> ## Documentation Index
> Fetch the complete documentation index at: https://infercrane.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Python SDK

> Submit and resume durable InferCrane operations from typed Python applications.

The zero-dependency Python SDK uses the authenticated control-plane API. It does not connect to
PostgreSQL, invoke infrastructure providers, or implement deployment lifecycle logic in the client.

## Install from the repository

<Warning>
  Install the SDK from the same checked-out InferCrane tag as your control plane until a matching
  published Python distribution is listed in the release notes.
</Warning>

```bash theme={"theme":{"light":"github-light-default","dark":"vesper"}}
python -m pip install ./sdk/python
```

## Submit durable work

```python theme={"theme":{"light":"github-light-default","dark":"vesper"}}
from infercrane import InferCrane

client = InferCrane(
    api_key="...",
    base_url="https://infercrane.internal",
)

operation = client.deploy(
    name="qwen-prod",
    model="Qwen/Qwen3-8B",
    cloud="runpod",  # use a provider qualified by your installation
    gpu="L40S",
    min_replicas=1,
    max_replicas=4,
    idempotency_key="qwen-prod-initial",
)

ready = client.wait(operation.id, timeout=900)
```

A wait timeout raises `OperationTimeout` and includes the operation ID. It does not call the cancel
endpoint. A later process can resume with `client.wait(operation_id)`.

<CodeGroup>
  ```python Synchronous stream theme={"theme":{"light":"github-light-default","dark":"vesper"}}
  for event in client.stream_chat(
      "qwen-prod",
      [{"role": "user", "content": "Summarize durable operations."}],
  ):
      print(event)
  ```

  ```python Async operation theme={"theme":{"light":"github-light-default","dark":"vesper"}}
  from infercrane import AsyncInferCrane

  client = AsyncInferCrane(api_key="...", base_url="https://infercrane.internal")
  operation = await client.deploy(
      name="qwen-prod",
      model="Qwen/Qwen3-8B",
      cloud="runpod",
      gpu="L40S",
  )
  ready = await client.wait(operation.id, timeout=900)
  ```
</CodeGroup>

The stream parser yields each SSE JSON event once and requires `[DONE]`. Closing the iterator closes
the HTTP response; it never replays a partially transmitted inference request.

## Evaluate inference evidence

```python theme={"theme":{"light":"github-light-default","dark":"vesper"}}
client.set_slo_policy(
    "qwen-prod",
    max_ttft_p95_ms=250,
    max_error_rate=0.01,
)
recommendation = client.recommend("qwen-prod")
history = client.recommendations("qwen-prod", limit=20)
```

The async client exposes the same SLO and recommendation methods. Recommendations are persisted,
advisory decisions; these calls never create or resize provider resources.

## Errors

| Error                | Meaning                                                                |
| -------------------- | ---------------------------------------------------------------------- |
| `APIError`           | Typed status, code, retryability, and remediation from the control API |
| `OperationFailed`    | Persisted durable operation reached `failed`                           |
| `OperationCancelled` | Persisted operation reached `cancelled`                                |
| `OperationTimeout`   | This client stopped waiting; server-side work continues                |
| `StreamError`        | Malformed or incomplete SSE stream                                     |

The generated low-level client is available as `client.api`. Generated files carry a do-not-edit
header and are checked against [`openapi.json`](/control-api) during qualification.