Deploy and run ML workloads on Modal's serverless GPU cloud platform.
npx skills add https://github.com/loteiron/ZeusAgent --skill modal-loteiron
Please help me install this Agent Skill. Skill: modal Source: https://github.com/loteiron/ZeusAgent/tree/main/optional-skills/mlops/modal Command: npx skills add https://github.com/loteiron/ZeusAgent --skill modal-loteiron
.map(), and run cron-scheduled workloads with persistent Volumes for model caching. - Use Case: Deploy a text-generation model as an auto-scaling REST API: define a @app.cls with @modal.enter() to load the model once per container, attach an A10G GPU, and run modal deploy to get a production HTTPS endpoint. ## Quick Start Ask the agent to write and deploy a Modal app that runs a GPU inference function and exposes it as a web endpoint.High-intent search queries and answers about installing and using this skill.
Define a Modal App with a function or class, attach a GPU and container image, then add the @modal.fastapi_endpoint or @modal.asgi_app decorator. Run modal deploy script.py to get a persistent HTTPS endpoint with zero-downtime updates.
Install the modal package, authenticate with modal setup, and decorate a function with @app.function(gpu="T4"). Calling .remote() executes it on Modal's cloud, and modal run script.py runs it end to end with pay-per-second billing.
Modal suits serverless, scale-to-zero jobs and auto-scaling APIs defined in Python. RunPod fits longer-running pods with persistent state, while SkyPilot targets multi-cloud orchestration and cost optimization across providers.
Modal offers T4, L4, A10G, L40S, A100-40GB, A100-80GB, H100, H200, and B200. L40S is recommended for inference cost-performance, A100-80GB for large model training, and you can specify fallbacks like gpu=["H100", "A100"].
Cold starts happen when no warm container exists. Increase scaledown_window to keep containers alive, set min_containers to keep instances warm, and load models in a @modal.enter() lifecycle hook so initialization runs once per container.
Switch to a larger GPU such as A100-80GB, enable gradient checkpointing, and use mixed precision with torch.autocast. You can also verify the allocated GPU at runtime by running nvidia-smi inside the function.
ChuaHanChongValidate ML project data paths, formats, and Python dependencies before training.
jvdominoDeploy and monitor model endpoints in Domino with Grafana and NVIDIA Triton support.
organvm-i-theoriaTrack ML experiments, version datasets, and manage model registries with MLflow, W&B, and DVC.
NzettodessOrchestrates ML workloads across multiple clouds with automatic cost optimization and spot recovery.