* 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 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>
Added NVIDEA's new "Align Your Steps" style scheduler in accordance with their
quick start guide. Currently has handling for SD1.5, SDXL, and SVD, using the
noise levels from their paper to generate the sigma values. Can be selected
using the --schedule ays command line switch. Updates the main.cpp help
message and README to reflect this option, also they now inform the user
of the --color switch as well.
---------
Co-authored-by: leejet <leejet714@gmail.com>
* apply pmid lora only once for multiple txt2img calls
* add better support for SDXL LoRA
* fix for some sdxl lora, like lcm-lora-xl
---------
Co-authored-by: bssrdf <bssrdf@gmail.com>
Co-authored-by: leejet <leejet714@gmail.com>
* Fixed a double free issue when running multiple backends on the CPU, eg: CLIP
and the primary backend, as this would result in the *_backend pointers both
pointing to the same thing resulting in a segfault when calling the
StableDiffusionGGML destructor.
* Improve logging to allow for a color switch on the command line interface.
Changed the base log_printf function to not bake the log level directly into
the log buffer as that information is already passed the logging function via
the level parameter and it's easier to add in there than strip it out.
* Added a fix for certain SDXL LoRAs that don't seem to follow the expected
naming convention, converts over the tensor name during the LoRA model
loading. Added some logging of useful LoRA loading information. Had to
increase the base size of the GGML graph as the existing size results in an
insufficient graph memory error when using SDXL LoRAs.
* small fixes
---------
Co-authored-by: leejet <leejet714@gmail.com>
* first efforts at implementing photomaker; lots more to do
* added PhotoMakerIDEncoder model in SD
* fixed soem bugs; now photomaker model weights can be loaded into their tensor buffers
* added input id image loading
* added preprocessing inpit id images
* finished get_num_tensors
* fixed a bug in remove_duplicates
* add a get_learned_condition_with_trigger function to do photomaker stuff
* add a convert_token_to_id function for photomaker to extract trigger word's token id
* making progress; need to implement tokenizer decoder
* making more progress; finishing vision model forward
* debugging vision_model outputs
* corrected clip vision model output
* continue making progress in id fusion process
* finished stacked id embedding; to be tested
* remove garbage file
* debuging graph compute
* more progress; now alloc buffer failed
* fixed wtype issue; input images can only be 1 because issue with transformer when batch size > 1 (to be investigated)
* added delayed subject conditioning; now photomaker runs and generates images
* fixed stat_merge_step
* added photomaker lora model (to be tested)
* reworked pmid lora
* finished applying pmid lora; to be tested
* finalized pmid lora
* add a few print tensor; tweak in sample again
* small tweak; still not getting ID faces
* fixed a bug in FuseBlock forward; also remove diag_mask op in for vision transformer; getting better results
* disable pmid lora apply for now; 1 input image seems working; > 1 not working
* turn pmid lora apply back on
* fixed a decode bug
* fixed a bug in ggml's conv_2d, and now > 1 input images working
* add style_ratio as a cli param; reworked encode with trigger for attention weights
* merge commit fixing lora free param buffer error
* change default style ratio to 10%
* added an option to offload vae decoder to CPU for mem-limited gpus
* removing image normalization step seems making ID fidelity much higher
* revert default style ratio back ro 20%
* added an option for normalizing input ID images; cleaned up debugging code
* more clean up
* fixed bugs; now failed with cuda error; likely out-of-mem on GPU
* free pmid model params when required
* photomaker working properly now after merging and adapting to GGMLBlock API
* remove tensor renaming; fixing names in the photomaker model file
* updated README.md to include instructions and notes for running PhotoMaker
* a bit clean up
* remove -DGGML_CUDA_FORCE_MMQ; more clean up and README update
* add input image requirement in README
* bring back freeing pmid lora params buffer; simply pooled output of CLIPvision
* remove MultiheadAttention2; customized MultiheadAttention
* added a WIN32 get_files_from_dir; turn off Photomakder if receiving no input images
* update docs
* fix ci error
* make stable-diffusion.h a pure c header file
This reverts commit 27887b630db6a92f269f0aef8de9bc9832ab50a9.
* fix ci error
* format code
* reuse get_learned_condition
* reuse pad_tokens
* reuse CLIPVisionModel
* reuse LoraModel
* add --clip-on-cpu
* fix lora name conversion for SDXL
---------
Co-authored-by: bssrdf <bssrdf@gmail.com>
Co-authored-by: leejet <leejet714@gmail.com>
* add controlnet to pipeline
* add cli params
* control strength cli param
* cli param keep controlnet in cpu
* add Textual Inversion
* add canny preprocessor
* refactor: change ggml_type_sizef to ggml_row_size
* process hint once time
* ignore the embedding name case
---------
Co-authored-by: leejet <leejet714@gmail.com>
