Senior Software Engineer - AI Inference Performance

About the role

NVIDIA is a technology company building GPU-accelerated platforms for AI applications. The Senior Software Engineer will advance LLM and VLM inference performance across models, serving software, distributed runtimes, communication, CUDA kernels, and GPU architecture, delivering measurable improvements in latency, throughput, efficiency, and scale. The role also involves profiling workloads, optimizing kernels and serving configurations, establishing repeatable benchmarks, and collaborating across engineering teams.

Responsibilities

• Lead end-to-end analysis of LLM/VLM inference processes.
• Define representative prefill and decode workloads.
• Optimize time to first token, inter-token latency, P99 end-to-end latency, processing efficiency, and key-value (KV) cache capacity.
• For multimodal models, isolate preprocessing, encoder, and decoder costs.
• Build speed-of-light and roofline models to quantify performance headroom.
• Connect arithmetic intensity, bandwidth, occupancy, memory hierarchy, and communication costs to clear optimization hypotheses.
• Profile workloads using NVIDIA Nsight Systems, Nsight Compute, PyTorch Profiler, and custom instrumentation.
• Eliminate bottlenecks in host code, CUDA kernels, memory, communication, and scheduling.
• Tune serving hyperparameters and techniques such as batching, KV-cache management, quantization, speculative decoding, CUDA Graphs, and model parallelism.
• Choose them based on workload, hardware, model quality, and service-level objectives.
• Build and optimize performance-critical kernels, including attention, matrix multiplication, mixture-of-experts routing, quantization, and data movement.
• Use CUDA, CUTLASS, Triton, or related technologies.
• Establish repeatable benchmarks, canonical run records, and performance regression gates.
• Manage aspects such as model, precision, hardware, topology, software, features, and workload; Balance between performance and accuracy.
• Collaborate across with various teams and contribute high-quality upgrades to TensorRT-LLM, vLLM, SGLang, or associated projects.