Task YAML

Complete reference for SkyPilot-compatible Task YAML.

The Mithril CLI uses the SkyPilot YAML task spec for defining and launching runs.

Because the spec is SkyPilot-compatible, existing SkyPilot tasks run without modification when targeting Mithril infrastructure.

To run on Mithril, specify:

resources:
  infra: mithril

Mithril extends the spec with a limit price declared in task config that caps what you’ll pay per hour for each instance. Clusters launch only when capacity is available at or below that price, and often run at lower cost.

config:
  mithril:
    limit_price: 32.0  # max $/hour/instance

If you’ve launched jobs with SkyPilot, the structure below will be familiar.

→ SkyPilot YAML reference

---

Spec structure

A task spec packages everything required to run your code on remote GPUs:

• Compute
• Environment setup
• Code sync
• Storage mounts
• Execution commands

Minimal example:

resources:
  accelerators: B200:8

run: python train.py

Execution lifecycle

Tasks execute in a fixed order:

Provision cluster
→ Sync workdir
→ Run setup (once)
→ Execute run commands

Subsequent runs:

• Skip provisioning if the cluster exists
• Skip setup
• Re-sync code
• Re-run commands

This creates a tight iteration loop for testing changes.

---

Top-level fields

Field	Type	Required	CLI override	Description
name	string	No	-n, --name	Task name (auto-generated if omitted)
resources	object	Yes	See resources	Compute requirements
run	string	Yes	Inline command arg	Main commands to execute
setup	string	No	—	Setup commands (run once per cluster)
workdir	string	No	--workdir	Local directory to sync
file_mounts	object	No	—	Remote storage mounts
volumes	object	No	—	Mithril volumes to mount
envs	object	No	-e KEY=VALUE	Environment variables
num_nodes	int	No	--num-nodes	Number of nodes (default: 1)
config	object	No	--config KEY=VALUE	Provider-specific config

Core SkyPilot fields

The following fields come directly from the SkyPilot spec.

resources

resources:
  infra: mithril
  accelerators: B200:8

Fields

Field	Type	CLI override	Description
infra	string	--infra	Infrastructure (mithril, mithril/region)
accelerators	string	--gpus	GPU type:count (e.g., B200:8)

→ Infra — multi-cloud support, spot bids, reservations, and hosted Kubernetes

run

Main task commands. Runs every time the task is launched.

run: python train.py

Multi-line:

run: |
  cd /app
  python train.py --epochs 100
  python eval.py

setup

One-time setup commands. Runs once when cluster is provisioned.

setup: |
  pip install -r requirements.txt
  apt-get update && apt-get install -y vim

workdir

Sync local directory to cluster.

workdir: .

Files are synced to ~/sky_workdir/ on the cluster.

file_mounts

Mount remote storage (S3, GCS, etc.).

file_mounts:
  /data: s3://my-bucket/data
  /models: gs://my-bucket/models

→ File Mounts for details.

volumes

Mount Mithril persistent disks

volumes:
  /data: my-volume-name

envs

Environment variables available in setup and run.

envs:
  WANDB_API_KEY: xxx
  BATCH_SIZE: 32
  DEBUG: "true"

num_nodes

Multi-node distributed training.

num_nodes: 2

SkyPilot sets up distributed environment variables automatically.

config

Provider-specific configuration.

config:
  mithril:
    limit_price: 8.0  # max $/hour/instance you'll pay

Default: $32.00/hour/instance if not specified.

→ Spot auction mechanics — how the spot auction, pricing, and preemption work

Complete example

name: llm-training

resources:
  infra: mithril
  accelerators: B200:8

config:
  mithril:
    limit_price: 32.0

num_nodes: 4

workdir: .

file_mounts:
  /data: s3://my-bucket/training-data

volumes:
  /checkpoints: my-checkpoints-volume

envs:
  WANDB_PROJECT: my-project
  WANDB_API_KEY: ${WANDB_API_KEY}

setup: |
  pip install -r requirements.txt
  wandb login

run: |
  torchrun \
    --nproc_per_node=8 \
    --nnodes=$SKYPILOT_NUM_NODES \
    train.py \
    --data /data \
    --output /checkpoints