Monitor Training Progress with AI Agents

Use MCP tools to check training logs, GPU utilization, disk space, and estimate time remaining — all through conversation.

Instead of SSH-ing in manually to check nvidia-smi and tail logs, ask your AI agent. It can read logs, parse metrics, check system health, and estimate completion time.

---

”How’s training going on my A100 instance?”

The agent checks logs and GPU state:

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

Training progress on training-run:

Epoch 3/10 | Step 4,200/14,000 | Loss: 0.3421 | LR: 1.8e-5
Epoch 3/10 | Step 4,220/14,000 | Loss: 0.3398 | LR: 1.8e-5

GPU health:

• Utilization: 96%
• Memory: 74,200 / 81,920 MiB (91%)
• Temperature: 72C (healthy, throttles at 83C)

Progress: 30% complete (4,200 / 14,000 steps). Estimated time remaining: ~4.5 hours at the current step rate (~36 steps/min). Loss is trending down steadily — training looks healthy.

---

”Check disk space and make sure I won’t run out mid-training.”

The agent inspects storage:

1. ssh_execute — df -h /
2. ssh_execute — du -sh /workspace/checkpoints/ 2>/dev/null
3. ssh_execute — ls -lhrt /workspace/checkpoints/ | tail -5

Disk usage on training-run:

Mount	Used	Available	Capacity
/	142 GB	58 GB	71%

Checkpoint directory: 89 GB across 3 checkpoints:

checkpoint-1400/  28 GB  (epoch 1)
checkpoint-2800/  30 GB  (epoch 2)
checkpoint-4200/  31 GB  (epoch 3, latest)

Each checkpoint is ~30 GB. You have 7 more epochs, which means ~210 GB of additional checkpoints — but you only have 58 GB free.

You’ll run out of disk at epoch 5. Options:

1. Delete old checkpoints: rm -rf checkpoint-1400 checkpoint-2800 (frees 58 GB)
2. Save only the latest checkpoint (add --save_total_limit 2 to your training args)
3. Write checkpoints to a mounted persistent volume instead

---

”Compare GPU utilization across all my running instances.”

The agent checks every instance:

1. list_instances — gets all running instances
2. ssh_execute on each — nvidia-smi --query-gpu=utilization.gpu,memory.used,memory.total --format=csv,noheader

Instance	GPU	Utilization	Memory	Status
training-run	A100 80GB	96%	74.2 / 80 GB	training actively
inference-prod	H100 80GB	28%	72.0 / 80 GB	serving (low traffic)
dev-box	RTX 4090	0%	0.4 / 24 GB	idle

training-run is fully utilized — healthy. inference-prod is underutilized at 28% — could downgrade to A100 if traffic stays low. dev-box is completely idle — consider terminating to save $0.35/hr.

---

Tools used in this workflow

Tool	Purpose
list_instances	Find all running instances
ssh_execute	Check GPU stats, read logs, inspect disk usage