ε-coresets for attention exist of size O(√d e^{ρ+o(ρ)}/ε) for unit-norm keys/values and queries of norm ≤ρ, nearly matching the Ω(√d e^ρ/ε) lower bound.
super hub Mixed citations
Fast Transformer Decoding: One Write-Head is All You Need
Mixed citation behavior. Most common role is background (67%).
109
Pith papers citing it
Background
67% of classified citations
abstract
Multi-head attention layers, as used in the Transformer neural sequence model, are a powerful alternative to RNNs for moving information across and between sequences. While training these layers is generally fast and simple, due to parallelizability across the length of the sequence, incremental inference (where such paralleization is impossible) is often slow, due to the memory-bandwidth cost of repeatedly loading the large "keys" and "values" tensors. We propose a variant called multi-query attention, where the keys and values are shared across all of the different attention "heads", greatly reducing the size of these tensors and hence the memory bandwidth requirements of incremental decoding. We verify experimentally that the resulting models can indeed be much faster to decode, and incur only minor quality degradation from the baseline.
hub tools
citation-role summary
background 17
method 6
dataset 1
citation-polarity summary
claims ledger
- abstract Multi-head attention layers, as used in the Transformer neural sequence model, are a powerful alternative to RNNs for moving information across and between sequences. While training these layers is generally fast and simple, due to parallelizability across the length of the sequence, incremental inference (where such paralleization is impossible) is often slow, due to the memory-bandwidth cost of repeatedly loading the large "keys" and "values" tensors. We propose a variant called multi-query attention, where the keys and values are shared across all of the different attention "heads", greatly
authors
co-cited works
representative citing papers
Prism Transformer uses a progressive increase in attention heads with depth to form a local-to-global hierarchy, reporting lower validation loss and gains on zero-shot benchmarks versus uniform-head baselines at 124M-757M scales.
CrossPool separates weights and KV-cache into distinct GPU pools plus a planner, virtualizer, and layer-wise scheduler to cut P99 time-between-tokens by up to 10.4x versus prior kvcached multi-LLM systems.
Under a polynomial context-truncation sensitivity assumption, suffix-only KV cache policies require per-token memory scaling as Θ(ε^{-1/α}) to achieve distortion ε.
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.
LLMForge is a NAS framework with Infinite-Head Attention, a Forge-Former surrogate, and Forge-DSE engine that discovers hardware-specific architectures for edge language models, yielding variants with improved accuracy, energy, or latency on different substrates.
GQLA exposes dual MQA-absorb and GQA decoding paths from identical parameters to enable hardware-adaptive LLM inference while preserving cache compression on one path and GQA-level traffic on the other.
RPA kernel for TPUs achieves 86% MBU in decode and 73% MFU in prefill on Llama 3 8B via tiling for ragged memory, fused pipelines, and specialized compilation for prefill/decode workloads.
HormoneT5 augments T5 with a hormone-inspired block that predicts six continuous emotion values and uses them to modulate responses, reporting over 85% per-hormone accuracy and human preference for emotional quality.
MMEE encodes dataflow decisions in matrix form for fast exhaustive search, delivering 40-69% lower latency and energy use than prior methods while running 64-343x faster.
SCOPE accelerates autoregressive video diffusion up to 4.73x by using a tri-modal cache-predict-recompute scheduler with Taylor extrapolation and selective active-frame computation while preserving output quality.
One of the Q, K or V weights in transformer self-attention is redundant and replaceable by the identity matrix under mild assumptions, reducing parameters by 25 percent with no loss in small-model performance.
DELTA partitions layers into full, delta, and sparse groups to select salient tokens via aggregated attention scores, matching full-attention accuracy on AIME and GPQA while cutting attended tokens up to 4.25x and achieving 1.54x speedup.
FlashAttention-3 achieves 1.5-2x speedup on H100 GPUs for attention, reaching 740 TFLOPs/s (75% utilization) in FP16 and near 1.2 PFLOPs/s in FP8 while cutting numerical error by 2.6x versus baseline FP8 attention.
Transformers and SSMs are unified through structured state space duality, producing a 2-8X faster Mamba-2 model that remains competitive with Transformers.
DeepSeek-V2 delivers top-tier open-source LLM performance using only 21B active parameters by compressing the KV cache 93.3% and cutting training costs 42.5% via MLA and DeepSeekMoE.
Griffin hybrid model matches Llama-2 performance while trained on over 6 times fewer tokens and offers lower inference latency with higher throughput.
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.
A new shared video-image tokenizer enables large language models to surpass diffusion models on standard visual generation benchmarks.
Speculative sampling accelerates LLM decoding 2-2.5x by letting a draft model propose short sequences that the target model scores in parallel, then applies modified rejection sampling to keep the exact target distribution.
Speculative decoding accelerates exact sampling from large autoregressive models by 2-3x on T5-XXL by running smaller approximation models in parallel to propose token sequences that the large model then verifies in batches while preserving the original output distribution.
KernelSight-LM simulates token-level LLM inference to predict per-kernel latencies and end-to-end metrics (TTFT, TPOT, throughput) with 12.1% and 3.8% kernel errors in cross-generation and target-measured tiers.
Experiments across code LLMs show no-review collapses fastest, human-gated filters slow collapse, and AI self-gates lose effect over time, degenerating to ungated self-training under self-confirming acceptance as proven via gated distributional reweighting and spectral analysis.
RedKnot decomposes the KV cache by attention heads to enable position-independent reuse, prefix compression, hot/cold separation, and distributed placement for long-context LLM serving without model changes.
citing papers explorer
-
Nearly Optimal Attention Coresets
ε-coresets for attention exist of size O(√d e^{ρ+o(ρ)}/ε) for unit-norm keys/values and queries of norm ≤ρ, nearly matching the Ω(√d e^ρ/ε) lower bound.
-
Prism Transformer: Progressive Head Schedules for Hierarchical Attention Processing
Prism Transformer uses a progressive increase in attention heads with depth to form a local-to-global hierarchy, reporting lower validation loss and gains on zero-shot benchmarks versus uniform-head baselines at 124M-757M scales.
-
CrossPool: Efficient Multi-LLM Serving for Cold MoE Models through KV-Cache and Weight Disaggregation
CrossPool separates weights and KV-cache into distinct GPU pools plus a planner, virtualizer, and layer-wise scheduler to cut P99 time-between-tokens by up to 10.4x versus prior kvcached multi-LLM systems.
-
Polynomial Context-Truncation Sensitivity in Autoregressive Language Models: Sequential Wyner-Ziv Bounds for KV Cache Compression
Under a polynomial context-truncation sensitivity assumption, suffix-only KV cache policies require per-token memory scaling as Θ(ε^{-1/α}) to achieve distortion ε.
-
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.
-
LLMForge: Multi-Backend Hardware-Aware Neural Architecture Search with Infinite-Head Attention for Edge Language Models
LLMForge is a NAS framework with Infinite-Head Attention, a Forge-Former surrogate, and Forge-DSE engine that discovers hardware-specific architectures for edge language models, yielding variants with improved accuracy, energy, or latency on different substrates.
-
GQLA: Group-Query Latent Attention for Hardware-Adaptive Large Language Model Decoding
GQLA exposes dual MQA-absorb and GQA decoding paths from identical parameters to enable hardware-adaptive LLM inference while preserving cache compression on one path and GQA-level traffic on the other.
-
Ragged Paged Attention: A High-Performance and Flexible LLM Inference Kernel for TPU
RPA kernel for TPUs achieves 86% MBU in decode and 73% MFU in prefill on Llama 3 8B via tiling for ragged memory, fused pipelines, and specialized compilation for prefill/decode workloads.
-
A Hormone-inspired Emotion Layer for Transformer language models (HELT)
HormoneT5 augments T5 with a hormone-inspired block that predicts six continuous emotion values and uses them to modulate responses, reporting over 85% per-hormone accuracy and human preference for emotional quality.
-
Fast Cross-Operator Optimization of Attention Dataflow
MMEE encodes dataflow decisions in matrix form for fast exhaustive search, delivering 40-69% lower latency and energy use than prior methods while running 64-343x faster.
-
Not All Frames Deserve Full Computation: Accelerating Autoregressive Video Generation via Selective Computation and Predictive Extrapolation
SCOPE accelerates autoregressive video diffusion up to 4.73x by using a tri-modal cache-predict-recompute scheduler with Taylor extrapolation and selective active-frame computation while preserving output quality.
-
Key and Value Weights Are Probably All You Need: On the Necessity of the Query, Key, Value weight Triplet in Self-Attention Transformers
One of the Q, K or V weights in transformer self-attention is redundant and replaceable by the identity matrix under mild assumptions, reducing parameters by 25 percent with no loss in small-model performance.
-
DELTA: Dynamic Layer-Aware Token Attention for Efficient Long-Context Reasoning
DELTA partitions layers into full, delta, and sparse groups to select salient tokens via aggregated attention scores, matching full-attention accuracy on AIME and GPQA while cutting attended tokens up to 4.25x and achieving 1.54x speedup.
-
FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision
FlashAttention-3 achieves 1.5-2x speedup on H100 GPUs for attention, reaching 740 TFLOPs/s (75% utilization) in FP16 and near 1.2 PFLOPs/s in FP8 while cutting numerical error by 2.6x versus baseline FP8 attention.
-
Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality
Transformers and SSMs are unified through structured state space duality, producing a 2-8X faster Mamba-2 model that remains competitive with Transformers.
-
DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
DeepSeek-V2 delivers top-tier open-source LLM performance using only 21B active parameters by compressing the KV cache 93.3% and cutting training costs 42.5% via MLA and DeepSeekMoE.
-
Griffin: Mixing Gated Linear Recurrences with Local Attention for Efficient Language Models
Griffin hybrid model matches Llama-2 performance while trained on over 6 times fewer tokens and offers lower inference latency with higher throughput.
-
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.
-
Language Model Beats Diffusion -- Tokenizer is Key to Visual Generation
A new shared video-image tokenizer enables large language models to surpass diffusion models on standard visual generation benchmarks.
-
Accelerating Large Language Model Decoding with Speculative Sampling
Speculative sampling accelerates LLM decoding 2-2.5x by letting a draft model propose short sequences that the target model scores in parallel, then applies modified rejection sampling to keep the exact target distribution.
-
Fast Inference from Transformers via Speculative Decoding
Speculative decoding accelerates exact sampling from large autoregressive models by 2-3x on T5-XXL by running smaller approximation models in parallel to propose token sequences that the large model then verifies in batches while preserving the original output distribution.
-
KernelSight-LM: A Kernel-Level LLM Inference Simulator
KernelSight-LM simulates token-level LLM inference to predict per-kernel latencies and end-to-end metrics (TTFT, TPOT, throughput) with 12.1% and 3.8% kernel errors in cross-generation and target-measured tiers.
-
When AI Reviews Its Own Code: Recursive Self-Training Collapse in Code LLMs
Experiments across code LLMs show no-review collapses fastest, human-gated filters slow collapse, and AI self-gates lose effect over time, degenerating to ungated self-training under self-confirming acceptance as proven via gated distributional reweighting and spectral analysis.
-
RedKnot: Efficient Long-Context LLM Serving with Head-Aware KV Reuse and SegPagedAttention
RedKnot decomposes the KV cache by attention heads to enable position-independent reuse, prefix compression, hot/cold separation, and distributed placement for long-context LLM serving without model changes.
-
Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories
Language models can use a two-stage sleep process of upward distillation for memory consolidation and RL-based dreaming for unsupervised self-improvement to enable continual learning.
-
Dynamic Short Convolutions Improve Transformers
Dynamic short convolutions applied to key/query/value projections and linear layers in Transformers yield consistent performance gains and 1.33-1.60x compute advantages over standard models on language modeling from 150M to 2B parameters.
-
Do Transformers Need Three Projections? Systematic Study of QKV Variants
Q-K=V projection sharing in transformers matches standard QKV performance with 50% KV cache reduction and combines with GQA/MQA for up to 96.9% reduction across vision and language tasks.
-
Do Value Vectors in Deep Layers Need Context from the Residual Stream?
Deeper transformer layers benefit from context-free token-specific value vectors in a Bank of Values lookup table, improving performance over standard attention with less compute.
-
Idleness is Relative: Exploiting Tool-Call Idle Windows for Offloading in Agentic Systems with MORI
MORI improves throughput 20-71% and TTFT 18-43% over baselines by ranking programs on a continuous idleness spectrum and shifting the GPU-CPU boundary to match capacity in agentic LLM serving.
-
DREAM-S: Speculative Decoding with Searchable Drafting and Target-Aware Refinement for Multimodal Generation
DREAM-S combines neural architecture search, target-aware supernet training, and attention-entropy-guided distillation to accelerate speculative decoding in VLMs, reporting up to 3.85x speedup over standard methods.
-
BlockBatch: Multi-Scale Consensus Decoding for Efficient Diffusion Language Model Inference
BlockBatch is a training-free framework that coordinates multiple block-size branches via token merging and synchronization to reduce denoising NFEs by 26.6% and achieve 1.33x speedup in dLLM inference.
-
Tensor Memory: Fixed-Size Recurrent State for Long-Horizon Transformers
Tensor Memory augments Transformers with a constant-size 3D voxel grid using differentiable soft writes at predicted locations, local interaction, and gated recurrent dynamics to decouple memory capacity from sequence length.
-
Optimus: Elastic Decoding for Efficient Diffusion LLM Serving
Optimus enables elastic decoding granularity adaptation in diffusion LLMs via chunked decoding and load-based scheduling to raise throughput under dynamic conditions.
-
Motion-Compensated Weight Compression
MCWC aligns permutation-symmetric blocks across layers to enable sequential prediction and residual entropy coding, improving rate-accuracy tradeoffs versus quantization and prior codecs on language and vision models.
-
Interdomain Attention: Beyond Token-Level Key-Value Memory
Interdomain Attention integrates SSMs into attention via finite feature maps and basis projections to enable query-conditioned attention over fixed states, showing gains over SSM baselines and matching softmax at 1.3B scale with length-flat scaling.
-
ArborKV: Structure-Aware KV Cache Management for Scaling Tree-based LLM Reasoning
ArborKV uses search-structure awareness to evict low-reuse KV states in Tree-of-Thoughts inference, delivering up to 4x memory savings with near-full accuracy retention.
-
ObjectCache: Layerwise Object-Storage Retrieval for KV Cache Reuse
ObjectCache enables KV cache storage in object storage via layerwise retrieval and custom scheduling, adding 5.6% latency for 64K contexts over local DRAM on a 100 Gbps RoCE cluster.
-
LRCP: Low-Rank Compressibility Guided Visual Token Pruning for Efficient LVLMs
LRCP prunes visual tokens in LVLMs by scoring projection residuals onto a PCA-estimated low-rank subspace, achieving 88.9% image token reduction with 94.7% performance retention and 87.5% video reduction with 97.8% accuracy retention.
-
When Attention Closes: How LLMs Lose the Thread in Multi-Turn Interaction
Attention to goal tokens declines in multi-turn LLM interactions while residual representations often retain decodable goal information, and the gap between these predicts whether goal-conditioned behavior survives.
-
Search Your Block Floating Point Scales!
ScaleSearch optimizes block floating point scales via fine-grained search to cut quantization error by 27% for NVFP4, improving PTQ by up to 15 points on MATH500 for Qwen3-8B and attention PPL by 0.77 on Llama 3.1 70B.
-
Nectar: Neural Estimation of Cached-Token Attention via Regression
Nectar fits small per-layer per-head neural networks via regression to predict attention outputs and normalizers, enabling constant-time inference independent of context length while preserving semantic generation quality.
-
Memory-Efficient Looped Transformer: Decoupling Compute from Memory in Looped Language Models
MELT decouples reasoning depth from memory in looped language models by sharing a single gated KV cache per layer and training it via chunk-wise distillation from Ouro starting models.
-
Revisiting Transformer Layer Parameterization Through Causal Energy Minimization
CEM recasts Transformer layers as energy minimization steps, enabling constrained parameterizations like weight sharing and low-rank interactions that match standard baselines in 100M-scale language modeling.
-
ZAYA1-8B Technical Report
ZAYA1-8B is a reasoning MoE model with 700M active parameters that matches larger models on math and coding benchmarks and reaches 91.9% on AIME'25 via Markovian RSA test-time compute.
-
The Impossibility Triangle of Long-Context Modeling
No model can achieve efficiency, compactness, and recall capacity scaling with sequence length at once, as any two imply a strict bound of O(poly(d)/log V) on recallable facts.
-
WindowQuant: Mixed-Precision KV Cache Quantization based on Window-Level Similarity for VLMs Inference Optimization
WindowQuant performs window-adaptive mixed-precision KV cache quantization guided by similarity to the text prompt, with reordering to enable efficient inference in VLMs.
-
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.
-
Graph-Guided Adaptive Channel Elimination for KV Cache Compression
GRACE reframes KV cache channel pruning as graph optimization to find a near-optimal subset, achieving 60% compression with negligible degradation and outperforming prior methods.
-
Open-TQ-Metal: Fused Compressed-Domain Attention for Long-Context LLM Inference on Apple Silicon
Fused compressed-domain int4 attention on Apple Silicon delivers 48x speedup and 3.2x KV cache compression for 128K-context 70B models while matching FP16 token predictions.
-
The Illusion of Equivalence: Systematic FP16 Divergence in KV-Cached Autoregressive Inference
FP16 KV caching in transformers causes deterministic token divergence versus cache-free inference due to non-associative floating-point accumulation orderings.