Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
72 changes: 50 additions & 22 deletions .github/workflows/comfyui-quick-start.yml
Original file line number Diff line number Diff line change
@@ -1,18 +1,12 @@
# ComfyUI quick start guard - project thin trigger.
#
# Calls the common engine .github/workflows/quick-start-template.yml;
# this file only declares what varies per project: schedule,
# concurrency, runner, container image, upstream repo, monitored doc
# and the test entry command. The guard loop, cache I/O, state relay
# and publishing all live in the engine. See the engine's
# `workflow_call.inputs` block for the full input contract.
#
# Engine-fixed behaviours this trigger relies on (declared in the
# engine, NOT here):
# * cache key prefix: monitor-state-comfyui- (from inputs.project)
# * artifact name: comfyui-quick-start-<run_id>
# * result.json is written by the engine and validated against
# schemas/result.schema.json before upload.
# Calls the common engine .github/workflows/quick-start-template.yml once per
# machine generation (A2/A3) and merges the per-machine result.json files into
# one artifact; this file only declares what varies per project: schedule,
# concurrency, runner, container image, upstream repo, monitored doc and the
# test entry command. The guard loop, cache I/O, state relay and publishing all
# live in the engine. See the engine's `workflow_call.inputs` block for the full
# input contract.
#
# ComfyUI note: upstream auto-detects NPU (torch_npu import +
# torch.npu.is_available() -> torch.device("npu")), so the doc's server
Expand Down Expand Up @@ -40,11 +34,11 @@ concurrency:

on:
schedule:
# Twice a day, 12 h apart: UTC 10:45 and 22:45
# (Beijing +8). All quick-start guards share this cadence; the
# (hour, minute) slot is staggered across the guards so the NPU
# runners do not all fire at the same time.
- cron: '45 10,22 * * *'
# Twice a day, 12 h apart: UTC 07:00 and 19:00 (Beijing +8).
# A2 and A3 legs fire together and the A3 pool runs one job at a
# time, so the guards take 2 h windows in sequence starting at
# deepspeed 00:00 UTC; this one owns 07:00.
- cron: '0 7,19 * * *'
workflow_dispatch:
# PR trigger: docs/tests changes get a guard run. paths filter avoids burning
# the self-hosted NPU runner on unrelated PRs. `pull_request` (not
Expand All @@ -54,18 +48,19 @@ on:
paths:
- 'sources/comfyui/**'
- 'tests/comfyui/**'
- '.github/workflows/comfyui-quick-start.yml'

permissions:
contents: read

jobs:
comfyui-quick-start:
qs-a2:
uses: ./.github/workflows/quick-start-template.yml
with:
# Namespaces cache keys (monitor-state-comfyui-*), artifacts
# (comfyui-quick-start-<run_id>) and the test working dir
# (tests/comfyui).
# Namespaces cache keys, artifacts and the test working dir
# (tests/comfyui) per machine generation.
project: comfyui
machine_type: a2
# Self-hosted NPU runner for the test job only; the engine pins
# the cache I/O jobs (restore-cache / publish-and-persist) to
# GitHub-hosted ubuntu-latest.
Expand All @@ -88,3 +83,36 @@ jobs:
# engine. All project env prep and installs live in the test
# subclass's prepare_environment hook.
test_command: python -m unittest tests.comfyui.test_quick_start_ascend -v 2>&1

qs-a3:
uses: ./.github/workflows/quick-start-template.yml
with:
project: comfyui
machine_type: a3
# Atlas 900 A3 (Ascend 910C) leg; same CANN line as A2.
test_runner: '["linux-aarch64-a3-800i-ascend-docs-fed-2"]'
image: swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-a3-ubuntu22.04-py3.12
# Cache binds for the A3 pool, same layout as the other A3 jobs:
# the 4.3 GB checkpoint comes from Hugging Face Hub.
container_options: >-
--volume=/data/ci-cache/hf/comfyui:/root/.cache/huggingface
--volume=/data/ci-cache/pip:/root/.cache/pip
timeout_minutes: 120
upstream_repo: comfyanonymous/ComfyUI
doc_url: 'https://raw.githubusercontent.com/Ascend/docs/{0}/sources/comfyui/quick_start.md'
doc_path: https://github.com/Ascend/docs/blob/main/sources/comfyui/quick_start.md
test_command: python -m unittest tests.comfyui.test_quick_start_ascend -v 2>&1

merge:
needs: [qs-a2, qs-a3]
if: always()
runs-on: ubuntu-latest
steps:
- name: Checkout repository (provides the merge action)
uses: actions/checkout@v6
- name: Merge per-machine results into one artifact
uses: ./.github/actions/merge-quick-start-results
with:
project: comfyui
result_a2: ${{ needs.qs-a2.outputs.result_json }}
result_a3: ${{ needs.qs-a3.outputs.result_json }}
57 changes: 47 additions & 10 deletions .github/workflows/evalscope-quick-start.yml
Original file line number Diff line number Diff line change
@@ -1,6 +1,8 @@
# evalscope quick start guard - project thin trigger.
#
# Calls the common engine .github/workflows/quick-start-template.yml.
# Calls the common engine .github/workflows/quick-start-template.yml once per
# machine generation (A2/A3) and merges the per-machine result.json files into
# one artifact.

name: evalscope-quick-start

Expand All @@ -18,11 +20,11 @@ concurrency:

on:
schedule:
# Twice a day, 12 h apart: UTC 04:05 and 16:05
# (Beijing +8). All quick-start guards share this cadence; the
# (hour, minute) slot is staggered across the guards so the NPU
# runners do not all fire at the same time.
- cron: '5 4,16 * * *'
# Twice a day, 12 h apart: UTC 03:00 and 15:00 (Beijing +8).
# A2 and A3 legs fire together and the A3 pool runs one job at a
# time, so the guards take 2 h windows in sequence starting at
# deepspeed 00:00 UTC; this one owns 03:00.
- cron: '0 3,15 * * *'
workflow_dispatch:
# PR trigger: docs/tests changes get a guard run. paths filter avoids burning
# the self-hosted NPU runner on unrelated PRs. `pull_request` (not
Expand All @@ -32,18 +34,19 @@ on:
paths:
- 'sources/evalscope/**'
- 'tests/evalscope/**'
- '.github/workflows/evalscope-quick-start.yml'

permissions:
contents: read

jobs:
evalscope-quick-start:
qs-a2:
uses: ./.github/workflows/quick-start-template.yml
with:
# Namespaces cache keys (monitor-state-evalscope-*), artifacts
# (evalscope-quick-start-<run_id>) and the test working dir
# (workflows/tests/evalscope).
# Namespaces cache keys, artifacts and the test working dir
# (tests/evalscope) per machine generation.
project: evalscope
machine_type: a2
# Self-hosted NPU runner for the test job only; the engine pins
# the cache I/O jobs (restore-cache / publish-and-persist) to
# GitHub-hosted ubuntu-latest.
Expand Down Expand Up @@ -73,3 +76,37 @@ jobs:
# engine. All project env prep and installs live in the test
# subclass's prepare_environment hook.
test_command: python -m unittest tests.evalscope.test_quick_start_ascend -v 2>&1

qs-a3:
uses: ./.github/workflows/quick-start-template.yml
with:
project: evalscope
machine_type: a3
# Atlas 900 A3 (Ascend 910C) leg; same CANN line as A2.
test_runner: '["linux-aarch64-a3-800i-ascend-docs-fed-2"]'
image: swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-a3-ubuntu22.04-py3.12
# Cache binds for the A3 pool, same layout as the other A3 jobs:
# the model and both dataset repos come from ModelScope.
container_options: >-
--volume=/data/ci-cache/modelscope/evalscope:/root/.cache/modelscope
--volume=/data/ci-cache/pip:/root/.cache/pip
--volume=/data/ci-cache/hf/evalscope:/root/.cache/huggingface
timeout_minutes: 120
upstream_repo: modelscope/evalscope
doc_url: 'https://raw.githubusercontent.com/Ascend/docs/{0}/sources/evalscope/quick_start.md'
doc_path: https://github.com/Ascend/docs/blob/main/sources/evalscope/quick_start.md
test_command: python -m unittest tests.evalscope.test_quick_start_ascend -v 2>&1

merge:
needs: [qs-a2, qs-a3]
if: always()
runs-on: ubuntu-latest
steps:
- name: Checkout repository (provides the merge action)
uses: actions/checkout@v6
- name: Merge per-machine results into one artifact
uses: ./.github/actions/merge-quick-start-results
with:
project: evalscope
result_a2: ${{ needs.qs-a2.outputs.result_json }}
result_a3: ${{ needs.qs-a3.outputs.result_json }}
73 changes: 51 additions & 22 deletions .github/workflows/minionerec-quick-start.yml
Original file line number Diff line number Diff line change
@@ -1,20 +1,16 @@
# MiniOneRec quick start guard - project thin trigger.
#
# Calls the common engine .github/workflows/quick-start-template.yml; this
# file only declares what varies per project: schedule, concurrency, runner,
# container image, upstream repo, the pinned ref, the monitored doc and the
# test entry command. The guard loop, cache I/O, state relay and publishing
# all live in the engine - see its `workflow_call.inputs` block for the full
# contract.
#
# Cache keys (quick-start-monitor-state-minionerec_*) and artifact names
# (minionerec-quick-start-<run_id>) are derived from project=minionerec.
# Calls the common engine .github/workflows/quick-start-template.yml once per
# machine generation (A2/A3) and merges the per-machine result.json files into
# one artifact; this file only declares what varies per project: schedule,
# concurrency, runner, container image, upstream repo, the pinned ref, the
# monitored doc and the test entry command.
#
# Upstream shape: MiniOneRec is research code on `main`, no release tags, so
# fixed_ref=main - the engine resolves it to the branch tip SHA on every run,
# which is also the monitor's change key.
#
# Scope: single card (910B), three stages on one 0.5B model - SFT (one epoch,
# Scope: single card, three stages on one 0.5B model - SFT (one epoch,
# sample=600 per sub-dataset), constrained-decoding evaluation (CC must be 0)
# and one GRPO epoch. Upstream PR #85 ships `npu_support.py` and imports it
# first in every entry script, so no injection or source edit is needed:
Expand All @@ -37,11 +33,11 @@ concurrency:

on:
schedule:
# Twice a day, 12 h apart: UTC 11:50 and 23:50
# (Beijing +8). All quick-start guards share this cadence; the
# (hour, minute) slot is staggered across the guards so the NPU
# runners do not all fire at the same time.
- cron: '50 11,23 * * *'
# Twice a day, 12 h apart: UTC 09:00 and 21:00 (Beijing +8).
# A2 and A3 legs fire together and the A3 pool runs one job at a
# time, so the guards take 2 h windows in sequence starting at
# deepspeed 00:00 UTC; this one owns 09:00.
- cron: '0 9,21 * * *'
workflow_dispatch:
# PR trigger: docs/tests changes get a guard run before merge. paths filter
# avoids burning the self-hosted NPU runner on unrelated PRs.
Expand All @@ -58,14 +54,15 @@ permissions:
contents: read

jobs:
minionerec-quick-start:
qs-a2:
uses: ./.github/workflows/quick-start-template.yml
with:
# Namespaces cache keys (quick-start-monitor-state-minionerec_*),
# artifacts (minionerec-quick-start-<run_id>) and the test working dir.
# Namespaces cache keys, artifacts and the test working dir
# (tests/minionerec) per machine generation.
project: minionerec
# Self-hosted NPU runner for the test job only; the engine pins the
# cache I/O jobs (restore-cache / publish-and-persist) to
machine_type: a2
# Self-hosted NPU runner for the test job only; the engine pins
# the cache I/O jobs (restore-cache / publish-and-persist) to
# GitHub-hosted ubuntu-latest. The quick start is single-card
# (ASCEND_RT_VISIBLE_DEVICES=0).
test_runner: '["linux-aarch64-a2-1"]'
Expand All @@ -84,9 +81,41 @@ jobs:
# lands on Ascend/docs post-merge.
doc_url: 'https://raw.githubusercontent.com/Ascend/docs/{0}/sources/minionerec/quick_start.md'
doc_path: https://github.com/Ascend/docs/blob/main/sources/minionerec/quick_start.md
# cwd is the repo root inside the `workflows` checkout (matches the
# engine template's `working-directory: workflows`); env contract
# cwd is the repo root inside the checkout; env contract
# (MONITORED_DOC_URL / UPSTREAM_REF / NPU_READY) is injected by the
# engine. The test subclass pins cwd to /root/minionerec-test and
# sources CANN; every business step is in the doc.
test_command: python -m unittest tests.minionerec.test_quick_start_ascend -v 2>&1

qs-a3:
uses: ./.github/workflows/quick-start-template.yml
with:
project: minionerec
machine_type: a3
# Atlas 900 A3 (Ascend 910C) leg; same CANN line as A2.
test_runner: '["linux-aarch64-a3-800i-ascend-docs-fed-2"]'
image: swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-a3-ubuntu22.04-py3.12
# Cache bind for the A3 pool, same layout as the other A3 jobs:
# the base model comes from Hugging Face Hub.
container_options: >-
--volume=/data/ci-cache/hf/minionerec:/root/.cache/huggingface
timeout_minutes: 120
upstream_repo: AkaliKong/MiniOneRec
fixed_ref: main
doc_url: 'https://raw.githubusercontent.com/Ascend/docs/{0}/sources/minionerec/quick_start.md'
doc_path: https://github.com/Ascend/docs/blob/main/sources/minionerec/quick_start.md
test_command: python -m unittest tests.minionerec.test_quick_start_ascend -v 2>&1

merge:
needs: [qs-a2, qs-a3]
if: always()
runs-on: ubuntu-latest
steps:
- name: Checkout repository (provides the merge action)
uses: actions/checkout@v6
- name: Merge per-machine results into one artifact
uses: ./.github/actions/merge-quick-start-results
with:
project: minionerec
result_a2: ${{ needs.qs-a2.outputs.result_json }}
result_a3: ${{ needs.qs-a3.outputs.result_json }}
Loading
Loading