Guanchen Li

Jinze Li, Yixing Xu, Guanchen Li et al.

Speculative decoding (SPD) accelerates large language model (LLM) inference by letting a smaller draft model propose multiple future tokens that are verified in parallel by a larger target model. The dominant SPD paradigm treats the target model as the sole reliable teacher, accepting a draft token only when it exactly matches the target prediction. This design implicitly assumes that the target is always the better choice at every position. In practice, this assumption does not hold. Although the draft is the weaker model overall, it is not uniformly inferior at the token level. In a meaningful fraction of cases where draft and target disagree, the draft's choice is the one that leads to the correct final answer. Inspired by this, we introduce \textbf{Collaborative Speculative Decoding (CoSpec)}, a generalization of SPD that no longer treats the target model as the sole token-level authority. CoSpec trains an arbitration policy via reinforcement learning to decide whether to accept tokens from the draft or target model, selectively accepting draft tokens at mismatches when doing so is likely to yield a correct final answer. Experimental results show that CoSpec maintains substantial speedups while surpassing target-only performance. By shifting the emphasis from imitation to collaboration, CoSpec suggests a new perspective on speculative decoding.

Guanchen Li, Xiandong Zhao, Lian Liu et al.

Pre-trained language models (PLMs) are engineered to be robust in contextual understanding and exhibit outstanding performance in various natural language processing tasks. However, their considerable size incurs significant computational and storage costs. Modern pruning strategies employ one-shot techniques to compress PLMs without the need for retraining on task-specific or otherwise general data; however, these approaches often lead to an indispensable reduction in performance. In this paper, we propose SDS, a Sparse-Dense-Sparse pruning framework to enhance the performance of the pruned PLMs from a weight distribution optimization perspective. We outline the pruning process in three steps. Initially, we prune less critical connections in the model using conventional one-shot pruning methods. Next, we reconstruct a dense model featuring a pruning-friendly weight distribution by reactivating pruned connections with sparse regularization. Finally, we perform a second pruning round, yielding a superior pruned model compared to the initial pruning. Experimental results demonstrate that SDS outperforms the state-of-the-art pruning techniques SparseGPT and Wanda under an identical sparsity configuration. For instance, SDS reduces perplexity by 9.13 on Raw-Wikitext2 and improves accuracy by an average of 2.05% across multiple zero-shot benchmarks for OPT-125M with 2:4 sparsity.

Guanchen Li, Yixing Xu, Zeping Li et al.

Structural pruning enhances hardware-agnostic inference efficiency for large language models (LLMs) yet often fails to maintain comparable performance. Local pruning performs efficient layer-by-layer compression but ignores global topology. Although global pruning aims to identify an optimal sparse model, intuitive methods typically adopt a two-stage paradigm that first evaluates substructure saliency and then applies global pruning, which ignores inter-structure dependencies and fails to achieve end-to-end optimization. To address these limitations, we propose Týr-the-Pruner, an efficient end-to-end search-based global structural pruning framework. This framework constructs a supernet by repeatedly applying local pruning across a range of sparsity ratios to each layer in an LLM, with the core goal of determining the optimal sparsity distribution under a target overall sparsity ratio. Concretely, we introduce an effective local pruning and an expectation error accumulation approach to improve supernet construction. Furthermore, we employ an iterative prune-and-search strategy with coarse-to-fine sparsity granularity to ensure efficient search convergence. Experimental results show that Týr-the-Pruner achieves state-of-the-art structural pruning, retaining 97% of the dense model's performance while removing a challenging 50% of Llama-3.1-70B's parameters. Code will be available at https://github.com/AMD-AGI/Tyr-the-Pruner.

Zekai Li, Ji Liu, Guanchen Li et al.

Structured sparsity has emerged as a popular model pruning technique, widely adopted in various architectures, including CNNs, Transformer models, and especially large language models (LLMs) in recent years. A promising direction to further improve post-pruning performance is weight permutation, which reorders model weights into patterns more amenable to pruning. However, the exponential growth of the permutation search space with the scale of Transformer architectures forces most methods to rely on greedy or heuristic algorithms, limiting the effectiveness of reordering. In this work, we propose a novel end-to-end learnable permutation framework. Our method introduces a learnable permutation cost matrix to quantify the cost of swapping any two input channels of a given weight matrix, a differentiable bipartite matching solver to obtain the optimal binary permutation matrix given a cost matrix, and a sparsity optimization loss function to directly optimize the permutation operator. We extensively validate our approach on vision and language Transformers, demonstrating that our method achieves state-of-the-art permutation results for structured sparsity.

Huanxuan Liao, Yixing Xu, Shizhu He et al.

Long-context inference in large language models (LLMs) is increasingly constrained by the KV cache bottleneck: memory usage grows linearly with sequence length, while attention computation scales quadratically. Existing approaches address this issue by compressing the KV cache along the temporal axis through strategies such as token eviction or merging to reduce memory and computational overhead. However, these methods often neglect fine-grained importance variations across feature dimensions (i.e., the channel axis), thereby limiting their ability to effectively balance efficiency and model accuracy. In reality, we observe that channel saliency varies dramatically across both queries and positions: certain feature channels carry near-zero information for a given query, while others spike in relevance. To address this oversight, we propose SPARK, a training-free plug-and-play method that applies unstructured sparsity by pruning KV at the channel level, while dynamically restoring the pruned entries during attention score computation. Notably, our approach is orthogonal to existing KV compression and quantization techniques, making it compatible for integration with them to achieve further acceleration. By reducing channel-level redundancy, SPARK enables processing of longer sequences within the same memory budget. For sequences of equal length, SPARK not only preserves or improves model accuracy but also reduces KV cache storage by over 30% compared to eviction-based methods. Furthermore, even with an aggressive pruning ratio of 80%, SPARK maintains performance with less degradation than 5% compared to the baseline eviction method, demonstrating its robustness and effectiveness. Our code will be available at https://github.com/Xnhyacinth/SparK.

Jinze Li, Yixing Xu, Guanchen Li et al.

Large language models (LLMs) achieve strong performance across diverse tasks but suffer from high inference latency due to their autoregressive generation. Speculative Decoding (SPD) mitigates this issue by verifying candidate tokens in parallel from a smaller draft model, yet its strict exact-match verification discards many semantically valid continuations. Moreover, existing training-based SPD methods often suffer from performance degradation on out-of-distribution (OOD) tasks. To this end, we propose Training-Free Loosely Speculative Decoding (FLy), a novel method that loosens the rigid verification criterion by leveraging the target model's self-corrective behavior to judge whether a draft-target mismatch remains semantically valid. FLy introduces a two-tier mechanism: an entropy-level gate that identifies whether the current token allows multiple plausible alternatives or is nearly deterministic, and a token-level deferred window that distinguishes genuine errors from differently worded yet semantically correct variants. To further reduce latency, we design a multi-level acceleration strategy that accelerates not only the target model but also the drafter itself. Owing to its training-free design, FLy composes seamlessly with arbitrary draft-target pairs and generalizes across models and domains without hyperparameter re-tuning. Experiments show that FLy preserves more than 99% of the target model's accuracy while achieving an average 2.81x speedup on Llama-3.1-70B-Instruct and 5.07x speedup on the 405B variant. Notably, on out-of-domain datasets, our method remains highly effective and outperforms the training-based method EAGLE-3 by 1.62x.

Zeping Li, Xinlong Yang, Ziheng Gao et al.

Large Language Models (LLMs) inherently use autoregressive decoding, which lacks parallelism in inference and results in significantly slow inference speed. While methods such as Medusa constructs parallelized heads, they lack adequate information interaction across different prediction positions. To overcome this limitation, we introduce Amphista, an enhanced speculative decoding framework that builds upon Medusa. Specifically, Amphista models an Auto-embedding Block capable of parallel inference, incorporating bi-directional attention to enable interaction between different drafting heads. Additionally, Amphista integrates Staged Adaptation Layers, which ensure a seamless transition of semantic information from the target model's autoregressive inference to the drafting heads' non-autoregressive inference, effectively achieving paradigm shift and feature fusion. Experimental results on Vicuna models using MT-Bench and Spec-Bench demonstrate that Amphista achieves substantial acceleration while maintaining generation quality. On MT-Bench, Amphista delivers up to 2.75$\times$ speedup over vanilla autoregressive decoding and 1.40$\times$ over Medusa on Vicuna 33B in wall-clock time.