precache fa3 kernel and font before gpu task

optimize cold start with local cache paths and font resolution

fix(cuda): prevent flash_sdp init on cpu

Preserve native FP8 quantization instead of un/re-quantizing

Pre-load InternVL embedding at startup to save GPU time

Refactor to use Float8 + torch.compile from FLUX-Kontext-fp8

Add Float8 quantization and torch.compile optimizations

Use fp32 for inference to fix CUBLAS errors on ZeroGPU

Always use fp16 for inference, convert bf16 to fp16

Add dtype parameter to fix CUDA bf16 compatibility issues

Fix: improve HF repo ID detection to avoid confusing local paths with repo IDs

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

perf: parallel loading of safetensors shards

perf: enable FlashAttention/MemEfficient SDPA backends instead of torch.compile

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

fix: use assign=True in load_state_dict to preserve checkpoint dtype

fix: ensure model dtype is bfloat16 after loading safetensors

feat: support sharded safetensors for faster model loading

fix: version compatibility and HF_TOKEN for private repo

fix: update checkpoint path to TSXu/Unicalli_Pro

feat: Add UniCalli Chinese calligraphy generator