Quantum-Accelerators/electrai#64 - 2-way sync via ghpr (https://github.com/runsascoded/ghpr)

electrai#64.md

Quantum-Accelerators/electrai#64 Add GPU CI: e2e tests and benchmarks

Summary

• Add deterministic e2e training test (tests/e2e_train.py) with platform-specific expected values
• Add GPU e2e workflow (gpu-e2e.yml) using ec2-gha on EC2 g6.xlarge (NVIDIA L4), runs on PRs targeting main
• Add GPU benchmark workflow (gpu-benchmark.yml) with configurable model size, weekly schedule, and manual dispatch
• Add WandB logging to benchmark: logs training metrics, model config, dataset version, and instance type
• Add scripts/s3_sync.py: reusable S3 data sync with size filtering and deterministic dataset hashing
• Add gen-expected.yml for regenerating expected values on GHA runners (macOS, Ubuntu)
• Add e2e_training_demo.ipynb notebook with training visualization
• Add .github/workflows/README.md documenting all workflows
• Rename tests/electrai/ → tests/test_electrai/ to fix import shadowing
• Add --gradient-checkpoint flag to e2e_train.py for large models on limited VRAM
• Benchmark summary includes linked commit SHA for traceability

Passing Runs

• GPU E2E Test #27 — GPU + CPU baseline, both pass (Feb 24)
• GPU Benchmark #27 — 50 S3 samples, WandB logging (Feb 24)

Required Setup

Secrets:

• GH_SA_TOKEN — GitHub PAT for runner registration
• WANDB_API_KEY — WandB API key (optional, for benchmark logging; currently Ryan's personal key)

IAM/OIDC:

• Trust policy configured in Open-Athena/ops for ec2-gha OIDC authentication

Screenshots

EC2 instances	EC2 tags
GitHub runners
WandB dashboard

Test plan

• GPU e2e test passes on EC2 g6.xlarge
• CPU e2e test passes on EC2
• GPU benchmark completes with production-size model (32ch/16 blocks, 128^3 grids)
• gen-expected.yml generates correct values on macOS and Ubuntu
• Expected values verified across all 3 platforms (darwin-arm64, linux, linux-gpu)
• WandB run logged with correct metadata (project, dataset version, instance type)
• scripts/s3_sync.py downloads correct samples and generates deterministic dataset hash