Hany Hassan Awadalla

Shufang Xie, Huishuai Zhang, Junliang Guo et al. · microsoft-research

Transformer networks have become the preferred architecture for many tasks due to their state-of-the-art performance. However, the optimal way to implement residual connections in Transformer, which are essential for effective training, is still debated. Two widely used variants are the Post-Layer-Normalization (Post-LN) and Pre-Layer-Normalization (Pre-LN) Transformers, which apply layer normalization after each residual block's output or before each residual block's input, respectively. While both variants enjoy their advantages, they also suffer from severe limitations: Post-LN causes gradient vanishing issue that hinders training deep Transformers, and Pre-LN causes representation collapse issue that limits model capacity. In this paper, we propose ResiDual, a novel Transformer architecture with Pre-Post-LN (PPLN), which fuses the connections in Post-LN and Pre-LN together and inherits their advantages while avoids their limitations. We conduct both theoretical analyses and empirical experiments to verify the effectiveness of ResiDual. Theoretically, we prove that ResiDual has a lower bound on the gradient to avoid the vanishing issue due to the residual connection from Pre-LN. Moreover, ResiDual also has diverse model representations to avoid the collapse issue due to the residual connection from Post-LN. Empirically, ResiDual outperforms both Post-LN and Pre-LN on several machine translation benchmarks across different network depths and data sizes. Thanks to the good theoretical and empirical performance, ResiDual Transformer can serve as a foundation architecture for different AI models (e.g., large language models). Our code is available at https://github.com/microsoft/ResiDual.

Young Jin Kim, Rawn Henry, Raffy Fahim et al. · microsoft-research

Large Language Models (LLMs) have achieved state-of-the-art performance across various language tasks but pose challenges for practical deployment due to their substantial memory requirements. Furthermore, the latest generative models suffer from high inference costs caused by the memory bandwidth bottleneck in the auto-regressive decoding process. To address these issues, we propose an efficient weight-only quantization method that reduces memory consumption and accelerates inference for LLMs. To ensure minimal quality degradation, we introduce a simple and effective heuristic approach that utilizes only the model weights of a pre-trained model. This approach is applicable to both Mixture-of-Experts (MoE) and dense models without requiring additional fine-tuning. To demonstrate the effectiveness of our proposed method, we first analyze the challenges and issues associated with LLM quantization. Subsequently, we present our heuristic approach, which adaptively finds the granularity of quantization, effectively addressing these problems. Furthermore, we implement highly efficient GPU GEMMs that perform on-the-fly matrix multiplication and dequantization, supporting the multiplication of fp16 or bf16 activations with int8 or int4 weights. We evaluate our approach on large-scale open source models such as OPT-175B and internal MoE models, showcasing minimal accuracy loss while achieving up to 3.65 times higher throughput on the same number of GPUs.

Rui Liu, Young Jin Kim, Alexandre Muzio et al. · microsoft-research

Sparsely activated transformers, such as Mixture of Experts (MoE), have received great interest due to their outrageous scaling capability which enables dramatical increases in model size without significant increases in computational cost. To achieve this, MoE models replace the feedforward sub-layer with Mixture-of-Experts sub-layer in transformers and use a gating network to route each token to its assigned experts. Since the common practice for efficient training of such models requires distributing experts and tokens across different machines, this routing strategy often incurs huge cross-machine communication cost because tokens and their assigned experts likely reside in different machines. In this paper, we propose \emph{Gating Dropout}, which allows tokens to ignore the gating network and stay at their local machines, thus reducing the cross-machine communication. Similar to traditional dropout, we also show that Gating Dropout has a regularization effect during training, resulting in improved generalization performance. We validate the effectiveness of Gating Dropout on multilingual machine translation tasks. Our results demonstrate that Gating Dropout improves a state-of-the-art MoE model with faster wall-clock time convergence rates and better BLEU scores for a variety of model sizes and datasets.

Amr Hendy, Mohamed Abdelrehim, Amr Sharaf et al. · microsoft-research

Generative Pre-trained Transformer (GPT) models have shown remarkable capabilities for natural language generation, but their performance for machine translation has not been thoroughly investigated. In this paper, we present a comprehensive evaluation of GPT models for machine translation, covering various aspects such as quality of different GPT models in comparison with state-of-the-art research and commercial systems, effect of prompting strategies, robustness towards domain shifts and document-level translation. We experiment with eighteen different translation directions involving high and low resource languages, as well as non English-centric translations, and evaluate the performance of three GPT models: ChatGPT, GPT3.5 (text-davinci-003), and text-davinci-002. Our results show that GPT models achieve very competitive translation quality for high resource languages, while having limited capabilities for low resource languages. We also show that hybrid approaches, which combine GPT models with other translation systems, can further enhance the translation quality. We perform comprehensive analysis and human evaluation to further understand the characteristics of GPT translations. We hope that our paper provides valuable insights for researchers and practitioners in the field and helps to better understand the potential and limitations of GPT models for translation.

Pengcheng He, Baolin Peng, Liyang Lu et al. · microsoft-research

This paper presents Z-Code++, a new pre-trained language model optimized for abstractive text summarization. The model extends the state of the art encoder-decoder model using three techniques. First, we use a two-phase pre-training process to improve model's performance on low-resource summarization tasks. The model is first pre-trained using text corpora for language understanding, and then is continually pre-trained on summarization corpora for grounded text generation. Second, we replace self-attention layers in the encoder with disentangled attention layers, where each word is represented using two vectors that encode its content and position, respectively. Third, we use fusion-in-encoder, a simple yet effective method of encoding long sequences in a hierarchical manner. Z-Code++ creates new state of the art on 9 out of 13 text summarization tasks across 5 languages. Our model is parameter-efficient in that it outperforms the 600x larger PaLM-540B on XSum, and the finetuned 200x larger GPT3-175B on SAMSum. In zero-shot and few-shot settings, our model substantially outperforms the competing models.

Hai Pham, Young Jin Kim, Subhabrata Mukherjee et al. · cmu, microsoft-research

Mixture-of-experts (MoE) architecture has been proven a powerful method for diverse tasks in training deep models in many applications. However, current MoE implementations are task agnostic, treating all tokens from different tasks in the same manner. In this work, we instead design a novel method that incorporates task information into MoE models at different granular levels with shared dynamic task-based adapters. Our experiments and analysis show the advantages of our approaches over the dense and canonical MoE models on multi-task multilingual machine translations. With task-specific adapters, our models can additionally generalize to new tasks efficiently.

Haoran Xu, Young Jin Kim, Amr Sharaf et al. · microsoft-research

Generative Large Language Models (LLMs) have achieved remarkable advancements in various NLP tasks. However, these advances have not been reflected in the translation task, especially those with moderate model sizes (i.e., 7B or 13B parameters), which still lag behind conventional supervised encoder-decoder translation models. Previous studies have attempted to improve the translation capabilities of these moderate LLMs, but their gains have been limited. In this study, we propose a novel fine-tuning approach for LLMs that is specifically designed for the translation task, eliminating the need for the abundant parallel data that traditional translation models usually depend on. Our approach consists of two fine-tuning stages: initial fine-tuning on monolingual data followed by subsequent fine-tuning on a small set of high-quality parallel data. We introduce the LLM developed through this strategy as Advanced Language Model-based trAnslator (ALMA). Based on LLaMA-2 as our underlying model, our results show that the model can achieve an average improvement of more than 12 BLEU and 12 COMET over its zero-shot performance across 10 translation directions from the WMT'21 (2 directions) and WMT'22 (8 directions) test datasets. The performance is significantly better than all prior work and even superior to the NLLB-54B model and GPT-3.5-text-davinci-003, with only 7B or 13B parameters. This method establishes the foundation for a novel training paradigm in machine translation.

Young Jin Kim, Raffy Fahim, Hany Hassan Awadalla · microsoft-research

Large Mixture of Experts (MoE) models could achieve state-of-the-art quality on various language tasks, including machine translation task, thanks to the efficient model scaling capability with expert parallelism. However, it has brought a fundamental issue of larger memory consumption and increased memory bandwidth bottleneck at deployment time. In this paper, we propose Mixture of Quantized Experts (MoQE) which is a simple weight-only quantization method applying ultra low-bit down to 2-bit quantizations only to expert weights for mitigating the increased memory and latency issues of MoE models. We show that low-bit quantization together with the MoE architecture delivers a reliable model performance while reducing the memory size significantly even without any additional training in most cases. In particular, expert layers in MoE models are much more robust to the quantization than conventional feedforward networks (FFN) layers. In our comprehensive analysis, we show that MoE models with 2-bit expert weights can deliver better model performance than the dense model trained on the same dataset. As a result of low-bit quantization, we show the model size can be reduced by 79.6% of the original half precision floating point (fp16) MoE model. Combined with an optimized GPU runtime implementation, it also achieves 1.24X speed-up on A100 GPUs.

Young Jin Kim, Rawn Henry, Raffy Fahim et al. · microsoft-research

Mixture of Experts (MoE) models with conditional execution of sparsely activated layers have enabled training models with a much larger number of parameters. As a result, these models have achieved significantly better quality on various natural language processing tasks including machine translation. However, it remains challenging to deploy such models in real-life scenarios due to the large memory requirements and inefficient inference. In this work, we introduce a highly efficient inference framework with several optimization approaches to accelerate the computation of sparse models and cut down the memory consumption significantly. While we achieve up to 26x speed-up in terms of throughput, we also reduce the model size almost to one eighth of the original 32-bit float model by quantizing expert weights into 4-bit integers. As a result, we are able to deploy 136x larger models with 27% less cost and significantly better quality compared to the existing solutions. This enables a paradigm shift in deploying large scale multilingual MoE transformers models replacing the traditional practice of distilling teacher models into dozens of smaller models per language or task.

Akiko Eriguchi, Shufang Xie, Tao Qin et al.

Multilingual Neural Machine Translation (MNMT) enables one system to translate sentences from multiple source languages to multiple target languages, greatly reducing deployment costs compared with conventional bilingual systems. The MNMT training benefit, however, is often limited to many-to-one directions. The model suffers from poor performance in one-to-many and many-to-many with zero-shot setup. To address this issue, this paper discusses how to practically build MNMT systems that serve arbitrary X-Y translation directions while leveraging multilinguality with a two-stage training strategy of pretraining and finetuning. Experimenting with the WMT'21 multilingual translation task, we demonstrate that our systems outperform the conventional baselines of direct bilingual models and pivot translation models for most directions, averagely giving +6.0 and +4.1 BLEU, without the need for architecture change or extra data collection. Moreover, we also examine our proposed approach in an extremely large-scale data setting to accommodate practical deployment scenarios.

Vikas Raunak, Hany Hassan Awadalla, Arul Menezes

Most of the recent work in leveraging Large Language Models (LLMs) such as GPT-3 for Machine Translation (MT) has focused on selecting the few-shot samples for prompting. In this work, we try to better understand the role of demonstration attributes for the in-context learning of translations through perturbations of high-quality, in-domain demonstrations. We find that asymmetric perturbation of the source-target mappings yield vastly different results. We show that the perturbation of the source side has surprisingly little impact, while target perturbation can drastically reduce translation quality, suggesting that it is the output text distribution that provides the most important learning signal during in-context learning of translations. We propose a method named Zero-Shot-Context to add this signal automatically in Zero-Shot prompting. We demonstrate that it improves upon the zero-shot translation performance of GPT-3, even making it competitive with few-shot prompted translations.

Muhammad ElNokrashy, Amr Hendy, Mohamed Maher et al.

This paper proposes a simple yet effective method to improve direct (X-to-Y) translation for both cases: zero-shot and when direct data is available. We modify the input tokens at both the encoder and decoder to include signals for the source and target languages. We show a performance gain when training from scratch, or finetuning a pretrained model with the proposed setup. In the experiments, our method shows nearly 10.0 BLEU points gain on in-house datasets depending on the checkpoint selection criteria. In a WMT evaluation campaign, From-English performance improves by 4.17 and 2.87 BLEU points, in the zero-shot setting, and when direct data is available for training, respectively. While X-to-Y improves by 1.29 BLEU over the zero-shot baseline, and 0.44 over the many-to-many baseline. In the low-resource setting, we see a 1.5~1.7 point improvement when finetuning on X-to-Y domain data.

Yifei He, Yang Liu, Chen Liang et al.

Mixture-of-Experts (MoE) models have become a key approach for scaling large language models efficiently by activating only a subset of experts during training and inference. Typically, the number of activated experts presents a trade-off: fewer experts reduce computational costs, while more experts improve performance. Recent studies reveal that not all activated experts contribute equally to model performance, with some providing minimal utility, particularly when finetuning pretrained MoE models for specialized downstream tasks. The co-existence of significant and redundant parameters in experts provides us an opportunity to reduce the number of activated experts while maintaining model performance. In this work, we propose the concept of compressed experts, lightweight modules that serve as compact representations of full experts. Our approach preserves the most important experts while replacing other auxiliary activated experts with compressed experts. The reduction of active parameters significantly lowers inference costs while achieving comparable performance. Extensive experiments on models including Phi-MoE and OLMoE demonstrate that compressed experts recover over 90% of full expert performance across various tasks while reducing more than 30% active parameters and saving 20% in inference costs. This approach enables efficient deployment of MoE models in resource-constrained settings and facilitates scaling to larger models with manageable overhead. Our code is available at https://github.com/yifei-he/Compressed-Experts.

Young Jin Kim, Ammar Ahmad Awan, Alexandre Muzio et al.

The Mixture of Experts (MoE) models are an emerging class of sparsely activated deep learning models that have sublinear compute costs with respect to their parameters. In contrast with dense models, the sparse architecture of MoE offers opportunities for drastically growing model size with significant accuracy gain while consuming much lower compute budget. However, supporting large scale MoE training also has its own set of system and modeling challenges. To overcome the challenges and embrace the opportunities of MoE, we first develop a system capable of scaling MoE models efficiently to trillions of parameters. It combines multi-dimensional parallelism and heterogeneous memory technologies harmoniously with MoE to empower 8x larger models on the same hardware compared with existing work. Besides boosting system efficiency, we also present new training methods to improve MoE sample efficiency and leverage expert pruning strategy to improve inference time efficiency. By combining the efficient system and training methods, we are able to significantly scale up large multitask multilingual models for language generation which results in a great improvement in model accuracy. A model trained with 10 billion parameters on 50 languages can achieve state-of-the-art performance in Machine Translation (MT) and multilingual natural language generation tasks. The system support of efficient MoE training has been implemented and open-sourced with the DeepSpeed library.

Zeyu Qin, Qingxiu Dong, Xingxing Zhang et al.

Large language models (LLMs) achieve strong performance across diverse tasks, largely driven by high-quality web data used in pre-training. However, recent studies indicate this data source is rapidly depleting. Synthetic data emerges as a promising alternative, but it remains unclear whether synthetic datasets exhibit predictable scalability comparable to raw pre-training data. In this work, we systematically investigate the scaling laws of synthetic data by introducing SynthLLM, a scalable framework that transforms pre-training corpora into diverse, high-quality synthetic datasets. Our approach achieves this by automatically extracting and recombining high-level concepts across multiple documents using a graph algorithm. Key findings from our extensive mathematical experiments on SynthLLM include: (1) SynthLLM generates synthetic data that reliably adheres to the rectified scaling law across various model sizes; (2) Performance improvements plateau near 300B tokens; and (3) Larger models approach optimal performance with fewer training tokens. For instance, an 8B model peaks at 1T tokens, while a 3B model requires 4T. Moreover, comparisons with existing synthetic data generation and augmentation methods demonstrate that SynthLLM achieves superior performance and scalability. Our findings highlight synthetic data as a scalable and reliable alternative to organic pre-training corpora, offering a viable path toward continued improvement in model performance.

Vikas Raunak, Arul Menezes, Matt Post et al.

Large Language Models (LLMs) such as GPT-3 have emerged as general-purpose language models capable of addressing many natural language generation or understanding tasks. On the task of Machine Translation (MT), multiple works have investigated few-shot prompting mechanisms to elicit better translations from LLMs. However, there has been relatively little investigation on how such translations differ qualitatively from the translations generated by standard Neural Machine Translation (NMT) models. In this work, we investigate these differences in terms of the literalness of translations produced by the two systems. Using literalness measures involving word alignment and monotonicity, we find that translations out of English (E-X) from GPTs tend to be less literal, while exhibiting similar or better scores on MT quality metrics. We demonstrate that this finding is borne out in human evaluations as well. We then show that these differences are especially pronounced when translating sentences that contain idiomatic expressions.

Amr Hendy, Esraa A. Gad, Mohamed Abdelghaffar et al.

This paper describes our submission to the constrained track of WMT21 shared news translation task. We focus on the three relatively low resource language pairs Bengali to and from Hindi, English to and from Hausa, and Xhosa to and from Zulu. To overcome the limitation of relatively low parallel data we train a multilingual model using a multitask objective employing both parallel and monolingual data. In addition, we augment the data using back translation. We also train a bilingual model incorporating back translation and knowledge distillation then combine the two models using sequence-to-sequence mapping. We see around 70% relative gain in BLEU point for English to and from Hausa, and around 25% relative improvements for both Bengali to and from Hindi, and Xhosa to and from Zulu compared to bilingual baselines.

Jian Yang, Shuming Ma, Haoyang Huang et al.

This report describes Microsoft's machine translation systems for the WMT21 shared task on large-scale multilingual machine translation. We participated in all three evaluation tracks including Large Track and two Small Tracks where the former one is unconstrained and the latter two are fully constrained. Our model submissions to the shared task were initialized with DeltaLM\footnote{\url{https://aka.ms/deltalm}}, a generic pre-trained multilingual encoder-decoder model, and fine-tuned correspondingly with the vast collected parallel data and allowed data sources according to track settings, together with applying progressive learning and iterative back-translation approaches to further improve the performance. Our final submissions ranked first on three tracks in terms of the automatic evaluation metric.

Shuming Ma, Li Dong, Shaohan Huang et al.

While pretrained encoders have achieved success in various natural language understanding (NLU) tasks, there is a gap between these pretrained encoders and natural language generation (NLG). NLG tasks are often based on the encoder-decoder framework, where the pretrained encoders can only benefit part of it. To reduce this gap, we introduce DeltaLM, a pretrained multilingual encoder-decoder model that regards the decoder as the task layer of off-the-shelf pretrained encoders. Specifically, we augment the pretrained multilingual encoder with a decoder and pre-train it in a self-supervised way. To take advantage of both the large-scale monolingual data and bilingual data, we adopt the span corruption and translation span corruption as the pre-training tasks. Experiments show that DeltaLM outperforms various strong baselines on both natural language generation and translation tasks, including machine translation, abstractive text summarization, data-to-text, and question generation. The code and pretrained models are available at \url{https://aka.ms/deltalm}.

Shuming Ma, Jian Yang, Haoyang Huang et al.

Multilingual machine translation enables a single model to translate between different languages. Most existing multilingual machine translation systems adopt a randomly initialized Transformer backbone. In this work, inspired by the recent success of language model pre-training, we present XLM-T, which initializes the model with an off-the-shelf pretrained cross-lingual Transformer encoder and fine-tunes it with multilingual parallel data. This simple method achieves significant improvements on a WMT dataset with 10 language pairs and the OPUS-100 corpus with 94 pairs. Surprisingly, the method is also effective even upon the strong baseline with back-translation. Moreover, extensive analysis of XLM-T on unsupervised syntactic parsing, word alignment, and multilingual classification explains its effectiveness for machine translation. The code will be at https://aka.ms/xlm-t.

Muhammad N. ElNokrashy, Amr Hendy, Mohamed Abdelghaffar et al.

This paper describes our submission to the WMT20 sentence filtering task. We combine scores from (1) a custom LASER built for each source language, (2) a classifier built to distinguish positive and negative pairs by semantic alignment, and (3) the original scores included in the task devkit. For the mBART finetuning setup, provided by the organizers, our method shows 7% and 5% relative improvement over baseline, in sacreBLEU score on the test set for Pashto and Khmer respectively.

Young Jin Kim, Hany Hassan Awadalla

Transformer-based models are the state-of-the-art for Natural Language Understanding (NLU) applications. Models are getting bigger and better on various tasks. However, Transformer models remain computationally challenging since they are not efficient at inference-time compared to traditional approaches. In this paper, we present FastFormers, a set of recipes to achieve efficient inference-time performance for Transformer-based models on various NLU tasks. We show how carefully utilizing knowledge distillation, structured pruning and numerical optimization can lead to drastic improvements on inference efficiency. We provide effective recipes that can guide practitioners to choose the best settings for various NLU tasks and pretrained models. Applying the proposed recipes to the SuperGLUE benchmark, we achieve from 9.8x up to 233.9x speed-up compared to out-of-the-box models on CPU. On GPU, we also achieve up to 12.4x speed-up with the presented methods. We show that FastFormers can drastically reduce cost of serving 100 million requests from 4,223 USD to just 18 USD on an Azure F16s_v2 instance. This translates to a sustainable runtime by reducing energy consumption 6.9x - 125.8x according to the metrics used in the SustaiNLP 2020 shared task.

Yiren Wang, ChengXiang Zhai, Hany Hassan Awadalla

While monolingual data has been shown to be useful in improving bilingual neural machine translation (NMT), effectively and efficiently leveraging monolingual data for Multilingual NMT (MNMT) systems is a less explored area. In this work, we propose a multi-task learning (MTL) framework that jointly trains the model with the translation task on bitext data and two denoising tasks on the monolingual data. We conduct extensive empirical studies on MNMT systems with 10 language pairs from WMT datasets. We show that the proposed approach can effectively improve the translation quality for both high-resource and low-resource languages with large margin, achieving significantly better results than the individual bilingual models. We also demonstrate the efficacy of the proposed approach in the zero-shot setup for language pairs without bitext training data. Furthermore, we show the effectiveness of MTL over pre-training approaches for both NMT and cross-lingual transfer learning NLU tasks; the proposed approach outperforms massive scale models trained on single task.

Junyoung Chung, Jacob Devlin, Hany Hassan Awadalla

In this paper, we explore different neural network architectures that can predict if a speaker of a given utterance is asking a question or making a statement. We com- pare the outcomes of regularization methods that are popularly used to train deep neural networks and study how different context functions can affect the classification performance. We also compare the efficacy of gated activation functions that are favorably used in recurrent neural networks and study how to combine multimodal inputs. We evaluate our models on two multimodal datasets: MSR-Skype and CALLHOME.