The impressive performance gains of modern language models currently rely on scaling parameters: larger models store more world knowledge and reason better. Yet compressing all world knowledge into parameters is unnecessary, as only a fraction is used per prompt, and impractical for edge devices with limited inference-time memory and compute. We address this shortcoming by a memory-augmented architecture and a pretraining strategy aligned with existing hardware paradigms. We introduce small language models that access large hierarchical parametric memory banks encoding world knowledge. During pretraining and inference, we fetch a small, context-dependent memory block and add it to the model. Our pretraining learns to store long-tail world knowledge in the memory parameters, while the small language model acts as an anchor capturing common knowledge and general reasoning abilities. Through trillion-token-scale experiments, we show significant gains: a 160M-parameters model augmented with an 18M-parameters memory fetched from a 4.6B memory bank obtains comparable performance to a regular model with more than 2x the parameters. Through extensive experiments, we study the optimal type and size of parametric memories in transformers, scaling them to over 21B parameters. We find that our proposed hierarchical feed-forward memories work robustly across transformer architectures, whether added during pretraining or post-hoc.

In today's Function-as-a-Service offerings, a programmer is usually responsible for configuring function memory for its successful execution, which allocates proportional function resources such as CPU and network. However, right-sizing the function memory force developers to speculate performance and make ad-hoc configuration decisions. Recent research has highlighted that a function's input characteristics, such as input size, type and number of inputs, significantly impact its resource demand, run-time performance and costs with fluctuating workloads. This correlation further makes memory configuration a non-trivial task. On that account, an input-aware function memory allocator not only improves developer productivity by completely hiding resource-related decisions but also drives an opportunity to reduce resource wastage and offer a finer-grained cost-optimised pricing scheme. Therefore, we present MemFigLess, a serverless solution that estimates the memory requirement of a serverless function with input-awareness. The framework executes function profiling in an offline stage and trains a multi-output Random Forest Regression model on the collected metrics to invoke input-aware optimal configurations. We evaluate our work with the state-of-the-art approaches on AWS Lambda service to find that MemFigLess is able to capture the input-aware resource relationships and allocate upto 82% less resources and save up to 87% run-time costs.