From 413cf54cb63a6eb8be014f3e5465055b575a357b Mon Sep 17 00:00:00 2001 From: Disty0 Date: Thu, 5 Jun 2025 18:11:36 +0300 Subject: [PATCH] Update changelog --- CHANGELOG.md | 9 ++++++++- modules/processing_vae.py | 7 +++---- 2 files changed, 11 insertions(+), 5 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 968bb8e1c..880f16846 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,13 @@ # Change Log for SD.Next +## Update for 2025-06-05 + +- **SDNQ Quantization** + - Add group size support for convolutional layers + - Add quantized matmul support for for convolutional layers + - Fix VAE with conv quant + + ## Update for 2025-06-02 ### Highlights for 2025-06-02 @@ -31,7 +39,6 @@ Take a look at [Docs](https://github.com/vladmandic/sdnext/wiki/Docs), [Hints](h - `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` - - Add group size support for convolutional layers - Set the default quant mode to `pre` - Use per token input quant with int8 and fp8 quantized matmul - Implement better layer hijacks diff --git a/modules/processing_vae.py b/modules/processing_vae.py index 15831a5a4..b8e5dc4ae 100644 --- a/modules/processing_vae.py +++ b/modules/processing_vae.py @@ -143,11 +143,10 @@ def full_vae_decode(latents, model): if getattr(model.vae, "post_quant_conv", None) is not None: if getattr(model.vae.post_quant_conv, "bias", None) is not None: latents = latents.to(model.vae.post_quant_conv.bias.dtype) + elif "VAE" in shared.opts.sdnq_quantize_weights: + latents = latents.to(devices.dtype_vae) else: - if "VAE" in shared.opts.sdnq_quantize_weights: - latents = latents.to(devices.dtype_vae) - else: - latents = latents.to(next(iter(model.vae.post_quant_conv.parameters())).dtype) + latents = latents.to(next(iter(model.vae.post_quant_conv.parameters())).dtype) else: latents = latents.to(model.vae.dtype)