nodus run

nodus run

Run a command as a Job: estimate, phases, logs and the final cost

Synopsis

Uploads the current directory (unless –no-source), creates a Job, prints its estimate and phases, streams its logs and ends with the cost line. The command’s exit code becomes nodus’s; 125 means a Nodus or API error, 124 a timeout and 130 an interrupt. On a terminal Ctrl+C asks before canceling; -d detaches.

nodus run [flags] -- COMMAND [ARGS...] | run APP.py[::FUNCTION] [ARGS...]

Examples

  nodus run --gpu L4 --image nodus/pytorch -- python train.py
  nodus run --gpu H100:8 --max-cost 40 -d -- torchrun train.py
  nodus run app.py::main

Options

      --allow-large-source                Upload source above 500 MiB compressed
      --checkpoint string                 Checkpoint path (default /nodus/state)
      --complete-by string                Finish-by time (RFC 3339) for placement
      --completions int                   Indexed completions
      --continuity string                 Recovery: Checkpointed, Restartable or Ephemeral
      --cpu string                        vCPUs
  -d, --detach                            Return once the Job is created
      --disk string                       Disk, such as 200Gi
      --dry-run                           Print the estimate only
      --env stringArray                   Environment variable K=V
      --expected-duration duration        Expected runtime, for the estimate
      --gpu string                        Accelerator TYPE[:COUNT], such as L4 or H100:8
      --gpus-per-node int                 Beta: GPUs per gang node
  -h, --help                              help for run
      --idempotency-key string            Override the generated idempotency key
      --image string                      Image, such as nodus/pytorch
      --interruptible string[="prefer"]   Interruptible capacity: allow, prefer or never
      --keep                              Keep the Job after it finishes (no 30-day TTL)
  -l, --label strings                     Labels k=v
      --launcher string                   Beta: plain, torchrun, ray or verl
      --max-cost string                   Spend cap in USD
      --memory string                     Memory, such as 64Gi
      --name string                       Job name (default: generated run-xxxxx)
      --network string                    Beta: colocated, regional or global
      --no-source                         Do not upload the current directory
      --no-warm                           Release capacity at once instead of keeping it warm for 60 s
      --nodes int                         Beta: gang nodes
      --output stringArray                Output NAME=/path
      --parallelism int                   Indexes running at once
      --profile string                    Placement profile: Balanced, Cost or Speed
      --region string                     Region class, such as us or eu
      --secret stringArray                Secret to expose as environment variables
      --startup-timeout duration          Beta: gang assembly budget
      --timeout duration                  Wall-clock limit, such as 2h
      --total-gpus int                    Beta: total GPUs across the gang
      --transport string                  Beta: direct or auto
      --volume stringArray                Volume NAME:/path[:ro]

Options inherited from parent commands

      --context string   Context from the config file to use
      --org string       Organization (selects the context for that org)
  -p, --project string   Project to work in
  -v, --verbose          Log each API request (never credentials)

SEE ALSO

• nodus - Run compute Jobs on Nodus

Tested examples

These invocations run in CI against the CLI’s fake API.

nodus run app.py::main --epochs 3
nodus run fail.py
nodus run --gpu L4 --image nodus/pytorch -- python train.py
nodus run --gpu L4 --no-source -- sh -c 'exit 3'
nodus run --gpu L4 --no-source --name late -- python deadline.py
nodus run -d --no-source --gpu L4 --keep --name detached -- python train.py
nodus run --dry-run --no-source --gpu L4 -- python train.py
nodus run -d --no-source --name gang --gpu H100 --gpus-per-node 8 --nodes 2 --launcher torchrun --network regional --max-cost 40 --env A=1 --interruptible -- torchrun train.py
nodus run --dry-run --no-source --gpu H100 --gpus-per-node 8 --nodes 2 -- torchrun train.py
nodus run --no-source --gpu none -- python train.py