* 2x speedup using memory efficient attention
* remove einops dependency
* Swap K, M in op instantiation
* Simplify code, remove unnecessary maybe_init call and function, remove unused self.scale parameter
* make xformers a soft dependency
* remove one-liner functions
* change one letter variable to appropriate names
* Remove Env variable dependency, remove MemoryEfficientCrossAttention class and use enable_xformers_memory_efficient_attention method
* Add memory efficient attention toggle to img2img and inpaint pipelines
* Clearer management of xformers' availability
* update optimizations markdown to add info about memory efficient attention
* add benchmarks for TITAN RTX
* More detailed explanation of how the mem eff benchmark were ran
* Removing autocast from optimization markdown
* import_utils: import torch only if is available
Co-authored-by: Nouamane Tazi <nouamane98@gmail.com>
* document cpu offloading method
* address review comments
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* Docs: refer to pre-RC version of PyTorch 1.13.0.
* Remove temporary workaround for unavailable op.
* Update comment to make it less ambiguous.
* Remove use of contiguous in mps.
It appears to not longer be necessary.
* Special case: use einsum for much better performance in mps
* Update mps docs.
* Minor doc update.
* Accept suggestion
Co-authored-by: Anton Lozhkov <anton@huggingface.co>
Co-authored-by: Anton Lozhkov <anton@huggingface.co>
* Update README.md
Additionally add FLAX so the model card can be slimmer and point to this page
* Find and replace all
* v-1-5 -> v1-5
* revert test changes
* Update README.md
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* Update docs/source/quicktour.mdx
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Update README.md
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Update docs/source/quicktour.mdx
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Update README.md
Co-authored-by: Suraj Patil <surajp815@gmail.com>
* Revert certain references to v1-5
* Docs changes
* Apply suggestions from code review
Co-authored-by: apolinario <joaopaulo.passos+multimodal@gmail.com>
Co-authored-by: anton-l <anton@huggingface.co>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
Co-authored-by: Suraj Patil <surajp815@gmail.com>
* mps: alt. implementation for repeat_interleave
* style
* Bump mps version of PyTorch in the documentation.
* Apply suggestions from code review
Co-authored-by: Suraj Patil <surajp815@gmail.com>
* Simplify: do not check for device.
* style
* Fix repeat dimensions:
- The unconditional embeddings are always created from a single prompt.
- I was shadowing the batch_size var.
* Split long lines as suggested by Suraj.
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Suraj Patil <surajp815@gmail.com>
* initial commit
* make UNet stream capturable
* try to fix noise_pred value
* remove cuda graph and keep NB
* non blocking unet with PNDMScheduler
* make timesteps np arrays for pndm scheduler
because lists don't get formatted to tensors in `self.set_format`
* make max async in pndm
* use channel last format in unet
* avoid moving timesteps device in each unet call
* avoid memcpy op in `get_timestep_embedding`
* add `channels_last` kwarg to `DiffusionPipeline.from_pretrained`
* update TODO
* replace `channels_last` kwarg with `memory_format` for more generality
* revert the channels_last changes to leave it for another PR
* remove non_blocking when moving input ids to device
* remove blocking from all .to() operations at beginning of pipeline
* fix merging
* fix merging
* model can run in other precisions without autocast
* attn refactoring
* Revert "attn refactoring"
This reverts commit 0c70c0e189cd2c4d8768274c9fcf5b940ee310fb.
* remove restriction to run conv_norm in fp32
* use `baddbmm` instead of `matmul`for better in attention for better perf
* removing all reshapes to test perf
* Revert "removing all reshapes to test perf"
This reverts commit 006ccb8a8c6bc7eb7e512392e692a29d9b1553cd.
* add shapes comments
* hardcore whats needed for jitting
* Revert "hardcore whats needed for jitting"
This reverts commit 2fa9c698eae2890ac5f8e367ca80532ecf94df9a.
* Revert "remove restriction to run conv_norm in fp32"
This reverts commit cec592890c32da3d1b78d38b49e4307aedf459b9.
* revert using baddmm in attention's forward
* cleanup comment
* remove restriction to run conv_norm in fp32. no quality loss was noticed
This reverts commit cc9bc1339c998ebe9e7d733f910c6d72d9792213.
* add more optimizations techniques to docs
* Revert "add shapes comments"
This reverts commit 31c58eadb8892f95478cdf05229adf678678c5f4.
* apply suggestions
* make quality
* apply suggestions
* styling
* `scheduler.timesteps` are now arrays so we dont need .to()
* remove useless .type()
* use mean instead of max in `test_stable_diffusion_inpaint_pipeline_k_lms`
* move scheduler timestamps to correct device if tensors
* add device to `set_timesteps` in LMSD scheduler
* `self.scheduler.set_timesteps` now uses device arg for schedulers that accept it
* quick fix
* styling
* remove kwargs from schedulers `set_timesteps`
* revert to using max in K-LMS inpaint pipeline test
* Revert "`self.scheduler.set_timesteps` now uses device arg for schedulers that accept it"
This reverts commit 00d5a51e5c20d8d445c8664407ef29608106d899.
* move timesteps to correct device before loop in SD pipeline
* apply previous fix to other SD pipelines
* UNet now accepts tensor timesteps even on wrong device, to avoid errors
- it shouldnt affect performance if timesteps are alrdy on correct device
- it does slow down performance if they're on the wrong device
* fix pipeline when timesteps are arrays with strides
* Initial version of `fp16` page.
* Fix typo in README.
* Change titles of fp16 section in toctree.
* PR suggestion
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* PR suggestion
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* Clarify attention slicing is useful even for batches of 1
Explained by @patrickvonplaten after a suggestion by @keturn.
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* Do not talk about `batches` in `enable_attention_slicing`.
* Use Tip (just for fun), add link to method.
* Comment about fp16 results looking the same as float32 in practice.
* Style: docstring line wrapping.
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Ilmari Heikkinen