1 year ago
Responsibilities
- Write high-performance GPU kernels for novel model architectures.
- Integrate kernels into PyTorch pipelines through custom operations, extensions, dispatch, and benchmarking.
- Profile and optimize training and inference workflows to eliminate bottlenecks.
- Build correctness tests and numerical checks.
- Build and maintain performance benchmarks and guardrails to prevent regressions.
- Collaborate with researchers to turn promising ideas into shipped, maintained performance improvements.
- Own the process from profiling ambiguous bottlenecks through kernel or integration changes, benchmarked results, and ongoing maintenance.
Requirements
- Authored custom CUDA kernels beyond merely calling cuDNN or cuBLAS.
- Strong understanding of GPU architecture and performance, including memory hierarchy, warps, shared memory, register pressure, bandwidth versus compute limits, occupancy, and tensor core utilization.
- Proficiency with low-level profiling using Nsight Systems and/or Nsight Compute.
- Strong C/C++ skills.
- CUTLASS experience and tensor core utilization strategies are preferred.
- Triton kernel experience and/or PyTorch custom operation integration is preferred.
- Experience building benchmark harnesses and performance regression tests is preferred.
Benefits
- Open to locations beyond preferred San Francisco and Boston.
- 100% of medical, dental, and vision premiums covered for employees and dependents.
- 401(k) matching up to 4% of base pay.
- Unlimited PTO and company-wide Refill Days.
- Competitive base salary with equity.
About Liquid AI
Liquid AI builds general-purpose AI systems that run efficiently from data center accelerators to on-device hardware, emphasizing low latency, memory efficiency, privacy, and reliability. The company partners with enterprises in consumer electronics, automotive, life sciences, and financial services to deploy and benchmark models for real-world workloads. Founded in 2023 out of MIT CSAIL and headquartered in Cambridge, Massachusetts, it is privately held.