Skip to content

FP8 RL Training and Rollout Weight Synchronization (#1898) #643

FP8 RL Training and Rollout Weight Synchronization (#1898)

FP8 RL Training and Rollout Weight Synchronization (#1898) #643

name: SkyRL-Train-GPU
on:
push:
branches:
- main
paths:
- 'ci/**'
- 'skyrl/backends/skyrl_train/**'
- 'skyrl/train/**'
- 'tests/backends/skyrl_train/**'
- 'pyproject.toml'
- '!docs/**'
- '!examples/**'
- '.github/workflows/**'
pull_request_target:
types: [labeled]
workflow_dispatch:
permissions:
checks: write # for status checks to appear
contents: read
jobs:
skyrl_train_tests:
# Gate on the label that fired this event. `types: [labeled]` delivers one event
# per label added, and `contains(...labels.*.name, ...)` tests the whole label
# set, so labelling in bulk launched one identical GPU run per label.
if: >
github.event_name == 'push' ||
github.event_name == 'workflow_dispatch' ||
(
github.event_name == 'pull_request_target' &&
!github.event.pull_request.draft &&
github.event.label.name == 'run_train_gpu_ci'
)
runs-on: ubuntu-latest
defaults:
run:
shell: bash
working-directory: .
steps:
- uses: actions/checkout@v4
with:
ref: ${{ github.event.pull_request.head.sha || github.ref }}
# The Anyscale job uploads this working dir, so we need the PR head.
# Fork code only reaches here after a maintainer applies the gating
# label above; don't leave the base repo token on disk for it.
allow-unsafe-pr-checkout: true
persist-credentials: false
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.12'
- name: Install the latest version of uv
uses: astral-sh/setup-uv@v6
with:
activate-environment: true
- name: Install basic dependencies
run: uv pip install anyscale==0.26.103 typer==0.9.0
- name: Install envsubst
run: sudo apt-get update && sudo apt-get install -y gettext-base
# Run tests
- name: GPU tests
env:
ANYSCALE_CLI_TOKEN: ${{ secrets.ANYSCALE_CLI_TOKEN }}
ANYSCALE_HOST: https://console.anyscale.com
HF_TOKEN: ${{ secrets.HF_TOKEN }}
run: |
envsubst < ci/anyscale_gpu_ci_skyrl_train.yaml > ci/anyscale_gpu_ci_skyrl_train_envsubst.yaml
COMMIT_SHA="${{ github.event.pull_request.head.sha || github.sha }}"
JOB_NAME="skyrl-train-gpu-ci-${COMMIT_SHA:0:7}-${{ github.run_id }}-${{ github.run_attempt }}"
bash ci/submit_anyscale_job.sh ci/anyscale_gpu_ci_skyrl_train_envsubst.yaml "$JOB_NAME" 12000
rm -f ci/anyscale_gpu_ci_skyrl_train_envsubst.yaml
# Cancelling the run stops the runner, not the cluster. GitHub still runs
# `if: cancelled()` steps, so this is where an orphaned GPU job gets reaped.
- name: Terminate Anyscale jobs on cancellation
if: cancelled()
env:
ANYSCALE_CLI_TOKEN: ${{ secrets.ANYSCALE_CLI_TOKEN }}
ANYSCALE_HOST: https://console.anyscale.com
run: bash ci/terminate_anyscale_jobs.sh