DL Performance Software Engineer - LLM Inference

NVIDIA•Toronto, ON

3d•CA$135,000 - CA$220,000•Hybrid

About The Position

We are seeking highly skilled and motivated software engineers to join us and build AI inference systems that serve large-scale models with extreme efficiency. You’ll architect and implement high-performance inference software, optimize GPU kernels, drive industry benchmarks, and work with state-of-the-art research techniques to improve serving efficiency. You’ll collaborate across inference performance, kernels, training, large-scale serving, and research teams to push the frontier of accelerated computing for AI.

Requirements

Bachelor’s, Master’s, or PhD degree in Computer Science (CS), Computer Engineering (CE) or Software Engineering (SE).
5+ years of industry experience in software engineering or equivalent research experience.
Strong programming skills in Python and one of C/C++, Go, or Rust.
Solid CS fundamentals: algorithms & data structures, operating systems, computer architecture, parallel programming, software engineering, distributed systems, deep learning theories.
Knowledgeable and passionate about performance engineering in ML frameworks (e.g., PyTorch) and inference engines (e.g., vLLM and SGLang).
Familiarity with GPU programming and performance: CUDA, memory hierarchy, streams, NCCL; proficiency with profiling/debug tools (e.g., Nsight Systems/Compute).
Excellent debugging, problem-solving, and communication skills; ability to excel in a fast-paced, multi-functional setting.

Nice To Haves

Experience developing major features and optimizations for LLM inference engines (e.g., vLLM, SGLang).
Hands-on work with LLM inference and training runtimes (deploying LLMs to production, large-scale LLM pre-training and RL), ML compilers and DSLs (e.g., Triton, CuTe, MLIR/LLVM, XLA), GPU libraries (e.g., CUTLASS) and features (e.g., CUDA Graph, Tensor Cores).
Experience with speculative decoding training and runtime features: tree-structured drafting, parallel drafting, diffusion LLMs, DFlash, EAGLE.
Contributions to open-source projects and/or publications; please include links to GitHub pull requests, published papers and artifacts.

Responsibilities

Contribute features to vLLM that empower the newest models with the latest NVIDIA GPU hardware features and serving runtime algorithms.
Profile and optimize the inference framework (vLLM) with methods like speculative decoding, 5D Parallelism, and prefill-decode disaggregation.
Architect novel frameworks and runtime optimizations for inference infrastructure, benchmarking, and kernels.
Conduct and publish original research that advances the Pareto frontier in ML Systems; survey recent publications and find a way to integrate research ideas and prototypes into production-grade, open-source software.
Develop, optimize, and benchmark GPU kernels (both hand-tuned and compiler-generated) using techniques such as fusion, autotuning, and memory/layout optimization.