fix: PCB use pcb3/bad009 (Wrong Position), candle prompt with numbered labels and total count

fix: verified examples - leather cut (precise bbox) + candle counting (4 perfect bboxes)

fix: PCB use pcb4/bad/001 (accurate missing part bbox), verified visualization

fix: use bottle broken_large for localization (accurate bbox), keep screw/000 for counting (5 individual bboxes)

fix: duration 5s/10s, stable screw prompt, parse bbox_label field

fix: load model directly to cuda like ltx-2-TURBO, enable flash_attention_2

fix: replace PCB with pcb3 (clearer defect), Run Long to 60s

feat: show GPU Load time (ZeroGPU weight transfer) separately from preprocess and generate

feat: show attn impl + timing breakdown; remove metal nut localization example

fix: remove timing display

fix: fallback to sdpa if flash-attn not available

perf: reduce GPU duration to 15s/30s with flash_attention_2

perf: enable flash_attention_2 for faster inference

feat: two run buttons - Run (30s) for normal tasks, Run Long (60s) for counting/detection

fix: example 5 change from duplicate cable to PCB localization

feat: replace capsule counting with screw_bag (2 screws + 3 nuts individually detected)

fix: capsule counting prompt - pure count instead of bbox (model not trained for per-object detection)

perf: compress example images from 12MB to 542KB (512px JPEG)

feat: show preprocessing vs inference timing in output

feat: single-image mode, 7 examples (discrimination, description, 3x localization, counting, OCR), bbox visualization

fix: reduce GPU duration to 30s

fix: reduce GPU duration from 120s to 60s to save quota

feat: auto-visualize bounding boxes on test image when model outputs bbox JSON

fix: simplify to single @spaces.GPU function, remove inference_mode decorator, fix type hint

fix: switch from ZeroGPU to dedicated a10g-small, remove @spaces.GPU to avoid TypeError

fix: pass file paths instead of PIL objects to ZeroGPU worker to avoid serialization TypeError

fix: replace @torch .inference_mode decorator with context manager for ZeroGPU compat

debug: return full traceback in output for diagnosis

debug: wrap predict with traceback logging to diagnose TypeError

fix: ZeroGPU TypeError - cast slider to int, use cuda device, copy images before resize

feat: 5 diverse examples - discrimination, description, localization, counting, OCR

feat: diverse examples showcasing 5 capabilities + real images

fix: add torchvision, remove device_map for ZeroGPU, fix processor args

fix: use AutoModelForImageTextToText, add spaces dep

Upload folder using huggingface_hub