Responsibilities
THE ROLE:
Drive the performance of post-training workloads on AMD Instinct™ GPUs. You’ll work across kernels, distributed training, compiler and framework integrations, and AI-assisted engineering workflows to deliver fast, stable, and reproducible fine-tuning and RL training pipelines on ROCm.
This role is centered on hands-on performance engineering, with the added expectation that you can apply agentic and AI-assisted workflows to accelerate profiling, root-cause analysis, experiment orchestration, regression triage, and optimization at scale.
THE PERSON:
The ideal candidate is passionate about software engineering and the craft of training performance. You lead sophisticated cross-stack issues—spanning data pipelines, kernels, distributed training, runtime behavior, and compilers—to clear resolution. You communicate crisply and collaborate effectively with framework, compiler, kernel, and model teams across AMD, driving measurable improvements with rigor, ownership, and reproducibility.
You are also comfortable using AI-assisted and agentic workflows to improve engineering productivity: automating repetitive analysis, generating performance hypotheses, accelerating regression triage, and helping teams move faster while maintaining strong validation standards.
KEY RESPONSIBILITIES:
- Lead performance for fine-tuning and RL training solutions on AMD GPUs.
- Improve throughput, memory efficiency, and stability across data, model, optimizer, and runtime steps.
- Optimize multi-GPU and multi-node training performance, including communication and scaling behavior.
- Contribute efficient kernels/ops and targeted graph-level or runtime-level optimizations where they deliver the highest impact.
- Profile, diagnose, and resolve bottlenecks using standard tooling; prevent regressions in CI and benchmarking pipelines.
- Build and apply agentic or AI-assisted workflows to accelerate profiling analysis, experiment generation, regression triage, and performance debugging.
- Develop scalable tooling and automation to improve reproducibility, performance reporting, and optimization velocity across teams.
- Ship reproducible pipelines and documentation adopted by internal teams and external developers.
- Collaborate with framework, compiler, kernel, and model teams to land durable performance improvements.
- PREFERRED EXPERIENCE: Proven GPU performance engineering for deep learning workloads (ROCm/HIP, Triton, or similar).
- Hands-on experience with SFT, LoRA, and RL-based training at scale.
- Strong PyTorch experience, including torch.distributed
- , FSDP/ZeRO, or equivalent distributed training approaches.
- Proficient in Python and C++; comfortable reading and writing kernels when needed.
- Experience with distributed systems and collective communication libraries.
- Track record of turning profiles into fixes, upstreaming changes, and documenting measurable results.
- Experience building or using AI-assisted developer workflows, agentic tooling, or automation for performance analysis, experiment management, or engineering productivity is a plus.
- Ability to balance deep technical investigation with practical automation and cross-functional execution.
- ACADEMIC CREDENTIALS: B.S., M.S., or Ph.D. in Computer Science, Computer Engineering, Electrical Engineering, or equivalent.
- LOCATION: San Jose, CA preferred. Other U.S.-based locations may be considered.
Qualifications
