Senior Software Engineer - Model Performance
Inference is a distributed GPU network for running AI models efficiently across a global infrastructure.
Funding
About Inference
Inference is a distributed GPU network for running AI models. Users can connect their devices to contribute computing power, access APIs to perform model inference, and monitor workloads through a dashboard. The platform offers an open infrastructure that helps scale machine learning workloads without relying on centralized servers.
Skills
About the Role
You will be responsible for making our inference stack as fast and efficient as possible. Your work spans from implementing known optimization techniques to experimenting with novel approaches, always with the goal of serving models faster and cheaper at scale. Your north star is inference performance: latency, throughput, cost efficiency, and how quickly new model architectures can be brought into production. You'll work across the full inference stack—from CUDA kernels to serving frameworks—to find and eliminate bottlenecks. This role reports directly to the founding team. You'll have the autonomy, a large compute budget, and technical support to push the limits of what's possible in model serving.
Requirements
- 2+ years of experience in ML systems, inference optimization, or GPU programming
- Strong proficiency in Python and familiarity with C++
- Hands-on experience with LLM inference frameworks (vLLM, SGLang, TensorRT-LLM, or similar)
- Deep understanding of GPU architecture and experience profiling GPU workloads
- Familiarity with LLM optimization techniques (quantization, speculative decoding, continuous batching, KV cache management)
- Experience with PyTorch and understanding of how models execute on hardware
- Track record of measurably improving system performance
Responsibilities
- Implement and productionize optimization techniques including quantization, speculative decoding, KV cache optimization, continuous batching, and LoRA serving
- Deep dive into inference frameworks (vLLM, SGLang, TensorRT-LLM) and underlying libraries to debug and improve performance
- Profile and optimize CUDA kernels and GPU utilization across serving infrastructure
- Add support for new model architectures, ensuring they meet performance standards before going to production
- Experiment with novel inference techniques and bring successful approaches into production
- Build tooling and benchmarks to measure and track inference performance across the fleet
- Collaborate with applied ML engineers to ensure trained models can be served efficiently
Benefits
- Equity
- Comprehensive benefits