SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Add FA2 and sdpa support for SigLIP (#31499)

Browse Source 
* Rebase to main

* Fix attention implementation autoset for tex and vision configs

* Fixup

* Minor fixes

* Fix copies

* Fix attention_mask for FA2

* Add eqvivalence tests for siglip

* Remove right padding test

* Uncomment flaky

* Fix import

* Add to docs

* Fix test message

* Add sdpa

* Add sdpa equivalence test

* Add siglip sdpa to docs

* Fix typing for attention output

* Add sdpa tests

* Fix signature of FA2

* Autoset attn_implementation in config

* Rename bsz -> batch_size

* Move back autoset attn method

* Mark as flaky

* Correct attention mask padding

* [run-slow] siglip

* Add FA2 and sdpa docs

* Style fix

* Remove flaky for FA2 test

* Change attention implementation set

* Change attn_implementaiton propogation

* Fix typos

* Add modality to assert message

* Add more sdpa backends in test

* [run slow] siglip

* Add math sdpa backend for all options

* [run slow] siglip
This commit is contained in:
Pavel Iakubovskii
2024-07-08 11:10:02 +01:00
committed by GitHub
parent 076e66e479
commit a177821b24
5 changed files with 680 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
docs/source/en/perf_infer_gpu_one.md
View File

@@ -70,6 +70,7 @@ FlashAttention-2 is currently supported for the following architectures:
* [OPT](https://huggingface.co/docs/transformers/model_doc/opt#transformers.OPTModel)
* [Phi](https://huggingface.co/docs/transformers/model_doc/phi#transformers.PhiModel)
* [Phi3](https://huggingface.co/docs/transformers/model_doc/phi3#transformers.Phi3Model)
* [SigLIP](https://huggingface.co/docs/transformers/model_doc/siglip)
* [StableLm](https://huggingface.co/docs/transformers/model_doc/stablelm#transformers.StableLmModel)
* [Starcoder2](https://huggingface.co/docs/transformers/model_doc/starcoder2#transformers.Starcoder2Model)
* [Qwen2](https://huggingface.co/docs/transformers/model_doc/qwen2#transformers.Qwen2Model)
@@ -231,6 +232,7 @@ For now, Transformers supports SDPA inference and training for the following arc
* [wav2vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2#transformers.Wav2Vec2Model)
* [Hubert](https://huggingface.co/docs/transformers/model_doc/hubert#transformers.HubertModel)
* [data2vec_audio](https://huggingface.co/docs/transformers/main/en/model_doc/data2vec#transformers.Data2VecAudioModel)
* [SigLIP](https://huggingface.co/docs/transformers/model_doc/siglip)
* [Sew](https://huggingface.co/docs/transformers/main/en/model_doc/sew#transformers.SEWModel)
* [UniSpeech](https://huggingface.co/docs/transformers/v4.39.3/en/model_doc/unispeech#transformers.UniSpeechModel)
* [unispeech_sat](https://huggingface.co/docs/transformers/v4.39.3/en/model_doc/unispeech-sat#transformers.UniSpeechSatModel)