GPU job

Source: demo/gpu/app.py in the repository. The same @job, placed on a GPU worker. Needs a pool with GPUs: on the Levers AI Lab cluster that is the L40S Spot pool, elsewhere a GPU box joined with the agent or a cloud pool. Without any GPU worker the task stays pending until one joins.

cd demo/gpu

# see what GPU capacity the cluster has
levers9 machine list
levers9 pool list
levers9 worker list

# run a 4096×4096 matmul on the first available of RTXPro6000, A10G, T4
python app.py
# {"device": "cuda", "gpu": "NVIDIA RTX PRO 6000 Blackwell..., 98304 MiB", "n": 4096, "seconds": 0.02}

# add a GPU machine (on that machine)
levers9 machine create --pool rtxpro6000       # writes the join command to ~/.levers9/machines/ID.sh: run it on the box

Notes

gpu=["RTXPro6000", "A10G", "T4"] is a preference order; gpu_count asks for more than one.
The blueprint installs torch once and is cached; the first run builds it, later runs start in seconds.
Available GPU names: T4, L4, A10G, L40S, A100-40, A100-80, H100, RTXPro6000, H200.
The dashboard’s Workers page shows GPU workers and their utilisation.

Get started

Build with the SDK

Dashboard

Data Lab

RL Arena

Operate