Longwei Zou

Longwei Zou, Lin Zhong

Linear attention has recently gained significant attention for long-context inference due to its constant decoding cost with respect to context length. However, existing serving systems typically serve linear attention by recurrently computing and updating a large linear attention state in every decoding step. Since the state is much larger than the per-token key and value, recurrent decoding incurs substantial memory access and becomes inefficient for serving linear attention. In this paper, we propose KVBuffer, an IO-aware serving mechanism for linear attention. By buffering recent keys and values, KVBuffer enables serving systems to compute linear attention outputs in more flexible and memory-efficient ways. For decoding, KVBuffer enables chunkwise computation, which reduces average memory access and decoding latency by deferring state updates and applying them in batch. For speculative decoding, KVBuffer verifies draft tokens in parallel and avoids storing temporary states. For short contexts, KVBuffer computes attention outputs directly from buffered keys and values, without creating or updating the linear attention state. We implement KVBuffer in SGLang for Qwen3-Next. Our evaluations show that KVBuffer can reduce linear attention decoding latency by up to 45.17% and increase the maximum number of serving requests by 5x for speculative decoding when verifying four draft tokens.

Longwei Zou, Yan Liu, Jiamu Kang et al.

The revolutionary capabilities of Large Language Models (LLMs) are attracting rapidly growing popularity and leading to soaring user requests to inference serving systems. Caching techniques, which leverage data reuse to reduce computation, offer opportunities to optimize the performance of LLM inference engines. On the one hand, the low-level key-value (KV) cache working at the token level is widely adopted, albeit it incurs significant overhead as request volume grows. On the other hand, instruction-level caching, which stores full instruction-response pairs, is expected to play an increasingly crucial role. However, the high variability in the content and length of instructions make it rare for identical instructions to recur within a short time window, presenting challenges for effective caching instruction-response pairs. To address this challenge, we propose InstCache, a predictive caching mechanism for LLM serving systems. Leveraging the capability of LLMs, we can effectively reorder the representation space of instruction texts and develop a sufficient level of spatial locality. Such spatial locality enables us to predict potential instructions located in a compact region in the space, resulting in an effective caching system at runtime. Experimental results demonstrate that InstCache achieves a 2.3x higher hit rate compared to the upper bound of traditional caching mechanisms on WildChat dataset and reduces the time per output token of vLLM by up to 42.0% and 50.0% on LMSys and Moss datasets, respectively.

Longwei Zou, Han Zhang, Yangdong Deng

The fast growing capabilities of large-scale deep learning models, such as Bert, GPT and ViT, are revolutionizing the landscape of NLP, CV and many other domains. Training such models, however, poses an unprecedented demand for computing power, which incurs exponentially increasing energy cost and carbon dioxide emissions. It is thus critical to develop efficient training solutions to reduce the training costs. Motivated by a set of key observations of inter- and intra-layer similarities among feature maps and attentions that can be identified from typical training processes, we propose a multi-level framework for training acceleration. Specifically, the framework is based on three basic operators, Coalescing, De-coalescing and Interpolation, which can be orchestrated to build a multi-level training framework. The framework consists of a V-cycle training process, which progressively down- and up-scales the model size and projects the parameters between adjacent levels of models via coalescing and de-coalescing. The key idea is that a smaller model that can be trained for fast convergence and the trained parameters provides high-qualities intermediate solutions for the next level larger network. The interpolation operator is designed to break the symmetry of neurons incurred by de-coalescing for better convergence performance. Our experiments on transformer-based language models (e.g. Bert, GPT) as well as a vision model (e.g. DeiT) prove that the proposed framework reduces the computational cost by about 20% on training BERT/GPT-Base models and up to 51.6% on training the BERT-Large model while preserving the performance.

Longwei Zou, Qingyang Wang, Han Zhao et al.

The fast-growing large scale language models are delivering unprecedented performance on almost all natural language processing tasks. However, the effectiveness of large language models are reliant on an exponentially increasing number of parameters. The overwhelming computation complexity incurs a high inference latency that negatively affects user experience. Existing methods to improve inference efficiency, such as tensor parallelism and quantization, target to reduce per-layer computing latency, yet overlook the cumulative latency due to the number of layers. Recent works on reducing the cumulative latency through layer removing, however, lead to significant performance drop. Motivated by the similarity of inputs among adjacent layers, we propose to identify quasi-independent layers, which can be concurrently computed to significantly decrease inference latency. We also introduce a bypassing technique to mitigate the effect of information loss. Empirical experiments of the proposed approach on the LLaMA models confirm that Concurrent Computation of Quasi-Independent Layers (CQIL) can reduce latency by up to 48.3% on LLaMA-33B, while maintaining a close level of performance.