SD.Next Release 2025-02-06

[vladmandic]

SD.Next Release 2025-06-02

Highlights

This release is all about quantization: with new SD.Next own quantization method: SDNQ
SDNQ is based on NNCF, but has been re-implemented, optimized and evolved enough to become its own quantization method!
It's fully cross-platform, supports all GPUs and includes tons of quantization methods:

• 8-bit, 6-bit, 4-bit, 2-bit and 1-bit int and uint
• 8-bit e5, e4 and fnuz float

Also unlike most traditional methods, its also applicable to nearly all model types

Hint: Even if you may not need quantization for your current model, it may be worth trying it out as it can significantly improve performance or capabilities of your existing workflow! For example, you may not have issues with SD15 or SDXL, but you may have been limited running at high resolutions or with multiple ControlNet due to VRAM requirements - this will significantly reduce memory requirements. And on-the-fly quantization takes just few seconds during model load, there is no need to have multiple quant models permanently saved.

On a different topic, SD.Next Wiki & Docs and its UI Hints and UI Localization system are community efforts and any contributions are welcome!
You dont need any coding experience, but if you learned something and you find documentation either wrong or insufficient, please do suggest edits!
Take a look at Docs, Hints and Localization contribution guides

ReadMe | ChangeLog | Docs | WiKi | Discord

Details

• SDNQ Quantization
  • Renamed NNCF to SDNQ
  • Renamed quantization scheme names to the underlying dtype names instead of NNCF names
    • INT8_SYM -> int8
    • INT8 -> uint8
    • INT4_SYM -> int4
    • INT4 -> uint4
  • Add float8_e4m3fn, float8_e5m2, float8_e4m3fnuz, float8_e5m2fnuz, int6, uint6, int2, uint2 and uint1 support
  • Add quantized matmul support for float8_e4m3fn and float8_e5m2
  • Set the default quant mode to pre
  • Use per token input quant with int8 and fp8 quantized matmul
  • Implement better layer hijacks
  • Add an option to toggle quantize with GPU
  • Fix conv quant and add support for conv quant with asym modes
  • Fix lora weight change
  • Fix high RAM usage with pre mode
  • Fix scale and zero_point not being offloaded
• IPEX
  • Disabe Dynamic Attention by default on PyTorch 2.7
  • Remove GradScaler hijack and use torch.amp.GradScaler instead
• Feature
  • TeaCache support for HiDream I1
• Changes
  • Set the default attention optimizer to Scaled-Dot-Product on all backends
  • Enable Dynamic attention for Scaled-Dot-Product with ROCm, DirectML, MPS and CPU backends
• Fixes
  • Gallery duplicate entries
  • Prompt enhancement args mismatch