Zhuorui Liu

GLM-4. 5 Team, Aohan Zeng, Xin Lv et al.

We present GLM-4.5, an open-source Mixture-of-Experts (MoE) large language model with 355B total parameters and 32B activated parameters, featuring a hybrid reasoning method that supports both thinking and direct response modes. Through multi-stage training on 23T tokens and comprehensive post-training with expert model iteration and reinforcement learning, GLM-4.5 achieves strong performance across agentic, reasoning, and coding (ARC) tasks, scoring 70.1% on TAU-Bench, 91.0% on AIME 24, and 64.2% on SWE-bench Verified. With much fewer parameters than several competitors, GLM-4.5 ranks 3rd overall among all evaluated models and 2nd on agentic benchmarks. We release both GLM-4.5 (355B parameters) and a compact version, GLM-4.5-Air (106B parameters), to advance research in reasoning and agentic AI systems. Code, models, and more information are available at https://github.com/zai-org/GLM-4.5.

Chen Zhang, Zhuorui Liu, Dawei Song

With the increasingly giant scales of (causal) large language models (LLMs), the inference efficiency comes as one of the core concerns along the improved performance. In contrast to the memory footprint, the latency bottleneck seems to be of greater importance as there can be billions of requests to a LLM (e.g., GPT-4) per day. The bottleneck is mainly due to the autoregressive innateness of LLMs, where tokens can only be generated sequentially during decoding. To alleviate the bottleneck, the idea of speculative execution, which originates from the field of computer architecture, is introduced to LLM decoding in a \textit{draft-then-verify} style. Under this regime, a sequence of tokens will be drafted in a fast pace by utilizing some heuristics, and then the tokens shall be verified in parallel by the LLM. As the costly sequential inference is parallelized, LLM decoding speed can be significantly boosted. Driven by the success of LLMs in recent couple of years, a growing literature in this direction has emerged. Yet, there lacks a position survey to summarize the current landscape and draw a roadmap for future development of this promising area. To meet this demand, we present the very first survey paper that reviews and unifies literature of speculative execution in LLMs (e.g., blockwise parallel decoding, speculative decoding, etc.) in a comprehensive framework and a systematic taxonomy. Based on the taxonomy, we present a critical review and comparative analysis of the current arts. Finally we highlight various key challenges and future directions to further develop the area.

Zhuorui Liu, Chen Zhang, Dawei Song

With the rapid development of large language models (LLMs), handling long context has become one of the vital abilities in LLMs. Such long-context ability is accompanied by difficulties in deployment, especially due to the increased consumption of KV cache. There is certain work aiming to optimize the memory footprint of KV cache, inspired by the observation that attention heads can be categorized into retrieval heads that are of great significance and streaming heads that are of less significance. Typically, identifying the streaming heads and and waiving the KV cache in the streaming heads would largely reduce the overhead without hurting the performance that much. However, since employing both retrieval and streaming heads in one layer decomposes one large round of attention computation into two small ones, it may unexpectedly bring extra latency on accessing and indexing tensors. Based on this intuition, we impose an important improvement to the identification process of retrieval and streaming heads, in which we design a criterion that enforces exclusively retrieval or streaming heads gathered in one unique layer. In this way, we further eliminate the extra latency and only incur negligible performance degradation. Our method named \textsc{ZigzagAttention} is competitive among considered baselines owing to reduced latency and comparable performance.

Youhui Zuo, Sibo Wei, Chen Zhang et al.

With the advancements in long-context inference capabilities of large language models (LLMs), the KV cache has become one of the foundational components. However, its substantial GPU memory consumption makes KV cache compression a key technique for enabling efficient LLM inference in industrial scenarios. While recent studies have focused on optimizing the memory occupied by the KV cache, they overlook two critical factors: preserving semantic coherence and considering task-specific characteristic during compression. To address these limitations, we propose a novel task-adaptive KV cache window selection method, WindowKV. WindowKV dynamically selects local semantic windows consisting of consecutive tokens, according to task-specific characteristics, ensuring the retained KV cache captures continuous, essential context. Additionally, we introduce an intra-group layer KV cache indices sharing strategy to reduce computational overhead, achieving a balance between performance and efficiency. We rigorously evaluate WindowKV on the LongBench benchmark, and the results demonstrate that it maintains a performance comparable to full KV cache retention while using only 12% of the original KV cache, significantly reducing memory requirements. Furthermore, our method also achieves state-of-the-art results in the Needle-in-a-Haystack evaluation, highlighting its effectiveness and robustness.