AWS - Veri Documentation

AWS is the default provider and the most reliable for NVIDIA training. Omit the provider field entirely to use it.

GPU types

GPU type	VRAM	Instance	GPUs per node
`L4-24GB`	24 GB	g6.xlarge (1), g6.12xlarge (4)	1 or 4
`A10G-24GB`	24 GB	g5.xlarge (1), g5.12xlarge (4)	1 or 4
`A100-80GB`	80 GB	p4de.24xlarge	8
`H100-80GB`	80 GB	p5.48xlarge	8

L4 and A10G come as 1-GPU or 4-GPU instances: set gpu_count to 1 or 4. A100 and H100 come only as whole 8-GPU nodes: set gpu_count to 8 (you rent and pay for all 8 cards). Any other gpu_count is rejected at submit with a 400 that names the providers selling that shape, for example Vast.ai for 1, 2, or 4 A100/H100 GPUs.

Submitting a job

job = client.training_jobs.create(
    base_model="Qwen/Qwen2.5-0.5B-Instruct",
    dataset_id=dataset.id,
    reward_source=open("reward.py").read(),
    gpu_type="L4-24GB",
    gpu_count=1,
    # provider defaults to "aws"
)

Notes

Billing is per-second, on on-demand instances.
Regions: AWS is the only provider with manual region selection — pass region at submit, or omit it for the default. Unknown regions 400 with the offered list.
Multi-node gangs (num_nodes >= 2, custom-script only) are AWS-only and require the 8-GPU A100/H100 shapes. Multi-node is not enabled yet: a num_nodes of 2 or more currently returns a 400.

Where to go next

All providers

Compare GPUs, billing, and status across providers.

Custom scripts (L3)

Bring your own training loop — runs on AWS NVIDIA GPUs.