Open to contribution: adding torch.nn.functional.scaled_dot_product_attention support for more architectures

[fxmarty]

Feature request

In Transformers 4.36, we started adding native support of torch.nn.functional.scaled_dot_product_attention (SDPA), enabled by default in Transformers: https://huggingface.co/docs/transformers/perf_infer_gpu_one#flashattention-and-memory-efficient-attention-through-pytorchs-scaleddotproductattention

SDPA allows to dispatch to memory-efficient attention, flash attention on supported GPUs (currently NVIDIA-only), and even on Intel CPUs.

For the record, here's a benchmark on some currently supported models:

Training benchmark, run on A100-SXM4-80GB.

Model	Batch size	Sequence length	Time per batch ("eager", s)	Time per batch ("sdpa", s)	Speedup	Peak memory ("eager", MB)	Peak memory ("sdpa", MB)	Memory savings
llama2 7b	4	1024	1.065	0.90	19.4%	73878.28	45977.81	60.7%
llama2 7b	4	2048	OOM	1.87	/	OOM	78394.58	SDPA does not OOM
llama2 7b	1	2048	0.64	0.48	32.0%	55557.01	29795.63	86.4%
llama2 7b	1	3072	OOM	0.75	/	OOM	37916.08	SDPA does not OOM
llama2 7b	1	4096	OOM	1.03	/	OOM	46028.14	SDPA does not OOM
llama2 7b	2	4096	OOM	2.05	/	OOM	78428.14	SDPA does not OOM

Inference benchmark, run on A100-SXM4-80GB.

Model	Batch size	Prompt length	Num new tokens	Per token latency "eager" (ms)	Per token latency "sdpa" (ms)	Speedup
llama2 13b	1	1024	1 (prefill)	178.66	159.36	12.11%
llama2 13b	1	100	100	40.35	37.62	7.28%
llama2 13b	8	100	100	40.55	38.06	6.53%
Whisper v3 large	1	/	62	20.05	18.90	6.10%
Whisper v3 large	8	/	77	25.42	24.77	2.59%
Whisper v3 large	16	/	77	28.51	26.32	8.34%

Previously, we had a partial support of SDPA in Optimum BetterTransformer but we are now looking to slowly deprecate it in favor of upstream support of SDPA directly in Transformers.

Here are the architectures for which support has been requested:

• Codegen (BetterTransformer not supporting CodeGen2 optimum#1050)
• LLAVA (Can optimum.bettertransformer supports LLAVA model? optimum#1592)
• Marian ( BetterTransforer not Support Marian optimum#1142)
• Mistral (Add support for mistral type Model to use Mistral and Zephyr optimum#1553)
• LongT5 (longT5 BetterTransformer implementation  optimum#1506)
• ViT (Add support for mistral type Model to use Mistral and Zephyr optimum#1553)

The integration could take inspiration from https://github.com/huggingface/optimum/blob/main/optimum/bettertransformer/models/decoder_models.py & https://github.com/huggingface/optimum/blob/main/optimum/bettertransformer/models/attention.py

Motivation

Faster training & inference, lower memory requirement

Your contribution

I may work on some at some point, but contributions are most welcome.

You should refer to #26572 to add the support of SDPA for a model, roughly following these steps:

• Create a XxxSdpaAttention class inheriting from XxxAttention and implement the attention logic using SDPA
• Use _prepare_4d_causal_attention_mask_for_sdpa instead of _prepare_4d_causal_attention_mask for SDPA
• Use _prepare_4d_attention_mask_for_sdpa instead of _prepare_4d_attention_mask for SDPA
• Add _supports_sdpa = True to XxxPreTrainedModel
• Add "sdpa" key to XXX_ATTENTION_CLASSES in the model modeling file

[ENate]

Hi @fxmarty I can take a look at this issue. Of I can ask questions if necessary. Or has anyone taken it already?

[davidan5]

does someone know if longT5 and all T5 models are blocked by bias support in flash attention ?

Dao-AILab/flash-attention#617

[ENate]

Hi @davidan5 are you working on the implementation?

[davidan5]

@ENate I was trying to understand the status and have an estimation of the code change to see if I can contribute.

[ENate]

I see.

[hackyon]

I'm interested in taking a look at this for the Mistral model if that's still needed. Otherwise, please let me know if there are any other models that still need some work. Thanks

[ENate]

Is LongT5 still open?

[ArthurZucker]

Mistral is already covered! LongT5 if it is like T5 and has attention bias that might not be supported

[hackyon]

Oh yea, looks like you added support for Mistral/Mixtral last month.

It doesn't seem to be supported for BERT yet (I think someone else is working on FA2 but not SDPA), so I'll take a crack at it. It looks like there is a config for relative position embeddings for BERT, so I'll just have it fallback to the original attention for configs using relative position embeddings.

@ArthurZucker - Please let me know if you know if someone else is already working on SDPA for BERT and I can look for something else to do. Thanks!

[ArthurZucker]

Not sure anyone is working on that but bert is already so small that I doubt it will have a lot of impact on perf!

[abdulfatir]

@ArthurZucker for the T5 family of models, attention bias is required, so flash-attention won't work for now but torch SDPA can still use the memory efficient kernel from xformers, right? I did some benchmarking with Chronos models (based on T5 architecture) here (amazon-science/chronos-forecasting#33) and there's a clear speedup when using torch SDPA.

[fxmarty]

@abdulfatir That's correct

[abdulfatir]

I can open a PR for T5 with SDPA then. Are there specific things that I should know of or a reference that can look at?

[fxmarty]

@abdulfatir For sure, some specific things that are good to know:

pytorch/pytorch#108108 (is_causal=True may not do what you expect)
pytorch/pytorch#110213 (You need

transformers/src/transformers/modeling_attn_mask_utils.py

Line 189 in 416711c

def _unmask_unattended(

)

example of a PR: #29108

[ArthurZucker]

FYI going forward we should rather use

transformers/src/transformers/models/llama/modeling_llama.py

Line 1058 in 416711c

def _update_causal_mask(self, attention_mask, input_tensor, cache_position):

as it is more self contained, easier to debug and maintain than the many paths in the atnn_mask utils

[sayakpaul]

Hey @abdulfatir just wanted to check in if you are still working on dropping a PR to add SDPA support for T5? It would tremendously help accelerating diffusion models that use T5.

[avishaiElmakies]

I can handle dinoV2

[avishaiElmakies]

I have some free time, I will add sdpa to Speech2text. will use bart as inspiration

[michaelshekasta]

@allenyummy Have you implement sdpa on debertav3? if not, is it sophisticated?

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[ArthurZucker]

It should not be super complicated, and a good refactoring is quite needed. Once #22105 is merged, we can easily add it!

[toneyfox]

ailab_OmniGen
Phi3Transformer does not support an attention implementation through torch.nn.functional.scaled_dot_product_attention yet. Please request the support for this architecture: #28005. If you believe this error is a bug, please open an issue in Transformers GitHub repository and load your model with the argument attn_implementation="eager" meanwhile. Example: model = AutoModel.from_pretrained("openai/whisper-tiny", attn_implementation="eager")

HELP ME。
The boss asked how to solve this problem？

[ENate]

Hi there. Are there any open models?

[michaelshekasta]

It should not be super complicated, and a good refactoring is quite needed. Once #22105 is merged, we can easily add it!

@ArthurZucker
Okay, do we have something like a roadmap? I understand that we should finish the DeBERTaV2 refactoring before implementing FA2 for DeBERTaV3. Do I understand it correctly

[ArthurZucker]

Yep! The refactor is now merged, I'll have a look at #34826!

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.