Inference Runtime Performance Engineer — GPU Kernels

iFrame Corporation; San Francisco (CA); Remote; USD 220,000 - 360,000; Full time

Job summary

iFrame Corporation is hiring for a role on the runtime team to own end-to-end performance for multiple model families on our managed inference runtime. You will write fused CUDA/Triton kernels and work across tokenizer through KV cache and decoding, targeting H100/H200 and other accelerators.

With 5+ years in systems-level performance and a track record of speed-ups, you will design long-context primitives and publish external write-ups quarterly while sharing on-call duties with SRE and

Qualifications

• Five-plus years of systems-level performance work, with GPU kernel experience.
• Strong CUDA and at least one of Triton, CUTLASS, or HIP.
• Ability to read PTX and SASS when profiling demands it.
• Track record of measurable speed-ups in production or open source.

Responsibilities

• Profile real customer workloads with Nsight and internal tracer to turn bottlenecks into benchmarks.
• Write fused, attention-aware kernels in CUDA and Triton (and HIP) for H100, H200, B200, B300.
• Own one model family end-to-end from tokenizer to KV cache and decode paths on all accelerators.
• Design and ship next round of long-context primitives: paged KV, ring attention, sliding-window cache eviction.
• Co-author external write-ups each quarter: blog, paper, or kernel release.
• Carry runtime on-call rotation alongside SRE and customer engineering ~one week every six weeks.

Skills

CUDA

Triton

HIP

Kernels

Transformer internals

Tools

CUTLASS

PTX/SASS

NVIDIA tools

Job description

iFrame Corporation is hiring for a role on the runtime team to own end-to-end performance for multiple model families on our managed inference runtime. You will write fused CUDA/Triton kernels and work across tokenizer through KV cache and decoding, targeting H100/H200 and other accelerators.

With 5+ years in systems-level performance and a track record of speed-ups, you will design long-context primitives and publish external write-ups quarterly while sharing on-call duties with SRE and