MLSys 2026 Schedule

Filter Events

SUN 17 MAY

2 p.m.

Registration Open

(ends 6:00 PM)

MON 18 MAY

8 a.m.

Registration

(ends 5:00 PM)

8:50 a.m.

Remarks:

Opening Remarks

(ends 9:00 AM)

9 a.m.

Invited Talk:

Rethinking Open Source Contribution in the Age of AI Agents

Roger Wang

(ends 9:25 AM)

9:25 a.m.

Invited Talk:

Beyond Model Serving: Cross-Stack Co-Design for Agentic Systems

Esha Choukse

(ends 9:50 AM)

9:50 a.m.

Invited Talk:

LMCache: An Efficient KV Cache Layer for Enterprise-Scale LLM Inference

Yuhan Liu

(ends 10:15 AM)

10:15 a.m.

Invited Talk:

Eliciting Language Model Behaviors with Investigator Agents

Lisa Li

(ends 10:40 AM)

10:40 a.m.

Break:

Coffee Break

(ends 11:00 AM)

11 a.m.

noon

Break:

Lunch Break

(ends 1:30 PM)

1:30 p.m.

Keynote Talk:

When AI Starts Writing Systems Code

Mark Saroufim

(ends 2:30 PM)

2:30 p.m.

4 p.m.

Panel:

Panel

(ends 5:00 PM)

5 p.m.

YPS Poster Session & YPS Reception [5:00-7:00]

(ends 7:00 PM)

TUE 19 MAY

8 a.m.

Registration

(ends 5:00 PM)

8:30 a.m.

Remarks:

Opening Remarks

(ends 8:45 AM)

8:45 a.m.

Best Paper Session [8:45-10:00]

Moderator: Dimitrios Skarlatos

StreamDiffusionV2: A Streaming System for Dynamic and Interactive Video Generation

LEANN: A Low-Storage Overhead Vector Index

BLASST: Dynamic BLocked Attention Sparsity via Softmax Thresholding

ExecuTorch - A Unified PyTorch Solution to Run ML Models On-Device

(ends 10:00 AM)

Overflow / Viewing [8:45-10:00]

(ends 10:00 AM)

10:30 a.m.

Keynote Talk:

The Next Horizon of Systems: From MLSys to System Intelligence

Lidong Zhou

(ends 11:30 AM)

Overflow / Viewing [10:30-11:30]

(ends 11:30 AM)

11:30 a.m.

Break:

Lunch Break

(ends 1:00 PM)

1 p.m.

Agentic AI 1 [1:00-2:15]

Moderator: Cheng Tan

OSWorld-Human: Benchmarking the Efficiency of Computer-Use Agents

VeriMoA: A Mixture-of-Agents Framework for Spec-to-HDL Generation

When Enough is Enough: Rank-Aware Early Termination for Vector Search

Optimizing PyTorch Inference with LLM-Based Multi-Agent Systems

Matrix: Peer-to-Peer Multi-Agent Synthetic Data Generation Framework

(ends 2:15 PM)

Multimodal and Generative Models [1:00-2:15]

Moderator: Haoran Qiu

G-HEMP: FAST MULTI-GPU PRIVATE INFERENCE FOR LARGE-SCALE GCNS WITH HOMOMORPHIC ENCRYPTION

db-SP: Accelerating Sparse Attention for Visual Generative Models with Dual-Balanced Sequence Parallelism

Flash3DGS: Algorithm and System Co-Optimization for Fast 3D Gaussian Splatting on GPUs

TriInfer: Hybrid EPD Disaggregation for Efficient Multimodal Large Language Model Inference

TiDAR: Think in Diffusion, Talk in Autoregression

(ends 2:15 PM)

2:30 p.m.

Break:

Coffee Break

(ends 3:00 PM)

Break:

Break

(ends 2:45 PM)

LLM Serving 1 [2:30-4:00]

Moderator: Esha Choukse

HELIOS : Adaptive Model And Early-Exit Selection for Efficient LLM Inference Serving

Stream2LLM: Overlap Context Streaming and Prefill for Reduced Time-to-First-Token

SuperInfer: SLO-Aware Rotary Scheduling and Memory Management for LLM Inference on Superchips

PLA-Serve: A Prefill-Length-Aware LLM Serving System

TokenBlend: Accelerating Tensor Parallelism LLM Inference Through Efficient Compute-Communication Overlap

From Tokens to Layers: Redefining Stall-Free Scheduling for LLM Serving with Layered Prefill

(ends 4:00 PM)

Agentic AI 2 [2:30-4:00]

Moderator: Yawen Wang

Hippocampus: An Efficient and Scalable Memory Module for Agentic AI

ContextPilot: Fast Long-Context Inference via Context Reuse

AgenticCache: Cache-Driven Asynchronous Planning for Embodied AI Agents

RagInfer: Efficient Retrieval-Augmented Generation Inference with Lookahead Retrieval

FlashAgents: Accelerating Multi-Agent LLM Systems via Streaming Prefill Overlap

Ontology-Guided Long-Term Agent Memory for Conversational RAG

(ends 4:00 PM)

4:15 p.m.

Break:

Break

(ends 4:30 PM)

4:30 p.m.

Security and Privacy [4:30-6:00]

Moderator: Yifan Qiao

CSLE: A Reinforcement Learning Platform for Autonomous Security Management

Privatar: Scalable Privacy-preserving Multi-user VR via Secure Offloading

ZK-APEX: ZERO-KNOWLEDGE APPROXIMATE PERSONALIZED UNLEARNING WITH EXECUTABLE PROOFS

Blueprint, Bootstrap, and Bridge: A Security Look at NVIDIA GPU Confidential Computing

Toward Principled LLM Safety Testing: Solving the Jailbreak Oracle Problem

(ends 6:00 PM)

Federated Learning [4:30-6:00]

Moderator: Runzhou Tao

SONAR: Benchmarking Topology and Collaboration in Decentralized Learning

ProToken: Token-Level Attribution for Federated Large Language Models

FLoRIST: Singular Value Thresholding for Efficient and Accurate Federated Fine-Tuning of Large Language Models

DisAgg: Distributed Aggregators for Efficient Secure Aggregation

PLayer-FL: A Principled Approach to Personalized Layer-wise Cross-Silo Federated Learning

Zero redundancy distributed learning with differential privacy

(ends 6:00 PM)

6 p.m.

Poster Session 1 & Opening Reception [6:00-8:00]

Posters -

AgenticCache: Cache-Driven Asynchronous Planning for Embodied AI Agents

Blueprint, Bootstrap, and Bridge: A Security Look at NVIDIA GPU Confidential Computing

ContextPilot: Fast Long-Context Inference via Context Reuse

CSLE: A Reinforcement Learning Platform for Autonomous Security Management

db-SP: Accelerating Sparse Attention for Visual Generative Models with Dual-Balanced Sequence Parallelism

DisAgg: Distributed Aggregators for Efficient Secure Aggregation

FlashAgents: Accelerating Multi-Agent LLM Systems via Streaming Prefill Overlap

FLoRIST: Singular Value Thresholding for Efficient and Accurate Federated Fine-Tuning of Large Language Models

From Tokens to Layers: Redefining Stall-Free Scheduling for MoE Serving with Layered Prefill

G-HEMP: FAST MULTI-GPU PRIVATE INFERENCE FOR LARGE-SCALE GCNS WITH HOMOMORPHIC ENCRYPTION

HELIOS : Adaptive Model And Early-Exit Selection for Efficient LLM Inference Serving

Hippocampus: An Efficient and Scalable Memory Module for Agentic AI

Matrix: Peer-to-Peer Multi-Agent Synthetic Data Generation Framework

Ontology-Guided Long-Term Agent Memory for Conversational RAG

Optimizing PyTorch Inference with LLM-Based Multi-Agent Systems

OSWorld-Human: Benchmarking the Efficiency of Computer-Use Agents

PLA-Serve: A Prefill-Length-Aware LLM Serving System

PLayer-FL: A Principled Approach to Personalized Layer-wise Cross-Silo Federated Learning

Privatar: Scalable Privacy-preserving Multi-user VR via Secure Offloading

ProToken: Token-Level Attribution for Federated Large Language Models

SONAR: Benchmarking Topology and Collaboration in Decentralized Learning

Stream2LLM: Overlap Context Streaming and Prefill for Reduced Time-to-First-Token

StreamDiffusionV2: A Streaming System for Dynamic and Interactive Video Generation

SuperInfer: SLO-Aware Rotary Scheduling and Memory Management for LLM Inference on Superchips

SwiftGS: Algorithm and System Co-Optimization for Fast 3D Gaussian Splatting on GPUs

TeleRAG: Efficient Retrieval-Augmented Generation Inference with Lookahead Retrieval

TiDAR: Think in Diffusion, Talk in Autoregression

TokenWeave: Efficient Compute-Communication Overlap for Distributed LLM Inference

Toward Principled LLM Safety Testing: Solving the Jailbreak Oracle Problem

TriInfer: Hybrid EPD Disaggregation for Efficient Multimodal Large Language Model Inference

VeriMoA: A Mixture-of-Agents Framework for Spec-to-HDL Generation

When Enough is Enough: Rank-Aware Early Termination for Vector Search

Zero redundancy distributed learning with differential privacy

ZK-APEX: ZERO-KNOWLEDGE APPROXIMATE PERSONALIZED UNLEARNING WITH EXECUTABLE PROOFS

(ends 8:00 PM)

WED 20 MAY

8 a.m.

Registration

(ends 5:00 PM)

8:30 a.m.

LLM Serving 2 [8:30-10:00]

Moderator: Juncheng Yang

BEAM: Joint Resource–Power Optimization for Energy-Efficient LLM Inference under SLO contraints

MorphServe: Efficient and Workload-Aware LLM Serving via Runtime Quantized Layer Swapping and KV Cache Resizing

BOute: Cost-Efficient LLM Serving with Heterogeneous LLMs and GPUs via Multi-Objective Bayesian Optimization

Locality-Aware Beam Scheduling for Efficient Test-Time Compute with a Consumer-grade GPU

Breaking the Ice: Analyzing Cold Start Latency in vLLM

FaaScale: Unlocking Fast LLM Scaling for Serverless Inference

(ends 10:00 AM)

LLM Training 1 [8:30-10:00]

Moderator: Zhihao Zhang

BOOST: BOttleneck-Optimized Scalable Training Framework for Low-Rank Large Language Models

Unleashing Scalable Context Parallelism for Foundation Models Pre-Training via FCP

ProTrain: Efficient LLM Training via Automatic Memory Management

MTraining: Distributed Dynamic Sparse Attention for Efficient Ultra-Long Context Training

NEST: Network- and Memory-Aware Device Placement for Distributed Deep Learning

Efficient Long-Context Language Model Training by Core Attention Disaggregation

(ends 10:00 AM)

10 a.m.

Break:

Break

(ends 10:30 AM)

10:30 a.m.

Overflow / Viewing [10:30-11:30]

(ends 11:30 AM)

Keynote Talk:

Keynote - Amin Vahdat

Amin Vahdat

(ends 11:30 AM)

11:30 a.m.

Break:

Lunch Break

(ends 1:00 PM)

1 p.m.

LLM Serving 3 [1:00-2:30]

Moderator: Yifan Qiao

SkipKV: Selective Skipping of KV Generation and Storage for Efficient Inference with Large Reasoning Models

OPKV: A High-Throughput Plugin-Driven Framework for Recallable Sparsity in Paged KV Cache Systems

FlexiCache: Leveraging Temporal Stability of Attention Heads for Efficient KV Cache Management

MAC-Attention: a Match--Amend--Complete scheme for fast and accurate attention computation

Using Span Queries to Optimize Cache and Attention Locality

Kitty: Accurate and Efficient 2-bit KV Cache Quantization with Dynamic Channel-wise Precision Boost

(ends 2:30 PM)

ML for Systems [1:00-2:30]

Moderator: Hui Guan

LLMInfer-Bench: Building the Virtuous Cycle for AI-driven LLM Systems

Virtual Machine NUMA Placement at Scale: Learning the Norm, Shielding the Tail

When Machine Learning Isn’t Sure: Building Resilient ML-Based Computer Systems by Embracing Uncertainty

Practical Adversarial Multi-Armed Bandits with Sublinear Runtime

Unified LLM Model for Power, Performance, and Area Prediction from Hardware Code

Automated Algorithm Design for Auto-Tuning Optimizers

(ends 2:30 PM)

2:30 p.m.

Break:

Break

(ends 2:45 PM)

2:45 p.m.

LLM Training 2 [2:45-4:15]

Moderator: Cheng Tan

NexSpec: Towards Optimizing Speculative Decoding in Reinforcement Learning Systems

HetRL: Efficient Reinforcement Learning for LLMs in Heterogeneous Environments

Pylo: Towards Accessible Learned Optimizers in PyTorch

Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes

FlexTrain: Scalable Hybrid-Parallel Training with Elastic Resource Utilization and Consistent Accuracy

Beat the long tail: Distribution-Aware Speculative Decoding for RL Training

(ends 4:15 PM)

LLM Serving 4 [2:45-4:15]

Moderator: Christina Giannoula

RaidServe: High-performance Resilient Serving

RDMA Point-to-Point Communication for LLM Systems

Demystifying the Mixture of Experts Serving Tax

FarSkip-Collectives: Unhobbling Blocking Communication in Mixture of Experts Models

GhostServe: A Lightweight Checkpointing System in the Shadow for Fault-Tolerant LLM Serving

CRAFT: Fine-Grained Cost-Aware Expert Replication For Efficient Mixture-of-Experts Serving

(ends 4:15 PM)

4:15 p.m.

Break:

Break

(ends 4:30 PM)

4:30 p.m.

LLM Training 4 [4:30-6:00]

Moderator: Yanghua Peng

Grolar: Efficient LLM Training on Heterogeneous Clusters

GriNNder: Breaking the Memory Capacity Wall in Full-Graph GNN Training with Storage Offloading

HexiScale: Facilitating Large Language Model Training over Heterogeneous Hardware

A Lightweight High-Throughput Collective-Capable NoC for Large-Scale ML Accelerators

DreamDDP: Accelerating Low-Bandwidth Geo-Distributed LLM Training with Layer-wise Partial Synchronization

FP8-Flow-MoE: A Casting-Free FP8 Recipe without Double Quantization Error

(ends 6:00 PM)

Model Compression [4:30-6:00]

Moderator: Sanjeev Singh

CAGE: Curvature-Aware Gradient Estimation For Accurate Quantization-Aware Training

Shannonic: Efficient Entropy-Optimal Compression for ML Workloads

Once-for-All Channel Mixers (HyperTinyPW): Generative Compression for TinyML

MixLLM: LLM Quantization with Global Mixed-precision between Output-features and Highly-efficient System Design

Search Your NVFP4 Scales!

(ends 6:00 PM)

6 p.m.

Poster Session 2 [6:00-8:00]

Posters -

A Lightweight High-Throughput Collective-Capable NoC for Large-Scale ML Accelerators

Automated Algorithm Design for Auto-Tuning Optimizers

BEAM: Joint Resource–Power Optimization for Energy-Efficient LLM Inference under SLO contraints

Beat the long tail: Distribution-Aware Speculative Decoding for RL Training

BOOST: BOttleneck-Optimized Scalable Training Framework for Low-Rank Large Language Models

BOute: Cost-Efficient LLM Serving with Heterogeneous LLMs and GPUs via Multi-Objective Bayesian Optimization

Breaking the Ice: Analyzing Cold Start Latency in vLLM

CAGE: Curvature-Aware Gradient Estimation For Accurate Quantization-Aware Training

CRAFT: Fine-Grained Cost-Aware Expert Replication For Efficient Mixture-of-Experts Serving

Demystifying the Mixture of Experts Serving Tax

DreamDDP: Accelerating Low-Bandwidth Geo-Distributed LLM Training with Layer-wise Partial Synchronization

Efficient Long-Context Language Model Training by Core Attention Disaggregation

FaaScale: Unlocking Fast LLM Scaling for Serverless Inference

fabric-lib: RDMA Point-to-Point Communication for LLM Systems

FarSkip-Collective: Unhobbling Blocking Communication in Mixture of Experts Models

FlashInfer-Bench: Building the Virtuous Cycle for AI-driven LLM Systems

FlexiCache: Leveraging Temporal Stability of Attention Heads for Efficient KV Cache Management

FlexTrain: Scalable Hybrid-Parallel Training with Elastic Resource Utilization and Consistent Accuracy

FP8-Flow-MoE: A Casting-Free FP8 Recipe without Double Quantization Error

GhostServe: A Lightweight Checkpointing System in the Shadow for Fault-Tolerant LLM Serving

GriNNder: Breaking the Memory Capacity Wall in Full-Graph GNN Training with Storage Offloading

HetRL: Efficient Reinforcement Learning for LLMs in Heterogeneous Environments

HexiScale: Facilitating Large Language Model Training over Heterogeneous Hardware

Kitty: Accurate and Efficient 2-bit KV Cache Quantization with Dynamic Channel-wise Precision Boost

Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes

Locality-Aware Beam Scheduling for Efficient Test-Time Compute with a Consumer-grade GPU

MAC-Attention: a Match--Amend--Complete scheme for fast and accurate attention computation

MixLLM: LLM Quantization with Global Mixed-precision between Output-features and Highly-efficient System Design

MorphServe: Efficient and Workload-Aware LLM Serving via Runtime Quantized Layer Swapping and KV Cache Resizing

MTraining: Distributed Dynamic Sparse Attention for Efficient Ultra-Long Context Training

NEST: Network- and Memory-Aware Device Placement for Distributed Deep Learning

Once-for-All Channel Mixers (HyperTinyPW): Generative Compression for TinyML

OPKV: A High-Throughput Plugin-Driven Framework for Recallable Sparsity in Paged KV Cache Systems

Practical Adversarial Multi-Armed Bandits with Sublinear Runtime

ProTrain: Efficient LLM Training via Automatic Memory Management

Pylo: Towards Accessible Learned Optimizers in PyTorch

RaidServe: High-performance Resilient Serving

ReSpec: Towards Optimizing Speculative Decoding in Reinforcement Learning Systems

Search Your Block Floating Point Scales!

Shannonic: Efficient Entropy-Optimal Compression for ML Workloads

SkipKV: Selective Skipping of KV Generation and Storage for Efficient Inference with Large Reasoning Models

Unified LLM Model for Power, Performance, and Area Prediction from Hardware Code

Unleashing Scalable Context Parallelism for Foundation Models Pre-Training via FCP

Using Span Queries to Optimize Cache and Attention Locality

Virtual Machine NUMA Placement at Scale: Learning the Norm, Shielding the Tail

When Machine Learning Isn’t Sure: Building Resilient ML-Based Computer Systems by Embracing Uncertainty

Zorse: Optimizing LLM Training Efficiency on Heterogeneous GPU Clusters

(ends 8:00 PM)

THU 21 MAY

8 a.m.

Registration

(ends 5:00 PM)

8:30 a.m.

LLM Serving 5 [8:30-10:00]

Moderator: Esha Choukse

PRISM: PARAMETRICALLY RESTRUCTURED INFERENCE FOR SPECULATIVE SAMPLING DRAFT MODELS

CDLM: CONSISTENCY DIFFUSION LANGUAGE MODELS FOR FASTER SAMPLING

Speculative Decoding: Performance or Illusion?

SpecDiff-2: Scaling Diffusion Drafter Alignment For Faster Speculative Decoding

Accelerating Large-Scale Reasoning Model Inference with Sparse Self-Speculative Decoding

(ends 10:00 AM)

Agentic AI/MLSys [8:30-10:00]

Moderator: Binhang Yuan

ADS: AN AGENTIC DETECTION SYSTEM FOR ENTERPRISE AGENTIC AI SECURITY

PROMPTS: PeRformance Optimization via Multi-Agent Planning for LLM Training and Serving

Agentic Operator Generation for ML ASICs

Cost-aware Duration Prediction for Software Upgrades in Datacenters

The OpenHands Software Agent SDK: A Composable and Extensible Foundation for Production Agents

(ends 10:00 AM)

10 a.m.

Break:

Break

(ends 10:30 AM)

10:30 a.m.

Keynote Talk:

Rethinking Pretraining: Data and Architecture

Luke Zettlemoyer

(ends 11:30 AM)

Overflow / Viewing [10:30-11:30]

(ends 11:30 AM)

11:30 a.m.

Break:

Lunch Break

(ends 1:00 PM)

1 p.m.

Edge and Mobile [1:00-2:30]

Moderator: Vijay Janapa Reddi

IntAttention: A Fully Integer Attention Pipeline for Efficient Edge Inference

Massive-Scale Out-Of-Core UMAP on the GPU

REPARO: LOSS-RESILIENT GENERATIVE CODEC FOR VIDEO CONFERENCING

Rethinking DVFS for Mobile LLMs: Unified Energy-Aware Scheduling with CORE

EarthSight: A Distributed Framework for Low-Latency Satellite Intelligence

(ends 2:30 PM)

LLM Serving 6 [1:00-2:30]

Moderator: Xiang Song

SHIP: SRAM-Based Huge Inference Pipelines for Fast LLM Serving

BatchLLM: Optimizing Large Batched LLM Inference with Global Prefix Sharing and Throughput-oriented Token Batching

Beyond the Buzz: A Pragmatic Take on Inference Disaggregation

Optimizing Deployment Configurations for LLM Inference

Meeting SLOs, Slashing Hours: Automated Enterprise LLM Optimization with OptiKIT

Scaling Up Large Language Models Serving Systems for Semantic Job Search

(ends 2:30 PM)

2:15 p.m.

Break:

Break

(ends 2:30 PM)

2:45 p.m.

Benchmarks 1 [2:45-4:15]

Moderator: Youjie Li

Charon: A Unified and Fine-Grained Simulator for Large-Scale LLM Training and Inference

Hawkeye: Reproducing GPU-Level Non-Determinism

DriftBench: Measuring and Predicting Infrastructure Drift in LLM Serving Systems

Parrot: Persuasion and Agreement Robustness Rating of Output Truth

(ends 4:15 PM)

LLM Training 3 [2:45-4:15]

Moderator: Mohamed Abdelfattah

Sparing Strategies to Minimize Reliability Impact On Large Training Jobs

FlexScale: Flexible and High-Performance FSDP at Scale

AXLearn: Modular, Hardware-Agnostic Large Model Training

NodeSweep: Practical Straggler Detection and Health Monitoring for Large-Scale Foundation Model Training

MoEBlaze: Breaking the Memory Wall for Efficient MoE Training on Modern GPUs

FreeScale: Distributed Training for Sequence Recommendation Models with Minimal Scaling Cost

(ends 4:15 PM)

4:15 p.m.

Break:

Break

(ends 4:30 PM)

4:30 p.m.

Benchmarks 2 [4:30-6:00]

Moderator: Juncheng Yang

XProf: An Open, Scalable, and Extensible Profiling System for the Modern ML Stack

AIRS: Scaling Live Inference in Resource Constrained Environments

SAKURAONE: An Open Ethernet–Based AI HPC System and Its Observed Workload Dynamics in a Single-Tenant LLM Development Environment

MLCommons Chakra: Advancing Performance Benchmarking and Co-design using Standardized Execution Traces

ProfInfer: An eBPF-based Fine-Grained LLM Inference Profiler

ML Fleet Efficiency: Improving TPU Systems at Scale with ML Productivity Goodput

(ends 6:00 PM)

Compilers and Kernels [4:30-6:00]

Moderator: Steven Lyubomirsky

ParallelKittens: Systematic and Practical Simplification of Multi-GPU AI Kernels

Event Tensor: A Unified Abstraction for Compiling Dynamic Megakernel

AccelOpt: A Self-Improving LLM Agentic System for AI Accelerator Kernel Optimization

SchedFlow: Transparent and Flexible Intra-Device Parallelism via Programmable Operator Scheduling

ApproxMLIR : Accuracy-Aware Compiler for Compound ML System

HipKittens: Fast and Furious AMD Kernels

(ends 6:00 PM)

6 p.m.

Poster Session 3 [6:00-8:00]

Posters -

Accelerating Large-Scale Reasoning Model Inference with Sparse Self-Speculative Decoding

AccelOpt: A Self-Improving LLM Agentic System for AI Accelerator Kernel Optimization

ADS: AN AGENTIC DETECTION SYSTEM FOR ENTERPRISE AGENTIC AI SECURITY

Agentic Operator Generation for ML ASICs

AIRS: Scaling Live Inference in Resource Constrained Environments

ApproxMLIR : Accuracy-Aware Compiler for Compound ML System

Attribution-based Sparse Activation in Large Language Models

AXLearn: Modular, Hardware-Agnostic Large Model Training

BatchLLM: Optimizing Large Batched LLM Inference with Global Prefix Sharing and Throughput-oriented Token Batching

Beyond the Buzz: A Pragmatic Take on Inference Disaggregation

BLASST: Dynamic BLocked Attention Sparsity via Softmax Thresholding

CATWILD: Compiler Autotuning for TPU workloads in the Wild

CDLM: Consistency Diffusion Language Models for Faster Sampling

Charon: A Unified and Fine-Grained Simulator for Large-Scale LLM Training and Inference

Cost-aware Duration Prediction for Software Upgrades in Datacenters

Dataflow Is All You Need

DriftBench: Measuring and Predicting Infrastructure Drift in LLM Serving Systems

DynaFlow: Transparent and Flexible Intra-Device Parallelism via Programmable Operator Scheduling

EarthSight: A Distributed Framework for Low-Latency Satellite Intelligence

Efficient, VRAM-Constrained xLM Inference on Clients

Event Tensor: A Unified Abstraction for Compiling Dynamic Megakernel

ExecuTorch - A Unified PyTorch Solution to Run ML Models On-Device

FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling

Flashlight: PyTorch Compiler Extensions to Accelerate Attention Variants

FreeScale: Distributed Training for Sequence Recommendation Models with Minimal Scaling Cost

GUARD: SCALABLE STRAGGLER DETECTION AND NODE HEALTH MANAGEMENT FOR LARGE-SCALE TRAINING

Hawkeye: Reproducing GPU-Level Non-Determinism

HipKittens: Fast and Furious AMD Kernels

IntAttention: A Fully Integer Attention Pipeline for Efficient Edge Inference

LEANN: A Low-Storage Overhead Vector Index

Massive-Scale Out-Of-Core UMAP on the GPU

Meeting SLOs, Slashing Hours: Automated Enterprise LLM Optimization with OptiKIT

MLCommons Chakra: Advancing Performance Benchmarking and Co-design using Standardized Execution Traces

ML Fleet Efficiency: Improving TPU Systems at Scale with ML Productivity Goodput

MoEBlaze: Breaking the Memory Wall for Efficient MoE Training on Modern GPUs

Optimizing Deployment Configurations for LLM Inference

ParallelKittens: Systematic and Practical Simplification of Multi-GPU AI Kernels

PARROT: Persuasion and Agreement Robustness Rating of Output Truth — A Sycophancy Robustness Benchmark for LLMs

PRISM: Parametrically Refactor Inference for Speculative Decoding Draft Models

ProfInfer: An eBPF-based Fine-Grained LLM Inference Profiler

PROMPTS: PeRformance Optimization via Multi-Agent Planning for LLM Training and Serving

REPARO: LOSS-RESILIENT GENERATIVE CODEC FOR VIDEO CONFERENCING

Rethinking DVFS for Mobile LLMs: Unified Energy-Aware Scheduling with CORE

SAKURAONE: An Open Ethernet–Based AI HPC System and Its Observed Workload Dynamics in a Single-Tenant LLM Development Environment

Scaling Up Large Language Models Serving Systems for Semantic Job Search

SHIP: SRAM-Based Huge Inference Pipelines for Fast LLM Serving

Sparing Strategies to Minimize Reliability Impact On Large Training Jobs

SpecDiff-2: Scaling Diffusion Drafter Alignment For Faster Speculative Decoding

Speculative Decoding: Performance or Illusion?

Spira: Exploiting Voxel Data Structural Properties for Efficient Sparse Convolution in Point Cloud Networks

The OpenHands Software Agent SDK: A Composable and Extensible Foundation for Production Agents

veScale-FSDP: Flexible and High-Performance FSDP at Scale

Wave: A Symbolic Python DSL And Compiler for High-Performance Machine Learning

XProf: An Open, Scalable, and Extensible Profiling System for the Modern ML Stack

(ends 8:00 PM)

FRI 22 MAY

8 a.m.

Registration

(ends 1:00 PM)

8:15 a.m.

Compilers/HW [8:15-9:30]

Moderator: Phitchaya Phothilimthana

Dataflow Is All You Need

Efficient, VRAM-Constrained xLM Inference on Clients

WAVE: A SYMBOLIC PYTHON DSL AND COMPILER FOR HIGH PERFORMANCE MACHINE LEARNING

CATWILD: Compiler Autotuning for TPU workloads in the Wild

(ends 9:30 AM)

Efficient Computation [8:15-9:30]

Moderator: Aniruddh Ramrakhyani

Spira: Exploiting Voxel Data Structural Properties for Efficient Sparse Convolution in Point Cloud Networks

Attribution-based Sparse Activation in Large Language Models

Flashlight: PyTorch Compiler Extensions to Accelerate Attention Variants

FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling

(ends 9:30 AM)

9:30 a.m.

Break:

Break

(ends 9:45 AM)

9:45 a.m.

Overflow / Viewing [9:45-10:45]

(ends 10:45 AM)

Keynote Talk:

Keynote - Christos Kozyrakis

Christos Kozyrakis

(ends 10:45 AM)

10:45 a.m.

Break:

Break

(ends 11:00 AM)

11 a.m.

Overflow / Viewing [11:00-1:00]

(ends 1:00 PM)

Competition:

Competition Track: Google / NVIDIA / Amazon

(ends 12:50 PM)