Spaces:
Paused
Apply for a GPU community grant: Academic project
Description of the app
PaDoc is a public academic demo of the open Longin-Yu/PaDoc document parsing checkpoint. It lets researchers and developers upload a document image and observe layout-aware parallel decoding in real time. The Gradio interface intentionally exposes only parallel execution and streams the main layout output, active content branches, scheduler state, detected layout boxes, and final structured JSON while generation is running.
Justification
The approximately 2.4B-parameter Qwen3-VL-based checkpoint uses PyTorch inference and requires CUDA for a practical interactive demo. This Space is implemented specifically for ZeroGPU with @spaces.GPU, module-scope CUDA model placement, one Gradio inference job at a time, and at most eight concurrently decoded branches. It does not use vLLM or Docker, and it requests GPU capacity only while a document is being parsed. The source, checkpoint integration, and demo are public and intended for research and education. A Community GPU Grant would make this non-commercial public demo usable for the Hugging Face community.