Self-GC governs agent context as indexed objects with planner-proposed actions, achieving 84.85% no-impact on future continuations on a hard set versus 54-70% for baselines.
hub Canonical reference
H$_2$O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models
Canonical reference. 80% of citing Pith papers cite this work as background.
36
Pith papers citing it
28
external citations · Pith
Background
80% of classified citations
abstract
Large Language Models (LLMs), despite their recent impressive accomplishments, are notably cost-prohibitive to deploy, particularly for applications involving long-content generation, such as dialogue systems and story writing. Often, a large amount of transient state information, referred to as the KV cache, is stored in GPU memory in addition to model parameters, scaling linearly with the sequence length and batch size. In this paper, we introduce a novel approach for implementing the KV cache which significantly reduces its memory footprint. Our approach is based on the noteworthy observation that a small portion of tokens contributes most of the value when computing attention scores. We call these tokens Heavy Hitters (H$_2$). Through a comprehensive investigation, we find that (i) the emergence of H$_2$ is natural and strongly correlates with the frequent co-occurrence of tokens in the text, and (ii) removing them results in significant performance degradation. Based on these insights, we propose Heavy Hitter Oracle (H$_2$O), a KV cache eviction policy that dynamically retains a balance of recent and H$_2$ tokens. We formulate the KV cache eviction as a dynamic submodular problem and prove (under mild assumptions) a theoretical guarantee for our novel eviction algorithm which could help guide future work. We validate the accuracy of our algorithm with OPT, LLaMA, and GPT-NeoX across a wide range of tasks. Our implementation of H$_2$O with 20% heavy hitters improves the throughput over three leading inference systems DeepSpeed Zero-Inference, Hugging Face Accelerate, and FlexGen by up to 29$\times$, 29$\times$, and 3$\times$ on OPT-6.7B and OPT-30B. With the same batch size, H2O can reduce the latency by up to 1.9$\times$. The code is available at https://github.com/FMInference/H2O.
hub tools
citation-role summary
background 4
baseline 1
citation-polarity summary
representative citing papers
NLL-guided layer selection identifies 1/4 of layers for full attention in hybrid models, matching periodic 1/2-FA baseline accuracy on LongMemEval with Qwen3-4B while halving the full-attention compute budget.
STAR-KV applies differentiable soft thresholding for per-head and per-block adaptive low-rank KV cache compression, combined with hybrid decomposition and low-rank-aware quantization, achieving up to 75% compression and 3.1x throughput gains.
AVMP separates KV and SSM cache pools behind unified virtual addressing with failure-triggered migration, cutting OOM events 7.6% and raising throughput 1.83-13.3x on synthetic loads and 2.36x on ShareGPT traces.
Tensor Cache augments sliding-window attention with an eviction-fed outer-product associative memory and a training correction to improve long-context performance under bounded memory.
Apple MPS transformer decoding shows abrupt latency spikes up to 21x in narrow decoding-budget intervals due to KV cache and execution regime shifts, absent on CPU and CUDA.
MISA routes to a small subset of indexer heads via block statistics, matching full DSA performance on LongBench with 4-8x fewer heads and 3.82x speedup while recovering over 92% of selected tokens.
Lighthouse Attention enables faster long-context pre-training via gradient-free symmetrical hierarchical compression of QKV while preserving causality, followed by a short full-attention recovery that yields lower loss than standard full-attention training.
Transformers need depth scaling as the product of ceil(k/s) and log n terms for k-hop pointer chasing under cache size s, with a conjectured lower bound, proved upper bound via windowed pointer doubling, and an adaptive-oblivious error separation.
Sparse prefix caching via dynamic programming for optimal checkpoint placement under overlap distributions improves the Pareto frontier for recurrent and hybrid LLM serving on shared-prefix data.
SnapStream deploys sparse KV attention in a production inference system on dataflow accelerators, delivering 4x on-chip memory savings for DeepSeek-671B at 128k context with up to 1832 tokens/sec and minimal accuracy loss on LongBench-v2, AIME24, and LiveCodeBench.
Medusa augments LLMs with multiple decoding heads and tree-based attention to predict and verify several tokens in parallel, yielding 2.2-3.6x inference speedup via two fine-tuning regimes.
KV-cache eviction, prompt compression, recurrent state bounding, and agent memory consolidation are unified as one rate-distortion problem with a shared lower bound, shared failure mode, and transferable mechanisms.
HOLA pairs a compressive delta-rule recurrent state with a residual-selected exact KV cache and decoupled RMSNorm-gamma read, yielding lower perplexity than both standard linear attention and full-attention baselines on Wikitext and LAMBADA plus stronger needle-in-haystack recall.
EntropyInfer adaptively allocates inference compute using per-head attention entropy for rigid/dynamic classification during prefilling and compresses KV cache with generated tokens, achieving up to 2.39x speedup on long contexts.
STS is a two-stage pruning framework that decouples structural diversity via repulsion sampling from semantic filtering via cross-attention to reduce redundancy in visual tokens for VLMs.
MomentKV maintains count, key mean, value mean, and value-key covariance over evicted tokens to guide selective eviction and provide a first-order approximation of their attention contribution, outperforming baselines on LongBench and RULER.
Spherical KV combines angle-domain attention using spherical key codes with rate-distortion retention to cut KV cache residency and HBM traffic while keeping a paged, fusion-friendly decode path.
Louver is a new index for LLM KV caches that guarantees zero false negatives for keys above a relevance threshold, runs faster than prior sparse and some dense attention methods, and integrates lightly into existing pipelines.
Sub-token routing in LoRA-adapted transformers adds a finer compression axis for KV caches, with query-independent and query-aware designs that improve efficiency under reduced budgets when combined with token-level selection.
StreamingVLM enables stable real-time understanding of infinite video streams at up to 8 FPS using a streaming KV cache and aligned SFT on overlapped chunks, with a 66.18% win rate over GPT-4O mini on a new two-hour video benchmark.
KIVI applies asymmetric 2-bit quantization to KV cache with per-channel keys and per-token values, reducing memory 2.6x and boosting throughput up to 3.47x with near-identical quality on Llama, Falcon, and Mistral.
FastGen adaptively compresses LLM KV caches via lightweight attention profiling: evicting long-range contexts on local heads, non-special tokens on special-token heads, and retaining full caches on broad-attention heads, yielding substantial memory savings with negligible quality loss.
GSRQ applies a gain-shape variant of K-means inside residual quantization to improve directional fidelity, raising LongBench accuracy from 11.34 to 33.54 at 1-bit on LLaMA-3-8B.
citing papers explorer
-
Self-GC: Self-Governing Context for Long-Horizon LLM Agents
Self-GC governs agent context as indexed objects with planner-proposed actions, achieving 84.85% no-impact on future continuations on a hard set versus 54-70% for baselines.
-
NLL-Guided Full-Attention Layer Selection for Training-Free Sliding-Window Adaptation
NLL-guided layer selection identifies 1/4 of layers for full attention in hybrid models, matching periodic 1/2-FA baseline accuracy on LongMemEval with Qwen3-4B while halving the full-attention compute budget.
-
STAR-KV: Low-Rank KV Cache Compression via Soft Thresholding for Adaptive Rank Control
STAR-KV applies differentiable soft thresholding for per-head and per-block adaptive low-rank KV cache compression, combined with hybrid decomposition and low-rank-aware quantization, achieving up to 75% compression and 3.1x throughput gains.
-
Asymmetric Virtual Memory Paging for Hybrid Mamba-Transformer Inference
AVMP separates KV and SSM cache pools behind unified virtual addressing with failure-triggered migration, cutting OOM events 7.6% and raising throughput 1.83-13.3x on synthetic loads and 2.36x on ShareGPT traces.
-
Tensor Cache: Eviction-conditioned Associative Memory for Transformers
Tensor Cache augments sliding-window attention with an eviction-fed outer-product associative memory and a training correction to improve long-context performance under bounded memory.
-
Non-Monotonic Latency in Apple MPS Decoding: KV Cache Interactions and Execution Regimes
Apple MPS transformer decoding shows abrupt latency spikes up to 21x in narrow decoding-budget intervals due to KV cache and execution regime shifts, absent on CPU and CUDA.
-
MISA: Mixture of Indexer Sparse Attention for Long-Context LLM Inference
MISA routes to a small subset of indexer heads via block statistics, matching full DSA performance on LongBench with 4-8x fewer heads and 3.82x speedup while recovering over 92% of selected tokens.
-
Long Context Pre-Training with Lighthouse Attention
Lighthouse Attention enables faster long-context pre-training via gradient-free symmetrical hierarchical compression of QKV while preserving causality, followed by a short full-attention recovery that yields lower loss than standard full-attention training.
-
How Much Cache Does Reasoning Need? Depth-Cache Tradeoffs in KV-Compressed Transformers
Transformers need depth scaling as the product of ceil(k/s) and log n terms for k-hop pointer chasing under cache size s, with a conjectured lower bound, proved upper bound via windowed pointer doubling, and an adaptive-oblivious error separation.
-
Sparse Prefix Caching for Hybrid and Recurrent LLM Serving
Sparse prefix caching via dynamic programming for optimal checkpoint placement under overlap distributions improves the Pareto frontier for recurrent and hybrid LLM serving on shared-prefix data.
-
SnapStream: Efficient Long Sequence Decoding on Dataflow Accelerators
SnapStream deploys sparse KV attention in a production inference system on dataflow accelerators, delivering 4x on-chip memory savings for DeepSeek-671B at 128k context with up to 1832 tokens/sec and minimal accuracy loss on LongBench-v2, AIME24, and LiveCodeBench.
-
Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads
Medusa augments LLMs with multiple decoding heads and tree-based attention to predict and verify several tokens in parallel, yielding 2.2-3.6x inference speedup via two fine-tuning regimes.
-
What to Keep, What to Forget: A Rate--Distortion View of Memory Compaction in LLMs and Agents
KV-cache eviction, prompt compression, recurrent state bounding, and agent memory consolidation are unified as one rate-distortion problem with a shared lower bound, shared failure mode, and transferable mechanisms.
-
A Hippocampus for Linear Attention: An Exact Memory for What the Recurrent State Forgets
HOLA pairs a compressive delta-rule recurrent state with a residual-selected exact KV cache and decoupled RMSNorm-gamma read, yielding lower perplexity than both standard linear attention and full-attention baselines on Wikitext and LAMBADA plus stronger needle-in-haystack recall.
-
From Rigid to Dynamic: Entropy-Guided Adaptive Inference for Long-Context LLMs
EntropyInfer adaptively allocates inference compute using per-head attention entropy for rigid/dynamic classification during prefilling and compresses KV cache with generated tokens, achieving up to 2.39x speedup on long contexts.
-
When Attention Collapses: Stage-Aware Visual Token Pruning from Structure to Semantics
STS is a two-stage pruning framework that decouples structural diversity via repulsion sampling from semantic filtering via cross-attention to reduce redundancy in visual tokens for VLMs.
-
MomentKV: Closing the Directional Gap in KV Cache Eviction for Long-Context Inference
MomentKV maintains count, key mean, value mean, and value-key covariance over evicted tokens to guide selective eviction and provide a first-order approximation of their attention contribution, outperforming baselines on LongBench and RULER.
-
SPHERICAL KV: Angle-Domain Attention and Rate-Distortion Retention for Efficient Long-Context Inference
Spherical KV combines angle-domain attention using spherical key codes with rate-distortion retention to cut KV cache residency and HBM traffic while keeping a paged, fusion-friendly decode path.
-
Sparse Attention as a Range Searching Problem: Towards an Inference-Efficient Index for KV Cache
Louver is a new index for LLM KV caches that guarantees zero false negatives for keys above a relevance threshold, runs faster than prior sparse and some dense attention methods, and integrates lightly into existing pipelines.
-
Sub-Token Routing in LoRA for Adaptation and Query-Aware KV Compression
Sub-token routing in LoRA-adapted transformers adds a finer compression axis for KV caches, with query-independent and query-aware designs that improve efficiency under reduced budgets when combined with token-level selection.
-
StreamingVLM: Real-Time Understanding for Infinite Video Streams
StreamingVLM enables stable real-time understanding of infinite video streams at up to 8 FPS using a streaming KV cache and aligned SFT on overlapped chunks, with a 66.18% win rate over GPT-4O mini on a new two-hour video benchmark.
-
KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache
KIVI applies asymmetric 2-bit quantization to KV cache with per-channel keys and per-token values, reducing memory 2.6x and boosting throughput up to 3.47x with near-identical quality on Llama, Falcon, and Mistral.
-
Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs
FastGen adaptively compresses LLM KV caches via lightweight attention profiling: evicting long-range contexts on local heads, non-special tokens on special-token heads, and retaining full caches on broad-attention heads, yielding substantial memory savings with negligible quality loss.
-
GSRQ: Gain-Shape Residual Quantization for Sub-1-bit KV Cache
GSRQ applies a gain-shape variant of K-means inside residual quantization to improve directional fidelity, raising LongBench accuracy from 11.34 to 33.54 at 1-bit on LLaMA-3-8B.
-
Position: LLM Inference Should Be Evaluated as Energy-to-Token Production
LLM inference should be reframed and evaluated as energy-to-token production with a Token Production Function that accounts for power, cooling, and efficiency ceilings.
-
KV-RM: Regularizing KV-Cache Movement for Static-Graph LLM Serving
KV-RM regularizes KV-cache movement via block paging and coalesced transfers to improve throughput, tail latency, and memory efficiency in static-graph LLM serving without changing the decoder interface.
-
StreamIndex: Memory-Bounded Compressed Sparse Attention via Streaming Top-k
Chunked streaming top-k enables CSA indexer execution at 1M sequence length with 6.21 GB peak memory and >=0.998 recall on synthetic V4-shaped inputs.
-
ShadowNPU: System and Algorithm Co-design for NPU-Centric On-Device LLM Inference
ShadowNPU presents shadowAttn, a co-designed sparse attention system that uses NPU pilot compute and techniques like graph bucketing and per-head sparsity to minimize CPU/GPU fallback during on-device LLM inference while maintaining accuracy.
-
Meta-Soft: Leveraging Composable Meta-Tokens for Context-Preserving KV Cache Compression
Meta-Soft dynamically synthesizes targeted soft tokens from a learnable meta-library using Gumbel-Softmax and applies attention-flow integration to compress KV cache while attempting to preserve evicted context information.
-
Protection Is (Nearly) All You Need: Structural Protection Dominates Scoring in Globally Capped KV Eviction
Structural protection of boundary tokens in globally capped KV cache eviction recovers 69-90% of full-cache quality at 13% retention and dominates differences among scoring policies.
-
Token-Operations-Oriented Inference Optimization Techniques for Large Models
A survey of large-model inference optimization, organized as a four-layer 'token-operations' taxonomy: multi-model fusion, model optimization, compute-model fusion, and compute-network-model fusion.
-
Personal LLM Agents: Insights and Survey about the Capability, Efficiency and Security
This survey discusses key components and challenges for Personal LLM Agents and reviews solutions for their capability, efficiency, and security.
-
The Hitchhiker's Guide to Agentic AI: From Foundations to Systems
A survey-style reference book mapping the full agentic-AI stack from transformer internals to production deployment, with no new research result.
- Fractal KV-Cache Archives: Lossless Symbolic Storage with In-Place Retrieval for Long-Context LLM Inference
- Alignment Collapse Under KV Cache Quantization: Diagnosis and Mitigation
- HieraSparse: Hierarchical Semi-Structured Sparse KV Attention