aws-instance-requirements

Daily automation to recommend AWS SageMaker GPU instance types for popular Hugging Face models and update the amazon-sagemaker/repository-metadata dataset (modal.json).

What it does

• Uses the Hugging Face Hub CLI (hf) to fetch the top 100 models by:
  • likes
  • downloads
  • trending_score
• Computes the union of these lists (up to 300 unique models/day).
• Downloads the current dataset config (modal.json) from amazon-sagemaker/repository-metadata.
• For models not present in modal.json, runs hf-mem to estimate memory requirements.
• Picks the cheapest allowed SageMaker instance that should not OOM using aggregate GPU VRAM.
• The set of “allowed” instances is defined by config/instance_catalog.json. For now we only consider the G5, G6, G6e, P4 (p4d/p4de), and P5 (p5/p5e) families.
• Appends a new entry with:
  • id
  • instanceType
  • numGpu (all GPUs on that instance)
  • containerStartupHealthCheckTimeout (hardcoded per instance type)

Repo layout

• scripts/sync_modal.py: main entrypoint (discover → diff → hf-mem → map → update → upload/PR)
• scripts/hf_mem.py: runs uvx hf-mem and parses output
• scripts/instance_selection.py: instance catalog loading and cheapest-instance selection logic
• scripts/modal_io.py: reads/writes modal.json config format
• config/instance_catalog.json: instance allowlist, VRAM, GPU counts, timeouts, and pricing
• .github/workflows/daily-sync.yml: scheduled GitHub Actions workflow

Running locally

Prereqs:

• Python 3.10+
• uv installed (so uvx hf-mem ... works)
• Hugging Face CLI available via pip install -U huggingface_hub (provides the hf command)

Auth (needed for gated models + dataset PRs):

export HF_TOKEN="hf_..."
hf auth login --token "$HF_TOKEN"

Dry-run (no upload):

python scripts/sync_modal.py --dry-run

Write + open PR on the dataset:

python scripts/sync_modal.py --write --create-pr

GitHub Actions

Set a repository secret named HF_TOKEN with write access to the dataset.