| """RunPod Serverless worker: one Jev/SystemOne request (or a list) per job, billed per second. |
| |
| Job input: {"input": <request>} or {"input": {"requests": [<request>, ...]}}. |
| The model loads once per worker from DECIDER_MODEL (repo id or folder). |
| """ |
| import os |
| import sys |
|
|
| import runpod |
| import torch |
| from huggingface_hub import snapshot_download |
|
|
| MODEL = os.environ.get("DECIDER_MODEL", "Jeetu456/decider-2b") |
| path = MODEL if os.path.isdir(MODEL) else snapshot_download(MODEL) |
| sys.path.insert(0, path) |
| from decider_model import SchemaError, load_release_model, systemone |
|
|
| model = load_release_model(path, device="cuda" if torch.cuda.is_available() else "cpu") |
|
|
|
|
| def handler(job): |
| request = job["input"] |
| try: |
| if isinstance(request, dict) and "requests" in request: |
| return [systemone(model, r) for r in request["requests"]] |
| return systemone(model, request) |
| except (SchemaError, ValueError) as error: |
| return {"error": str(error)} |
|
|
|
|
| if __name__ == "__main__": |
| runpod.serverless.start({"handler": handler}) |
|
|
