON)
else()
option(LLMODEL_KOMPUTE "llmodel: use Kompute" ON)
option(LLMODEL_VULKAN "llmodel: use Vulkan" OFF)
option(LLMODEL_CUDA "llmodel: use CUDA" ON)
option(LLMODEL_ROCM "llmodel: use ROCm" OFF)
endif()
if (APPLE)
if (BUILD_UNIVERSAL)
# Build a Universal binary on macOS
# This requires that the found Qt library is compiled as Universal binaries.
set(CMAKE_OSX_ARCHITECTURES "arm64;x86_64" CACHE STRING "" FORCE)
else()
# Build for the host architecture on macOS
if(NOT CMAKE_OSX_ARCHITECTURES)
if (NOT CMAKE_OSX_ARCHITECTURES)
set(CMAKE_OSX_ARCHITECTURES "${CMAKE_HOST_SYSTEM_PROCESSOR}" CACHE STRING "" FORCE)
endif()
endif()
@@ -39,11 +47,35 @@ else()
message(STATUS "Interprocedural optimization support detected")
endif()
set(DIRECTORY llama.cpp-mainline)
include(llama.cpp.cmake)
set(BUILD_VARIANTS default avxonly)
if (${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
set(BUILD_VARIANTS ${BUILD_VARIANTS} metal)
set(BUILD_VARIANTS)
set(GPTJ_BUILD_VARIANT cpu)
if (APPLE)
list(APPEND BUILD_VARIANTS metal)
endif()
if (LLMODEL_KOMPUTE)
list(APPEND BUILD_VARIANTS kompute kompute-avxonly)
set(GPTJ_BUILD_VARIANT kompute)
else()
list(PREPEND BUILD_VARIANTS cpu cpu-avxonly)
endif()
if (LLMODEL_VULKAN)
list(APPEND BUILD_VARIANTS vulkan vulkan-avxonly)
endif()
if (LLMODEL_CUDA)
include(CheckLanguage)
check_language(CUDA)
if (NOT CMAKE_CUDA_COMPILER)
message(WARNING "CUDA Toolkit not found. To build without CUDA, use -DLLMODEL_CUDA=OFF.")
endif()
enable_language(CUDA)
list(APPEND BUILD_VARIANTS cuda cuda-avxonly)
endif()
if (LLMODEL_ROCM)
enable_language(HIP)
list(APPEND BUILD_VARIANTS rocm rocm-avxonly)
endif()
set(CMAKE_VERBOSE_MAKEFILE ON)
@@ -51,24 +83,34 @@ set(CMAKE_VERBOSE_MAKEFILE ON)
# Go through each build variant
foreach(BUILD_VARIANT IN LISTS BUILD_VARIANTS)
# Determine flags
if (BUILD_VARIANT STREQUAL avxonly)
set(GPT4ALL_ALLOW_NON_AVX NO)
if (BUILD_VARIANT MATCHES avxonly)
set(GPT4ALL_ALLOW_NON_AVX OFF)
else()
set(GPT4ALL_ALLOW_NON_AVX YES)
set(GPT4ALL_ALLOW_NON_AVX ON)
endif()
set(LLAMA_AVX2 ${GPT4ALL_ALLOW_NON_AVX})
set(LLAMA_F16C ${GPT4ALL_ALLOW_NON_AVX})
set(LLAMA_FMA ${GPT4ALL_ALLOW_NON_AVX})
if (BUILD_VARIANT STREQUAL metal)
set(LLAMA_METAL YES)
else()
set(LLAMA_METAL NO)
set(LLAMA_METAL OFF)
set(LLAMA_KOMPUTE OFF)
set(LLAMA_VULKAN OFF)
set(LLAMA_CUDA OFF)
set(LLAMA_ROCM OFF)
if (BUILD_VARIANT MATCHES metal)
set(LLAMA_METAL ON)
elseif (BUILD_VARIANT MATCHES kompute)
set(LLAMA_KOMPUTE ON)
elseif (BUILD_VARIANT MATCHES vulkan)
set(LLAMA_VULKAN ON)
elseif (BUILD_VARIANT MATCHES cuda)
set(LLAMA_CUDA ON)
elseif (BUILD_VARIANT MATCHES rocm)
set(LLAMA_HIPBLAS ON)
endif()
# Include GGML
set(LLAMA_K_QUANTS YES)
include_ggml(llama.cpp-mainline -mainline-${BUILD_VARIANT} ON)
include_ggml(-mainline-${BUILD_VARIANT})
# Function for preparing individual implementations
function(prepare_target TARGET_NAME BASE_LIB)
@@ -93,11 +135,15 @@ foreach(BUILD_VARIANT IN LISTS BUILD_VARIANTS)
LLAMA_VERSIONS=>=3 LLAMA_DATE=999999)
prepare_target(llamamodel-mainline llama-mainline)
if (NOT LLAMA_METAL)
if (BUILD_VARIANT MATCHES ${GPTJ_BUILD_VARIANT})
add_library(gptj-${BUILD_VARIANT} SHARED
gptj.cpp utils.h utils.cpp llmodel_shared.cpp llmodel_shared.h)
prepare_target(gptj llama-mainline)
endif()
if (BUILD_VARIANT STREQUAL cuda)
set(CUDAToolkit_BIN_DIR ${CUDAToolkit_BIN_DIR} PARENT_SCOPE)
endif()
endforeach()
add_library(llmodel