Resources

Resources documentation for Harbor.

Tasks declare resources in task.toml. Harbor applies CPU and memory using enforcement policies; storage, GPU, and TPU requests are passed through when the provider supports them.

Task fields

[environment]
cpus = 2
memory_mb = 4096
storage_mb = 10240
gpus = 1
gpu_types = ["H100", "A100"]

[environment.tpu]   # optional; GKE only
type = "v6e"
topology = "2x4"

Field	Description
cpus	CPU count
memory_mb	RAM in MB
storage_mb	Ephemeral disk in MB
gpus	GPU count
gpu_types	Acceptable GPU types (optional)
tpu.type	TPU accelerator type — alias (v6e, trillium, v4) or GKE label (tpu-v6e-slice)
tpu.topology	TPU topology as NxM or NxMxK (required; chip count = product of dimensions)

All fields are optional. Omitted fields use the provider’s default sizing.

Separate verifier sandboxes can set their own values under [verifier.environment].

Enforcement policies

CPU and memory each get an independent policy. Set them via --cpus / --memory, or cpu_enforcement_policy / memory_enforcement_policy in job or trial config.

Policy	Meaning	Requires cpus / memory_mb?
auto	Use the provider’s default mode	No
limit	Hard ceiling only	Yes
request	Reservation only, no ceiling	Yes
guarantee	Both reservation and hard ceiling	Yes
ignore	Do not pass the value to the provider	No

CLI

harbor run -p "<path/to/dataset>" -m "<model>" -a "<agent>" \
  -e docker --cpus limit --memory guarantee

Config

environment:
  type: docker
  cpu_enforcement_policy: limit
  memory_enforcement_policy: auto

You can override resource values at runtime:

CLI

harbor run -p "<path/to/dataset>" -m "<model>" -a "<agent>" \
  --override-cpus 4 \
  --override-memory-mb 8192 \
  --override-storage-mb 20480 \
  --override-tpu v6e=2x4

Config

environment:
  override_cpus: 4
  override_memory_mb: 8192
  override_storage_mb: 20480
  override_tpu:
    type: v6e
    topology: 2x4

Provider support

Harbor validates policies at the start of a job. Unsupported combinations fail before trials run. limit and guarantee require limit support; request and guarantee require request support.

Storage, GPUs, and TPUs do not have enforcement policies. Harbor passes their declared values to providers that support them.

Capability	Environments
CPU and memory limits	apple-container, cwsandbox, docker, ec2, gke, modal, opensandbox, openshift, podman, skypilot, wandb
CPU requests	beam, cwsandbox, daytona, e2b, gke, hyperbrowser, islo, modal, novita, openshift, runloop, skypilot, tensorlake, vercel, wandb
Memory requests	beam, blaxel, cwsandbox, daytona, e2b, gke, hyperbrowser, islo, modal, novita, openshift, runloop, skypilot, tensorlake, vercel, wandb
Storage sizing	ack, daytona, gke, hyperbrowser, islo, langsmith, runloop, tensorlake, use-computer
GPU allocation	beam, daytona, gke, modal, opensandbox
TPU allocation	gke

Validation

Check	When
Policy vs provider	Job creation
Missing value for non-auto/ignore policy	Environment start
GPU / TPU / internet requirements	Environment start
GPU and TPU both set (GKE)	Environment start