AMD Showcases AI Advancements at PyTorch Conference 2026

At the PyTorch Conference 2026, held October 20–21 in San Jose, AMD demonstrated its commitment to advancing the open AI ecosystem through innovations in training, inference, and performance optimization. The company emphasized its contributions to the ROCm™ software stack and its support for scalable, efficient AI applications across diverse platforms.

Expanding PyTorch Integration Across Platforms

AMD showcased native PyTorch integration across its ROCm software stack, enabling seamless development on cloud, data center, workstation, and client platforms. Key highlights included the MLIR-native FlyDSL backend for TorchInductor, reinforcement learning post-training and rollout, and PyTorch inference on AMD XDNA™ NPU architecture. Hands-on learning sessions through AMD Skills further empowered developers to optimize their AI workflows.

Optimizing Performance Through Kernel and Framework Enhancements

Performance optimization was a central theme of AMD's contributions. The company presented advancements such as PerfModel for Triton kernel optimization, Grouped GEMM on ROCm, TorchAO kernels with MXFP8, FlexAttention and KV cache optimization, and inference analysis using TraceLens. Other efforts included agentic workload optimization with Hyperloom and efficient masked diffusion language model inference.

Scaling AI Workloads with AMD Instinct GPUs

AMD highlighted its progress in scaling AI training and inference workloads. This included MXFP8 pretraining across over 1,000 AMD Instinct™ MI355X GPUs, distributed training with TorchTitan, and innovations like rocSHMEM symmetric memory and RCCL GPU-initiated networking. Additional advancements included MoRI and vLLM for expert parallelism and disaggregated Mixture of Experts serving, as well as large-scale reinforcement learning training and rollout.

Technical Contributions and Developer Engagement

AMD engineers delivered breakout sessions, lightning talks, and posters at the conference. Notable presentations included:

• Mahdi Ghodsi: "Building Cost-Efficient Mixture of Models on AMD GPUs with vLLM Semantic Router"
• Jiahui Cao: "Extending TorchInductor with FlyDSL: A New MLIR Native Backend for High-Performance GEMMs"
• Prachi Gupta: "rocSHMEM Symmetric Memory in PyTorch for AMD GPUs"
• Shekhar Pandey: "MXFP8/4 Pretraining at Scale with TorchTitan on AMD Instinct GPUs"
• Anshu Raina and Peyman Razaghi: "Primus Tuning: Hybrid Projection and Agentic Search for Distributed Training"

Poster contributions included topics such as Grouped GEMM on ROCm, efficient inference of masked diffusion language models, and wide expert parallelism with MoRI and vLLM.

Interactive Demonstrations at the AMD Booth

AMD’s booth offered hands-on demonstrations of AI workflows, including local and cloud-based development with ROCm and AMD Skills. Attendees explored model serving with vLLM, Hyperloom for workload optimization, and live demonstrations of robotic reinforcement learning and on-device world models. These experiences showcased AMD’s ability to support AI development across client and data center platforms.

Recognizing Leadership in the PyTorch Community

Jeff Daily, one of three program chairs for PyTorch Conference 2026, was recognized for his role in shaping the event’s technical program. AMD’s contributions underscored its growing influence within the PyTorch ecosystem, driven by its focus on performance, scalability, and open-source collaboration.