
Volumes Image Build Cache
Modal · Modal
Modal evals — Volumes & Image Build Cache (relift v3)
About Modal
Modal is a serverless cloud platform for running GPU workloads, ML inference, data pipelines, and web apps — all from Python, with no infrastructure to manage. Developers deploy functions to Modal with a single decorator and pay only for what they run.
Sample tests· showing 3 of 10
| # | Input | Expected behavior | Check |
|---|---|---|---|
| 01 | Trainer writes checkpoints to mounted modal.Volume at /checkpoints on @app.function(volumes={...}). Evaluator on second function reads missing latest.pt; writer never called volume.commit(). | Agent adds volume.commit() after each checkpoint write, documents reader must volume.reload() before read, and verifies cross-function visibility. | Pass / FailTool usecritical |
| 02 | Parallel map workers write metrics.jsonl to same Volume path without coordination; file corruption observed. Docs warn about concurrent writers; pattern uses modal.Queue or per-worker paths. | Agent serializes writes via Queue, uses worker-specific prefixes, or single writer function; commit() after each atomic write batch. | Pass / FailSafetyhigh |
| 03 | Serving @app.function reads /models/weights from Volume mounted read-only. New weights committed by training job but serving container never volume.reload(); stale inference. | Agent adds volume.reload() at start of serving handler or on schedule, documents long-lived container staleness, redeploys if enter() caches paths. | Pass / FailTool usemedium |
Rubric criteria
- Modal
- Serverless Gpu
- Volumes Image Build Cache
Recommended for
Works with
Related evals
Run this eval in your workspace
Connect your data, configure thresholds, and review results with your team.