a100x8-jobs-guide — 8× Nvidia A100 on Hugging Face Jobs

Everything you need to run workloads on the a100x8 hardware flavor: 8× Nvidia A100 (640 GB total VRAM).

Flavor specification

Property	Value
Flavor	a100x8
GPU	8× Nvidia A100 (640 GB total, 80 GB each)
vCPU	96
RAM	1136 GB
Disk	8000 GB
Price	$0.3333 / hour ($20.00 / day)

Verify with hf jobs hardware — flavors and pricing change over time.

Quickstart

hf auth login

hf jobs run --name a100x8-check --flavor a100x8 \
  pytorch/pytorch:2.6.0-cuda12.4-cudnn9-devel \
  python -c "import torch; print(torch.cuda.device_count(), torch.cuda.get_device_name(0))"

Expected output: 8 NVIDIA A100-SXM4-80GB

Multi-GPU training in one command

hf jobs run --name ddp-train --flavor a100x8 \
  pytorch/pytorch:2.6.0-cuda12.4-cudnn9-devel \
  torchrun --nproc_per_node=8 train.py

Python API

from huggingface_hub import run_job

run_job(
    image="pytorch/pytorch:2.6.0-cuda12.4-cudnn9-devel",
    command=["torchrun", "--nproc_per_node=8", "train.py"],
    flavor="a100x8",
    timeout="6h",
)

Monitoring

hf jobs ps
hf jobs logs <job_id> --follow
hf jobs inspect <job_id>
hf jobs cancel <job_id>

Cost control

a100x8 is the most expensive flavor at $0.3333/hour. Always:

• set --timeout so a stuck job cannot run forever;
• benchmark on a10g-small before scaling out;
• prefer --detach only when you will inspect logs.

Related repositories

• Sinadavinc80/hf-jobs-docker-like — the general hf jobs (Docker-like CLI) reference.

References

• Jobs Configuration
• Run and manage Jobs

License

Apache-2.0