deepsparse VS transformer-deploy

Interesting company. Yannic Kilcher interviewed Nir Shavit last year and they went into some depth: https://www.youtube.com/watch?v=0PAiQ1jTN5k DeepSparse is on GitHub: https://github.com/neuralmagic/deepsparse

sparsity https://github.com/neuralmagic/deepsparse

For 1), what is the easiest way to speed up inference (assume only PyTorch and primarily GPU but also some CPU)? I have been using ONNX and Torchscript but there is a bit of a learning curve and sometimes it can be tricky to get the model to actually work. Is there anything else worth trying? I am enthused by things like TorchDynamo (although I have not tested it extensively) due to its apparent ease of use. I also saw the post yesterday about Kernl using (OpenAI) Triton kernels to speed up transformer models which also looks interesting. Are things like SageMaker Neo or NeuralMagic worth trying? My only reservation with some of these is they still seem to be pretty model/architecture specific. I am a little reluctant to put much time into these unless I know others have had some success first.

You should look into deepsparse, they are working on delivering GPU level performance on consumer CPUs with some great results: https://github.com/neuralmagic/deepsparse. There is a great interview with the founder, Nir Shavit here: https://piped.kavin.rocks/watch?v=0PAiQ1jTN5k

Utilizing the oBERT research we published at Neural Magic and some further iteration, we’ve enabled an increase in NLP performance of 175X while retaining 99% accuracy on the question-answering task in MLPerf. A combination of distillation, layer dropping, quantization, and unstructured pruning with oBERT enabled these large performance gains through the DeepSparse Engine. All of our contributions and research are open-sourced or free to use. Read through the oBERT paper on arxiv, try out the research in SparseML, and dive into the writeup to learn more about how we achieved these impressive results and utilize them for your own use cases!

Open-source projects leveraged by nebullvm include OpenVINO, TensorRT, Intel Neural Compressor, SparseML and DeepSparse, Apache TVM, ONNX Runtime, TFlite and XLA. A huge thank you to the open-source community for developing and maintaining these amazing projects.

BERT-Large (345 million parameters) is now faster than the much smaller DistilBERT (66 million parameters) all while retaining the accuracy of the much larger BERT-Large model! We made this possible with Intel Labs by applying cutting-edge sparsification and quantization research from their Prune Once For All paper and utilizing it in the DeepSparse engine. It makes BERT-Large 12x smaller while delivering 8x latency speedup on commodity CPUs. We open-sourced the research in SparseML; run through the overview here and give it a try!

And benchmark/deploy with 8X better performance in DeepSparse!

Hi _Arsenie, this runs the deepsparse.server command for multiple models. and btw, we recently updated the READMEs for the Deepsparse Engine https://github.com/neuralmagic/deepsparse

For 2), I am aware of a few options. Triton inference server is an obvious one as is the ‘transformer-deploy’ version from LDS. My only reservation here is that they require the model compilation or are architecture specific. I am aware of others like Bento, Ray serving and TorchServe. Ideally I would have something that allows any (PyTorch model) to be used without the extra compilation effort (or at least optionally) and has some convenience things like ease of use, easy to deploy, easy to host multiple models and can perform some dynamic batching. Anyway, I am really interested to hear people's experience here as I know there are now quite a few options! Any help is appreciated! Disclaimer - I have no affiliation or are connected in any way with the libraries or companies listed here. These are just the ones I know of. Thanks in advance.

We work for Lefebvre Sarrut, a leading European legal publisher. Several of our products include transformer models in latency sensitive scenarios (search, content recommendation). So far, ONNX Runtime and TensorRT served us well, and we learned interesting patterns along the way that we shared with the community through an open-source library called transformer-deploy. However, recent changes in our environment made our needs evolve:

here you will find a notebook for T5 on GPU with some tricks to make it fast: https://github.com/ELS-RD/transformer-deploy/blob/main/demo/generative-model/t5.ipynb

Check the notebook https://github.com/ELS-RD/transformer-deploy/blob/main/demo/TorchDynamo/benchmark.ipynb for detailed results, but what we will keep in mind:

notebook: https://github.com/ELS-RD/transformer-deploy/blob/main/demo/generative-model/t5.ipynb (Onnx Runtime only)

Regarding CPU inference, quantization is very easy, and supported by Transformer-deploy , however performance on transformer are very low outside corner cases (like no batch, very short sequence and distilled model), and last Intel generation CPU based instance like C6 or M6 on AWS are quite expensive compared to a cheap GPU like Nvidia T4, to say it otherwise, on transformer, until you are ok with slow inference and takes a small instance (for a PoC for instance), CPU inference is probably not a good idea.

The end to end tutorial: https://github.com/ELS-RD/transformer-deploy/blob/main/demo/quantization_end_to_end.ipynb

Want to try it 👉 https://github.com/ELS-RD/transformer-deploy