Test-time interventions for language models can enhance factual accuracy, mitigate harmful outputs, and improve model efficiency without costly retraining. But despite a flood of new methods, different types of interventions are largely developing independently. In practice, multiple interventions must be applied sequentially to the same model, yet we lack standardized ways to study how interventions interact. We fill this gap by introducing composable interventions, a framework to study the effects of using multiple interventions on the same language models, featuring new metrics and a unified codebase. Using our framework, we conduct extensive experiments and compose popular methods from three emerging intervention categories---knowledge editing, model compression, and machine unlearning. Our results over 417 different compositions uncover meaningful interactions: compression hinders editing and unlearning, composing interventions hinges on their order of application, and popular general-purpose metrics are inadequate for assessing composability. Taken together, our findings showcase clear gaps in composability, suggesting a need for new multi-objective interventions.

Memorization in language models is typically treated as a homogenous phenomenon, neglecting the specifics of the memorized data. We instead model memorization as the effect of a set of complex factors that describe each sample and relate it to the model and corpus. To build intuition around these factors, we break memorization down into a taxonomy: recitation of highly duplicated sequences, reconstruction of inherently predictable sequences, and recollection of sequences that are neither. We demonstrate the usefulness of our taxonomy by using it to construct a predictive model for memorization. By analyzing dependencies and inspecting the weights of the predictive model, we find that different factors have different influences on the likelihood of memorization depending on the taxonomic category.

How do large language models (LLMs) develop and evolve over the course of training? How do these patterns change as models scale? To answer these questions, we introduce Pythia, a suite of 16 LLMs all trained on public data seen in the exact same order and ranging in size from 70M to 12B parameters. We provide public access to 154 checkpoints for each one of the 16 models, alongside tools to download and reconstruct their exact training dataloaders for further study. We intend Pythia to facilitate research in many areas, and we present several case studies including novel results in memorization, term frequency effects on few-shot performance, and reducing gender bias. We demonstrate that this highly controlled setup can be used to yield novel insights toward LLMs and their training dynamics. Trained models, analysis code, training code, and training data can be found at https://github.com/EleutherAI/pythia.

Demand for high performance deep learning (DL) inference in software applications is growing rapidly. DL workloads run on myriad platforms, including general purpose processors (CPU), system-on-chip (SoC) with accelerators, graphics processing units (GPU), and neural processing unit (NPU) addin cards. DL software engineers typically must choose between relatively slow general hardware (e.g., CPUs, SoCs) or relatively expensive, large, power-hungry hardware (e.g., GPUs, NPUs). This paper describes Centaur Technology's Ncore, the industry's first high-performance DL coprocessor technology integrated into an x86 SoC with server-class CPUs. Ncore's 4096 byte-wide SIMD architecture supports INT8, UINT8, INT16, and BF16 datatypes, with 20 tera-operations-per-second compute capability. Ncore shares the SoC ring bus for low-latency communication and work sharing with eight 64-bit x86 cores, offering flexible support for new and evolving models. The x86 SoC platform can further scale out performance via multiple sockets, systems, or third-party PCIe accelerators. Ncore's software stack automatically converts quantized models for Ncore consumption and leverages existing DL frameworks. In MLPerf's Inference v0.5 closed division benchmarks, Ncore achieves 1218 IPS throughput and 1.05ms latency on ResNet-50v1.5 and achieves lowest latency of all Mobilenet-V1 submissions (329μs). Ncore yields 23x speedup over other x86 vendor percore throughput, while freeing its own x86 cores for other work. Ncore is the only integrated solution among the memory intensive neural machine translation (NMT) submissions.

Even though their capacities are still orders of magnitude lower than those of hard disks, flash storage systems are rapidly gaining importance in energy-constrained systems. This paper focuses on USB flash drives, which can provide portable storage to mobile systems or storage to systems that otherwise do not have persistent storage opportunities (e.g., low-power sensor devices). The paper presents studies relating to power consumption, energy overheads and benefits, and performance impacts of USB flash drives. The key insights obtained from these experiments are that (i) read/write costs are not significantly greater than idle costs and (ii) the size of the flash itself has only limited bearing on energy consumption.