Language models (LMs) are pretrained to imitate text from large and diverse datasets that contain content that would violate human preferences if generated by an LM: falsehoods, offensive comments, personally identifiable information, low-quality or buggy code, among others. Here, we explore alternative objectives for pretraining LMs in a way that also guides them to generate text aligned with human preferences. We benchmark five objectives for pretraining with human feedback across three tasks and study how they affect the alignment and capabilities of pretrained LMs. We find a Pareto-optimal and simple approach among those we explored: conditional training, or learning distribution over tokens conditional on their human preference scores. Conditional training reduces the rate of undesirable content by up to an order of magnitude, both when generating without a prompt and with an adversarially-chosen prompt. Moreover, conditional training maintains the downstream task performance of standard LM pretraining, both before and after task-specific finetuning. Pretraining with human feedback results in much better preference satisfaction than standard LM pretraining followed by finetuning with feedback, i.e., learning and then unlearning undesirable behavior. Our results suggest that we should move beyond imitation learning when pretraining LMs and incorporate human preferences from the start of training.

Large pre-trained transformers have been receiving explosive attention in the past few years, due to their acculturation for numerous downstream applications via fine-tuning, but their exponentially increasing parameter counts are becoming a primary hurdle to even just fine-tune them without industry-standard hardware. Recently, Lottery Ticket Hypothesis (LTH) and its variants, have been exploited to prune these large pre-trained models generating subnetworks which can achieve similar performance as their dense counterparts, but LTH pragmatism is enormously inhibited by repetitive full training and pruning routine of iterative magnitude pruning (IMP) which worsens with increasing model size. Motivated by the recent observations of model soups, which suggest that fine-tuned weights of multiple models can be merged to a better minima, we propose **Instant Soup Pruning (ISP)** to generate lottery ticket quality subnetworks, using a fraction of the original IMP cost by replacing the expensive intermediate pruning stages of IMP with computationally efficient weak mask generation and aggregation routine. More specifically, during the mask generation stage, ISP takes a small handful of iterations using varying training protocols and data subsets to generate many weak and noisy subnetworks, and superpose them to average out the noise creating a high-quality denoised subnetwork. Our extensive experiments and ablation on two popular large-scale pre-trained models: $\texttt{CLIP} (unexplored in pruning till date)$ and $\texttt{BERT}$ across multiple benchmark vision $\texttt{\{MNIST, SVHN, Cars, GTSRB, CIFAR-10, CIFAR-100\}}$ and language datasets $\texttt{\{MNLI, QNLI, QQP, SST, ...\}}$ validate the effectiveness of ISP compared to several state-of-the-art pruning methods. Additionally, we show that ISP can be easily modified with minimal overhead to produce benefits comparable to model soups, without the prerequisite to generate multiple candidates fine-tuned models. Codes are available at: https://github.com/VITA-Group/instant_soup.

The surprising ability of Large Language Models (LLMs) to perform well on complex reasoning with only few-shot chain-of-thought prompts is believed to emerge only in very large-scale models. We show that such abilities can, in fact, be distilled down from GPT-3.5 (≥ 175B) to T5 variants (≤ 11B). We propose model specialization, to specialize the model’s ability towards a target task. The hypothesis is that large models (commonly viewed as larger than 100B) have strong modeling power such that they can perform a large spectrum of tasks. Small models (commonly viewed as smaller than 10B) have limited model capacity, but if we specialize their capacity towards a target task, the model can achieve decent performance improvements. We use multi-step math reasoning as our testbed because it is a very typical emergent ability. We show two important aspects of model abilities: (1) balancing language model’s performance on multiple tasks is a delicate matter, as improvements on one task may compromise other tasks; (2) yet by intentionally paying the price of decreased generic ability, we can clearly improve across different model scales smaller than 10B towards a specialized multi-step math reasoning ability. We further give comprehensive discussions about important design choices for better generalization, including the data format mixture and the start model checkpoint. We hope our practice and discoveries can serve as an important attempt towards specialized smaller models in the new research paradigm set by LLMs.

We show for the first time that large-scale generative pretrained transformer (GPT) family models can be pruned to at least 50% sparsity in one-shot, without any retraining, at minimal loss of accuracy. This is achieved via a new pruning method called SparseGPT, specifically designed to work efficiently and accurately on massive GPT-family models. We can execute SparseGPT on the largest available open-source models, OPT-175B and BLOOM-176B, in under 4.5 hours, and can reach 60% unstructured sparsity with negligible increase in perplexity: remarkably, more than 100 billion weights from these models can be ignored at inference time. SparseGPT generalizes to semi-structured (2:4 and 4:8) patterns, and is compatible with weight quantization approaches. The code is available at: https://github.com/IST-DASLab/sparsegpt.

Decoding methods for large language models often trade-off between diversity of outputs and parallelism of computation. Methods such as beam search and Gumbel top-k sampling can guarantee a different output for each element of the beam, but are not easy to parallelize. Alternatively, methods such as temperature sampling and its modifications (top-k sampling, nucleus sampling, typical decoding, and others), are embarrassingly parallel, but have no guarantees about duplicate samples. We present a framework for sampling according to an arithmetic code book implicitly defined by a large language model, compatible with common sampling variations, with provable beam diversity under certain conditions, as well as being embarrassingly parallel and providing unbiased and consistent expectations from the original model. We demonstrate the effectiveness of our approach on WMT machine translation, more than halving the standard deviation when estimating expected BLEU score reward, and closing the BLEU score gap between independent sampling and beam search by up to 63%.

The high computational and memory requirements of large language model (LLM) inference make it feasible only with multiple high-end accelerators. Motivated by the emerging demand for latency-insensitive tasks with batched processing, this paper initiates the study of high-throughput LLM inference using limited resources, such as a single commodity GPU. We present FlexGen, a high-throughput generation engine for running LLMs with limited GPU memory. FlexGen can be flexibly configured under various hardware resource constraints by aggregating memory and computation from the GPU, CPU, and disk. By solving a linear programming problem, it searches for efficient patterns to store and access tensors. FlexGen further compresses the weights and the attention cache to 4 bits with negligible accuracy loss. These techniques enable FlexGen to have a larger space of batch size choices and thus significantly increase maximum throughput. As a result, when running OPT-175B on a single 16GB GPU, FlexGen achieves significantly higher throughput compared to state-of-the-art offloading systems, reaching a generation throughput of 1 token/s for the first time with an effective batch size of 144. On the HELM benchmark, FlexGen can benchmark a 30B model with a 16GB GPU on 7 representative sub-scenarios in 21 hours. The code is available at https://github.com/FMInference/FlexGen.

Pipeline parallelism is a key technique for training large language models within GPU clusters. However, it often leads to a memory imbalance problem, where certain GPUs face high memory pressure while others underutilize their capacity. This imbalance results in suboptimal training performance, even when the overall GPU memory capacity is sufficient for more efficient setups. To address this inefficiency, we propose BPipe, a novel approach for achieving memory balance in pipeline parallelism. BPipe employs an activation balancing method to transfer intermediate activations between GPUs during training, enabling all GPUs to utilize comparable amounts of memory. With balanced memory utilization, BPipe enhances the training efficiency of large language models like GPT-3 by eliminating redundant recomputations or increasing the micro-batch size. Our evaluation conducted on 48 A100 GPUs across six nodes interconnected with HDR InfiniBand shows that BPipe accelerates the training of GPT-3 96B and GPT-3 134B models by 1.25x-2.17x compared to Megatron-LM, a state-of-the-art framework for training large language models.

Despite the success of autoregressive large language models in text generation, it remains a major challenge to generate text that satisfies complex constraints: sampling from the conditional distribution ${\Pr}(\text{text} | \alpha)$ is intractable for even the simplest lexical constraints $\alpha$. To overcome this challenge, we propose to use tractable probabilistic models (TPMs) to impose lexical constraints in autoregressive text generation models, which we refer to as GeLaTo (Generating Language with Tractable Constraints). To demonstrate the effectiveness of this framework, we use distilled hidden Markov models, where we can efficiently compute ${\Pr}(\text{text} | \alpha)$, to guide autoregressive generation from GPT2. GeLaTo achieves state-of-the-art performance on challenging benchmarks for constrained text generation (e.g., CommonGen), beating various strong baselines by a large margin. Our work not only opens up new avenues for controlling large language models but also motivates the development of more expressive TPMs.

Designing machine learning architectures for processing neural networks in their raw weight matrix form is a newly introduced research direction. Unfortunately, the unique symmetry structure of deep weight spaces makes this design very challenging. If successful, such architectures would be capable of performing a wide range of intriguing tasks, from adapting a pre-trained network to a new domain to editing objects represented as functions (INRs or NeRFs). As a first step towards this goal, we present here a novel network architecture for learning in deep weight spaces. It takes as input a concatenation of weights and biases of a pre-trained MLP and processes it using a composition of layers that are equivariant to the natural permutation symmetry of the MLP's weights: Changing the order of neurons in intermediate layers of the MLP does not affect the function it represents. We provide a full characterization of all affine equivariant and invariant layers for these symmetries and show how these layers can be implemented using three basic operations: pooling, broadcasting, and fully connected layers applied to the input in an appropriate manner. We demonstrate the effectiveness of our architecture and its advantages over natural baselines in a variety of learning tasks.