fix local cache path embedder type detection

optimize cold start with local cache paths and font resolution

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

Fix float8 noise generation and fix gpu container download cache miss

Pre-load InternVL embedding at startup to save GPU time

Enable FA3 by default for ZeroGPU H200

Add Flash Attention 3 support (optional)

fp32

Fix CUBLAS errors: enforce dtype consistency at model entry points

Use pure fp32 for ZeroGPU - disable autocast entirely

Disable autocast for MLPEmbedder and Modulation to fix CUBLAS errors

Use torch.autocast for automatic mixed precision inference

Disable TF32 to fix CUBLAS fp16 errors

Use fp32 for MLPEmbedder to avoid CUBLAS errors

Switch to full model with fp16/bf16 inference for better performance

Add NF4 4-bit inference with bitsandbytes

UI improvements: move status bar to right side, simplify layout, update defaults to Wang Xizhi

fix: use float32 instead of bfloat16 for compatibility

Add batch generation, torch.compile acceleration, fix dtype issues

fix: ensure consistent dtype in prepare function (use img.dtype)

feat: Add UniCalli Chinese calligraphy generator