Sail builds the world's most efficient software for inference (processing LLM tokens) and agent hosting (cloud VMs). Together, our technologies allow our customers to deploy AI agents at large scale to do the most challenging work.
In this role, you'll own token processing down to the lowest layers of the stack. You'll optimize kernel performance, develop new request scheduling and parallelism strategies at the engine level, and help us use a heterogenous mix of hardware at max efficiency.
What you’ll do
- Modify and extend state-of-the-art inference engines like vLLM and SGLang.
- Understand every microsecond of GPU time spent during a forward pass. You'll be able to explain every kernel launch on an NSys profile.
- Design and implement exotic parallelism schemes to work with "interesting" hardware topologies.
- Write custom GPU kernels to excel in specific regimes, such as cascade attention.
What we’re looking for
- Strong understanding of core LLM mechanics, like KV cache, mixture-of-experts, prefill vs. decode phases.
- Interest in MLSys research - great ideas like speculative decoding and sparse attention come from research, that we need to follow closely.
- Familiarity with modern, tile-based GPU programming, e.g. Triton, CUTLASS, ThunderKittens, etc. Or an interest in learning these!
- Great interpersonal communication - please don't use LLMs to write prose. We desk-reject most LLM-generated cover letters and resumes.
#J-18808-Ljbffr