* repair flash attention in _ext
this does not fix the currently broken fa behind the define, which is only used by VAE
Co-authored-by: FSSRepo <FSSRepo@users.noreply.github.com>
* make flash attention in the diffusion model a runtime flag
no support for sd3 or video
* remove old flash attention option and switch vae over to attn_ext
* update docs
* format code
---------
Co-authored-by: FSSRepo <FSSRepo@users.noreply.github.com>
Co-authored-by: leejet <leejet714@gmail.com>
* first attempt at updating to photomaker v2
* continue adding photomaker v2 modules
* finishing the last few pieces for photomaker v2; id_embeds need to be done by a manual step and pass as an input file
* added a name converter for Photomaker V2; build ok
* more debugging underway
* failing at cuda mat_mul
* updated chunk_half to be more efficient; redo feedforward
* fixed a bug: carefully using ggml_view_4d to get chunks of a tensor; strides need to be recalculated or set properly; still failing at soft_max cuda op
* redo weight calculation and weight*v
* fixed a bug now Photomaker V2 kinds of working
* add python script for face detection (Photomaker V2 needs)
* updated readme for photomaker
* fixed a bug causing PMV1 crashing; both V1 and V2 work
* fixed clean_input_ids for PMV2
* fixed a double counting bug in tokenize_with_trigger_token
* updated photomaker readme
* removed some commented code
* improved reconstructing class word free prompt
* changed reading id_embed to raw binary using existing load tensor function; this is more efficient than using model load and also makes it easier to work with sd server
* minor clean up
---------
Co-authored-by: bssrdf <bssrdf@gmail.com>
* mmdit-x
* add support for sd3.5 medium
* add skip layer guidance support (mmdit only)
* ignore slg if slg_scale is zero (optimization)
* init out_skip once
* slg support for flux (expermiental)
* warn if version doesn't support slg
* refactor slg cli args
* set default slg_scale to 0 (oops)
* format code
---------
Co-authored-by: leejet <leejet714@gmail.com>
* fix and improve: VAE tiling
- properly handle the upper left corner interpolating both x and y
- refactor out lerp
- use smootherstep to preserve more detail and spend less area blending
* actually fix vae tile merging
Co-authored-by: stduhpf <stephduh@live.fr>
* remove the now unused lerp function
---------
Co-authored-by: stduhpf <stephduh@live.fr>
* Fix includes and init vulkan the same as llama.cpp
* Add Windows Vulkan CI
* Updated ggml submodule
* support epsilon as a parameter for ggml_group_norm
---------
Co-authored-by: Cloudwalk <cloudwalk@icculus.org>
Co-authored-by: Oleg Skutte <00.00.oleg.00.00@gmail.com>
Co-authored-by: leejet <leejet714@gmail.com>
* add flux support
* avoid build failures in non-CUDA environments
* fix schnell support
* add k quants support
* add support for applying lora to quantized tensors
* add inplace conversion support for f8_e4m3 (#359)
in the same way it is done for bf16
like how bf16 converts losslessly to fp32,
f8_e4m3 converts losslessly to fp16
* add xlabs flux comfy converted lora support
* update docs
---------
Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com>
This change makes checkpoints load significantly faster by optimizing
pkzip's cyclic redundancy check. This code was developed by Intel and
Google and Mozilla. See Chromium's zlib codebase for further details.
