Rigel reverse-engineers the Metal 4.1 tensor compute path on M4 Max, finding fp8 matmul2d is emulated on GPU shader cores at 0.94x fp16 throughput with an 8x8 fragment layout and no ANE involvement.
hub Canonical reference
FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness
Canonical reference. 94% of citing Pith papers cite this work as background.
69
Pith papers citing it
Background
94% of classified citations
abstract
Transformers are slow and memory-hungry on long sequences, since the time and memory complexity of self-attention are quadratic in sequence length. Approximate attention methods have attempted to address this problem by trading off model quality to reduce the compute complexity, but often do not achieve wall-clock speedup. We argue that a missing principle is making attention algorithms IO-aware -- accounting for reads and writes between levels of GPU memory. We propose FlashAttention, an IO-aware exact attention algorithm that uses tiling to reduce the number of memory reads/writes between GPU high bandwidth memory (HBM) and GPU on-chip SRAM. We analyze the IO complexity of FlashAttention, showing that it requires fewer HBM accesses than standard attention, and is optimal for a range of SRAM sizes. We also extend FlashAttention to block-sparse attention, yielding an approximate attention algorithm that is faster than any existing approximate attention method. FlashAttention trains Transformers faster than existing baselines: 15% end-to-end wall-clock speedup on BERT-large (seq. length 512) compared to the MLPerf 1.1 training speed record, 3$\times$ speedup on GPT-2 (seq. length 1K), and 2.4$\times$ speedup on long-range arena (seq. length 1K-4K). FlashAttention and block-sparse FlashAttention enable longer context in Transformers, yielding higher quality models (0.7 better perplexity on GPT-2 and 6.4 points of lift on long-document classification) and entirely new capabilities: the first Transformers to achieve better-than-chance performance on the Path-X challenge (seq. length 16K, 61.4% accuracy) and Path-256 (seq. length 64K, 63.1% accuracy).
hub tools
citation-role summary
background 15
method 1
citation-polarity summary
representative citing papers
RoundPipe achieves near-zero-bubble pipeline parallelism for LLM training on consumer GPUs by dynamically dispatching computation stages round-robin, yielding 1.48-2.16x speedups and enabling 235B model fine-tuning on 8x RTX 4090.
Tessera performs kernel-granularity disaggregation on heterogeneous GPUs, achieving up to 2.3x throughput and 1.6x cost efficiency gains for large model inference while generalizing beyond prior methods.
Pythia releases 16 identically trained LLMs with full checkpoints and data tools to study training dynamics, scaling, memorization, and bias in language models.
Garment Particles is a 5D point cloud representation jointly encoding 2D sewing patterns and 3D geometry, supporting rectified flow generation from high-level inputs and diffusion-based editing of patterns or shapes.
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.
LlamaWeb is a WebGPU backend for llama.cpp that uses static memory planning, tunable kernels, and templated multi-precision support to cut memory use by 29-33% and raise decode throughput by 45-69% versus prior browser frameworks on tested hardware.
PilotWiMAE pretrains an encoder on noisy pilots with factorized attention, 99% masking, patch-normalized reconstruction, scale loss, and AWGN curriculum to outperform supervised baselines in cross-frequency beam selection and channel tasks from 3.5 GHz pretraining to 28 GHz evaluation.
CUDAHercules benchmark demonstrates that leading LLMs generate functional CUDA code but fail to recover expert-level optimization strategies needed for peak performance on Ampere, Hopper, and Blackwell GPUs.
Kerncap automates extraction of faithful, self-contained GPU kernel reproducers from AMD HIP and Triton workloads via HSA interception and address-space closure, delivering 13.6x faster isolated tuning.
Gaussian Kernel Attention replaces learned QKV projections with a Gaussian RBF kernel on per-head token features, using 0.42x parameters and 0.49x FLOPs while showing competitive language modeling performance at depth 20.
CacheFlow cuts TTFT by 10-62% in batched LLM serving via 3D-parallel KV cache restoration and a two-pointer scheduler that overlaps recompute and I/O.
Transactional Attention uses semantic sponsorship from anchor patterns to retain dormant critical tokens in KV caches, achieving 100% credential retrieval at 16 tokens where all prior methods fail.
FFM finds optimal fused mappings for tensor accelerators over 10,000 times faster than prior mappers while cutting energy-delay product by up to 1.8x versus hand-tuned designs.
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.
A recurrent-depth architecture enables language models to improve reasoning performance by iterating computation in latent space, achieving gains equivalent to much larger models on benchmarks.
DuoAttention identifies retrieval heads requiring full KV cache and streaming heads using constant-length cache to reduce memory and latency in long-context LLM inference.
GPTQ quantizes 175B-parameter GPT models to 3-4 bits per weight in one shot using approximate second-order information, achieving negligible accuracy degradation and 3-4x inference speedups.
ABot-M0.5 proposes a unified mobility-and-manipulation world action model using three alignment strategies that achieves state-of-the-art performance on mobile and fine-grained manipulation benchmarks.
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.
MoA framework derives a denotational normal form for attention that eliminates all intermediate arrays by algebraic construction, yielding O(n_dk + n_dv) memory traffic with numerical verification against PyTorch.
TRI trains LLMs on goal-conditioned fill-in-the-middle tasks via PSM token rearrangement and symbolic verification to surgically repair erroneous CoT segments.
KForge uses dual LLM agents for cross-platform kernel generation, reporting 2.12% throughput gain on NVIDIA B200 vs TensorRT-LLM and 5.13x geometric mean speedup on Intel Arc B580 vs PyTorch on 37 workloads.
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.
citing papers explorer
-
Rigel: Reverse-Engineering the Metal 4.1 Tensor Compute Path on the Apple M4 Max GPU
Rigel reverse-engineers the Metal 4.1 tensor compute path on M4 Max, finding fp8 matmul2d is emulated on GPU shader cores at 0.94x fp16 throughput with an 8x8 fragment layout and no ANE involvement.
-
Efficient Training on Multiple Consumer GPUs with RoundPipe
RoundPipe achieves near-zero-bubble pipeline parallelism for LLM training on consumer GPUs by dynamically dispatching computation stages round-robin, yielding 1.48-2.16x speedups and enabling 235B model fine-tuning on 8x RTX 4090.
-
Tessera: Unlocking Heterogeneous GPUs through Kernel-Granularity Disaggregation
Tessera performs kernel-granularity disaggregation on heterogeneous GPUs, achieving up to 2.3x throughput and 1.6x cost efficiency gains for large model inference while generalizing beyond prior methods.
-
Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling
Pythia releases 16 identically trained LLMs with full checkpoints and data tools to study training dynamics, scaling, memorization, and bias in language models.
-
Garment Particles: A 2D--3D Symmetric Garment Representation for Generation and Editing
Garment Particles is a 5D point cloud representation jointly encoding 2D sewing patterns and 3D geometry, supporting rectified flow generation from high-level inputs and diffusion-based editing of patterns or shapes.
-
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.
-
Llamas on the Web: Memory-Efficient, Performance-Portable, and Multi-Precision LLM Inference with WebGPU
LlamaWeb is a WebGPU backend for llama.cpp that uses static memory planning, tunable kernels, and templated multi-precision support to cut memory use by 29-33% and raise decode throughput by 45-69% versus prior browser frameworks on tested hardware.
-
PilotWiMAE: Pilot-Native Representation Learning for Wireless Channels
PilotWiMAE pretrains an encoder on noisy pilots with factorized attention, 99% masking, patch-normalized reconstruction, scale loss, and AWGN curriculum to outperform supervised baselines in cross-frequency beam selection and channel tasks from 3.5 GHz pretraining to 28 GHz evaluation.
-
CUDAHercules: Benchmarking Hardware-Aware Expert-level CUDA Optimization for LLMs
CUDAHercules benchmark demonstrates that leading LLMs generate functional CUDA code but fail to recover expert-level optimization strategies needed for peak performance on Ampere, Hopper, and Blackwell GPUs.
-
Kerncap: Automated Kernel Extraction and Isolation for AMD GPUs
Kerncap automates extraction of faithful, self-contained GPU kernel reproducers from AMD HIP and Triton workloads via HSA interception and address-space closure, delivering 13.6x faster isolated tuning.
-
Projection-Free Transformers via Gaussian Kernel Attention
Gaussian Kernel Attention replaces learned QKV projections with a Gaussian RBF kernel on per-head token features, using 0.42x parameters and 0.49x FLOPs while showing competitive language modeling performance at depth 20.
-
CacheFlow: Efficient LLM Serving with 3D-Parallel KV Cache Restoration
CacheFlow cuts TTFT by 10-62% in batched LLM serving via 3D-parallel KV cache restoration and a two-pointer scheduler that overlaps recompute and I/O.
-
Transactional Attention: Semantic Sponsorship for KV-Cache Retention
Transactional Attention uses semantic sponsorship from anchor patterns to retain dormant critical tokens in KV caches, achieving 100% credential retrieval at 16 tokens where all prior methods fail.
-
Fast and Fusiest: An Optimal Fusion-Aware Mapper for Accelerator Design
FFM finds optimal fused mappings for tensor accelerators over 10,000 times faster than prior mappers while cutting energy-delay product by up to 1.8x versus hand-tuned designs.
-
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.
-
Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
A recurrent-depth architecture enables language models to improve reasoning performance by iterating computation in latent space, achieving gains equivalent to much larger models on benchmarks.
-
DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads
DuoAttention identifies retrieval heads requiring full KV cache and streaming heads using constant-length cache to reduce memory and latency in long-context LLM inference.
-
GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers
GPTQ quantizes 175B-parameter GPT models to 3-4 bits per weight in one shot using approximate second-order information, achieving negligible accuracy degradation and 3-4x inference speedups.
-
ABot-M0.5: Unified Mobility-and-Manipulation World Action Model
ABot-M0.5 proposes a unified mobility-and-manipulation world action model using three alignment strategies that achieves state-of-the-art performance on mobile and fine-grained manipulation benchmarks.
-
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.
-
Attention at the Theoretical Minimum: A Mathematics of Arrays Framework for Memory-Optimal Transformer Kernels
MoA framework derives a denotational normal form for attention that eliminates all intermediate arrays by algebraic construction, yielding O(n_dk + n_dv) memory traffic with numerical verification against PyTorch.
-
Imbuing Large Language Models with Bidirectional Logic for Robust Chain Repair
TRI trains LLMs on goal-conditioned fill-in-the-middle tasks via PSM token rearrangement and symbolic verification to surgically repair erroneous CoT segments.
-
KForge: LLM-Driven Cross-Platform Kernel Generation for AI Accelerators
KForge uses dual LLM agents for cross-platform kernel generation, reporting 2.12% throughput gain on NVIDIA B200 vs TensorRT-LLM and 5.13x geometric mean speedup on Intel Arc B580 vs PyTorch on 37 workloads.
-
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.
-
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.
-
HyperParallel-MoE: Multi-Core Interleaved Scheduling for Fast MoE Training on Ascend NPUs
HyperParallel-MoE reduces Dispatch-to-Combine MoE-FFN latency by up to 1.58x on Ascend A3 clusters via tile-level heterogeneous scheduling that overlaps communication, matrix, and vector computation inside a single kernel launch.
-
Adaptive Mass-Segmented KV Compression for Long-Context Reasoning
AMS KV compression adaptively partitions the cache by attention mass regions and assigns quotas to protect contiguous reasoning blocks during long-context LLM inference.
-
Towards Understanding Self-Pretraining for Sequence Classification
Self-pretraining improves Transformer sequence classification by enabling learning of proximity-biased attention from positional encodings that label supervision alone cannot easily acquire from random starts.
-
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.
-
Remember to Forget: Gated Adaptive Positional Encoding
GAPE augments RoPE with query- and key-dependent gates to stabilize attention and improve long-context performance in language models.
-
Dooly: Configuration-Agnostic, Redundancy-Aware Profiling for LLM Inference Simulation
Dooly reduces LLM inference profiling GPU-hours by 56.4% across 12 models while keeping simulation MAPE under 5% for TTFT and 8% for TPOT by making profiling configuration-agnostic and redundancy-aware.
-
Piper: Efficient Large-Scale MoE Training via Resource Modeling and Pipelined Hybrid Parallelism
Piper introduces resource modeling and pipelined hybrid parallelism for MoE training, delivering 2-3.5X higher MFU than prior frameworks and 1.2-9X better all-to-all bandwidth.
-
Leveraging LLMs for Multi-File DSL Code Generation: An Industrial Case Study
Fine-tuning 7B code LLMs on a custom multi-file DSL dataset achieves structural fidelity of 1.00, high exact-match accuracy, and practical utility validated by expert survey and execution checks.
-
HubRouter: A Pluggable Sub-Quadratic Routing Primitive for Hybrid Sequence Models
HubRouter is a sub-quadratic routing primitive using learned hubs that replaces attention layers in hybrid models while delivering competitive perplexity and large throughput gains.
-
HypEHR: Hyperbolic Modeling of Electronic Health Records for Efficient Question Answering
HypEHR is a hyperbolic embedding model for EHR data that uses Lorentzian geometry and hierarchy-aware pretraining to answer clinical questions nearly as well as large language models but with much smaller size.
-
TiledAttention: a CUDA Tile SDPA Kernel for PyTorch
TiledAttention is a cuTile-based SDPA kernel that balances performance with Python-level customizability for attention research in PyTorch.
-
SweetSpot: An Analytical Model for Predicting Energy Efficiency of LLM Inference
SweetSpot is an analytical model from Transformer computational and memory complexity that identifies energy minima at short-to-moderate inputs and medium outputs, achieving 1.79% MAPE on H100 GPU measurements across multiple LLMs.
-
Flashlight: PyTorch Compiler Extensions to Accelerate Attention Variants
Flashlight is a compiler-native PyTorch framework that generates efficient fused kernels for arbitrary and data-dependent attention variants, supporting more cases than FlexAttention with competitive performance.
-
OjaKV: Context-Aware Online Low-Rank KV Cache Compression
OjaKV introduces hybrid full-rank storage for key tokens combined with online low-rank KV cache compression via Oja's algorithm to support memory-efficient long-context LLM inference.
-
Amoeba: Runtime Tensor Parallel Transformation for LLM Inference Services
Amoeba adaptively adjusts tensor parallelism at runtime for LLM inference services to handle mixed short and long context requests, delivering 1.75x-6.57x throughput gains over prior solutions in real-world trace evaluations.
-
Positional Encoding via Token-Aware Phase Attention
TAPA adds a learnable phase function to attention to preserve long-range token interactions, enabling direct continual pretraining, length extrapolation, lower perplexity, and stronger retrieval than RoPE-style methods.
-
Hits to Higgs: Hit-Level Higgs Classification from Raw LHC Detector Data Using Higgsformer
Higgsformer achieves AUC 0.855 on t tbar H vs t tbar classification from raw hits, matching a Delphes-based Particle Transformer at ~40% b-tagging efficiency.
-
TabICL: A Tabular Foundation Model for In-Context Learning on Large Data
TabICL scales in-context learning to large tabular data via column-then-row attention for row embeddings followed by a transformer, matching TabPFNv2 speed and performance while outperforming it and CatBoost on datasets over 10K samples.
-
Deep Learning Alternatives of the Kolmogorov Superposition Theorem
ActNet is a new KST-based neural network that outperforms KANs and competes with MLPs in PINN benchmarks for PDE simulation tasks.
-
FlashNorm: Fast Normalization for Transformers
FlashNorm is an exact algebraic reformulation of RMSNorm plus linear projection that folds weights and defers normalization to allow parallel execution, plus scale-invariance simplifications that remove redundant norms in certain architectures.
-
RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval
RAPTOR introduces a tree-organized retrieval method using recursive abstractive summaries, achieving a 20% absolute accuracy improvement on the QuALITY benchmark when paired with GPT-4.
-
Efficient Streaming Language Models with Attention Sinks
StreamingLLM lets finite-window LLMs generalize to infinite-length sequences by retaining initial-token KV states as attention sinks, enabling stable streaming inference up to 4M tokens.
-
GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints
Uptraining multi-head transformer checkpoints to grouped-query attention models achieves near multi-head quality at multi-query inference speeds using 5% additional compute.
-
Token Merging: Your ViT But Faster
Token Merging (ToMe) doubles the throughput of large Vision Transformers on images, video, and audio by merging similar tokens with a fast matching algorithm, incurring only 0.2-0.4% accuracy loss.
-
Simplified State Space Layers for Sequence Modeling
S5 uses a single MIMO state space model with S4-derived initialization to match S4 efficiency and reach 87.4% average accuracy on the Long Range Arena benchmark.