DESCRIPTION: 5-track coverage + Resilient inference (BF16 merged + NF4 fallback) + expanded artifacts row

Add BF16->NF4 OOM fallback for tier-mismatch defense

Surgical refresh — align A4B cloud demo with production helpers

Update notice: anonymous/free users get 1 query (2 min/day quota)

Increase ZeroGPU duration to 120s for multi-round tool calling

Handle empty model response after tool calling

Fix empty response after tool calling

Fix KeyError and improve error handling for ZeroGPU

Fix model loading for ZeroGPU per official docs

Add GPU error diagnostics and update notice for ZeroGPU

Add styled notice banner with usage limits and disclaimer

Revert to merged model with runtime NF4 quantization

Fix: use device_map=auto for ZeroGPU compatibility

Switch to pre-quantized NF4 model repo for faster loading

Add ZeroGPU fallback: serve live API data for anonymous users

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload requirements.txt with huggingface_hub

Upload folder using huggingface_hub

Fix chat template mismatch: use fine-tuned tokenizer + tool_responses format

Show full traceback on inference errors for debugging

Remove theme param — not supported in Gradio 6.x ChatInterface

Remove type='messages' — default in Gradio 6.x

Use AutoProcessor from base google/gemma-4-E4B-it (matches official docs)

Fix: use Gemma4ImageProcessorPil, pass None for audio/video processors

Build processor manually — bypass AutoProcessor audio/video dependency

Remove torchvision — not needed after stripping audio/video from processor config

Add torchvision dependency (required by Gemma4VideoProcessor)

Add Gradio app with Gemma 4 E4B + native function calling

initial commit