Parel SDK for Python

Official Python SDK for Parel. 100+ AI models, GPU rental (BYOM), compare, and credits via a single OpenAI-compatible API.

Feature parity with @parel-cloud/node (v0.1.1). Both SDKs wrap the same gateway and share a machine-readable parity.json.

Install

pip install parel-cloud
# Optional: enable the `parel.openai` lazy client (streaming, tools, vision)
pip install "parel-cloud[openai]"

Requires Python 3.10+.

Quick start (sync)

import os
from parel_cloud import Parel

parel = Parel(api_key=os.environ["PAREL_API_KEY"])

snapshot = parel.credits.get()
print(f"remaining: ${snapshot['remaining_usd']:.2f}")

# Drop-in OpenAI client pointed at Parel
chat = parel.openai.chat.completions.create(
    model="qwen3.5-72b",
    messages=[{"role": "user", "content": "merhaba"}],
)
print(chat.choices[0].message.content)

Quick start (async)

import asyncio
from parel_cloud import AsyncParel

async def main():
    async with AsyncParel() as parel:
        models = await parel.models.list()
        print(len(models["data"]))

asyncio.run(main())

Namespaces

Typed error handling

from parel_cloud import (
    Parel,
    ParelBudgetExceededError,
    ParelRateLimitError,
    ParelTaskNotCancellableError,
    ParelCapacityExhaustedError,
)

parel = Parel()

try:
    parel.images.generate(model="flux-schnell", prompt="a cat")
except ParelBudgetExceededError as err:
    print(f"top up: {err.message} (request_id={err.request_id})")
except ParelRateLimitError as err:
    print(f"slow down for {err.retry_after} seconds")
except ParelCapacityExhaustedError:
    print("no GPU capacity right now")
except ParelTaskNotCancellableError:
    print("task already terminal, nothing to cancel")

Every error subclasses ParelError and carries message, code, status, request_id, and param.

BYOM (bring-your-own-model) example

from parel_cloud import Parel

parel = Parel()

# 1. Validate a HuggingFace model (VRAM needs, recommended tier)
info = parel.gpu.validate_huggingface("meta-llama/Llama-3.1-8B-Instruct")
print(info["recommended_gpu_tier"])

# 2. Deploy
dep = parel.gpu.create(
    huggingface_id="meta-llama/Llama-3.1-8B-Instruct",
    gpu_tier="rtx4090",
    idle_timeout_minutes=15,
)

# 3. Wait for it to come online
running = parel.gpu.wait_for_running(dep["id"])

# 4. Inference
resp = parel.gpu.chat(dep["id"], {
    "messages": [{"role": "user", "content": "hi"}],
})
print(resp["choices"][0]["message"]["content"])

# 5. Tear down when done
parel.gpu.delete(dep["id"])

Configuration

Retries

GET, HEAD, OPTIONS, PUT, DELETE are retried on 429, 5xx, and transport errors with jittered exponential backoff (500ms base, capped at 8s). POST and PATCH are not auto-retried (the gateway considers them non-idempotent at v0.1).

JavaScript / TypeScript

Same SDK shape, different language: @parel-cloud/node.

Roadmap

• LangChain, LlamaIndex, CrewAI, DSPy integrations (as separate packages)
• Idempotency-Key support for retryable POSTs
• SSE streaming helpers for long-running generations
• Multipart file upload for audio.transcribe

License

MIT © Aleonis Teknoloji