Reiner Pope, Sholto Douglas, Aakanksha Chowdhery
Jacob Devlin, James Bradbury, Anselm Levskaya, Jonathan Heek, Kefan Xiao, Shivani Agrawal, Jeff Dean
2022 · arXiv
We study the problem of efficient generative inference for Transformer models, in one of its most challenging settings: large deep models, with tight latency targets and long sequence lengths. Better understanding of the engineering tradeoffs for inference for large Transformer-based models is important as use cases of these models are growing rapidly throughout application areas. We develop a simple analytical model for inference efficiency to select the best multi-dimensional partitioning techniques optimized for TPU v4 slices based on the application requirements. We combine these with a suite of low-level optimizations to achieve a new Pareto frontier on the latency and model FLOPS utilization (MFU) tradeoffs on 500B+ parameter models that outperforms the FasterTransformer suite of benchmarks. We further show that with appropriate partitioning, the lower memory requirements of multiquery attention (i.e. multiple query heads share single key/value head) enables scaling up to 32x larger context lengths. Finally, we achieve a low-batch-size latency of 29ms per token during generation (using int8 weight quantization) and a 76% MFU during large-batch-size processing of input tokens, while supporting a long 2048-token context length on the PaLM 540B parameter model.
Metadata: arXiv. Abstracts retain their original authors’ rights.
Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels
3 selected excerpts · 3 total mentions
> GPUs Groq will still come out on top because you can never reduce latency by adding more parallel compute :) You literally can, in fact that’s the entire reason to use multiple chips. See eg the TPU group’s paper: https://arxiv.org/abs/2211.05102
One possible explanation is that they hit the teraflops number during the prefill stage, where you can process all tokens at once, and are generally more operationally intensive, so you can use more compute. Utilization usually drops during the token generation stage. The utilization of the TPU during token generation is 3% when batch size 16. (https://arxiv.org/pdf/2211.05102.pdf, Table on the last page).
> Though inference for the 8B model is almost definitely not capable of near real time inference yet Google previously showed you could get the fullsized 540b-parameter PaLM-1 model down to "a low-batch-size latency of 29ms per token during generation (with int8 weight quantization)" https://arxiv.org/abs/2211.05102#google . How many tokens per 1000ms do humans speak? I'm guessing fewer than 34. The real question is who wants to pay for it.
More from the same topics.
Tim Kraska, Alex Beutel, Ed H. Chi +2
Indexes are models: a B-Tree-Index can be seen as a model to map a key to the position of a record within a sorted array, a Hash-Index as a model to map a key to a position of a record within an unsorted array, and a BitMap-Index as a model to indicate if a data record exists or not.
Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan +11
Artificial intelligence (AI) researchers have been developing and refining large language models (LLMs) that exhibit remarkable capabilities across a variety of domains and tasks, challenging our understanding of learning and cognition.
Martin Kleppmann
The CAP Theorem is a frequently cited impossibility result in distributed systems, especially among NoSQL distributed databases.
Tom B. Brown, Benjamin Mann, Nick Ryder +28
Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task.