Papers of the day   All papers

Q8BERT: Quantized 8Bit BERT

Comments

Nov 03 2019 (((ل()(ل() 'yoav))))

This effective 8bit BERT from Intel shows that we can remove a lot out of large transformer LMs, while remaining accurate. Eagerly waiting the efficient inference code+hardware to go along with it! #greenAI https://arxiv.org/abs/1910.06188
8 replies, 486 likes


Nov 03 2019 Peter Izsak

🚀8bit BERT is part of our open source library NLP Architect https://github.com/NervanaSystems/nlp-architect
1 replies, 160 likes


Nov 04 2019 hardmaru 😷

BERT coming to small compute devices? “We show how quantization-aware training during the fine-tuning phase of BERT can compress it down by 4×. The quantized model can accelerate inference speed if it is optimized for 8bit Integer supporting hardware.” https://arxiv.org/abs/1910.06188
1 replies, 104 likes


Nov 03 2019 (((ل()(ل() 'yoav))))

in the picture: 8bit, but just as energetic. https://t.co/WUlSfrrPVe
0 replies, 56 likes


Nov 04 2019 Rasa

Large pre-trained models are becoming the standard, but there's a high cost in memory and compute resources. Read @IntelAI 's paper on compressing BERT: https://arxiv.org/abs/1910.06188 And then check out related research from #Rasa's Sam Sucik: https://blog.rasa.com/compressing-bert-for-faster-prediction-2/
0 replies, 17 likes


Oct 18 2019 arXiv CS-CL

Q8BERT: Quantized 8Bit BERT http://arxiv.org/abs/1910.06188
0 replies, 7 likes


Nov 04 2019 Hanlin Tang

The code is open-sourced as part of our NLP Architect library: https://github.com/NervanaSystems/nlp-architect. Work by @MosheWasserblat, @peter_izsak and others. #IntelAI
0 replies, 6 likes


Oct 20 2019 arXiv CS-CL

Q8BERT: Quantized 8Bit BERT http://arxiv.org/abs/1910.06188
0 replies, 3 likes


Oct 15 2019 arXiv CS-CL

Q8BERT: Quantized 8Bit BERT http://arxiv.org/abs/1910.06188
0 replies, 1 likes


Oct 15 2019 arXiv CS-CL

Q8BERT: Quantized 8Bit BERT http://arxiv.org/abs/1910.06188
0 replies, 1 likes


Oct 19 2019 arXiv CS-CL

Q8BERT: Quantized 8Bit BERT http://arxiv.org/abs/1910.06188
0 replies, 1 likes


Content