Publications | Shuo Yang

Publications

My work spans the stack, from reusable GPU primitives to algorithm–system co-design to complete serving systems, for both language models and multimodal generation. Browse it as a research tree that shows how the projects build on each other, or as a list of all papers.

FreeToken

13.8k stars Edge Serving MoE

FreeToken

Efficient edge-native MoE serving with bandwidth-adaptive execution.

FlashLib

GPU Library Classical ML Operators Agent-Native API

FlashLib

Bringing flash magic to classical machine learning operators.

Flash-KMeans

NeurIPS 2026 Exact K-Means Kernel Optimization

Flash-KMeans

Fast and memory-efficient exact K-Means.

Quant VideoGen

ICML 2026 Long Video KV Cache Quantization

Quant VideoGen

Auto-regressive long video generation via 2-bit KV-cache quantization.

BlendServe

ASPLOS 2026 Offline Inference LLM Serving

BlendServe

Optimizing offline inference for autoregressive large models with resource-aware batching.

StreamDiffusionV2

MLSys 2026 Best Paper Interactive Video Streaming System

StreamDiffusionV2

A streaming system for dynamic and interactive video generation.

vAttention

ICLR 2026 Verified Sparsity Sparse Attention

vAttention

Verified sparse attention.

SLA

ICLR 2026 Sparse-Linear Attention Diffusion Transformers

SLA

Beyond sparsity in diffusion transformers via fine-tunable sparse-linear attention.

Sparse VideoGen2

NeurIPS 2025 Spotlight Semantic Permutation Video Generation

Sparse VideoGen2

Accelerating video generation with sparse attention via semantic-aware permutation.

Radial Attention

NeurIPS 2025 Long Video Sparse Attention

Radial Attention

O(n log n) sparse attention with energy decay for long video generation.

Sparse VideoGen

ICML 2025 Sparse Attention Video Generation

Sparse VideoGen

Accelerating video diffusion transformers with spatial-temporal sparsity.

Prism

OSDI 2026 GPU Sharing Multi-LLM Serving

Prism

Unleashing GPU sharing for cost-efficient multi-LLM serving.

UCCL

OSDI 2026 GPU Networking Transport Layer

UCCL

An extensible software transport layer for GPU networking.

WorldModelBench

NeurIPS 2025 Benchmark World Models

WorldModelBench

Judging video generation models as world models.

Twilight

NeurIPS 2025 Spotlight Adaptive Sparsity Long Context

Twilight

Adaptive attention sparsity with hierarchical top-p pruning.

HashAttention

ICML 2025 Semantic Sparsity Sparse Attention

HashAttention

Semantic sparsity for faster inference.

Post-Training Sparse Attention with Double Sparsity

Sparse Attention KV Cache LLM Inference

Post-Training Sparse Attention with Double Sparsity

Sparse attention for reducing KV-cache bandwidth in LLM inference.

S-LoRA

MLSys 2024 LoRA Serving CUDA Kernels

S-LoRA

Serving thousands of concurrent LoRA adapters.

Rethinking Benchmark and Contamination for Language Models with Rephrased Samples

Data Quality Benchmark Contamination Evaluation

Rethinking Benchmark and Contamination for Language Models with Rephrased Samples

Decontamination and benchmark overlap analysis for language models.