Senior Deep Learning Engineer (Cosmos World Foundation Models)

NVIDIA is hiring for its Cosmos team—a platform purpose-built for "Physical AI." While ChatGPT understands text, Cosmos understands the physical world (physics, gravity, motion) to train Autonomous Vehicles and Robots. Your job is to bring these massive World Foundation Models (WFMs) from research into efficient, production-grade systems. You sit at the intersection of Deep Learning and Systems, focusing heavily on Inference Optimization.

Key Responsibilities

• Inference Optimization: Improve inference speed for Cosmos models on GPU platforms (Latency & Throughput).
• Production Deployment: Carry out the production deployment of these models, ensuring they run efficiently in the real world.
• Profiling: Profile and analyze deep learning workloads to identify bottleneck kernels (Memory vs Compute bound).
• Collaboration: Work closely with research scientists and hardware experts to bridge the gap between theory and silicon.

Requirements

• Experience: 5+ years of experience with an MSc or PhD in CS/EE.
• Core Stack: Strong programming skills in Python and PyTorch.
• Inference Tech: Experience with Quantization and optimization frameworks like TensorRT, TensorRT-LLM, vLLM, or SGLang.
• Deep Learning: Strong background in Deep Learning fundamentals.

Nice to Have

• CUDA: Experience writing custom CUDA kernels.
• Deployment: Familiarity with Docker and Triton Inference Server.
• Model Types: Familiarity with Diffusion Models (likely used for video generation in Cosmos).
• Performance Tuning: Proven experience tuning GPU workloads for both inference and training.