GitLab

* proposal for flaky tests

* more precision fixes

* move more tests to use cosine distance

* more test fixes

* clean up

* use default attn

* clean up

* update expected value

* make style

* make style

* Apply suggestions from code review

* Update src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_img2img.py

* make style

* fix failing tests

---------
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* [LoRA Attn] Refactor LoRA attn

* correct for network alphas

* fix more

* fix more tests

* fix more tests

* Move below

* Finish

* better version

* correct serialization format

* fix

* fix more

* fix more

* fix more

* Apply suggestions from code review

* Update src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_img2img.py

* deprecation

* relax atol for slow test slighly

* Finish tests

* make style

* make style

* sdxl lora changes.

* better name replacement.

* better replacement.

* debugging

* debugging

* debugging

* debugging

* debugging

* remove print.

* print state dict keys.

* print

* distingisuih better

* debuggable.

* fxi: tyests

* fix: arg from training script.

* access from class.

* run style

* debug

* save intermediate

* some simplifications for SDXL LoRA

* styling

* unet config is not needed in diffusers format.

* fix: dynamic SGM block mapping for SDXL kohya loras (#4322)

* Use lora compatible layers for linear proj_in/proj_out (#4323)

* improve condition for using the sgm_diffusers mapping

* informative comment.

* load compatible keys and embedding layer maaping.

* Get SDXL 1.0 example lora to load

* simplify

* specif ranks and hidden sizes.

* better handling of k rank and hidden

* debug

* debug

* debug

* debug

* debug

* fix: alpha keys

* add check for handling LoRAAttnAddedKVProcessor

* sanity comment

* modifications for text encoder SDXL

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* denugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* up

* up

* up

* up

* up

* up

* unneeded comments.

* unneeded comments.

* kwargs for the other attention processors.

* kwargs for the other attention processors.

* debugging

* debugging

* debugging

* debugging

* improve

* debugging

* debugging

* more print

* Fix alphas

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* clean up

* clean up.

* debugging

* fix: text

---------
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Batuhan Taskaya <batuhan@python.org>

* Support to load Kohya-ss style LoRA file format (without restrictions)
Co-Authored-By: Takuma Mori <takuma104@gmail.com>
Co-Authored-By: Sayak Paul <spsayakpaul@gmail.com>

* tmp: add sdxl to mlp_modules

---------
Co-authored-by: Takuma Mori <takuma104@gmail.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

Resolve bf16 error as mentioned in this [issue](https://github.com/huggingface/diffusers/issues/4139#issuecomment-1639977304) (#4214)

* resolve bf16 error

* resolve bf16 error

* resolve bf16 error

* resolve bf16 error

* resolve bf16 error

* resolve bf16 error

* resolve bf16 error

fix inner_dim

* refactor to support patching LoRA into T5

instantiate the lora linear layer on the same device as the regular linear layer

get lora rank from state dict

tests

fmt

can create lora layer in float32 even when rest of model is float16

fix loading model hook

remove load_lora_weights_ and T5 dispatching

remove Unet#attn_processors_state_dict

docstrings

* text encoder monkeypatch class method

* fix test

---------
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Add new text encoder

* add transformers depth

* More

* Correct conversion script

* Fix more

* Fix more

* Correct more

* correct text encoder

* Finish all

* proof that in works in run local xl

* clean up

* Get refiner to work

* Add red castle

* Fix batch size

* Improve pipelines more

* Finish text2image tests

* Add img2img test

* Fix more

* fix import

* Fix embeddings for classic models (#3888)

Fix embeddings for classic SD models.

* Allow multiple prompts to be passed to the refiner (#3895)

* finish more

* Apply suggestions from code review

* add watermarker

* Model offload (#3889)

* Model offload.

* Model offload for refiner / img2img

* Hardcode encoder offload on img2img vae encode

Saves some GPU RAM in img2img / refiner tasks so it remains below 8 GB.

---------
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* correct

* fix

* clean print

* Update install warning for `invisible-watermark`

* add: missing docstrings.

* fix and simplify the usage example in img2img.

* fix setup for watermarking.

* Revert "fix setup for watermarking."

This reverts commit 491bc9f5a640bbf46a97a8e52d6eff7e70eb8e4b.

* fix: watermarking setup.

* fix: op.

* run make fix-copies.

* make sure tests pass

* improve convert

* make tests pass

* make tests pass

* better error message

* fiinsh

* finish

* Fix final test

---------
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* feat: add lora attention processor for pt 2.0.

* explicit context manager for SDPA.

* switch to flash attention

* make shapes compatible to work optimally with SDPA.

* fix: circular import problem.

* explicitly specify the flash attention kernel in sdpa

* fall back to efficient attention context manager.

* remove explicit dispatch.

* fix: removed processor.

* fix: remove optional from type annotation.

* feat: make changes regarding LoRAAttnProcessor2_0.

* remove confusing warning.

* formatting.

* relax tolerance for PT 2.0

* fix: loading message.

* remove unnecessary logging.

* add: entry to the docs.

* add: network_alpha argument.

* relax tolerance.

* add _convert_kohya_lora_to_diffusers

* make style

* add scaffold

* match result: unet attention only

* fix monkey-patch for text_encoder

* with CLIPAttention

While the terrible images are no longer produced,
the results do not match those from the hook ver.
This may be due to not setting the network_alpha value.

* add to support network_alpha

* generate diff image

* fix monkey-patch for text_encoder

* add test_text_encoder_lora_monkey_patch()

* verify that it's okay to release the attn_procs

* fix closure version

* add comment

* Revert "fix monkey-patch for text_encoder"

This reverts commit bb9c61e6faecc1935c9c4319c77065837655d616.

* Fix to reuse utility functions

* make LoRAAttnProcessor targets to self_attn

* fix LoRAAttnProcessor target

* make style

* fix split key

* Update src/diffusers/loaders.py

* remove TEXT_ENCODER_TARGET_MODULES loop

* add print memory usage

* remove test_kohya_loras_scaffold.py

* add: doc on LoRA civitai

* remove print statement and refactor in the doc.

* fix state_dict test for kohya-ss style lora

* Apply suggestions from code review
Co-authored-by: Takuma Mori <takuma104@gmail.com>

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

Make sure we also change the config when setting `encoder_hid_dim_type=="text_proj"` and allow xformers (#3615)

* fix if

* make style

* make style

* add tests for xformers

* make style

* update

* Fix temb attention

* Apply suggestions from code review

* make style

* Add tests and fix docker

* Apply suggestions from code review

* add attnprocessor to docs

* fix path to class

* create separate page for attnprocessors

* fix path

* fix path for real

* fill in docstrings

* apply feedback

* apply feedback

add kandinsky2.1

---------
Co-authored-by: yiyixuxu <yixu310@gmail,com>
Co-authored-by: Ayush Mangal <43698245+ayushtues@users.noreply.github.com>
Co-authored-by: ayushmangal <ayushmangal@microsoft.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* Cross-attention masks

prefer qualified symbol, fix accidental Optional

prefer qualified symbol in AttentionProcessor

prefer qualified symbol in embeddings.py

qualified symbol in transformed_2d

qualify FloatTensor in unet_2d_blocks

move new transformer_2d params attention_mask, encoder_attention_mask to the end of the section which is assumed (e.g. by functions such as checkpoint()) to have a stable positional param interface. regard return_dict as a special-case which is assumed to be injected separately from positional params (e.g. by create_custom_forward()).

move new encoder_attention_mask param to end of CrossAttn block interfaces and Unet2DCondition interface, to maintain positional param interface.

regenerate modeling_text_unet.py

remove unused import

unet_2d_condition encoder_attention_mask docs
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

versatile_diffusion/modeling_text_unet.py encoder_attention_mask docs
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

transformer_2d encoder_attention_mask docs
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

unet_2d_blocks.py: add parameter name comments
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

revert description. bool-to-bias treatment happens in unet_2d_condition only.

comment parameter names

fix copies, style

* encoder_attention_mask for SimpleCrossAttnDownBlock2D, SimpleCrossAttnUpBlock2D

* encoder_attention_mask for UNetMidBlock2DSimpleCrossAttn

* support attention_mask, encoder_attention_mask in KCrossAttnDownBlock2D, KCrossAttnUpBlock2D, KAttentionBlock. fix binding of attention_mask, cross_attention_kwargs params in KCrossAttnDownBlock2D, KCrossAttnUpBlock2D checkpoint invocations.

* fix mistake made during merge conflict resolution

* regenerate versatile_diffusion

* pass time embedding into checkpointed attention invocation

* always assume encoder_attention_mask is a mask (i.e. not a bias).

* style, fix-copies

* add tests for cross-attention masks

* add test for padding of attention mask

* explain mask's query_tokens dim. fix explanation about broadcasting over channels; we actually broadcast over query tokens

* support both masks and biases in Transformer2DModel#forward. document behaviour

* fix-copies

* delete attention_mask docs on the basis I never tested self-attention masking myself. not comfortable explaining it, since I don't actually understand how a self-attn mask can work in its current form: the key length will be different in every ResBlock (we don't downsample the mask when we downsample the image).

* review feedback: the standard Unet blocks shouldn't pass temb to attn (only to resnet). remove from KCrossAttnDownBlock2D,KCrossAttnUpBlock2D#forward.

* remove encoder_attention_mask param from SimpleCrossAttn{Up,Down}Block2D,UNetMidBlock2DSimpleCrossAttn, and mask-choice in those blocks' #forward, on the basis that they only do one type of attention, so the consumer can pass whichever type of attention_mask is appropriate.

* put attention mask padding back to how it was (since the SD use-case it enabled wasn't important, and it breaks the original unclip use-case). disable the test which was added.

* fix-copies

* style

* fix-copies

* put encoder_attention_mask param back into Simple block forward interfaces, to ensure consistency of forward interface.

* restore passing of emb to KAttentionBlock#forward, on the basis that removal caused test failures. restore also the passing of emb to checkpointed calls to KAttentionBlock#forward.

* make simple unet2d blocks use encoder_attention_mask, but only when attention_mask is None. this should fix UnCLIP compatibility.

* fix copies

* add: debugging to enabling memory efficient processing

* add: better warning message.

Release large tensors in attention (as soon as they're no longer required). Reduces peak VRAM by nearly 2 GB for 1024x1024 (even after slicing), and the savings scale up with image size.

* Replace `AttentionBlock` with `Attention`

* use _from_deprecated_attn_block check re: @patrickvonplaten

* add: a warning message when using xformers in a PT 2.0 env.

* Apply suggestions from code review
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

---------
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* update IF stage I pipelines

add fixed variance schedulers and lora loading

* added kv lora attn processor

* allow loading into alternative lora attn processor

* make vae optional

* throw away predicted variance

* allow loading into added kv lora layer

* allow load T5

* allow pre compute text embeddings

* set new variance type in schedulers

* fix copies

* refactor all prompt embedding code

class prompts are now included in pre-encoding code
max tokenizer length is now configurable
embedding attention mask is now configurable

* fix for when variance type is not defined on scheduler

* do not pre compute validation prompt if not present

* add example test for if lora dreambooth

* add check for train text encoder and pre compute text embeddings

* fix more

* Fix more

* fix more

* Apply suggestions from code review

* fix

* make style

* make fix-copies

* fix

* make sure torch compile

* Clean

* fix test

* diffusers==0.14.0 update

* custom diffusion update

* custom diffusion update

* custom diffusion update

* custom diffusion update

* custom diffusion update

* custom diffusion update

* custom diffusion

* custom diffusion

* custom diffusion

* custom diffusion

* custom diffusion

* apply formatting and get rid of bare except.

* refactor readme and other minor changes.

* misc refactor.

* fix: repo_id issue and loaders logging bug.

* fix: save_model_card.

* fix: save_model_card.

* fix: save_model_card.

* add: doc entry.

* refactor doc,.

* custom diffusion

* custom diffusion

* custom diffusion

* apply style.

* remove tralining whitespace.

* fix: toctree entry.

* remove unnecessary print.

* custom diffusion

* custom diffusion

* custom diffusion test

* custom diffusion xformer update

* custom diffusion xformer update

* custom diffusion xformer update

---------
Co-authored-by: Nupur Kumari <nupurkumari@Nupurs-MacBook-Pro.local>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Nupur Kumari <nupurkumari@nupurs-mbp.wifi.local.cmu.edu>

add AttnAddedKVProcessor2_0 block

add group norm type to attention processor cross attention norm

This lets the cross attention norm use both a group norm block and a
layer norm block.

The group norm operates along the channels dimension
and requires input shape (batch size, channels, *) where as the layer norm with a single
`normalized_shape` dimension only operates over the least significant
dimension i.e. (*, channels).

The channels we want to normalize are the hidden dimension of the encoder hidden states.

By convention, the encoder hidden states are always passed as (batch size, sequence
length, hidden states).

This means the layer norm can operate on the tensor without modification, but the group
norm requires flipping the last two dimensions to operate on (batch size, hidden states, sequence length).

All existing attention processors will have the same logic and we can
consolidate it in a helper function `prepare_encoder_hidden_states`

prepare_encoder_hidden_states -> norm_encoder_hidden_states re: @patrickvonplaten

move norm_cross defined check to outside norm_encoder_hidden_states

add missing attn.norm_cross check

* add only cross attention to simple attention blocks

* add test for only_cross_attention re: @patrickvonplaten

* mid_block_only_cross_attention better default

allow mid_block_only_cross_attention to default to
`only_cross_attention` when `only_cross_attention` is given
as a single boolean

* `AttentionProcessor.group_norm` num_channels should be `query_dim`

The group_norm on the attention processor should really norm the number
of channels in the query _not_ the inner dim. This wasn't caught before
because the group_norm is only used by the added kv attention processors
and the added kv attention processors are only used by the karlo models
which are configured such that the inner dim is the same as the query
dim.

* add_{k,v}_proj should be projecting to inner_dim

* rename file

* rename attention

* fix more

* rename more

* up

* more deprecation imports

* fixes

* fix AttnProcessor2_0

Fix use of AttnProcessor2_0 for cross attention with mask

* added scale_qk and out_bias flags

* fixed for xformers

* check if it has scale argument

* Update cross_attention.py

* check torch version

* fix sliced attn

* style

* set scale

* fix test

* fixed addedKV processor

* revert back AttnProcessor2_0

* if missing if

* fix inner_dim

---------
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

Remove explicit message argument

Fix typo in AttnProcessor2_0 symbol.

* add sdpa processor

* don't use it by default

* add some checks and style

* typo

* support torch sdpa in dreambooth example

* use torch attn proc by default when available

* typo

* add attn mask

* fix naming

* being doc

* doc

* Apply suggestions from code review

* polish

* torctree

* Apply suggestions from code review
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* better name

* style

* add benchamrk table

* Update docs/source/en/optimization/torch2.0.mdx

* up

* fix example

* check if processor is None

* Apply suggestions from code review
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* add fp32 benchmakr

* Apply suggestions from code review
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

replace torch.concat by torch.cat

* Fix running LoRA with xformers

* support disabling xformers

* reformat

* Add test