MLSys 2025 Schedule

Filter Events

MON 12 MAY

8:45 a.m.

Registration and coffee

(ends 10:45 AM)

9:20 a.m.

Opening Remarks:

Opening Remarks

(ends 9:30 AM)

9:30 a.m.

Invited Talk:

Keynote: Soumith Chintala

Soumith Chintala

(ends 10:30 AM)

10:30 a.m.

Break:

Break

(ends 10:45 AM)

10:45 a.m.

Talk:

YPS - Speaker 1

(ends 11:05 AM)

11:05 a.m.

Talk:

YPS - Speaker 2

(ends 11:25 AM)

11:25 a.m.

Talk:

YPS - Speaker 3

(ends 11:45 AM)

11:45 a.m.

Talk:

YPS - Speaker 4

(ends 12:05 PM)

12:05 p.m.

Break:

Lunch Break

(ends 1:30 PM)

1:30 p.m.

Industry Lightning Talks [1:30-2:30]

(ends 2:30 PM)

2:30 p.m.

Panel Discussion:

Panel Discussion

(ends 3:30 PM)

4 p.m.

Poster Session and Reception [4:00-6:00]

(ends 6:00 PM)

TUE 13 MAY

7 a.m.

Registration and Check-in

(ends 8:30 AM)

8:30 a.m.

Opening Remarks:

Opening Remarks

(ends 8:45 AM)

8:45 a.m.

Session 1: LLM and Diffusion Model Serving [8:45-10:00]

DiffServe: Efficiently Serving Text-to-Image Diffusion Models with Query-Aware Model Scaling

FastTree: Optimizing Attention Kernel and Runtime for Tree-Structured LLM Inference

FlashInfer: Efficient and Customizable Attention Engine for LLM Inference Serving

LeanAttention: Hardware-Aware Scalable Attention Mechanism for the Decode-Phase of Transformers

Rethinking Key-Value Cache Compression Techniques for Large Language Model Serving

(ends 10:00 AM)

10 a.m.

Break:

Break

(ends 10:30 AM)

10:30 a.m.

Invited Talk:

Keynote: Ion Stoica (UC Berkeley and Databricks)

Ion Stoica

(ends 11:30 AM)

11:30 a.m.

Break:

Lunch Break

(ends 1:15 PM)

1:15 p.m.

Session 2: Parallel and Distributed Systems [1:15-2:40]

AdaParse: An Adaptive Parallel PDF Parsing and Resource Scaling Engine

Context Parallelism for Scalable Million-Token Inference

PipeFill: Using GPUs During Bubbles in Pipeline-parallel LLM Training

Rubick: Exploiting Job Reconfigurability for Deep Learning Cluster Scheduling

SPA: SCALING GRAPH NEURAL NETWORK TRAINING ON LARGE GRAPHS VIA PROBABILISTIC SPLITTING

(ends 2:40 PM)

2:40 p.m.

Session 3: Quantization and Sparsity [2:40-4:00]

Enabling Unstructured Sparse Acceleration on Structured Sparse Accelerators

MiLo: Efficient Quantized MoE Inference with Mixture of Low-Rank Compensators

QServe:W4A8KV4 Quantization and System Co-design for Efficient LLM Serving

Radius: Range-based Gradient Sparsity for Large Foundation Model Pre-training

Self-Data Distillation for Recovering Quality in Pruned Large Language Models

(ends 4:00 PM)

4 p.m.

Break:

Break

(ends 4:45 PM)

4:45 p.m.

Session 4: Measurement and Analysis [4:45-6:00]

AI Metropolis: Scaling Large Language Model-based Multi-Agent Simulation with Out-of-order Execution

AIOpsLab: A Holistic Framework to Evaluate AI Agents for Enabling Autonomous Clouds

Interference-aware Edge Runtime Prediction with Conformal Matrix Completion

Know Where You’re Uncertain When Planning with Multimodal Foundation Models: A Formal Framework

The Hidden Bloat in Machine Learning Systems

(ends 6:00 PM)

6 p.m.

Poster Session and Reception [6:00-8:00]

(ends 8:00 PM)

WED 14 MAY

7 a.m.

Registration and Check-in

(ends 8:40 AM)

8:30 a.m.

Session 5: LLM training and fine-tuning [8:30-10:00]

APOLLO: SGD-like Memory, AdamW-level Performance

HyC-LoRA: Memory Efficient LoRA Fine-tuning with Hybrid Activation Compression

Lumos: Efficient Performance Modeling and Estimation for Large-scale LLM Training

ReaL: Efficient RLHF Training of Large Language Models with Parameter Reallocation

Training Ultra Long Context Language Model with Fully Pipelined Distributed Transformer

Youmu: Efficient Columnar Data Pipeline for LLM Training

(ends 10:00 AM)

10 a.m.

Break:

Break

(ends 10:30 AM)

10:30 a.m.

Invited Talk:

Hardware-aware training and inference for large-scale AI

Animashree Anandkumar

(ends 11:30 AM)

11:30 a.m.

Break:

Lunch Break

(ends 1:15 PM)

1:15 p.m.

Session 6: Edge and Cloud Systems [1:15-2:40]

A Practical Cross-Layer Approach for ML-Driven Storage Placement in Warehouse-Scale Computers

Efficient On-Device Machine Learning with a Biologically-Plausible Forward-Only Algorithm

Optimizing LLM Queries in Relational Data Analytics Workloads

ProtoRAIL: A Risk-cognizant Imitation Agent for Adaptive vCPU Oversubscription In the Cloud

SwiftVI: Time-Efficient Planning and Learning with MDPs

(ends 2:40 PM)

2:40 p.m.

Session 7: Quantization and Sparsity [2:40-4:00]

Efficient LLM Inference using Dynamic Input Pruning and Cache-Aware Masking

Lightweight Software Kernels and Hardware Extensions for Efficient Sparse Deep Neural Networks on Microcontrollers

LServe: Efficient Long-sequence LLM Serving with Unified Sparse Attention

SampleAttention: Near-Lossless Acceleration of Long Context LLM Inference with Adaptive Structured Sparse Attention

SparseTransX: Efficient Training of Translation-Based Knowledge Graph Embeddings Using Sparse Matrix Operations

(ends 4:00 PM)

4 p.m.

Break:

Break

(ends 4:45 PM)

4:30 p.m.

Session 8: LLM and Diffusion Model Serving [4:30-6:00]

FlexInfer: Flexible LLM Inference with CPU Computations

ScaleFusion: Scalable Inference of Spatial-Temporal Diffusion Transformers for High-Resolution Long Video Generation

Seesaw: High-throughput LLM Inference via Model Re-sharding

SOLA: Optimizing SLO Attainment for Large Language Model Serving with State-Aware Scheduling

TurboAttention: Efficient attention approximation for high throughputs llm

(ends 6:00 PM)

6 p.m.

Poster Session [6:00-8:00]

(ends 8:00 PM)

THU 15 MAY

7 a.m.

Registration and Check-in

(ends 8:40 AM)

8:30 a.m.

Session 9: Parallel and Distributed Systems [8:30-10:00]

Balancing Pipeline Parallelism with Vocabulary Parallelism

COMET: Fine-grained Computation-communication Overlapping for Mixture-of-Experts

On Distributed Larger-Than-Memory Subset Selection With Pairwise Submodular Functions

Scaling Deep Learning Training with MPMD Pipeline Parallelism

TileLink: Generating Efficient Compute-Communication Overlapping Kernels using Tile-Centric Primitives

(ends 10:00 AM)

10 a.m.

Break:

Break

(ends 10:30 AM)

10:30 a.m.

Invited Talk:

Responsible Finetuning of Large Language Models

Ling Liu

(ends 11:30 AM)

11:30 a.m.

Break:

Lunch Break

(ends 1:15 PM)

1:15 p.m.

Session 10: LLM and Diffusion Model Serving [1:15-2:40]

FlexAttention: A Programming Model for Generating Fused Attention Variants.

Marconi: Prefix Caching for the Era of Hybrid LLMs

NEO: Saving GPU Memory Crisis with CPU Offloading for Online LLM Inference

ThunderServe: High-performance and Cost-efficient LLM Serving in Cloud Environments

XGrammar: Flexible and Efficient Structured Generation Engine for Large Language Models

(ends 2:40 PM)

2:40 p.m.

Session 11: Federated Learning [2:40-4:00]

FedProphet: Memory-Efficient Federated Adversarial Training via Robust and Consistent Cascade Learning

FLStore: Efficient Federated Learning Storage for non-training workloads

MAS-ATTENTION: MEMORY-AWARE STREAM PROCESSING FOR ATTENTION ACCELERATION ON RESOURCE-CONSTRAINED EDGE DEVICES

Photon: Federated LLM Pre-Training

Venn: Resource Management For Collaborative Learning Jobs

(ends 4:00 PM)

4 p.m.

Break:

Break

(ends 4:30 PM)

4:30 p.m.

Session 12: Edge and Cloud Systems [4:30-6:00]

Graph Learning at Scale: Characterizing and Optimizing Pre-Propagation GNNs

LAVA: Lifetime-Aware VM Allocation with Learned Distributions and Adaptation to Mispredictions

MEADOW: Memory-efficient Dataflow and Data Packing for Low Power Edge LLMs

Supply-Chain Attacks in Machine Learning Frameworks

VoLUT: Efficient Volumetric streaming enhanced by LUT-based super-resolution

(ends 6:00 PM)

6 p.m.

Closing Remarks:

Closing Remarks

(ends 6:05 PM)