Principal Software Engineer – PyTorch Training Frameworks

WHAT YOU DO AT AMD CHANGES EVERYTHING   At AMD, our mission is to build great products that accelerate next-generation computing experiences—from AI and data centers, to PCs, gaming and embedded systems. Grounded in a culture of innovation and collaboration, we believe real progress comes from bold ideas, human ingenuity and a shared passion to create something extraordinary. When you join AMD, you’ll discover the real differentiator is our culture. We push the limits of innovation to solve the world’s most important challenges—striving for execution excellence, while being direct, humble, collaborative, and inclusive of diverse perspectives. Join us as we shape the future of AI and beyond.   Together, we advance your career.

AMD is looking for a Principal-level PyTorch training framework expert to help drive performance, scalability, and correctness of large-scale AI training on AMD Instinct™ accelerators. You will work at the intersection of PyTorch internals, distributed training, and hardware-aware optimization, partnering closely with compiler, kernel, driver, and architecture teams to deliver industry-leading training performance and developer experience.       THE PERSON:   The ideal candidate is deeply hands-on with PyTorch training and thrives on solving complex systems problems (performance, scaling, memory efficiency, distributed communication). You bring strong technical leadership, can influence architecture across teams, and are comfortable driving ambiguity to crisp execution. You communicate clearly with both engineers and stakeholders and can represent AMD credibly in upstream/open-source discussions.

Act as a technical authority for PyTorch training at AMD, setting direction for performance, scalability, and reliability Drive optimization of key PyTorch training workloads (LLMs/foundation models) across single-node and multi-node systems Improve and debug training performance in areas such as DDP/FSDP, gradient checkpointing, mixed precision, memory planning, and communication/computation overlap Partner with ROCm compiler/runtime, kernel, and driver teams to resolve performance bottlenecks and correctness issues across the full stack Contribute to and influence upstream PyTorch (design discussions, code contributions, performance fixes, CI/debug) Develop and maintain representative training benchmarks, profiling workflows, and performance regression detection for key models Lead deep-dive investigations of performance regressions and hard correctness issues; drive cross-team resolution to closure Mentor engineers and raise the bar on framework-quality code, performance engineering practices, and technical rigor Engage with strategic customers/partners on training enablement, root-cause analysis, and best-practices for AMD platforms   PREFERRED EXPERIENCE:   Deep experience with PyTorch internals and training systems (Autograd, optimizers, dataloading, compilation paths, runtime behavior) Strong distributed training expertise: DDP, FSDP, tensor/pipeline parallel concepts, collectives (NCCL/RCCL), multi-node debugging Proven track record in performance engineering (profiling, tracing, kernel/runtime analysis, memory optimization, scaling studies) Strong programming skills in Python and C/C++ (ability to land clean, maintainable changes in large codebases) Familiarity with PyTorch ecosystem components such as TorchInductor / torch.compile, Triton, CUDA/HIP-style programming models, and performance tooling Experience working across OS/hardware boundaries in Linux-based environments (containers, CI, drivers/runtimes are a plus) Clear technical communication: design docs, code reviews, stakeholder updates, and cross-team coordination Demonstrated ability to lead through influence (principal-level impact, mentoring, and architectural decision-making)   ACADEMIC CREDENTIALS:   Bachelor’s or Master's degree in Computer Science, Computer