Skip to main content

Getting Started with Managed Inference

All models listed in the overview can be queried via the api.crusoe.ai path.

Retrieving your Inference API token​

You can retrieve your Inference API token via the Crusoe Cloud console by following the steps below.

- Visit the [Crusoe Cloud console](https://console.crusoecloud.com/)
- Click the "Security" tab in the left nav
- Select the "Inference API Key" tab in the top bar
- Click the "Create Inference API Key" button on the page
- Optionally provide an alias or an expiration date
- Click the "Create" button to view and save your API key

Querying Text models​

After retrieving an API key from the Crusoe Console, you can use the OpenAI SDK to make requests. The example below uses the meta-llama/Llama-3.3-70B-Instruct model.

import os
from openai import OpenAI

CRUSOE_API_KEY = os.getenv("CRUSOE_API_KEY")
client = OpenAI(
    api_key=CRUSOE_API_KEY,
    base_url="https://api.crusoe.ai/v1",
)

completion = client.chat.completions.create(
    model="meta-llama/Llama-3.3-70B-Instruct",
    messages=[
        {"role": "system", "content": "You are a helpful, concise assistant."},
        {"role": "user", "content": "Who is Robinson Crusoe?"},
    ],
)

print(completion.choices[0].message.content)

Interacting with Video to Video models (Batch)​

There are four main steps involved in submitting a request to the Decart Mirage - Batch model, namely:

  • Upload the input video via the files API
  • Submit the request to the queue API
  • Monitor the status of the job via the queue API
  • Download the output video via the files API

We break this down via the code snippets below.

import requests
import time

INPUT_VIDEO_PATH = "<INPUT_VIDEO_PATH>"
OUTPUT_VIDEO_PATH = "<OUTPUT_VIDEO_PATH>"

AUTH_TOKEN = "<YOUR_API_KEY>"
PROJECT_ID = "<YOUR_PROJECT_ID>"

# Step 1: Upload the video file.
with open(INPUT_VIDEO_PATH, "rb") as f:
    upload_response = requests.post(
        "https://api-video.crusoe.ai/v1/files",
        headers={
            "Authorization": f"Bearer {AUTH_TOKEN}",
            "Crusoe-Project-Id": PROJECT_ID,
        },
        files={"file": f},
        data={"purpose": "video"},
    )
    upload_response.raise_for_status()
upload_data = upload_response.json()
file_id = upload_data["id"]
if not file_id:
    raise ValueError("Could not find 'id' in upload response.")

print(f"File uploaded successfully. id: {file_id}")

# Step 2: Submit the inference request.
inference_response = requests.post(
    "https://api-video.crusoe.ai/v1/queue/decart/miragelsd-1-batch/enhanced",
    headers={
        "Authorization": f"Bearer {AUTH_TOKEN}",
        "Crusoe-Project-Id": PROJECT_ID,
        "Content-Type": "application/json",
    },
    json={
        "file_id": file_id,
        "prompt": "<YOUR_PROMPT_DESCRIBING_THE_DESIRED_VIDEO>",
    },
)
inference_response.raise_for_status()
inference_data = inference_response.json()
request_id = inference_data["request_id"]
status_url = inference_data["status_url"]
if not request_id or not status_url:
    raise ValueError("Could not find 'request_id' or 'status_url' in enqueue response.")

print(f"Request enqueued. request_id: {request_id}")

# Step 3: Poll for job completion.
is_complete = False
result_file_id = None
while not is_complete:
    status_response = requests.get(
        status_url,
        headers={
            "Authorization": f"Bearer {AUTH_TOKEN}",
            "Crusoe-Project-Id": PROJECT_ID,
        },
    )

    status_response.raise_for_status()
    status_data = status_response.json()
    status = status_data["status"]
    if status == 200:
        is_complete = True
        result_file_id = status_data["result_file_id"]
    else:
        time.sleep(2)

print(f"Request complete. result_file_id: {result_file_id}")

# Step 4: Download the result.
download_response = requests.get(
    f"https://api-video.crusoe.ai/v1/files/{result_file_id}",
    headers={
        "Authorization": f"Bearer {AUTH_TOKEN}",
        "Crusoe-Project-Id": PROJECT_ID,
    },
    stream=True,
)
download_response.raise_for_status()
with open(OUTPUT_VIDEO_PATH, "wb") as f:
    for chunk in download_response.iter_content(chunk_size=8192):
        f.write(chunk)

print(f"File downloaded successfully to: {OUTPUT_VIDEO_PATH}")