Speculative decoding has emerged as a powerful method to improve latency and throughput in hosting large language models. However, most existing implementations focus on generating a single sequence. Real-world generative AI applications often require multiple responses and how to perform speculative decoding in a batched setting while preserving its latency benefits poses non-trivial challenges. This paper describes a system of batched speculative decoding that sets a new state of the art in multi-sequence generation latency and that demonstrates superior GPU utilization as well as quality of generations within a time budget. For example, for a 7.8B-size model on a single A100 GPU and with a batch size of 8, each sequence is generated at an average speed of 5.8ms per token, the overall throughput being 1.1K tokens per second. These results represent state-of-the-art latency and a 2.15× speed-up over optimized regular decoding. Within a time budget that regular decoding does not finish, our system is able to generate sequences with HumanEval Pass@First of 43% and Pass@All of 61%, far exceeding what’s feasible with single-sequence speculative decoding. Our peak GPU utilization during decoding reaches as high as 15.8%, more than 3× the highest of that of regular decoding and around 10× of single-sequence speculative decoding.

Generative models, widely utilized in various applications, can often struggle with prompts corresponding to partial tokens. This struggle stems from tokenization, where partial tokens fall out of distribution during inference, leading to incorrect or nonsensical outputs. This paper examines a technique to alleviate the tokenization artifact on text completion in generative models, maintaining performance even in regular non-subword cases. The method, termed token alignment, involves backtracking to the last complete tokens and ensuring the model’s generation aligns with the prompt. This approach showcases marked improvement across many partial token scenarios, including nuanced cases like space-prefix and partial indentation, with only a minor time increase. The technique and analysis detailed in this paper contribute to the continuous advancement of generative models in handling partial inputs, bearing relevance for applications like code completion and text.

In this paper, we describe a cross-lingual information retrieval (CLIR) system that, given a query in English, and a set of audio and text documents in a foreign language, can return a scored list of relevant documents, and present findings in a summary form in English. Foreign audio documents are first transcribed by a state-of-the-art pretrained multilingual speech recognition model that is finetuned to the target language. For text documents, we use multiple multilingual neural machine translation (MT) models to achieve good translation results, especially for low/medium resource languages. The processed documents and queries are then scored using a probabilistic CLIR model that makes use of the probability of translation from GIZA translation tables and scores from a Neural Network Lexical Translation Model (NNLTM). Additionally, advanced score normalization, combination, and thresholding schemes are employed to maximize the Average Query Weighted Value (AQWV) scores. The CLIR output, together with multiple translation renderings, are selected and translated into English snippets via a summarization model. Our turnkey system is language agnostic and can be quickly trained for a new low-resource language in few days.