* Support q8_0 quantization for image model loading
q4_0 may degrade quality significantly, especially for smaller
models like SD 1.5 and SDXL. q8_0 provides a middle-ground,
giving half the memory savings of q4_0 but loading faster and
with less quality loss.
* Accept --sdquant with no parameters
* Use numerical values for the sdquant option
* Update stable-diffusion.cpp to 5900ef6605c6 (new API)
* Clean up pending LoRA code and simplify LoRA changes to upstream
* Move VAE tiling disabling for TAESD to sdtype_adapter.cpp
* Move auxiliary ctx functions to sdtype_adapter.cpp
* Use ref_images parameter for Kontext images
* Drop clip skip workaround (fixed upstream)
* Workaround for flash attention with img2img
leejet/stable-diffusion.cpp#756
* Workaround for Chroma with flash attention, debug prints
* Disable forcing CLIP weights to F32 for reduced memory usage
The recent ggml update added another quant type, GGML_TYPE_MXFP4,
which got the same value as SD_TYPE_COUNT. That made the embedded
sd.cpp quantize to GGML_TYPE_MXFP4 by default.
Photomaker in particular ends up crashing due to
"Missing CPY op for types: f32 mxfp4".
* add backend support for changing the VAE tiling threshold
* trigger VAE tiling by image area instead of dimensions
I've tested with GGML_VULKAN_MEMORY_DEBUG all resolutions with
the same 768x768 area (even extremes like 64x9216), and many
below that: all consistently allocate 6656 bytes per image pixel.
As tiling is primarily useful to avoid excessive memory usage, it
seems reasonable to enable VAE tiling based on area rather than
maximum image side.
However, as there is currently no user interface option to change
it back to a lower value, it's best to maintain the default
behavior for now.
* replace the notile option with a configurable threshold
This allows selecting a lower threshold value, reducing the
peak memory usage.
The legacy sdnotile parameter gets automatically converted to
the new parameter, if it's the only one supplied.
* simplify tiling checks, 768 default visible in launcher
---------
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Concedo
Wagner Bruna