About the Role
NVIDIA is seeking exceptional software engineers to join the TensorRT-LLM team. This group is responsible for defining how generative AI performs at a global scale on NVIDIA GPUs. In this role, you will be crucial in building and optimizing the core inference infrastructure for large language models, focusing on maximizing throughput, memory efficiency, and scalability of modern model runtimes. Your contributions will directly influence the frameworks behind state-of-the-art LLM inference used across NVIDIA and the broader AI community, pushing the boundaries of what is considered fast for LLM inference.
What you'll be doing:
- Design, implement, and optimize high-performance inference pipelines for large language models running on GPUs.
- Profile and tune model execution across the entire stack, from scheduler design to kernel fusions.
- Design and experiment with memory management strategies to improve memory bandwidth optimization and cache efficiency.
- Innovate and implement cutting-edge techniques such as Speculative Decoding, Context Caching, and FP8/INT4 quantization to enhance tokens-per-second-per-watt performance.
- Develop and maintain benchmarking and testing systems to quantify latency, utilization, and efficiency.
What we need to see:
- Bachelor's, Master's, or higher degree in Computer Engineering, Computer Science, Applied Mathematics, or a related computing-focused degree (or equivalent experience).
- 5+ years of relevant software development experience.
- Excellent Python programming skills, along with strong software design and engineering capabilities.
- Experience working with deep learning frameworks like PyTorch and HuggingFace.
- Proficiency in profiling and debugging performance at all levels, including Python runtime, PyTorch internals, and GPU utilization metrics.
- Awareness of the latest developments in LLM architectures and inference techniques.
- Proactive work ethic and ability to work without direct supervision.
- Excellent written and oral communication skills in English.
Ways to stand out from the crowd:
- Contributions to inference frameworks such as TensorRT-LLM, vLLM, SGLang, or similar systems.
- Demonstrated expertise in performance modeling, memory optimization, distributed model execution, or GPU execution workflows.
- Hands-on experience with NVIDIA profiling tools (Nsight Systems, PyTorch Profiler, custom benchmarking harnesses).
- A strong grasp of the trade-offs that influence inference efficiency: compute vs. memory, scheduling vs. batching, latency vs. throughput.