Yao Fu#
Dr. Yao Fu is a Director of Software Development in the AMD AI Group, where he leads efforts in distributed training solutions and performance optimization. Prior to focusing exclusively on training solutions, he spearheaded strategic AI solutions and competitive performance enhancements across AMD platforms—including AI PCs, data center CPUs, and embedded devices. With over six years of experience as a system architect for data center solutions at Xilinx/AMD, Dr. Fu has pioneered foundational methodologies that optimize FPGA-based deep learning solutions for higher throughput, lower latency, and improved power efficiency. His INT8 optimization methodology is widely adopted across deep learning inference solutions on UltraScale and UltraScale+ devices.
Posts by Yao Fu
MoE Training Optimization with Primus
Learn how to speed up MoE training on AMD Instinct GPUs with Primus: a fused MoE megakernel, FP8/MXFP8 expert GEMMs, and training-at-scale tuning.
Pre-Training Hybrid LLMs from Scratch in Primus on AMD Instinct™ GPUs
Pre-train hybrid LLMs mixing MLA attention with Gated DeltaNet, Kimi Delta Attention and Mamba2 mixers on AMD Instinct MI355X GPUs, with open recipes.
Implementing a High-Performance Custom Diffusion Attention Kernel with FlyDSL
Learn how to implement and optimize flexible, high-performance diffusion attention kernels with FlyDSL.
veRL on AMD: Production-Ready RL Post-Training on ROCm
Run veRL RL post-training on AMD Instinct GPUs with a turnkey ROCm container, AITER-accelerated rollout, and validated accuracy on MI300 and MI355.
Enabling DeepSeek-V4-Flash Training on AMD Instinct MI355X GPUs with Primus
DeepSeek-V4-Flash training on AMD Instinct GPUs with Primus: model architecture introduction, performance projection, kernel optimizations, and how to reproduce.
Scaling RL with verl on AMD Instinct MI355X: Async Walkthrough and Sync Benchmark
Scale RLHF on AMD Instinct MI355X with verl's fully async trainer. Hands-on GRPO + DAPO examples.
Using ODC to Accelerate AMD SFT Training
Learn how we ported ODC on-demand P2P communication to AMD Instinct MI300X to cut FSDP bubbles and speed up variable-length SFT training.
Primus Tuning Agent: Closing the Configuration-Search Loop
Use the Primus Tuning Agent to automatically find optimal LLM training configurations on AMD Instinct GPUs.
Primus Projection: Estimate Memory and Performance Before You Train
Learn how to use the Primus projection tool to estimate memory and performance for large-scale LLM training on AMD Instinct™ accelerator platforms.
MaxText-Slurm: Production-Grade LLM Training with Built-In Observability
MaxText-Slurm: A unified launch system for production-grade LLM training with observability on AMD GPU clusters.
Streamlining Recommendation Model Training on AMD Instinct™ GPUs
Explore how the ROCm training docker can be used for recommendation model training on Instinct GPUs, along with a guide on configuring the workload.
Primus-Pipeline: A More Flexible and Scalable Pipeline Parallelism Implementation
Learn how to use our flexible and scalable pipeline parallelism framework with Primus backend and AMD hardware.
Resilient Large-Scale Training: Integrating TorchFT with TorchTitan on AMD GPUs
Achieve resilient, checkpoint-less distributed training on AMD GPUs by integrating TorchFT with TorchTitan on Primus-SaFE.
Deep Dive into Primus: High-Performance Training for Large Language Models
Learn how to achieve peak dense LLM training performance on AMD Instinct™ GPUs using Primus’s unified CLI and optimized backend presets.
MoE Training Best Practices on AMD GPUs
Learn how to optimize Mixture-of-Experts (MoE) model training on AMD Instinct GPUs with ROCm. Maximize your AI training performance now!
Optimizing LLM Workloads: AMD Instinct MI355X GPUs Drive Competitive Performance
Explore ROCm 7.0’s AI training boost! See how MI355X accelerates JAX and PyTorch frameworks to unlock faster and efficient LLM scaling.
Stability at Scale: AMD’s Full‑Stack Platform for Large‑Model Training
Primus streamlines LLM training on AMD GPUs with unified configs, multi-backend support, preflight validation, and structured logging.
An Introduction to Primus-Turbo: A Library for Accelerating Transformer Models on AMD GPUs
Primus streamlines training on AMD ROCm, from fine-tuning to massive pretraining on MI300X GPUs—faster, safer, and easier to debug
Day 0 Developer Guide: Running the Latest Open Models from OpenAI on AMD AI Hardware
Day 0 support across our AI hardware ecosystem from our flagship AMD InstinctTM MI355X and MI300X GPUs, AMD Radeon™ AI PRO R700 GPUs and AMD Ryzen™ AI Processors
Optimized ROCm Docker for Distributed AI Training
AMD updated Docker images incorporate torchtune finetuning, FP8 support, single node performance boost, bug fixes & updated benchmarking for stable, efficient distributed training