Open position
Inference optimization
Make every sampler step cheaper without losing quality. This work decides what each output costs us, and so what it costs our customers.
What you'll do
- Profile image and video generation end to end and remove the slowest parts.
- Lower the cost where quality allows it: precision, fewer steps, better schedulers.
- Write and tune kernels for the GPUs we run, and bring up new hardware.
- Prove each change on accepted outputs per server-hour, not on a single benchmark.
What you bring
- Experience optimizing deep-learning inference on GPUs.
- CUDA, Triton or similar, and a good feel for memory and compute limits.
- Care for quality, checking that a faster model still makes the same outputs.
Good to have
- Quantization or distillation of diffusion models.
- Contributions to inference engines or compilers.