Skip to content

Commit b077aa8

Browse files
ko3n1gko3n1g
committed
test
Signed-off-by: oliver könig <okoenig@nvidia.com>
1 parent c100060 commit b077aa8

File tree

1 file changed

+121
-121
lines changed

1 file changed

+121
-121
lines changed

‎.github/workflows/_build.yml‎

Lines changed: 121 additions & 121 deletions
Original file line numberDiff line numberDiff line change
@@ -74,48 +74,48 @@ jobs:
7474
with:
7575
swap-size-gb: 10
7676

77-
- name: Install CUDA ${{ inputs.cuda-version }}
78-
if: ${{ inputs.cuda-version != 'cpu' }}
79-
uses: Jimver/cuda-toolkit@v0.2.26
80-
id: cuda-toolkit
81-
with:
82-
cuda: ${{ inputs.cuda-version }}
83-
linux-local-args: '["--toolkit"]'
84-
# default method is "local", and we're hitting some error with caching for CUDA 11.8 and 12.1
85-
# method: ${{ (inputs.cuda-version == '11.8.0' || inputs.cuda-version == '12.1.0') && 'network' || 'local' }}
86-
method: "network"
87-
sub-packages: '["nvcc"]'
88-
89-
- name: Install PyTorch ${{ inputs.torch-version }}+cu${{ inputs.cuda-version }}
90-
run: |
91-
pip install --upgrade pip
92-
# With python 3.13 and torch 2.5.1, unless we update typing-extensions, we get error
93-
# AttributeError: attribute '__default__' of 'typing.ParamSpec' objects is not writable
94-
pip install typing-extensions==4.12.2
95-
# We want to figure out the CUDA version to download pytorch
96-
# e.g. we can have system CUDA version being 11.7 but if torch==1.12 then we need to download the wheel from cu116
97-
# see https://github.com/pytorch/pytorch/blob/main/RELEASE.md#release-compatibility-matrix
98-
# This code is ugly, maybe there's a better way to do this.
99-
export TORCH_CUDA_VERSION=$(python -c "from os import environ as env; \
100-
minv = {'2.4': 118, '2.5': 118, '2.6': 118, '2.7': 118, '2.8': 126}[env['MATRIX_TORCH_VERSION']]; \
101-
maxv = {'2.4': 124, '2.5': 124, '2.6': 126, '2.7': 128, '2.8': 129}[env['MATRIX_TORCH_VERSION']]; \
102-
print(minv if int(env['MATRIX_CUDA_VERSION']) < 120 else maxv)" \
103-
)
104-
if [[ ${{ inputs.torch-version }} == *"dev"* ]]; then
105-
# pip install --no-cache-dir --pre torch==${{ inputs.torch-version }} --index-url https://download.pytorch.org/whl/nightly/cu${TORCH_CUDA_VERSION}
106-
# Can't use --no-deps because we need cudnn etc.
107-
# Hard-coding this version of pytorch-triton for torch 2.6.0.dev20241001
108-
pip install jinja2
109-
pip install https://download.pytorch.org/whl/nightly/pytorch_triton-3.1.0%2Bcf34004b8a-cp${MATRIX_PYTHON_VERSION}-cp${MATRIX_PYTHON_VERSION}-linux_x86_64.whl
110-
pip install --no-cache-dir --pre https://download.pytorch.org/whl/nightly/cu${TORCH_CUDA_VERSION}/torch-${{ inputs.torch-version }}%2Bcu${TORCH_CUDA_VERSION}-cp${MATRIX_PYTHON_VERSION}-cp${MATRIX_PYTHON_VERSION}-linux_x86_64.whl
111-
else
112-
pip install --no-cache-dir torch==${{ inputs.torch-version }} --index-url https://download.pytorch.org/whl/cu${TORCH_CUDA_VERSION}
113-
fi
114-
nvcc --version
115-
python --version
116-
python -c "import torch; print('PyTorch:', torch.__version__)"
117-
python -c "import torch; print('CUDA:', torch.version.cuda)"
118-
python -c "from torch.utils import cpp_extension; print (cpp_extension.CUDA_HOME)"
77+
# - name: Install CUDA ${{ inputs.cuda-version }}
78+
# if: ${{ inputs.cuda-version != 'cpu' }}
79+
# uses: Jimver/cuda-toolkit@v0.2.26
80+
# id: cuda-toolkit
81+
# with:
82+
# cuda: ${{ inputs.cuda-version }}
83+
# linux-local-args: '["--toolkit"]'
84+
# # default method is "local", and we're hitting some error with caching for CUDA 11.8 and 12.1
85+
# # method: ${{ (inputs.cuda-version == '11.8.0' || inputs.cuda-version == '12.1.0') && 'network' || 'local' }}
86+
# method: "network"
87+
# sub-packages: '["nvcc"]'
88+
89+
# - name: Install PyTorch ${{ inputs.torch-version }}+cu${{ inputs.cuda-version }}
90+
# run: |
91+
# pip install --upgrade pip
92+
# # With python 3.13 and torch 2.5.1, unless we update typing-extensions, we get error
93+
# # AttributeError: attribute '__default__' of 'typing.ParamSpec' objects is not writable
94+
# pip install typing-extensions==4.12.2
95+
# # We want to figure out the CUDA version to download pytorch
96+
# # e.g. we can have system CUDA version being 11.7 but if torch==1.12 then we need to download the wheel from cu116
97+
# # see https://github.com/pytorch/pytorch/blob/main/RELEASE.md#release-compatibility-matrix
98+
# # This code is ugly, maybe there's a better way to do this.
99+
# export TORCH_CUDA_VERSION=$(python -c "from os import environ as env; \
100+
# minv = {'2.4': 118, '2.5': 118, '2.6': 118, '2.7': 118, '2.8': 126}[env['MATRIX_TORCH_VERSION']]; \
101+
# maxv = {'2.4': 124, '2.5': 124, '2.6': 126, '2.7': 128, '2.8': 129}[env['MATRIX_TORCH_VERSION']]; \
102+
# print(minv if int(env['MATRIX_CUDA_VERSION']) < 120 else maxv)" \
103+
# )
104+
# if [[ ${{ inputs.torch-version }} == *"dev"* ]]; then
105+
# # pip install --no-cache-dir --pre torch==${{ inputs.torch-version }} --index-url https://download.pytorch.org/whl/nightly/cu${TORCH_CUDA_VERSION}
106+
# # Can't use --no-deps because we need cudnn etc.
107+
# # Hard-coding this version of pytorch-triton for torch 2.6.0.dev20241001
108+
# pip install jinja2
109+
# pip install https://download.pytorch.org/whl/nightly/pytorch_triton-3.1.0%2Bcf34004b8a-cp${MATRIX_PYTHON_VERSION}-cp${MATRIX_PYTHON_VERSION}-linux_x86_64.whl
110+
# pip install --no-cache-dir --pre https://download.pytorch.org/whl/nightly/cu${TORCH_CUDA_VERSION}/torch-${{ inputs.torch-version }}%2Bcu${TORCH_CUDA_VERSION}-cp${MATRIX_PYTHON_VERSION}-cp${MATRIX_PYTHON_VERSION}-linux_x86_64.whl
111+
# else
112+
# pip install --no-cache-dir torch==${{ inputs.torch-version }} --index-url https://download.pytorch.org/whl/cu${TORCH_CUDA_VERSION}
113+
# fi
114+
# nvcc --version
115+
# python --version
116+
# python -c "import torch; print('PyTorch:', torch.__version__)"
117+
# python -c "import torch; print('CUDA:', torch.version.cuda)"
118+
# python -c "from torch.utils import cpp_extension; print (cpp_extension.CUDA_HOME)"
119119

120120
- name: Restore build cache
121121
uses: actions/cache/restore@v4
@@ -136,82 +136,82 @@ jobs:
136136
137137
ls -al ./
138138
139-
- name: Build wheel
140-
id: build_wheel
141-
run: |
142-
cat > build.sh <<'EOF'
143-
#!/usr/bin/env bash
144-
set -euxo pipefail
145-
146-
echo "Running build.sh"
147-
# We want setuptools >= 49.6.0 otherwise we can't compile the extension if system CUDA version is 11.7 and pytorch cuda version is 11.6
148-
# https://github.com/pytorch/pytorch/blob/664058fa83f1d8eede5d66418abff6e20bd76ca8/torch/utils/cpp_extension.py#L810
149-
# However this still fails so I'm using a newer version of setuptools
150-
pip install setuptools==75.8.0
151-
pip install ninja packaging wheel
152-
export PATH=/usr/local/nvidia/bin:/usr/local/nvidia/lib64:$PATH
153-
export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
154-
# Limit MAX_JOBS otherwise the github runner goes OOM
155-
# nvcc 11.8 can compile with 2 jobs, but nvcc 12.3 goes OOM
156-
157-
export MAX_JOBS=$([ "$MATRIX_CUDA_VERSION" == "129" ] && echo 1 || echo 2)
158-
export NVCC_THREADS=2
159-
export FLASH_ATTENTION_FORCE_BUILD="TRUE"
160-
export FLASH_ATTENTION_FORCE_CXX11_ABI=${{ inputs.cxx11-abi }}
161-
162-
python setup.py bdist_wheel --dist-dir=dist
163-
164-
tmpname=cu${WHEEL_CUDA_VERSION}torch${MATRIX_TORCH_VERSION}cxx11abi${{ inputs.cxx11-abi }}
165-
wheel_name=$(ls dist/*whl | xargs -n 1 basename | sed "s/-/+$tmpname-/2")
166-
ls dist/*whl |xargs -I {} mv {} dist/${wheel_name}
167-
echo "wheel_name=${wheel_name}" >> $GITHUB_ENV
168-
EOF
169-
170-
chmod +x build.sh
171-
172-
timeout 20m ./build.sh || EXIT_CODE=$?
173-
174-
# Store exit code in GitHub env for later steps
175-
echo "build_exit_code=$EXIT_CODE" | tee -a "$GITHUB_OUTPUT"
176-
177-
# Do not fail the job if timeout killed the build
178-
exit 0
179-
180-
- name: Log build logs
181-
if: always()
182-
run: |
183-
ls -al ./
184-
185-
tar -cvf build-cache.tar -C build . --atime-preserve=replace
186-
187-
- name: Save build cache
188-
if: always()
189-
uses: actions/cache/save@v4
190-
with:
191-
key: build-${{ inputs.release-version }}-${{ inputs.python-version }}-${{ inputs.cuda-version }}-${{ inputs.torch-version }}-${{ inputs.cxx11-abi }}-${{ github.run_number }}-${{ github.run_attempt }}
192-
path: build-cache.tar
193-
194-
- name: Log Built Wheels
195-
run: |
196-
ls dist
197-
198-
- name: Get Release with tag
199-
id: get_current_release
200-
if: steps.build_wheel.outputs.build_exit_code == 0
201-
uses: joutvhu/get-release@v1
202-
with:
203-
tag_name: ${{ inputs.release-version }}
204-
env:
205-
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
206-
207-
- name: Upload Release Asset
208-
id: upload_release_asset
209-
if: inputs.upload-to-release && steps.build_wheel.outputs.build_exit_code == 0
210-
uses: actions/upload-release-asset@v1
211-
env:
212-
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
213-
with:
214-
upload_url: ${{ steps.get_current_release.outputs.upload_url }}
215-
asset_path: ./dist/${{env.wheel_name}}
216-
asset_name: ${{env.wheel_name}}
217-
asset_content_type: application/*
139+
# - name: Build wheel
140+
# id: build_wheel
141+
# run: |
142+
# cat > build.sh <<'EOF'
143+
# #!/usr/bin/env bash
144+
# set -euxo pipefail
145+
146+
# echo "Running build.sh"
147+
# # We want setuptools >= 49.6.0 otherwise we can't compile the extension if system CUDA version is 11.7 and pytorch cuda version is 11.6
148+
# # https://github.com/pytorch/pytorch/blob/664058fa83f1d8eede5d66418abff6e20bd76ca8/torch/utils/cpp_extension.py#L810
149+
# # However this still fails so I'm using a newer version of setuptools
150+
# pip install setuptools==75.8.0
151+
# pip install ninja packaging wheel
152+
# export PATH=/usr/local/nvidia/bin:/usr/local/nvidia/lib64:$PATH
153+
# export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
154+
# # Limit MAX_JOBS otherwise the github runner goes OOM
155+
# # nvcc 11.8 can compile with 2 jobs, but nvcc 12.3 goes OOM
156+
157+
# export MAX_JOBS=$([ "$MATRIX_CUDA_VERSION" == "129" ] && echo 1 || echo 2)
158+
# export NVCC_THREADS=2
159+
# export FLASH_ATTENTION_FORCE_BUILD="TRUE"
160+
# export FLASH_ATTENTION_FORCE_CXX11_ABI=${{ inputs.cxx11-abi }}
161+
162+
# python setup.py bdist_wheel --dist-dir=dist
163+
164+
# tmpname=cu${WHEEL_CUDA_VERSION}torch${MATRIX_TORCH_VERSION}cxx11abi${{ inputs.cxx11-abi }}
165+
# wheel_name=$(ls dist/*whl | xargs -n 1 basename | sed "s/-/+$tmpname-/2")
166+
# ls dist/*whl |xargs -I {} mv {} dist/${wheel_name}
167+
# echo "wheel_name=${wheel_name}" >> $GITHUB_ENV
168+
# EOF
169+
170+
# chmod +x build.sh
171+
172+
# timeout 20m ./build.sh || EXIT_CODE=$?
173+
174+
# # Store exit code in GitHub env for later steps
175+
# echo "build_exit_code=$EXIT_CODE" | tee -a "$GITHUB_OUTPUT"
176+
177+
# # Do not fail the job if timeout killed the build
178+
# exit 0
179+
180+
# - name: Log build logs
181+
# if: always()
182+
# run: |
183+
# ls -al ./
184+
185+
# tar -cvf build-cache.tar -C build . --atime-preserve=replace
186+
187+
# - name: Save build cache
188+
# if: always()
189+
# uses: actions/cache/save@v4
190+
# with:
191+
# key: build-${{ inputs.release-version }}-${{ inputs.python-version }}-${{ inputs.cuda-version }}-${{ inputs.torch-version }}-${{ inputs.cxx11-abi }}-${{ github.run_number }}-${{ github.run_attempt }}
192+
# path: build-cache.tar
193+
194+
# - name: Log Built Wheels
195+
# run: |
196+
# ls dist
197+
198+
# - name: Get Release with tag
199+
# id: get_current_release
200+
# if: steps.build_wheel.outputs.build_exit_code == 0
201+
# uses: joutvhu/get-release@v1
202+
# with:
203+
# tag_name: ${{ inputs.release-version }}
204+
# env:
205+
# GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
206+
207+
# - name: Upload Release Asset
208+
# id: upload_release_asset
209+
# if: inputs.upload-to-release && steps.build_wheel.outputs.build_exit_code == 0
210+
# uses: actions/upload-release-asset@v1
211+
# env:
212+
# GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
213+
# with:
214+
# upload_url: ${{ steps.get_current_release.outputs.upload_url }}
215+
# asset_path: ./dist/${{env.wheel_name}}
216+
# asset_name: ${{env.wheel_name}}
217+
# asset_content_type: application/*

0 commit comments

Comments
 (0)