At PyTorch Conference 2026, held October 20-21 in San Jose, AMD showcased how it is advancing the open AI ecosystem across training and inference. Through continued contributions to the ROCm™ software, performance optimization, and emerging AI workloads, AMD is helping developers build more efficient, scalable AI applications while strengthening support for open-source innovation across the PyTorch community.
How AMD Enables Native PyTorch Development Across Platforms
AMD demonstrated broad PyTorch support across cloud, data center, workstation, and client platforms. Highlights included native PyTorch integration across ROCm software stack, TorchInductor with the MLIR native FlyDSL backend, reinforcement learning post training and rollout, and PyTorch inference on AMD XDNA™ NPU architecture. AMD Skills also provided hands on learning experiences to help developers move from prompts to performance.
Driving Performance Through Kernel and Framework Optimization
Performance was a major theme across the technical program. Topics included PerfModel for validation driven Triton kernel optimization, Grouped GEMM on the ROCm software, TorchAO kernels and MXFP8, FlexAttention and KV cache optimization, and inference performance analysis using TraceLens. Other work explored agentic workload optimization with Hyperloom and efficient masked diffusion language model inference.
Scaling AI Training and Inference on AMD Instinct GPUs
AMD highlighted progress in scaling demanding AI workloads, including MXFP8 pretraining across more than 1,000 AMD Instinct™ MI355X GPUs, distributed training with TorchTitan, rocSHMEM symmetric memory, and RCCL GPU initiated networking. Additional work covered MoRI and vLLM for expert parallel and disaggregated Mixture of Experts serving, large scale reinforcement learning training and rollout, and resilient multimodal inference routing.
AMD Technical Sessions and Poster Contributions at PyTorch Conference 2026
AMD engineers contributed through breakout sessions, lightning talks, and posters.
- Mahdi Ghodsi: Building Cost Efficient Mixture of Models on AMD GPUs with vLLM Semantic Router, exploring intelligent model routing to balance quality, cost, and latency.
- 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, covering device side communication primitives. View technical talk.
- 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.
Additional posters highlighted advances across AI training and inference, including:
- Grouped GEMM on AMD ROCm in PyTorch by Jagadish Krishnamoorthy. View poster.
- One Policy, Two Engines: Closing Training and Rollout Drift in DeepSeek V4 Flash RL by Xinyu Kang and Yuankai Chen. View poster.
- Efficient Inference of Masked Diffusion Language Models on AMD XDNA 2 NPU by Ravi Gupta and Rishi Teja Madduri. View poster.
- Inference Performance Analysis Using TraceLens by Deval Shah. View poster.
- MoRI + vLLM: Wide Expert Parallelism and RDMA KV Cache Transfer for Disaggregated MoE Serving on AMD by Shiksha Patel and Rishi Teja Madduri. View poster
Experience AMD AI in Action Through Booth Demonstrations
Beyond the technical sessions and posters, AMD showcased practical AI development experiences through hands on demonstrations at the conference booth. Attendees could explore local and cloud AI workflows through ROCm and AMD Skills, including Lemonade related developer workflows and model serving with vLLM. Other demonstrations highlighted local AI inference and fine tuning across AMD platforms, giving developers a closer look at building and running AI applications on local hardware.
The booth also featured Hyperloom for agentic workload optimization and live demonstrations of robotic reinforcement learning and on device world models. Together, these experiences illustrated how developers can experiment with AI workloads across client and data center platforms, from local model development to advanced agentic and robotics applications.
Recognizing Leadership in the PyTorch Community
Jeff Daily was honored as one of three program chairs for PyTorch Conference 2026. He served on the committee responsible for reviewing conference submissions and helping shape the technical program.
The conference highlighted the growing momentum behind open AI development and the expanding role of AMD in the PyTorch ecosystem. From native integration and kernel optimization to large scale training and inference, AMD teams are advancing performance, scalability, and developer flexibility across a wide range of AI workloads. Continued collaboration with the PyTorch community and investment in ROCm are helping make high performance AI development more accessible.
Ready to build with PyTorch on AMD?
Explore the AMD AI Developer Program to connect with the developer community and access AI development resources. Discover practical workflows through AMD AI Playbooks, and get hands on with AMD Skills to build and enhance your PyTorch development expertise.
Get started with the tools, resources, and learning experiences to bring your next AI project to life on AMD platforms.
