Pass VideoMetadata so timestamps match real time

Switch live demo to uAI-NEXUS-MedVLM-1.0b-4B-RL (Qwen3-VL-4B)

update

Update HF model link to uAI-NEXUS-MedVLM-1.0a-7B-RL

update

update

Add Model link to header

update demo

Fix: use unified processor() call for text+video tokenization

Fix: load model to CPU, move to CUDA inside @spaces.GPU context

Load model with .to('cuda') for ZeroGPU compatibility

Add detailed logging to run_inference for debugging

Align inference pipeline with LlamaFactory Qwen3VL flow

Unpack video_kwargs list values to scalars for processor

Load model to CPU, move to GPU inside @spaces.GPU context

Add offload_folder for disk weight offloading during model load

Add HF_TOKEN debug log for auth troubleshooting

Large thumbnail cards with both image and button clickable

Make entire example card clickable using Button with icon

Fix: return PIL Images instead of paths for Gallery update

Fix live example: use non-interactive Gallery, robust frame parsing

Fix live example loading: use gr.update and factory function for proper state update

Rename Max Tokens to Max Response Length, hide in accordion

Replace dropdown with clickable thumbnail cards for live examples

Add 2 clickable live examples (TAL, Dense Captioning) for live inference tab

Extract at 1fps, support up to 60 frames (1 min video)

Remove unused sample_fps (no effect on PIL image inputs)

Remove flash-attn, use sdpa attention instead

Add live inference tab with ZeroGPU support

Initial MedVidQA demo: 8 medical video tasks with Qwen3.5-VL predictions