GPU Training Nodes

GPU Training Nodes

Provision a raw GPU VM (a DigitalOcean droplet), run a training or fine-tuning job on it, then destroy it. Unlike Dedicated GPU Deployments, which serve a model steady-state, a training node is burst compute you spin up, train on, and tear down.

Base path: /api/v1/public/deployments/gpu-training

---

Inference vs Training — which do I want?

	Dedicated Deployment	GPU Training Node
Purpose	Serve a model	Train / fine-tune a model
Mental model	Kept warm, latency-optimized	Spin up → train → tear down
You interact via	chat/completions proxy	SSH or cloud-init on a raw VM
Backing	DO dedicated inference	DO GPU droplet
Best for	Production / low-latency serving	Fine-tunes, experiments, batch jobs

Rule of thumb: if you want to send requests to it, use a Dedicated Deployment. If you want to run python train.py on it, use a GPU Training Node.

---

Endpoints

Method	Path	Auth	Description
GET	/deployments/gpu-training/sizes	No	List GPU droplet sizes + live pricing
POST	/deployments/gpu-training	Yes	Provision a node
GET	/deployments/gpu-training/{droplet_id}	Yes	Get status + public IP
DELETE	/deployments/gpu-training/{droplet_id}	Yes	Destroy the node (stops billing)

---

List GPU sizes

No authentication required.

curl https://api.ainative.studio/api/v1/public/deployments/gpu-training/sizes

Response (200):

{
  "sizes": [
    { "slug": "gpu-mi300x8-1536gb", "label": "AMD MI300X ×8 (1536GB VRAM)", "hourly_cost_usd": 15.92, "vram_gb": 1536 },
    { "slug": "gpu-h100x8-640gb",   "label": "NVIDIA H100 ×8 (640GB VRAM)",  "hourly_cost_usd": 23.92, "vram_gb": 640 },
    { "slug": "gpu-mi300x1-192gb",  "label": "AMD MI300X ×1 (192GB VRAM)",   "hourly_cost_usd": 1.99,  "vram_gb": 192 },
    { "slug": "gpu-h100x1-80gb",    "label": "NVIDIA H100 ×1 (80GB VRAM)",   "hourly_cost_usd": 3.39,  "vram_gb": 80 }
  ]
}

Prices come from DigitalOcean's live catalog at request time. The gpu-mi300x8-1536gb node (1,536 GB VRAM) is large enough to fine-tune GLM-5-class models.

---

Provision a node

import requests

API_KEY = "sk_your_key"
BASE = "https://api.ainative.studio/api/v1/public"
H = {"Authorization": f"Bearer {API_KEY}", "Content-Type": "application/json"}

node = requests.post(
    f"{BASE}/deployments/gpu-training",
    headers=H,
    json={
        "name": "glm5-finetune",
        "size_slug": "gpu-mi300x8-1536gb",   # default
        "region": "atl1",                     # default
        "ssh_key_ids": [12345678],            # OR user_data — at least one required
    },
).json()

print(node["droplet_id"], node["hourly_cost_usd"])

Request body:

Field	Type	Required	Default	Description
name	string	Yes	—	Node name
size_slug	string	No	gpu-mi300x8-1536gb	GPU droplet size
region	string	No	atl1	DigitalOcean region
ssh_key_ids	list[int]	No*	—	DO SSH key IDs to inject (interactive access)
user_data	string	No*	—	cloud-init script run on first boot (hands-off jobs)

At least one access method is required

You must supply ssh_key_ids, user_data, or both. A node with neither would be unreachable and would bill idle — the API rejects it with 400.

Response (201):

{
  "droplet_id": 987654321,
  "name": "glm5-finetune",
  "status": "new",
  "size_slug": "gpu-mi300x8-1536gb",
  "hourly_cost_usd": 15.92
}

---

Check status

curl https://api.ainative.studio/api/v1/public/deployments/gpu-training/987654321 \
  -H "Authorization: Bearer $TOKEN"

Response (200):

{ "droplet_id": 987654321, "status": "active", "public_ip": "203.0.113.42" }

Poll until status is active and public_ip is populated before connecting.

---

Destroy the node

curl -X DELETE https://api.ainative.studio/api/v1/public/deployments/gpu-training/987654321 \
  -H "Authorization: Bearer $TOKEN"

Response (200):

{ "destroyed": true, "droplet_id": 987654321 }

Billing stops the moment the node is destroyed.

---

Cost model

Billing runs from boot (POST) to destroy (DELETE) — the whole wall-clock window, including boot, model download, and weight load, not just the training loop. On large 8×GPU nodes, cold-start is a real fraction of the bill.

Batch experiments on one hot node to amortize cold-start: pull weights once, then iterate, rather than provisioning a fresh node per run.

---

Hands-off pattern with cloud-init

Pass a cloud-init script as user_data to run a fully unattended job. A robust pattern is: train → merge → push the adapter → self-destruct. The node destroys itself on completion by calling the same DELETE /gpu-training/{id} endpoint with a short-lived AINative bearer token, so it never bills idle after the job finishes.

cloud_init = """#!/bin/bash
set -euo pipefail

# 1. train + merge + push the adapter to Hugging Face
python train.py && python merge_and_push.py

# 2. self-destruct — stop billing immediately on completion
curl -X DELETE \
  "https://api.ainative.studio/api/v1/public/deployments/gpu-training/${DROPLET_ID}" \
  -H "Authorization: Bearer ${AINATIVE_TOKEN}"
"""

node = requests.post(
    f"{BASE}/deployments/gpu-training",
    headers=H,
    json={"name": "unattended-finetune", "user_data": cloud_init},
).json()

Self-destruct guarantee

When the job finishes (success or failure), the node calls DELETE /deployments/gpu-training/{droplet_id} on itself. Combined with the ainative-gpu-training tag applied to every node on DigitalOcean, this makes idle spend easy to prevent and orphaned nodes easy to find and clean up.

Only a scoped AINative token needs to reach the node — you never place a raw DigitalOcean API key on the VM.

---

Next steps

• Dedicated GPU Deployments — serve a model on reserved GPU capacity
• Inference Overview — chat completions, embeddings, rerank, audio
• Authentication — API keys and bearer tokens