Spaces:
Running on Zero
Running on Zero
Commit History
-
fix local cache path embedder type detection
Tianshuo-Xud84e45a -
optimize cold start with local cache paths and font resolution
Tianshuo-Xue51b773 -
Speed up Space by initializing globally and keeping on GPU, remove manual offload
Tianshuo-Xuc49775d -
Fix float8 noise generation and fix gpu container download cache miss
Tianshuo-Xuce4bbb3
-
Pre-load InternVL embedding at startup to save GPU time
- Added _preload_embedding() to load embedding/tokenizer to CPU at startup - Modified XFluxSampler to accept preloaded_embedding and preloaded_tokenizer - Modified CalligraphyGenerator to pass through preloaded resources - This saves ~5s of embedding loading time in GPU session4c08c35 -
Enable FA3 by default for ZeroGPU H200
1b5453a -
Add Flash Attention 3 support (optional)
- Modified src/flux/math.py to support FA3 when USE_FA3=1 - Uses kernels library to load vllm-flash-attn3 from HuggingFace - Registered as custom op for torch.export compatibility - Falls back to PyTorch SDPA (FA2) when FA3 not available - Added kernels to requirements.txt To enable FA3: export USE_FA3=18fc8d44
-
fp32
974a879 -
Fix CUBLAS errors: enforce dtype consistency at model entry points
- Add dtype conversion at Flux model forward() entry point - Add dtype conversion in LastLayer to match weight dtype - Remove torch.compile (not compatible with ZeroGPU) - This ensures all inputs match model weights dtype regardless of ZeroGPU auto-conversion82e509e -
Use pure fp32 for ZeroGPU - disable autocast entirely
b2bfb8e -
Disable autocast for MLPEmbedder and Modulation to fix CUBLAS errors
aecc9f1 -
Use torch.autocast for automatic mixed precision inference
8af673c -
Disable TF32 to fix CUBLAS fp16 errors
e7ca422 -
Use fp32 for MLPEmbedder to avoid CUBLAS errors
6d6e01a -
Switch to full model with fp16/bf16 inference for better performance
- Use sharded safetensors model instead of NF4 quantized version - Enable fp16/bfloat16 inference for reduced memory and faster speed - Keep VAE decode in float32 for image quality stability - Auto-detect model dtype and convert inputs accordingly39fa408
-
Add NF4 4-bit inference with bitsandbytes
414150e -
UI improvements: move status bar to right side, simplify layout, update defaults to Wang Xizhi
89e2699 -
fix: use float32 instead of bfloat16 for compatibility
e7cbbce -
Add batch generation, torch.compile acceleration, fix dtype issues
- Add batch generation option (1-4 images) with gallery output - Enable torch.compile() for inference acceleration - Fix bfloat16/float32 dtype mismatches in pipeline - Change default inference steps to 25 - Update examples with batch_size parameterd3ccd4b -
fix: ensure consistent dtype in prepare function (use img.dtype)
a800c99 -
feat: Add UniCalli Chinese calligraphy generator
- Add inference.py with CalligraphyGenerator class - Add Gradio app with support for 1-7 Chinese characters - Add 90+ historical calligraphers (王羲之, 颜真卿, 赵佶, etc.) - Support Regular (楷), Running (行), Cursive (草) scripts - Add dataset files for author styles and fonts - Add flux model source code5c86cdc