”Deploy 4 RTX 4090 instances for a hyperparameter sweep.”
The agent provisions the fleet:list_gpu_types: confirms RTX 4090 availability and pricingcreate_instance: createssweep-1with RTX 4090create_instance: createssweep-2with RTX 4090create_instance: createssweep-3with RTX 4090create_instance: createssweep-4with RTX 4090instance_status: polls all four until running
Deployed 4 instances:All 4 are ready for commands, each billed at the RTX 4090’s live rate while it runs.
”Install dependencies on all of them and start training with different learning rates.”
The agent runs commands across the fleet:ssh_executeon each instance:pip install torch transformers datasetsfile_uploadon each instance: uploadstrain.pyssh_executeon each instance: starts training with different args:- sweep-1:
python train.py --lr 1e-4 - sweep-2:
python train.py --lr 5e-5 - sweep-3:
python train.py --lr 1e-5 - sweep-4:
python train.py --lr 5e-6
- sweep-1:
All 4 instances are running training with different learning rates:Check back in ~30 minutes for results, or ask me to monitor progress.
”Check training status on all instances and tear down any that are done.”
The agent checks each instance and cleans up:ssh_executeon each:tail -5 /root/train.logssh_executeon each:nvidia-smi --query-gpu=utilization.gpu --format=csv,noheader
sweep-2 (lr=5e-5) has the best loss so far at 0.287. Terminating finished instances to save money.
file_downloadon sweep-1 and sweep-2: downloads resultsdelete_instance: terminatessweep-1andsweep-2
Downloaded results from sweep-1 and sweep-2. Terminated both. 2 instances remain, still billing.