GitLab

Add extra performance features for EMAModel, torch._foreach operations and better support for non-blocking CPU offloading (#7685)

* Add support for _foreach operations and non-blocking to EMAModel

* default foreach to false

* add non-blocking EMA offloading to SD1.5 T2I example script

* fix whitespace

* move foreach to cli argument

* linting

* Update README.md re: EMA weight training

* correct args.foreach_ema

* add tests for foreach ema

* code quality

* add foreach to from_pretrained

* default foreach false

* fix linting

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Co-authored-by: drhead <a@a.a>

Co-authored-by: YiYi Xu <yixu310@gmail,com>

* refactor lora conversion utility.

* remove error raises.

* add onetrainer support too.

* fix

* add check

* key present is checked before

* test case draft

* aply suggestions

* changed testing repo, back to old class

* forgot docstring

---------
Co-authored-by: YiYi Xu <yixu310@gmail.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* get rid of the legacy lora remnants and make our codebase lighter

* fix depcrecated lora argument

* fix

* empty commit to trigger ci

* remove print

* empty

fix
Co-authored-by: yiyixuxu <yixu310@gmail,com>

* fix typo in __call__ of pipeline_stable_diffusion_3.py

* fix typo in __call__ of pipeline_stable_diffusion_3_img2img.py

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* Update train_dreambooth_lora_sd3.py

* Update train_dreambooth_lora_sd3.py

* Update train_dreambooth_sd3.py

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* add support for lora fusion in sd3

* add test to ensure fused lora and effective lora produce same outpouts

reduce tolerance for a flax test

fix single file

* change to logit_normal as the weighting scheme

* sensible default mote

* sd3 controlnet



---------
Co-authored-by: haofanwang <haofanwang.ai@gmail.com>

[SD3] Fix mis-matched shape when num_images_per_prompt > 1 using without T5 (text_encoder_3=None) (#8558)

* fix shape mismatch when num_images_per_prompt > 1 and text_encoder_3=None

* style

* fix copies

---------
Co-authored-by: YiYi Xu <yixu310@gmail.com>
Co-authored-by: yiyixuxu <yixu310@gmail,com>

Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Co-authored-by: YiYi Xu <yixu310@gmail.com>

* add sd1.5 compatibility to controlnet-xs

* set use_linear_projection by base_block

* refine code style

Making self.upsample actually be Upsample1D

[SD3 Docs] Corrected title about loading model with T5

Corrected the documentation title to "Loading the single file checkpoint with T5" Previously, it incorrectly stated "Loading the single file checkpoint without T5" which contradicted the code snippet showing how to load the SD3 checkpoint with the T5 model

* [LoRA] text encoder: read the ranks for all the attn modules

 * In addition to out_proj, read the ranks of adapters for q_proj, k_proj, and  v_proj

 * Allow missing adapters (UNet already supports this)

* ruff format loaders.lora

* [LoRA] add tests for partial text encoders LoRAs

* [LoRA] update test_simple_inference_with_partial_text_lora to be deterministic

* [LoRA] comment justifying test_simple_inference_with_partial_text_lora

* style

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* shift factor argument to tiny

* remove shift factor rejigging from the sd3 docs

* max_sequence_length for the T5

* updated img2img

* apply suggestions

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Co-authored-by: YiYi Xu <yixu310@gmail.com>

* Fix sharding when no device_map is passed

* style

* add tests

* align

* add docstring

* format

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

[Single File Loading] Handle unexpected keys in CLIP models when `accelerate` isn't installed.  (#8462)

* update

* update

* update

* update

* update

---------
Co-authored-by: YiYi Xu <yixu310@gmail.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* tased3 docs

* apply suggestion

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

fix the position of param casting when loading them

refactor the density and weighting utilities.

Co-authored-by: bghira <bghira@users.github.com>
Co-authored-by: Kashif Rasul <kashif.rasul@gmail.com>

* initial commit

* default back to 77

* better text

* text correction

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

Update controlnet.md

Syntax error pipe -> pipeline

temp pin numpy

* fix

* up

---------
Co-authored-by: yiyixuxu <yixu310@gmail,com>

Fix the deletion of SD3 text encoders for Dreambooth/LoRA training if the text encoders are not being trained (#8536)

* Update train_dreambooth_sd3.py to fix TE garbage collection

* Update train_dreambooth_lora_sd3.py to fix TE garbage collection

---------
Co-authored-by: Kashif Rasul <kashif.rasul@gmail.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* Add lognorm and cosmap weighting

* Implement mode sampling

* Update examples/dreambooth/train_dreambooth_lora_sd3.py

* Update examples/dreambooth/train_dreambooth_lora_sd3.py

* Update examples/dreambooth/train_dreambooth_sd3.py

* Update examples/dreambooth/train_dreambooth_sd3.py

* Update examples/dreambooth/train_dreambooth_sd3.py

* Update examples/dreambooth/train_dreambooth_lora_sd3.py

* Update examples/dreambooth/train_dreambooth_sd3.py

* Update examples/dreambooth/train_dreambooth_sd3.py

* Update examples/dreambooth/train_dreambooth_lora_sd3.py

* keep timestamp sampling fully on cpu

---------
Co-authored-by: Kashif Rasul <kashif.rasul@gmail.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* pin accelerate to 0.31.0

* update dep table

* empty

* add `sentencepiece` requirement for SD3

add `sentencepiece` requirement

* Empty-Commit

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

thanks!

post release