惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

量
量子位
F
Fortinet All Blogs
J
Java Code Geeks
Y
Y Combinator Blog
Stack Overflow Blog
Stack Overflow Blog
V
Visual Studio Blog
M
MIT News - Artificial intelligence
腾
腾讯CDC
Last Week in AI
Last Week in AI
The Cloudflare Blog
酷 壳 – CoolShell
酷 壳 – CoolShell
Jina AI
Jina AI
Microsoft Security Blog
Microsoft Security Blog
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报
P
Proofpoint News Feed
博
博客园 - 叶小钗
Recent Announcements
Recent Announcements
T
Tailwind CSS Blog
Engineering at Meta
Engineering at Meta
Cyber Security Advisories - MS-ISAC
Cyber Security Advisories - MS-ISAC
人人都是产品经理
人人都是产品经理
L
LangChain Blog
博
博客园 - 司徒正美
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻

cs.DC updates on arXiv.org

DUAL-BLADE: Dual-Path NVMe-Direct KV-Cache Offloading for Edge LLM Inference Progressive Semantic Communication for Efficient Edge-Cloud Vision-Language Models Efficient, VRAM-Constrained xLM Inference on Clients Folding Tensor and Sequence Parallelism for Memory-Efficient Transformer Training & Inference DORA: A Scalable Asynchronous Reinforcement Learning System for Language Model Training AMMA: A Multi-Chiplet Memory-Centric Architecture for Low-Latency 1M Context Attention Serving RaMP: Runtime-Aware Megakernel Polymorphism for Mixture-of-Experts Spark Policy Toolkit: Semantic Contracts and Scalable Execution for Policy Learning in Spark Internet of Everything in the 6G Era: Paradigms, Enablers, Potentials and Future Directions PolyKV: A Shared Asymmetrically-Compressed KV Cache Pool for Multi-Agent LLM Inference A Survey on Split Learning for LLM Fine-Tuning: Models, Systems, and Privacy Optimizations ITAS: A Multi-Agent Architecture for LLM-Based Intelligent Tutoring Latency and Cost of Multi-Agent Intelligent Tutoring at Scale TACO: Efficient Communication Compression of Intermediate Tensors for Scalable Tensor-Parallel LLM Training FreeScale: Distributed Training for Sequence Recommendation Models with Minimal Scaling Cost CommFuse: Hiding Tail Latency via Communication Decomposition and Fusion for Distributed LLM Training A Taxonomy and Resolution Strategy for Client-Level Disagreements in Federated Learning Usable Agent Discovery for Decentralized AI Systems Cloud to Edge: Benchmarking LLM Inference On Hardware-Accelerated Single-Board Computers Data-Free Contribution Estimation in Federated Learning using Gradient von Neumann Entropy Shard the Gradient, Scale the Model: Serverless Federated Aggregation via Gradient Partitioning Promoting Simple Agents: Ensemble Methods for Event-Log Prediction GraphLeap: Decoupling Graph Construction and Convolution for Vision GNN Acceleration on FPGA AGNT2: Autonomous Agent Economies on Interaction-Optimized Layer 2 Infrastructure FedSIR: Spectral Client Identification and Relabeling for Federated Learning with Noisy Labels Stream-CQSA: Avoiding Out-of-Memory in Attention Computation via Flexible Workload Scheduling A Delta-Aware Orchestration Framework for Scalable Multi-Agent Edge Computing Federated Learning over Blockchain-Enabled Cloud Infrastructure Optimal Routing for Federated Learning over Dynamic Satellite Networks: Tractable or Not? Sherpa.ai Privacy-Preserving Multi-Party Entity Alignment without Intersection Disclosure for Noisy Identifiers
A Case Study in CUDA Kernel Fusion: Implementing FlashAtt...
Ganesh Bikshandi, Jay Shah · 2023-12-19 · via cs.DC updates on arXiv.org

We provide an optimized implementation of the forward pass of FlashAttention-2, a popular memory-aware scaled dot-product attention algorithm, as a custom fused CUDA kernel targeting NVIDIA Hopper architecture and written using the open-source CUTLASS library. In doing so, we explain the challenges and techniques involved in fusing online-softmax with back-to-back GEMM kernels, utilizing the Hopper-specific Tensor Memory Accelerator (TMA) and Warpgroup Matrix-Multiply-Accumulate (WGMMA) instructions, defining and transforming CUTLASS Layouts and Tensors, overlapping copy and GEMM operations, and choosing optimal tile sizes for the Q, K and V attention matrices while balancing the register pressure and shared memory utilization. In head-to-head benchmarks on a single H100 PCIe GPU for some common choices of hyperparameters, we observe 20-50% higher FLOPs/s over a version of FlashAttention-2 optimized for last-generation NVIDIA Ampere architecture.

此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。