Archives/gpt4all
2
0
Fork 0
mirror of https://github.com/nomic-ai/gpt4all synced 2024-11-08 07:10:32 +00:00
Code Issues Packages Projects Releases Wiki Activity

backend(build): set CUDA arch defaults before enable_language(CUDA) (#2855)

Browse Source 
Signed-off-by: Jared Van Bortel <jared@nomic.ai>
This commit is contained in:
Jared Van Bortel 2024-08-13 14:47:48 -04:00 committed by GitHub
parent 971c83d1d3
commit 7463b2170b
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 20 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
gpt4all-backend/CMakeLists.txt
View File

@ -63,6 +63,24 @@ if (LLMODEL_VULKAN)
list(APPEND BUILD_VARIANTS vulkan vulkan-avxonly)
endif()
if (LLMODEL_CUDA)
cmake_minimum_required(VERSION 3.18) # for CMAKE_CUDA_ARCHITECTURES
# Defaults must be set before enable_language(CUDA).
# Keep this in sync with the arch list in ggml/src/CMakeLists.txt.
if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
# 52 == lowest CUDA 12 standard
# 60 == f16 CUDA intrinsics
# 61 == integer CUDA intrinsics
# 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75") # needed for f16 CUDA intrinsics
else()
set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75") # lowest CUDA 12 standard + lowest for integer intrinsics
#set(CMAKE_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
endif()
endif()
message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
include(CheckLanguage)
check_language(CUDA)
if (NOT CMAKE_CUDA_COMPILER)

14
gpt4all-backend/llama.cpp.cmake
View File

@ -378,19 +378,7 @@ function(include_ggml SUFFIX)
find_package(CUDAToolkit REQUIRED)
set(CUDAToolkit_BIN_DIR ${CUDAToolkit_BIN_DIR} PARENT_SCOPE)
if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
# 52 == lowest CUDA 12 standard
# 60 == f16 CUDA intrinsics
# 61 == integer CUDA intrinsics
# 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75") # needed for f16 CUDA intrinsics
else()
set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75") # lowest CUDA 12 standard + lowest for integer intrinsics
#set(CMAKE_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
endif()
endif()
message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
# architectures are set in gpt4all-backend/CMakeLists.txt
set(GGML_HEADERS_CUDA ${DIRECTORY}/ggml/include/ggml-cuda.h)
file(GLOB GGML_HEADERS_CUDA "${DIRECTORY}/ggml/src/ggml-cuda/*.cuh")

1
gpt4all-chat/CHANGELOG.md
View File

@ -9,6 +9,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
### Fixed
- Do not initialize Vulkan driver when only using CPU ([#2843](https://github.com/nomic-ai/gpt4all/pull/2843))
- Fix a potential crash on exit when using only CPU on Linux with NVIDIA (does not affect X11) ([#2843](https://github.com/nomic-ai/gpt4all/pull/2843))
- Fix default CUDA architecture list after [#2802](https://github.com/nomic-ai/gpt4all/pull/2802) ([#2855](https://github.com/nomic-ai/gpt4all/pull/2855))
## [3.2.0] - 2024-08-12