Peng Sun#
Peng Sun is Sr. Director of AI Software Development at AMD, overseeing the GPU AI software upper layers for ROCm. As one of the first engineers on ROCm, he has helped shape AMD’s AI software stack from inception to global deployment. Peng leads teams delivering AI frameworks, inference engines, and high-performance GPU operators for large-scale AI workloads. His work spans enabling industry-leading models, optimizing open-source AI software, and collaborating with major ecosystem partners. Holding a Ph.D. in Computer Science, Peng combines deep technical expertise with proven leadership in building scalable AI solutions that power cutting-edge research and enterprise AI applications worldwide. Peng’s linkedin profile: https://www.linkedin.com/in/pengsun86/
Posts by Peng Sun
SGLang-ATOM: Bring ROCm-Native Acceleration to SGLang Serving
Explore how SGLang-ATOM connects SGLang serving applications with ROCm-native ATOM execution to accelerate LLM inference on AMD Instinct GPUs.
Accelerating LLM Inference on AMD GPUs with Low-Latency GEMMs
Learn how FlyDSL low-latency GEMMs speed up LLM decode on AMD GPUs with Split-K, K-slice parallelism, and an LDS-based pipeline.
OpenXLA and JAX - ROCm Support and the State of CI
Learn how OpenXLA and JAX run on AMD ROCm: what landed this year, how every PR is gated on real Instinct hardware, and how to get started.
DP Attention and TBO for DeepSeek-V4 on MI355X
Learn how ATOM improves DeepSeek-V4 inference on AMD Instinct MI355X GPUs with DP Attention scheduling and Two-Batch Overlap.
ATOM: Unlocking Extreme AMD Instinct Inference with Software-Hardware Co-Optimization
A technical walkthrough of ATOM on AMD Instinct GPUs, covering architecture, feature scope, model coverage, and practical benchmark dashboard usage.
From Naive to Near-Peak: Building High-Performance GEMM Kernels with Gluon
Learn how a Gluon GEMM tutorial teaches profiling-driven AMD GPU optimization from FP16 baseline to BF8 and MXFP4 kernels.
vLLM-ATOM: Unlocking Native AMD Performance in the vLLM Ecosystem
Use ATOM as an out-of-tree vLLM plugin to keep vLLM compatibility while enabling AMD-optimized attention, model execution, and multi-model support including Kimi-K2.5.
Getting Started with FlyDSL Nightly Wheels on ROCm
A practical guide to installing and using FlyDSL nightly wheels on ROCm for fast, Python-native GPU kernel development
Accelerating Kimi-K2.5 on AMD Instinct™ MI300X: Optimizing Fused MoE with FlyDSL
Optimize Kimi-K2.5 on AMD MI300X using FlyDSL for fused MoE kernel acceleration. Achieve faster TTFT, TPOT, and throughput with our step-by-step optimization guide.
FlyDSL: Expert GPU Kernel Development with the Ease of MLIR Python Native DSL on AMD GPUs
FlyDSL is a Python-first, MLIR-native DSL for expert GPU kernel development and tuning on AMD GPUs.
ROCm Becomes a First-Class Platform in the vLLM Ecosystem
ROCm is now a first-class vLLM platform: official wheels + Docker, stronger CI, and faster LLM & multimodal inference on AMD Instinct GPUs.
Accelerating Multimodal Inference in vLLM: The One-Line Optimization for Large Multimodal Models
Learn how to optimize multimodal model inference with batch-level data parallelism for vision encoders in vLLM, achieving up to 45% throughput gains on AMD MI300X.
The vLLM MoE Playbook: A Practical Guide to TP, DP, PP and Expert Parallelism
Learn how to combine TP, DP, PP, and EP for MoE models. Discover proven strategies to maximize performance on your vLLM deployments.
Practical, Fault‑Robust Distributed Inference for DeepSeek on AMD MI300X
Learn how a small-radius expert parallel design with prefill–decode disaggregation enables scalable, fault-isolated LLM inference on AMD Instinct™ MI300X clusters.
Empowering Developers to Build a Robust PyTorch Ecosystem on AMD ROCm™ with Better Insights and Monitoring
Production ROCm support for N-1 to N+1 PyTorch releases is in progress. The AI Software Head-Up Dashboard shows status of PyTorch on ROCm.
Matrix Core Programming on AMD CDNA™3 and CDNA™4 architecture
This blog post explains how to use Matrix Cores on CDNA3 and CDNA4 architecture, with a focus on low-precision data types such as FP16, FP8, and FP4
Accelerated LLM Inference on AMD Instinct™ GPUs with vLLM 0.9.x and ROCm
vLLM v0.9.x is here with major ROCm™ optimizations—boosting LLM performance, reducing latency, and expanding model support on AMD Instinct™ GPUs.
Unleash Full GPU Potential: Overlap Communication and Computation with Triton-Distributed
Unlock the full power of AMD GPUs—write portable, efficient kernels with Triton-Distributed, overlapping computation and communication with ease and flexibility
Supercharge DeepSeek-R1 Inference on AMD Instinct MI300X
Learn how to optimize DeepSeek-R1 on AMD MI300X with SGLang, AITER kernels and hyperparameter tuning for up to 5× throughput and 60% lower latency over Nvidia H200