Mitigating the Impact of Outlier Channels for Language Model Quantization with Activation Regularization
This work addresses the problem of efficient deployment of large language models for practitioners by mitigating activation quantization issues, though it is incremental as it builds on existing quantization techniques.
The paper tackles the challenge of accurately quantizing language models to 4 bits per parameter by addressing outlier channels in activations, which hinder low-bitwidth quantization, and proposes a regularization strategy that enables W4A4 models to perform competitively with W16A16 baselines.
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.