Aniruddha Nrusimha

Lucas Torroba Hennigen, Shannon Shen, Aniruddha Nrusimha et al. · mit

LLMs are vulnerable to hallucinations, and thus their outputs generally require laborious human verification for high-stakes applications. To this end, we propose symbolically grounded generation (SymGen) as a simple approach for enabling easier manual validation of an LLM's output. SymGen prompts an LLM to interleave its regular output text with explicit symbolic references to fields present in some conditioning data (e.g., a table in JSON format). The references can be used to display the provenance of different spans of text in the generation, reducing the effort required for manual verification. Across a range of data-to-text and question-answering experiments, we find that LLMs are able to directly output text that makes use of accurate symbolic references while maintaining fluency and factuality. In a human study we further find that such annotations can streamline human verification of machine-generated text. Our code will be available at http://symgen.github.io.

William Brandon, Aniruddha Nrusimha, Kevin Qian et al.

To help address the growing demand for ever-longer sequence lengths in transformer models, Liu et al. recently proposed Ring Attention, an exact attention algorithm capable of overcoming per-device memory bottle- necks by distributing self-attention across multiple devices. In this paper, we study the performance characteristics of Ring Attention in the important special case of causal transformer models, and identify a key workload imbal- ance due to triangular structure of causal attention computations. We propose a simple extension to Ring Attention, which we call Striped Attention to fix this imbalance. Instead of devices having contiguous subsequences, each device has a subset of tokens distributed uniformly throughout the sequence, which we demonstrate leads to more even workloads. In experiments running Striped Attention on A100 GPUs and TPUv4s, we are able to achieve up to 1.45x end-to-end throughput improvements over the original Ring Attention algorithm on causal transformer training at a sequence length of 256k. Furthermore, on 16 TPUv4 chips, we were able to achieve 1.65x speedups at sequence lengths of 786k. We release the code for our experiments as open source

Zachary Ankner, Rishab Parthasarathy, Aniruddha Nrusimha et al.

To combat the memory bandwidth-bound nature of autoregressive LLM inference, previous research has proposed the speculative decoding frame-work. To perform speculative decoding, a small draft model proposes candidate continuations of the input sequence that are then verified in parallel by the base model. One way to specify the draft model, as used in the recent Medusa decoding framework, is as a collection of lightweight heads, called draft heads, that operate on the base model's hidden states. To date, all existing draft heads have been sequentially independent, meaning that they speculate tokens in the candidate continuation independently of any preceding tokens in the candidate continuation. In this work, we propose Hydra heads: a sequentially-dependent drop-in replacement for standard draft heads that significantly improves the accuracy of draft head speculation. We further explore the design space of Hydra head training objectives and architectures, and propose a carefully tuned Hydra head recipe, which we call Hydra++, that improves decoding throughput by up to 1.31x and 2.70x compared to Medusa decoding and autoregressive de-coding respectively. Overall, Hydra heads are a simple and well-motivated intervention on standard draft heads that significantly improve the end-to-end speed of draft head-based speculative decoding. We make our code publicly available at https://github.com/zankner/Hydra.

Paras Jain, Ajay Jain, Aniruddha Nrusimha et al.

We formalize the problem of trading-off DNN training time and memory requirements as the tensor rematerialization optimization problem, a generalization of prior checkpointing strategies. We introduce Checkmate, a system that solves for optimal rematerialization schedules in reasonable times (under an hour) using off-the-shelf MILP solvers or near-optimal schedules with an approximation algorithm, then uses these schedules to accelerate millions of training iterations. Our method scales to complex, realistic architectures and is hardware-aware through the use of accelerator-specific, profile-based cost models. In addition to reducing training cost, Checkmate enables real-world networks to be trained with up to 5.1x larger input sizes. Checkmate is an open-source project, available at https://github.com/parasj/checkmate.

William Brandon, Mayank Mishra, Aniruddha Nrusimha et al.

Key-value (KV) caching plays an essential role in accelerating decoding for transformer-based autoregressive large language models (LLMs). However, the amount of memory required to store the KV cache can become prohibitive at long sequence lengths and large batch sizes. Since the invention of the transformer, two of the most effective interventions discovered for reducing the size of the KV cache have been Multi-Query Attention (MQA) and its generalization, Grouped-Query Attention (GQA). MQA and GQA both modify the design of the attention block so that multiple query heads can share a single key/value head, reducing the number of distinct key/value heads by a large factor while only minimally degrading accuracy. In this paper, we show that it is possible to take Multi-Query Attention a step further by also sharing key and value heads between adjacent layers, yielding a new attention design we call Cross-Layer Attention (CLA). With CLA, we find that it is possible to reduce the size of the KV cache by another 2x while maintaining nearly the same accuracy as unmodified MQA. In experiments training 1B- and 3B-parameter models from scratch, we demonstrate that CLA provides a Pareto improvement over the memory/accuracy tradeoffs which are possible with traditional MQA, enabling inference with longer sequence lengths and larger batch sizes than would otherwise be possible

Aniruddha Nrusimha, Mayank Mishra, Naigang Wang et al.

We consider the problem of accurate quantization for language models, where both the weights and activations are uniformly quantized to 4 bits per parameter, the lowest bitwidth format natively supported by GPU hardware. In this context, the key challenge is activation quantization: it is known that language models contain outlier channels whose values on average are orders of magnitude higher than than other channels, which prevents accurate low-bitwidth quantization with known techniques. We systematically study this phenomena and find that these outlier channels emerge early in training, and that they occur more frequently in layers with residual streams. We then propose a simple strategy which regularizes a layer's inputs via quantization-aware training (QAT) and its outputs via activation kurtosis regularization. We show that regularizing both the inputs and outputs is crucial for preventing a model's "migrating" the difficulty in input quantization to the weights, which makes post-training quantization (PTQ) of weights more difficult. When combined with weight PTQ, we show that our approach can obtain a W4A4 model that performs competitively to the standard-precision W16A16 baseline.

Aniruddha Nrusimha, William Brandon, Mayank Mishra et al.

The size and compute characteristics of modern large language models have led to an increased interest in developing specialized kernels tailored for training and inference. Existing kernels primarily optimize for compute utilization, targeting the large-batch training and inference settings. However, low-batch inference, where memory bandwidth and kernel launch overheads contribute are significant factors, remains important for many applications of interest such as in edge deployment and latency-sensitive applications. This paper describes FlashFormer, a proof-of-concept kernel for accelerating single-batch inference for transformer-based large language models. Across various model sizes and quantizations settings, we observe nontrivial speedups compared to existing state-of-the-art inference kernels.

Sehoon Kim, Amir Gholami, Zhewei Yao et al.

End-to-end neural network models achieve improved performance on various automatic speech recognition (ASR) tasks. However, these models perform poorly on edge hardware due to large memory and computation requirements. While quantizing model weights and/or activations to low-precision can be a promising solution, previous research on quantizing ASR models is limited. In particular, the previous approaches use floating-point arithmetic during inference and thus they cannot fully exploit efficient integer processing units. Moreover, they require training and/or validation data during quantization, which may not be available due to security or privacy concerns. To address these limitations, we propose an integer-only, zero-shot quantization scheme for ASR models. In particular, we generate synthetic data whose runtime statistics resemble the real data, and we use it to calibrate models during quantization. We apply our method to quantize QuartzNet, Jasper, and Conformer and show negligible WER degradation as compared to the full-precision baseline models, even without using any data. Moreover, we achieve up to 2.35x speedup on a T4 GPU and 4x compression rate, with a modest WER degradation of <1% with INT8 quantization.