Performance Engineer, Kernels

About the role

As a member of Sarvam’s Performance Engineering team, you will own the kernel layer that powers our large‑scale LLM serving fleet. You will design, implement and ship custom CUDA, DSL‑based and hand‑written PTX kernels that close the performance gap left by stock libraries and directly improve production latency.

Key responsibilities

• Author and maintain high‑performance GPU kernels (CUDA, PTX, Triton) for attention and communication primitives.
• Analyze roofline plots with Nsight Compute, propose and apply optimizations that move production p99 latency.
• Collaborate with the inference team to integrate kernels into the serving stack and own the performance explanations.

Required profile

• 5+ years of experience in machine‑learning systems with at least 2 years authoring production CUDA kernels.
• Demonstrated kernel in production that outperforms a published baseline.
• Deep understanding of thread‑block sizing, shared‑memory layout, warp primitives, async copies and MMA selection.

Required skills

• CUDA kernel development (including async copies, TMA, MMA).
• PTX debugging and hand‑written insertion.
• CUTLASS / CuTe DSL for kernel modification.
• Nsight Compute performance analysis.
• Attention kernels (FlashAttention family, sliding‑window, sparse).
• Communication kernels (NCCL, NVSHMEM, custom collectives) – strong plus.