fix: tighten dependency pins

chore: pin dependencies

fix: resolve dependency conflict

Fix empty gallery by restoring stable Gradio gallery layout

Fix gallery collapsing issue with min-height and height

Fix gradual gr.Examples length mapping

UI: Add vertical scrollbar for generated gallery list

Update UI with dynamic + button for adding combinations

Update UI to support up to 4 calligrapher/style pairs

adjust gpu duration to 24 plus 3 per image

precache fa3 kernel and font before gpu task

fix local cache path embedder type detection

fix duration callback signature for gradio progress arg

optimize cold start with local cache paths and font resolution

improve gradio progress stages and percentages

fix zerogpu cuda fork init

fix(zerogpu): sync internal device pointer to cuda for noise generation

fix(cuda): prevent flash_sdp init on cpu

fix(zerogpu): prevent cuda init before fork

Speed up Space by initializing globally and keeping on GPU, remove manual offload

Set default font for Wen Zhengming to Running Script

Set Wen Zhengming and his text as default input

Update examples with 4 steps and add new case

Revert to BF16 inference to fix pytorch float8 backward compatibility

Enable torchao float8 wrap for pytorch matmul compatibility

Fix float8 noise generation and fix gpu container download cache miss

Update download patterns and set dtype to None for native FP8 inference

Preserve native FP8 quantization instead of un/re-quantizing

Move PyTorch nightly installation from runtime to build requirements.txt

Fix glob pattern to allow internvl safetensors

Update checkpoint to FP8 e4m3fn, force 4 steps, disable cfg

Add debug prints to diagnose startup

Fix: Remove if __name__ check so demo.launch() runs on HF Space

Fix dropdown default value warning

Pre-load InternVL embedding at startup to save GPU time

Split model loading and generation for better progress visibility

Fix T5/CLIP pre-download patterns

updating examples

20s -> 40s

Fix T5 pre-download to include all safetensors files

Remove AOT compilation code, keep FA3 + FP8 only

Dynamic GPU duration: 20s + 1.5s * steps * images

FA3 + FP8 quantization (no AOT cache)

Use correct case for XLabs-AI/xflux_text_encoders

Pre-download T5, CLIP, VAE at startup (no GPU needed)

FA3 + bf16 (clean version)

Enable FA3 by default for ZeroGPU H200

Add Flash Attention 3 support (optional)

Fix torch/torchvision version compatibility

Load compiled graph in each GPU session