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

Add sdpa for DistilBert (#33724)

Browse Source 
* Add sdpa for DistilBert

* [run_slow] distilbert

* [run_slow] distilbert

* [run_slow] distilbert

* Try without slow tests

* [run_slow] distilbert

* [run_slow] distilbert
This commit is contained in:
Omar Salman
2024-10-02 17:55:19 +05:00
committed by GitHub
parent 614c79a9b0
commit e7c8af7f33
3 changed files with 144 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

47
docs/source/en/model_doc/distilbert.md
View File

@@ -66,6 +66,53 @@ contributed by [kamalkraj](https://huggingface.co/kamalkraj). The original code
* predicting the masked tokens correctly (but no next-sentence objective)
* a cosine similarity between the hidden states of the student and the teacher model
### Using Scaled Dot Product Attention (SDPA)
PyTorch includes a native scaled dot-product attention (SDPA) operator as part of `torch.nn.functional`. This function
encompasses several implementations that can be applied depending on the inputs and the hardware in use. See the
[official documentation](https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html)
or the [GPU Inference](https://huggingface.co/docs/transformers/main/en/perf_infer_gpu_one#pytorch-scaled-dot-product-attention)
page for more information.
SDPA is used by default for `torch>=2.1.1` when an implementation is available, but you may also set
`attn_implementation="sdpa"` in `from_pretrained()` to explicitly request SDPA to be used.
```
from transformers import DistilBertModel
model = DistilBertModel.from_pretrained("distilbert-base-uncased", torch_dtype=torch.float16, attn_implementation="sdpa")
```
For the best speedups, we recommend loading the model in half-precision (e.g. `torch.float16` or `torch.bfloat16`).
On a local benchmark (NVIDIA GeForce RTX 2060-8GB, PyTorch 2.3.1, OS Ubuntu 20.04) with `float16` and the `distilbert-base-uncased` model with
a MaskedLM head, we saw the following speedups during training and inference.
#### Training
| num_training_steps | batch_size | seq_len | is cuda | Time per batch (eager - s) | Time per batch (sdpa - s) | Speedup (%) | Eager peak mem (MB) | sdpa peak mem (MB) | Mem saving (%) |
|--------------------|------------|---------|---------|----------------------------|---------------------------|-------------|---------------------|--------------------|----------------|
| 100 | 1 | 128 | False | 0.010 | 0.008 | 28.870 | 397.038 | 399.629 | -0.649 |
| 100 | 1 | 256 | False | 0.011 | 0.009 | 20.681 | 412.505 | 412.606 | -0.025 |
| 100 | 2 | 128 | False | 0.011 | 0.009 | 23.741 | 412.213 | 412.606 | -0.095 |
| 100 | 2 | 256 | False | 0.015 | 0.013 | 16.502 | 427.491 | 425.787 | 0.400 |
| 100 | 4 | 128 | False | 0.015 | 0.013 | 13.828 | 427.491 | 425.787 | 0.400 |
| 100 | 4 | 256 | False | 0.025 | 0.022 | 12.882 | 594.156 | 502.745 | 18.182 |
| 100 | 8 | 128 | False | 0.023 | 0.022 | 8.010 | 545.922 | 502.745 | 8.588 |
| 100 | 8 | 256 | False | 0.046 | 0.041 | 12.763 | 983.450 | 798.480 | 23.165 |
#### Inference
| num_batches | batch_size | seq_len | is cuda | is half | use mask | Per token latency eager (ms) | Per token latency SDPA (ms) | Speedup (%) | Mem eager (MB) | Mem BT (MB) | Mem saved (%) |
|-------------|------------|---------|---------|---------|----------|-----------------------------|-----------------------------|-------------|----------------|--------------|---------------|
| 50 | 2 | 64 | True | True | True | 0.032 | 0.025 | 28.192 | 154.532 | 155.531 | -0.642 |
| 50 | 2 | 128 | True | True | True | 0.033 | 0.025 | 32.636 | 157.286 | 157.482 | -0.125 |
| 50 | 4 | 64 | True | True | True | 0.032 | 0.026 | 24.783 | 157.023 | 157.449 | -0.271 |
| 50 | 4 | 128 | True | True | True | 0.034 | 0.028 | 19.299 | 162.794 | 162.269 | 0.323 |
| 50 | 8 | 64 | True | True | True | 0.035 | 0.028 | 25.105 | 160.958 | 162.204 | -0.768 |
| 50 | 8 | 128 | True | True | True | 0.052 | 0.046 | 12.375 | 173.155 | 171.844 | 0.763 |
| 50 | 16 | 64 | True | True | True | 0.051 | 0.045 | 12.882 | 172.106 | 171.713 | 0.229 |
| 50 | 16 | 128 | True | True | True | 0.096 | 0.081 | 18.524 | 191.257 | 191.517 | -0.136 |
## Resources

1
docs/source/en/perf_infer_gpu_one.md
View File

@@ -219,6 +219,7 @@ For now, Transformers supports SDPA inference and training for the following arc
* [Dbrx](https://huggingface.co/docs/transformers/model_doc/dbrx#transformers.DbrxModel)
* [DeiT](https://huggingface.co/docs/transformers/model_doc/deit#transformers.DeiTModel)
* [Dinov2](https://huggingface.co/docs/transformers/en/model_doc/dinov2)
* [DistilBert](https://huggingface.co/docs/transformers/model_doc/distilbert#transformers.DistilBertModel)
* [Dpr](https://huggingface.co/docs/transformers/model_doc/dpr#transformers.DprReader)
* [Falcon](https://huggingface.co/docs/transformers/model_doc/falcon#transformers.FalconModel)
* [Gemma](https://huggingface.co/docs/transformers/model_doc/gemma#transformers.GemmaModel)