nanakura/moeka-project
1
0
Fork 0
Code Issues Pull Requests 1 Actions 1 Packages Projects Releases Wiki Activity

fix(ci): bring back CUDA to build system, but with better mechanism and settings

Browse Source
This commit is contained in:
Neko Ayaka
2025-07-09 19:33:55 +08:00
parent 6ed10eb48c
commit 86ecc6c784
1 changed files with 42 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/release-tamagotchi.yml
+42 -7
View File
@@ -77,11 +77,11 @@ jobs:
sudo apt-get update
sudo apt-get install -y libwebkit2gtk-4.1-dev libappindicator3-dev librsvg2-dev patchelf
# - name: Install CUDA Toolkit (Linux and Windows Only)
# uses: Jimver/cuda-toolkit@v0.2.24
# if: matrix.os == 'ubuntu-latest' || matrix.os == 'ubuntu-24.04-arm' || matrix.os == 'windows-latest'
# with:
# cuda: 12.5.0
- name: Install CUDA Toolkit (Linux and Windows Only)
uses: Jimver/cuda-toolkit@v0.2.24
if: matrix.os == 'ubuntu-latest' || matrix.os == 'ubuntu-24.04-arm' || matrix.os == 'windows-latest'
with:
cuda: 12.5.0
- name: Install dependencies
run: pnpm install --frozen-lockfile
@@ -95,8 +95,43 @@ jobs:
- name: Build Application (Linux and Windows Only)
if: matrix.os == 'ubuntu-latest' || matrix.os == 'ubuntu-24.04-arm' || matrix.os == 'windows-latest'
# run: cd apps/stage-tamagotchi && pnpm tauri build --target ${{ matrix.target }} --features cuda
run: cd apps/stage-tamagotchi && pnpm tauri build --target ${{ matrix.target }}
run: cd apps/stage-tamagotchi && pnpm tauri build --target ${{ matrix.target }} --features cuda
env:
# Call to `nvcc` (which part of the CUDA Toolkit) doesn't require a physical GPU, yet bindgen_cuda requires `nvidia-smi`
# to be available in order to determine the compute capability of the GPU.[^1]
#
# And bindgen_cuda is what candle depends on, to bypass the Compute compatibility here,
# we need to explicitly setting CUDA_COMPUTE_CAP as one of the environment varaible in order to
# build without any actual physical GPUs inside of the GitHub Actions runner.[^2]
#
# About how bindgen_cuda depends on nvidia-smi, check the actual source code here:
# https://github.com/Narsil/bindgen_cuda/blob/a6b0c891be8ebefb55600d46d8a77a358adc3913/src/lib.rs#L477-L500
#
# Similar information were seen from TEI (text-embedding-inference) documentations
# https://huggingface.co/docs/text-embeddings-inference/en/custom_container
#
# To understand what compute capability is required, visit [CUDA GPU Compute Capability](https://developer.nvidia.com/cuda-gpus)
#
# In short:
# | Compute Capability | GeForce / RTX |
# |---------------------|----------------|
# | 7.5 | GeForce GTX 1650 Ti, NVIDIA TITAN RTX, GeForce RTX 2000 series... |
# | 8.6 | RTX A series, GeForce RTX 3000 series |
# | 8.9 | GeForce RTX 4000 series, RTX A Ada series |
# | 9.0 | GeForce RTX 5000 series, RTX Pro Blackwell series |
#
# Also, the meaning of how the different ranges of numeric values will impact on which compilers, dependencies, to be used,
# TEI (text-embedding-inference) got a Dockerfile specified that well
# https://github.com/huggingface/text-embeddings-inference/blob/6e900afba71821fdf250e380d7da1f5a6e5e7e27/Dockerfile-cuda#L50-L63
#
# Thanks
# - https://github.com/Narsil/bindgen_cuda/issues/4
# - https://github.com/Narsil/bindgen_cuda/issues/8
# - https://github.com/huggingface/candle/issues/1516#issuecomment-1875440701
#
# [^1]: compilation - Can I compile a cuda program without having a cuda device - Stack Overflow https://stackoverflow.com/a/20196425
# [^2]: https://huggingface.github.io/candle/guide/installation.html
CUDA_COMPUTE_CAP: '75'
- name: Rename Artifacts (Nightly)
if: ${{ github.event_name == 'schedule' }}