Senior Software Engineer - AI Inference Performance

NVIDIA•Us, CA

About The Position

NVIDIA is seeking a Senior Software Engineer – AI Inference Performance to advance innovative LLM and VLM inference. You will push workloads toward practical performance limits on NVIDIA GPU-accelerated systems. Your work will span models, serving software, distributed runtimes, communication, CUDA kernels, and GPU architecture. Deliver measurable gains in latency, throughput, efficiency, and scale. This is a hands-on role for an engineer who turns performance models and profiler data into working code. You will collaborate with model, framework, kernel, networking, and GPU architecture teams. You will contribute improvements to open-source inference engines and develop methods that others can reproduce. Your work will improve production deployments and help build future NVIDIA platforms.

Requirements

More than 6 years of experience in full-stack LLM/VLM inference performance involving models, serving, distributed runtimes, kernels, and hardware.
Your efforts result in measurable gains in production or production-representative environments.
Strong programming skills in Python, Rust and/or C++, plus hands-on experience with CUDA or another GPU programming environment.
Demonstrated expertise in speed-of-light analysis, roofline models, microbenchmarks, and tools including NVIDIA Nsight Systems and Nsight Compute.
You convert profiles into testable hypotheses and validated progress.
Deep understanding of GPU architecture, including Tensor Cores, memory hierarchy, caches, occupancy, synchronization, and numerical formats across hardware generations.
Practical experience optimizing inference servers and model execution.
You can choose techniques for the workload, including batching, scheduling, KV-cache management, quantization, speculative decoding, and various parallelism strategies
Understanding of distributed systems and networking for accelerated computing.
You can reason about collectives, topology, and scale-up versus scale-out performance.
BS or MS in Computer Science, Computer Engineering, or a related field, or equivalent experience.

Nice To Haves

Contributions to one or more high-performance AI projects. Examples include TensorRT-LLM, vLLM, SGLang, PyTorch, CUDA, Triton, or NCCL.
Experience developing AI-agent-supported performance workflows that automatically gather and analyze profiles, identify bottlenecks, explore serving configurations, or produce optimized runtime and kernel code.
You validate generated changes through reproducible, human-reviewed tests for performance, model quality, and correctness.
Published research, conference presentations, technical talks, or blog posts that clearly explain inference performance methods and results.
Delivered advancements for new LLM or VLM architectures, long-context inference, mixture-of-experts models, multimodal pipelines, or large-scale distributed serving.

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 techniques 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.