Skip to content

[Performance] Improve performance of compiled ReplayBuffer #5912

[Performance] Improve performance of compiled ReplayBuffer

[Performance] Improve performance of compiled ReplayBuffer #5912

name: Habitat Tests on Linux
on:
pull_request:
push:
branches:
- nightly
- main
- release/*
workflow_dispatch:
concurrency:
# Documentation suggests ${{ github.head_ref }}, but that's only available on pull_request/pull_request_target triggers, so using ${{ github.ref }}.
# On master, we want all builds to complete even if merging happens faster to make it easier to discover at which point something broke.
group: ${{ github.workflow }}-${{ github.ref == 'refs/heads/main' && format('ci-master-{0}', github.sha) || format('ci-{0}', github.ref) }}
cancel-in-progress: true
jobs:
tests:
strategy:
matrix:
python_version: ["3.9"]
cuda_arch_version: ["12.1"]
fail-fast: false
uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
with:
runner: linux.g5.4xlarge.nvidia.gpu
repository: pytorch/rl
docker-image: "nvidia/cuda:12.1.1-devel-ubuntu22.04"
gpu-arch-type: cuda
gpu-arch-version: ${{ matrix.cuda_arch_version }}
timeout: 90
script: |
if [[ "${{ github.ref }}" =~ release/* ]]; then
export RELEASE=1
export TORCH_VERSION=stable
else
export RELEASE=0
export TORCH_VERSION=nightly
fi
# Set env vars from matrix
export PYTHON_VERSION=${{ matrix.python_version }}
# Commenting these out for now because the GPU test are not working inside docker
export CUDA_ARCH_VERSION=${{ matrix.cuda_arch_version }}
export CU_VERSION="cu${CUDA_ARCH_VERSION:0:2}${CUDA_ARCH_VERSION:3:1}"
# Remove the following line when the GPU tests are working inside docker, and uncomment the above lines
#export CU_VERSION="cpu"
export TD_GET_DEFAULTS_TO_NONE=1
bash .github/unittest/linux_libs/scripts_habitat/run_all.sh