Skip to content

[booster] use the accelerator API for the TorchFSDP device id #54

[booster] use the accelerator API for the TorchFSDP device id

[booster] use the accelerator API for the TorchFSDP device id #54

name: CI / Layer 2 / Trusted PR GPU
on:
pull_request:
types: [opened, synchronize, reopened, ready_for_review, converted_to_draft]
branches: [main]
paths:
- .github/scripts/ci/**
- .github/workflows/ci_layer_2_gpu_pr.yml
- .github/workflows/ci_reusable_gpu_batches.yml
- colossalai/**
- extensions/**
- docker/torch213/**
- requirements/**
- tests/**
- pytest.ini
- setup.py
workflow_dispatch:
permissions:
contents: read
concurrency:
group: ci-layer-2-${{ github.event.pull_request.number || github.ref }}
cancel-in-progress: true
jobs:
container-smoke:
# This is a separate check, not a successful full Layer 2 regression.
name: Torch 2.13 container smoke (batch 4a)
if: >-
github.repository == 'hpcaitech/ColossalAI' &&
github.event_name == 'pull_request' &&
github.event.pull_request.draft == true &&
github.event.pull_request.head.repo.full_name == github.repository &&
contains(fromJSON('["OWNER", "MEMBER", "COLLABORATOR"]'), github.event.pull_request.author_association)
uses: ./.github/workflows/ci_reusable_gpu_batches.yml
with:
layer: draft-smoke
source_sha: ${{ github.sha }}
batches: "4a"
gpu_pool_variable: COLOSSAL_CI_SMOKE_GPU_POOL
fast_mode: false
timeout_minutes: 15
artifact_retention_days: 14
gpu-regression:
# Preserve the existing check name until branch-protection rules migrate.
# Batch 4a is also included below for the FP8 compatibility regression.
name: Torch 2.13 GPU core (batches 2 and 5)
if: >-
github.repository == 'hpcaitech/ColossalAI' &&
(github.event_name == 'workflow_dispatch' ||
(github.event_name == 'pull_request' &&
github.event.pull_request.draft == false &&
github.event.pull_request.head.repo.full_name == github.repository &&
contains(fromJSON('["OWNER", "MEMBER", "COLLABORATOR"]'), github.event.pull_request.author_association)))
uses: ./.github/workflows/ci_reusable_gpu_batches.yml
with:
layer: layer-2
source_sha: ${{ github.sha }}
batches: "2 4a 5"
fast_mode: true
timeout_minutes: 240
artifact_retention_days: 14