GPU compute in three lines of Python

from openlink import OpenLink
ol = OpenLink(api_key="...")

# cheapest 8x H200 with real uptime history
offers = ol.search_offers(
    query="gpu_name=H200 num_gpus=8 reliability>0.99",
    order="dph",
    limit=3,
)

# launch with vLLM
job = ol.launch_instance(
    offer_id=offers[0]["id"],
    image="vllm/vllm-openai:latest",
    disk=128,
    ssh=True,
    env={"HF_TOKEN": os.environ["HF_TOKEN"]},
)
print(f"instance {job['instance_id']} created")

# wait for the box to come up, then stream logs
inst = ol.wait_for(job["instance_id"], status="running", timeout=300)
print(inst["ssh_host"], inst["ssh_port"], inst["dph"])

for line in ol.stream_logs(inst["id"]):
    if "loss" in line:
        print(line)

# scale out when the queue backs up, and tear down when it drains
if queue.depth() > 500:
    ol.launch_instance(offer_id=offers[1]["id"], image=IMAGE)

if queue.empty():
    ol.destroy_instance(inst["id"])  # billing stops here

# the same client also fronts 300+ models
resp = ol.chat.completions.create(
    model="deepseek/deepseek-v3",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(resp.choices[0].message.content)

# or pin a routing policy per call
resp = ol.chat.completions.create(
    model="qwen/qwen3-235b",
    messages=messages,
    route={"order": "latency", "fallback": True},
)

# spend, per model, this month
for row in ol.usage(period="month"):
    print(row["model"], row["tokens"], row["cost"])

# push a dataset up before the job starts
ol.copy("./data/shard-*.parquet", f"{inst['id']}:/workspace/data/")

# pull checkpoints back down
ol.copy(f"{inst['id']}:/workspace/ckpt/", "./ckpt/")

# or move straight between two instances, no local hop
ol.copy(f"{src['id']}:/workspace/ckpt/",
        f"{dst['id']}:/workspace/ckpt/")

# persistent volumes survive instance teardown
vol = ol.create_volume(size_gb=2000, region="US-West")
ol.launch_instance(offer_id=offer["id"], volumes=[vol["id"]])