Performance Engineer, Kernels

Overview

Performance Engineer (Kernels) in Sarvam AI’s Performance Engineering team, focused on authoring custom GPU kernel code to improve production performance.

What you'll do

• Own the kernel layer performance.
• Author custom CUDA, DSL-based, and PTX kernels when stock libraries fall short.
• Deliver kernel changes that move production p99 and provide explanations for the improvements.
• Create kernels that beat published baselines on real workloads.
• Work with attention kernel implementations, including modifications or authorship.

What you'll need

• 5+ years in ML systems.
• 2+ years authoring production CUDA kernels.
• Have a production kernel that beats the prior baseline by a measurable margin.
• CUDA kernel-authoring depth covering thread-block sizing, shared memory layout, warp primitives, async copies (cp.async, TMA), and MMA selection.
• CUTLASS/CuTe DSL experience at a modify-and-extend level with comfort in layout algebra.
• PTX debugging/modification experience, including inserting hand-written PTX when needed.
• Nsight Compute and systems fluency to interpret roofline plots and propose fixes.

Details

• Location: Bengaluru.