HN 표시: Willitrun – ML 모델이 모든 기기에서 실행되는지 확인(벤치마크 지원)

CLI to tell you if an ML model will fit and run on your device, using real benchmarks + lightweight estimation. - Real data first: SQLite DB of measured benchmarks (NVIDIA, Ultralytics, MLPerf, community). - Edge to cloud: Jetson, Apple Silicon, desktop GPUs, SBCs. - Vision + LLMs: Looks up exact matches, falls back to FLOPs/TFLOPS scaling with memory fit checks. - CLI-first: Fast, scriptable, offline-capable once the DB is installed. The interactive UI is the primary entrypoint: # recommended: isolated CLI install pipx install willitrun willitrun # or with pip if you prefer pip install willitrun willitrun Pick a model (name, HF ID, or local file), then pick a device. You’ll get a verdict and context instantly. Power users (non-interactive): willitrun --device [--json] Install extras as needed: # PyTorch/ONNX profiling (local .pt / .onnx parsing) pip install willitrun[profiling,onnx] - Raw (immutable cache): ingest scripts fetch and store HTTP responses in data/raw/{source}/{timestamp}.ext with TTL (configurable viatool.willitrun.raw_ttl_days ). - Normalized: the same scripts parse raw into JSONL at data/normalized/{source}.jsonl using the singleBenchmarkRecord schema (willitrun/pipeline/schema.py ). - Serving: make build_db (orpython scripts/run_pipeline.py build ) validates, deduplicates by source priority, and writesdata/benchmarks.db +data/benchmarks.meta.json . The packaged copy inwillitrun/data/ is refreshed bymake wheel . - Tier 1 (lookup): exact model+device benchmarks from SQLite. - Tier 2 (estimate): FLOPs/TFLOPS scaling plus memory fit with 20% overhead; emits ranges and a confidence marker. - Model name from DB ( yolov8n ,resnet50 , …) - Local PyTorch / ONNX file - HuggingFace model ID User CLI: willitrun --list-devices willitrun --list-models willitrun --device [--json] Pipeline (for data contributors/maintainers): make fetch # run all ingests (respects cache TTL) make build_db # normalized -> SQLite + metadata make status # coverage summary + gaps make wheel # rebuild DB, sync into package, build wheel - Add benchmarks: append JSON lines to data/normalized/.jsonl followingBenchmarkRecord . - Add devices/models: edit data/devices.yaml /data/models.yaml . - Rebuild DB: make build_db (regenerates DB + meta). - Run tests: pytest . - Open a PR with the updated normalized file(s), DB, and metadata. pip install -e ".[dev]" make build_db pytest make build_db # refresh DB + metadata make wheel # sync into package and build CI (GitHub Actions) runs make build_db + pytest on push/PR to guard against stale DBs and failing tests. - Ingest scripts need network access; cached raw files avoid re-fetching within TTL. - Tier 2 estimates rely on available reference benchmarks; uncommon models/devices may show wider ranges. - More benchmark data (community PRs welcome!) - Web frontend for shareability - Training memory estimation - Multi-GPU / model parallelism awareness - Auto-detect local hardware MIT