Files

T

Eric Rollei ff8dd5d691 Add VRAM management improvements and documentation

- Enhanced Force Unload with 'nuke_orphaned_tensors' option for post-OOM cleanup
- Improved HunyuanModelCache.clear() to handle meta tensors gracefully
- Added Soft Unload node with quantized/meta tensor detection
- Added auto-offload override for INT8 at high resolutions
- Fixed _resolve_loader_reserve AttributeError
- Updated VRAM estimation formula to empirical 12*MP^1.4

Documentation:
- Added PERFORMANCE_NOTES.md with benchmarks and recommendations
- Added feature request drafts for bitsandbytes and accelerate
- Added example workflows and images

2025-11-30 16:56:01 -08:00

290 B

Raw Permalink Blame History

Example Gallery

Use this directory to store curated renders that demonstrate different loader/generator combos and prompt rewriting modes.

Recommended naming: workflow-name_resolution_seed.png with a matching .md snippet describing prompt, steps, sampler, and any notable settings.

290 B Raw Permalink Blame History

Example Gallery

290 B

Raw Permalink Blame History