AIARCOASC Docs
Sign inStart free

Training

Fine-tuning on single- or multi-node GPU clusters. NCCL + InfiniBand, checkpoint anywhere, spot-safe.

What it is

Training is the managed product on top of Atlas. You provide a training script and configuration; we schedule single-node or multi-node distributed training (NCCL over InfiniBand for multi-node), stream checkpoints to object storage, and auto-resume from the last checkpoint on spot reclaim or node-health eviction.

When to use it

  • Single-node fine-tune of a 7B–70B model.
  • Multi-node distributed training across 8 to 128 GPUs.
  • Spot-aware training with automatic checkpoint + resume.

Quickstart

asc training submit my-run --image my/train:latest --gpu h100 --nodes 4 --spot --auto-resume --checkpoint-uri asc://my-bucket/ckpt
run = client.training.submit(
    name='my-run',
    image='my/train:latest',
    gpu='h100',
    nodes=4,
    spot=True,
    auto_resume=True,
    checkpoint_uri='asc://my-bucket/ckpt',
)
const run = await client.training.submit({ name: 'my-run', image: 'my/train:latest', gpu: 'h100', nodes: 4, spot: true, autoResume: true, checkpointUri: 'asc://my-bucket/ckpt' });

Limits & quotas

LimitDefaultBurstNotes
Max nodes per run1281,024Lift via sales
InterconnectIB-400G / IB-200G—
Checkpoint cadenceConfigurable—Per step or per wall-time

Pricing

See pricing. Pay-as-you-go, billed monthly via Stripe.

API surface

  • POST /v1/training/runs
  • GET /v1/training/runs/{id}

Required scope(s): pods:write, storage:write. See Scopes.

Security

All access is authenticated and scoped. See Auth & scopes and Network controls.