MiniCPM-SALA: Hybridizing Sparse and Linear Attention for Efficient Long-Context Modeling

The evolution of large language models (LLMs) towards applications with ultra-long contexts faces challenges posed by the high computational and memory costs of the Transformer architecture. While existing sparse and linear attention mechanisms attempt to mitigate these issues, they typically involve a trade-off between memory efficiency and model performance. This paper introduces MiniCPM-SALA, a 9B-parameter hybrid architecture that integrates the high-fidelity long-context modeling of sparse attention (InfLLM-V2) with the global efficiency of linear attention (Lightning Attention). By employing a layer selection algorithm to integrate these mechanisms in a 1:3 ratio and utilizing a hybrid positional encoding (HyPE), the model maintains efficiency and performance for long-context tasks. Furthermore, we introduce a cost-effective continual training framework that transforms pre-trained Transformer-based models into hybrid models, which reduces training costs by approximately 75% compared to training from scratch. Extensive experiments show that MiniCPM-SALA maintains general capabilities comparable to full-attention models while offering improved efficiency. On a single NVIDIA A6000D GPU, the model achieves up to 3.5x the inference speed of the full-attention model at the sequence length of 256K tokens and supports context lengths of up to 1M tokens, a scale where traditional full-attention 8B models fail because of memory constraints.

TMPQ-DM: Joint Timestep Reduction and Quantization Precision Selection for Efficient Diffusion Models

Diffusion models have emerged as preeminent contenders in the realm of generative models. Distinguished by their distinctive sequential generative processes, characterized by hundreds or even thousands of timesteps, diffusion models progressively reconstruct images from pure Gaussian noise, with each timestep necessitating full inference of the entire model. However, the substantial computational demands inherent to these models present challenges for deployment, quantization is thus widely used to lower the bit-width for reducing the storage and computing overheads. Current quantization methodologies primarily focus on model-side optimization, disregarding the temporal dimension, such as the length of the timestep sequence, thereby allowing redundant timesteps to continue consuming computational resources, leaving substantial scope for accelerating the generative process. In this paper, we introduce TMPQ-DM, which jointly optimizes timestep reduction and quantization to achieve a superior performance-efficiency trade-off, addressing both temporal and model optimization aspects. For timestep reduction, we devise a non-uniform grouping scheme tailored to the non-uniform nature of the denoising process, thereby mitigating the explosive combinations of timesteps. In terms of quantization, we adopt a fine-grained layer-wise approach to allocate varying bit-widths to different layers based on their respective contributions to the final generative performance, thus rectifying performance degradation observed in prior studies. To expedite the evaluation of fine-grained quantization, we further devise a super-network to serve as a precision solver by leveraging shared quantization results. These two design components are seamlessly integrated within our framework, enabling rapid joint exploration of the exponentially large decision space via a gradient-free evolutionary search algorithm.