TorchBench Nightly #15

	name: TorchBench Nightly

	on:
	workflow_dispatch:

	env:
	TRITON_USE_ASSERT_ENABLED_LLVM: "TRUE"

	jobs:
	TorchBench-Nvidia:
	runs-on: [self-hosted, A100]

	steps:
	- name: Checkout TorchBench
	uses: actions/checkout@v4
	with:
	repository: 'pytorch/benchmark'
	ref: '7dc8e8113f8d5dc97aaa2bef94f9d389047cfcde'
	path: 'torchbench'
	submodules: 'recursive'

	- name: Update PATH
	run: \|
	echo "PATH=${HOME}/.local/bin:${PATH}" >> "${GITHUB_ENV}"
	df -h
	du --max-depth=1 -h

	# We don't want to use to use TorchBench nightlies here
	# Force user install, TorchBench expects a conda environment
	- name: Install TorchBench and models
	run: \|
	cd python
	ls
	python3 -m pip install --upgrade pip
	python3 -m pip install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
	df -h
	du --max-depth=1 -h
	cd ../torchbench
	echo -e "[install]\nprefix=${HOME}/.local\n" > ${HOME}/.pydistutils.cfg
	python3 install.py detectron2_fasterrcnn_r_101_c4
	df -h
	du --max-depth=1 -h

	- name: Install the latest Triton wheel
	run: \|
	python3 -m pip install --upgrade triton-nightly --index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/
	df -h
	du --max-depth=1 -h

	- name: Tune Nvidia GPU
	run: \|
	sudo nvidia-smi -pm 1
	sudo nvidia-smi --lock-gpu-clocks=1280,1280
	nvidia-smi

	- name: Run TorchBench
	run: \|
	cd torchbench
	python3 test.py -v TestBenchmark.test_detectron2_fasterrcnn_r_101_c4_train_cuda

	- name: Reset GPU clocks
	run: \|
	sudo nvidia-smi -i 0 -rgc

Provide feedback