About the job
We’re looking for Research Engineers to help our world models train faster and run more efficiently, without compromising what they can do. You will profile, optimize, and rearchitect the systems that turn research ideas into models that run at scale and in real time — directly shaping what is computationally possible and, by extension, what capabilities we can build.
Responsibilities
- Optimize training throughput across large GPU clusters — improving MFU through custom kernels, mixed-precision strategies (FP8, BF16), memory-efficient attention, and activation checkpointing
- Design and maintain distributed training infrastructure: tensor parallelism, context parallelism, FSDP, and fault-tolerant multi-node setups
- Profile and accelerate inference pipelines for real-time multimodal generation — CUDA graph compilation, KV cache optimization, operator fusion, and latency reduction
- Optimize and scale our training infrastructure to improve efficiency and reliability
- Contribute to the entire stack, from low-level kernel optimizations to high-level model design
Qualifications
Minimum
- 4+ years of experience in systems engineering, ML infrastructure, or performance optimization for deep learning
- Familiarity with GPU kernel development (CUDA, Triton, CUTLASS) and distributed systems (NCCL, collective communication, model parallelism)
- Experience with ML framework internals (PyTorch, JAX) and mixed-precision / low-precision techniques (FP8, INT8)
- Experience building and operating large-scale training infrastructure, including fault tolerance and cluster orchestration
- Excitement about building AI that simulates the world — and making it performant enough to run in real time
Preferred
- Bonus if you have experience with torch’s compilation feature