- London, England, United Kingdom
- On-site
- Full-time · Staff
- Visa sponsorship
This role explores how model architectures and inference algorithms can be adapted to exploit heterogeneous hardware, creating specialised variants with different trade-offs between capability, latency, and cost.
Compensation
Competitive Salary, determined by skills and experience
Responsibilities
- Develop hardware-aware transformations that adapt models to different accelerator architectures and execution environments.
- Explore numerical, structural, and memory-related optimisations to create model variants with different capability, latency, and cost trade-offs.
- Investigate new inference algorithms, including speculative decoding, designed to exploit heterogeneous computational resources.
- Develop automated methods for discovering, characterising, and optimising hardware-specialised model variants.
- Build evaluation methods that measure model quality and establish the trade-offs introduced by hardware-aware adaptations.
Requirements
- Nice to have: mixture-of-experts
- Nice to have: diffusion
- Nice to have: JEPA
- Nice to have: alternative attention mechanisms
