Senior AI Performance Engineer

About Modular

About the role:

What you will do:

• Design and implement high-performance attention kernels for prefill and decode, including variants such as MHA, GQA, paged attention, sliding-window/global attention, and fused FlashAttention-style algorithms.
• Build and optimize general AI kernels—including matrix multiplication, softmax, normalization, activation, embedding, reduction, and data-movement primitives—needed to bring up modern models on new accelerators.
• Contribute to the design and implementation of the Attention framework, with reusable interfaces that balance portability, composability, tunability, and hardware-specific optimization.
• Profile workloads, identify bottlenecks, establish performance models, and tune implementations against hardware limits and competitive baselines.
• Develop benchmarks, correctness tests, regression tests, and automated performance analysis to ensure kernels remain reliable and fast across model shapes and hardware generations.
• Partner with the rest of Modular teams to contribute to overall software stack improvements and share learnings.
• Work directly with hardware partners when needed to understand new architectures, validate toolchains, diagnose low-level issues, and influence hardware/software interfaces.
• Write design documents, participate in technical reviews, and share architecture and performance insights across the team.
• Mentor teammates and raise the engineering bar for kernel quality, performance methodology, and maintainable low-level software.

What you bring to the table:

• 5+ years of relevant industry or research experience in high-performance computing, AI kernel development, accelerator programming, compiler engineering, or a closely related field.
• Demonstrated experience writing and optimizing production-quality GPU or accelerator kernels, with a strong understanding of parallel algorithms, numerical behavior, memory access patterns, synchronization, and data movement.
• Strong knowledge of modern attention algorithms and their implementation tradeoffs, including the distinct performance characteristics of prefill and decode.
• Strong understanding of architectures beyond conventional GPUs—such as NPUs, DSPs, TPUs, or custom ASICs—and the ability to reason about unfamiliar compute units, memory hierarchies, and execution models.
• Experience with at least one heterogeneous programming model or kernel ecosystem such as CUDA, Triton, SYCL, OpenCL, HIP, or a vendor accelerator SDK.
• Strong performance-analysis skills and proficiency with relevant profiling, tracing, benchmarking, and debugging tools.
• Ability to read architecture manuals, compiler output, and low-level generated code; form hypotheses from data; and systematically close performance gaps.
• A collaborative, team-oriented approach; clear written and verbal communication; intellectual curiosity; and comfort owning ambiguous technical problems from investigation through productization.

Helpful but not required

• Knowledge of the Mojo programming language or experience authoring kernels and performance-sensitive libraries in Mojo.
• Knowledge of MLIR, LLVM, and related AI compiler technologies, including dialect design, lowering pipelines, code generation, scheduling, or autotuning.
• Experience implementing FlashAttention or other fused attention algorithms, custom attention operators, paged KV-cache kernels, or inference-specific attention optimizations.
• Experience with modern kernel DSLs and libraries such as CuTe/CUTLASS, Triton, Pallas, or similar systems.
• Familiarity with AI framework internals and operator integration in systems such as PyTorch, JAX, TensorFlow, vLLM, SGLang, or TensorRT-LLM.
• Experience bringing up a model or kernel library on a new hardware platform, including working through compiler, runtime, driver, and hardware constraints.
• Experience with performance modeling, autotuning, numerical validation, or benchmarking infrastructure.
• Familiarity with LLM architectures such as GPT and Gemma, including grouped-query attention, mixture-of-experts, speculative decoding, and long-context inference.
• An advanced degree in Computer Science, Electrical Engineering, Computer Engineering, or a related field.

What Modular brings to the table:

• Amazing Team. We are a progressive and agile team with some of the industry’s best engineering and product leaders.
• World-class Benefits. In order to attract the best, we need to offer the best. Your benefits package may include comprehensive healthcare coverage, retirement and savings programs, employee stock purchase opportunities, paid time off, wellbeing resources, family support programs, and learning and development opportunities. Please note that specific benefit packages may vary based on your location, you can read more about benefits offered by Qualcomm here.
• Competitive Compensation. We offer very strong compensation packages, including RSU grants. We want people to be focused on their best work and believe in tailoring compensation plans to meet the needs of our workforce.
• Team Building Events. We organize regular team onsites and local meetups in Los Altos, CA as well as different cities. Traveling 2-4 times a year is expected for all roles.