Filtered by: LLM × Clear all

WIDE: Boosting Adaptive LLM Inference via Token-level Dynamic Width Pruning

Haozhe Hu, Hao Wu, Peiran Yin, Chao Han 2026-07-31
Attention Inference LLM × Training Hardware

WIDE introduces the first end-to-end differentiable token-level dynamic width pruning framework for LLMs, enabling fine-grained computation allocation by letting each token select attention-head and FFN-channel groups. The method uses a two-stage training pipeline and a pruning–kernel co-design that decomposes acceleration into mask reordering and block-level skipping for practical execution. At 50% sparsity, WIDE achieves a 55.1% performance boost over state-of-the-art dynamic depth pruning in calibration-only settings, with kernel-level speedups up to 1.98x for prefill and 4.95x for decoding, and end-to-end accelerations of 1.68x and 1.55x. This matters because it makes fine-grained dynamic pruning hardware-efficient, closing the gap between accuracy retention and real-world inference speedups for LLMs.

PDF

GyRot: Leveraging Hidden Synergy between Rotation and Fine-grained Group Quantization for Low-bit LLM Inference

Sangjin Kim, Yuseon Choi, Byeongcheol Kim, Jungjun Oh 2026-07-31
Inference LLM × Quantization Accelerator Hardware Low Bit Tensor

Problem: Low-bit quantization for LLM inference struggles when combining rotation and fine-grained group quantization due to a mismatch between global rotation and localized group scaling, causing accuracy loss or hardware overhead. Method: GyRot proposes an algorithm-hardware co-design framework with Coarse Rotation, Fine Grouping (CoRFiG) and Harmonic-Aligned Permutation (HAP) to integrate rotation and group quantization, plus a zero-point rounding strategy for fully integer dequantization. Finding: On an INT4-based tensor PE architecture, GyRot achieves state-of-the-art 4-bit accuracy across LLaMA-family models, delivering up to 3.4x speedup and 3.6x energy efficiency over baseline LLM accelerators. Why it matters: This validates GyRot's practical effectiveness for scalable and energy-efficient LLM deployment, addressing a critical bottleneck in low-bit inference hardware.

PDF

A Photonic-CXL Memory Appliance for Scalable KV Cache Management in LLM Inference

Jing Ding, Yash Nishant, Chandrish Ambati, Jyothsna Kamati 2026-07-31
Cache LLM × Inference GPU Architecture Simulation Workload

The paper addresses the memory wall in LLM inference, where KV cache demands tens of terabytes at hundreds of GB/s exceed current memory tier capabilities. The proposed Marvell Photonic Fabric Memory Appliance replaces electrical switches with a passive fiber shuffle in a switch-free full-crossbar topology, delivering 32 TB shared memory across 16 hosts via photonic-CXL hybrid architecture. Emulation results show over 50% latency reduction versus electrical CXL pools, while simulation demonstrates a 6.6x improvement in time-to-first-token by eliminating cache eviction cliffs for multi-turn workloads. This work matters because it enables practical TB-scale shared memory for concurrent long-context users, overcoming the scalability limits of electrical CXL pooling in real deployments.

PDF

InferScale: GPU-Native KV Injection for Personalized LLM Serving

Peter Li, Prashant Pandey 2026-07-31
Memory Microarchitecture Simulation GPU Cache LLM × Serving

InferScale is a GPU-native LLM memory system that replaces repeated prompt prefilling with reusable KV state, addressing the TTFT increase caused by injecting persistent personalized context into prompts. It precomputes KV representations for memory facts, stores them with semantic embeddings on the GPU, and injects them directly into vLLM's paged cache, using Chunked RoPE for dynamic memory assembly and Context-Window Encoding to mitigate the loss of cross-fact context. On LoCoMo with three open-weight models, InferScale keeps TTFT nearly constant as retrieval budget grows, reducing TTFT by 72-79% (3.6-4.8x) at k=50, achieving 60.3% accuracy versus 63.3% for Mem0, and delivering 3.7-4.5x throughput under concurrent load. This decouples memory-conditioned serving latency from retrieved-context size while preserving application quality, enabling efficient personalized LLM serving without engine modifications or fine-tuning.

PDF