GrandMasterPomidor/qwen-omni-endpoint-handler

Qwen Omni Hugging Face Inference Endpoint Handler

This directory contains a reusable custom handler for deploying Qwen 3 Omni models (via the Hugging Face Inference Endpoints service). The handler mirrors the multi-modal interaction blueprint from the official Qwen audio/visual dialogue cookbook and supports text, image, and audio turns in a single payload.

Files

• handler.py – entry-point loaded by the Inference Endpoint runtime.
• requirements.txt – Python dependencies installed before the handler is imported.

Usage

1. Upload the contents of this directory (handler.py, requirements.txt) to a Hugging Face model repository that you control (defaults to GrandMasterPomidor/qwen-omni-endpoint-handler via the provided Makefile).
2. Provision a custom Inference Endpoint that references that repository and the Qwen Omni model weights you wish to serve. Set environment variables such as MODEL_ID to point at your chosen checkpoint (e.g. Qwen/Qwen2.5-Omni-Mini).
3. Send JSON payloads to the endpoint as documented in the header docstring of handler.py.

Refer to the accompanying Makefile for convenience targets to package and push these assets.