Skip to content
Model Compression Pipeline

Model Compression Pipeline #125

Sign in to view logs

Model Compression Pipeline

Model Compression Pipeline #125

Workflow file for this run

.github/workflows/model_pipeline.yml at 0733da1
name: Model Compression Pipeline
# Daily run at 06:00 UTC and manual trigger via the GitHub UI.
on:
schedule:
- cron: '0 6 * * *'
workflow_dispatch:
inputs:
dry_run:
description: 'Dry run (no external calls, no writes)'
type: boolean
default: true
jobs:
pipeline:
runs-on: macos-14
steps:
# ── Checkout ─────────────────────────────────────────────────────────────
- name: Checkout repository
uses: actions/checkout@v4
# ── Python ───────────────────────────────────────────────────────────────
- name: Set up Python 3.12
uses: actions/setup-python@v5
with:
python-version: '3.12'
# ── Install ──────────────────────────────────────────────────────────────
- name: Install squish
run: pip install -e ".[dev]"
# ── Job 1: watch ─────────────────────────────────────────────────────────
# Poll HuggingFace for new candidate models and write models.json.
# Always run with --dry-run in CI so no HF_TOKEN is required.
- name: Watch — discover candidate models
run: |
python dev/scripts/model_pipeline.py \
--job watch \
--dry-run
env:
HF_TOKEN: ${{ secrets.HF_TOKEN }}
# ── Job 2: compress + validate ───────────────────────────────────────────
# Compress each candidate (INT4) and run the accuracy gate.
# --validate enables the manifest check; --dry-run skips real subprocess
# calls so no GPU or large downloads are needed in CI.
- name: Compress and validate — accuracy gate
run: |
python dev/scripts/model_pipeline.py \
--job compress \
--validate \
--dry-run
env:
HF_TOKEN: ${{ secrets.HF_TOKEN }}