Reduce default GPU duration for High Detail mode from 60s to 30s

Pad declared ZeroGPU duration with a cold-start buffer

Sync visible GPU duration slider default to Fast mode's 15s

Fix step-done log ordering vs text_encoder offload/pre-VAE-decode

Reduce default GPU duration for Fast mode from 30s to 15s

Add VAE decode logging: pre-decode mem snapshot, tiling prediction, peak reset

Raise VAE tile threshold to HIGH_DETAIL.max_dim (2048px) to skip tiling

Bump bitsandbytes to 0.49.2 for CUDA 13.0 support

Load text_encoder as int8 (bitsandbytes) when TEXT_ENCODER_INT8_REPO is set

Skip text_encoder CPU offload in fast mode; type mode as an enum

Add per-phase timing to pipeline call and text_encoder offload

Speed up fast mode: lower resolution cap, skip needless VAE tiling, fewer steps

Log exception type/repr and traceback in infer() error path

Correct data retention language to reflect automatic pruning

Replace vllm-flash-attn3 kernel with native SDPA cuDNN backend; wire up attn processor

Free text encoder before VAE decode and enable VAE tiling to fix OOM

Patch RMSNorm for fp8-resident weights; fix crash in prior fp8 commit

Keep transformer weights fp8-resident instead of upcasting at load

Refactor _infer_gpu into single-responsibility helper functions

Fix misleading "includes compile" log label; add first-touch instrumentation

Pin accelerate and peft to release versions instead of git HEAD

Revert NF4 quantization: this checkpoint's weights are stored fp8-native

Pin diffusers to 0.39.0 release to fix hub dependency conflict

Quantize transformer to NF4 via bitsandbytes to fix root-cause OOM

Gate full-GPU fast path on slice size; it OOMs on this Space's 47GB MIG

Try full pipeline GPU residency before falling back to cpu offload

Use sequential CPU offload instead of moving full pipeline to GPU at once

Use cudaMallocAsync allocator to fix MIG GPU CUDACachingAllocator crash

Add startup environment diagnostics (_log_env)

Load model on CPU at startup to bypass slow ZeroGPU disk packing

Pin spaces==0.50.4 to restore fast ZeroGPU inference

Set default GPU duration for high-detail mode to 40s

Set default GPU duration for high-detail mode to 30s

Fix Japanese IME exits on iPhone by checking e.isComposing

Handle empty dataset repo in _list_existing_files

Add user responsibility disclaimer to notice (English)

Revert FA3 and torch.compile: incompatible with this GPU environment

Fix torch.compile crash on torch.device ConstantVariable bug

Enable FA3 attention and torch.compile for faster inference

Add header link to Selectable Face Swap space

Ignore .claude/ directory

Add Python .gitignore

Batch log uploads to avoid HF 256-commits/hour rate limit

Prune log files before uploading, not after

Switch log pruning from time-based to count-based

Move GPU Duration to top of Advanced Settings

Move Bokeh Blur to front of Quick Prompts

Log gpu_duration and mode at infer start

Add user-adjustable GPU Duration slider to Advanced Settings

Raise ZeroGPU durations to 20s fast, 60s high-detail