Deploy and Train Models with AI Agents

Walk through a full training workflow without leaving your editor. Your AI agent handles GPU provisioning, file transfers, training execution, and cleanup.

---

​

”Deploy an A100, upload my dataset, and start training.”

The agent orchestrates the entire setup:

1. list_gpu_types — confirms A100 80GB availability and pricing ($1.60/hr)
2. create_instance — deploys an instance named fine-tune-run with an A100
3. instance_status — polls every 10 seconds until the instance is running
4. file_upload — uploads train.py to /workspace/train.py
5. file_upload — uploads dataset.jsonl to /workspace/dataset.jsonl
6. ssh_execute — runs pip install torch transformers datasets accelerate
7. ssh_execute — runs nohup python /workspace/train.py > /root/train.log 2>&1 &

Deployed fine-tune-run (1x A100 80GB, $1.60/hr). Uploaded train.py and dataset.jsonl. Training started — check progress with “how’s the training going?"

---

​

"How’s the training going? Show me GPU usage and the last few log lines.”

The agent checks two things in parallel:

1. ssh_execute — runs nvidia-smi --query-gpu=utilization.gpu,memory.used,memory.total --format=csv,noheader
2. ssh_execute — runs tail -20 /root/train.log

GPU utilization: 97% | Memory: 68,420 / 81,920 MiB (83%)

Epoch 2/5 | Step 1240/3100 | Loss: 0.4812 | LR: 2.1e-5
Epoch 2/5 | Step 1260/3100 | Loss: 0.4756 | LR: 2.0e-5
Epoch 2/5 | Step 1280/3100 | Loss: 0.4701 | LR: 2.0e-5

Loss is decreasing steadily. At the current rate, training will finish in ~2.5 hours.

---

​

”Training is done. Download the model checkpoint and tear down the instance.”

The agent wraps up the workflow:

1. ssh_execute — runs ls -lh /workspace/output/ to confirm checkpoint files
2. file_download — downloads /workspace/output/adapter_model.safetensors
3. file_download — downloads /workspace/output/adapter_config.json
4. file_download — downloads /workspace/output/training_args.json
5. billing_usage — checks total spend for this session
6. delete_instance — terminates fine-tune-run

Downloaded 3 files (adapter weights, config, training args). Total cost for this training run: $6.40 (4 hours on A100). Instance fine-tune-run terminated. Billing stopped.

---

​

Tools used in this workflow