Member of Technical Staff – Training Performance Engineer

About the role

As a Performance Engineer on the Pre‑Training team, you will optimise the throughput and accelerator utilisation of Cohere’s large language models. You will work closely with researchers and engineers to remove bottlenecks and build tooling that makes training faster and more reliable.

Key responsibilities

• Design and implement high‑performance, scalable software for model training.
• Write low‑level CUDA and Triton kernels to maximise accelerator efficiency.
• Research, prototype and evaluate ideas on Cohere’s super‑compute and data infrastructure.
• Collaborate with researchers to understand architectural changes and their impact on training quality.

Required profile

• Extremely strong software engineering background.
• Proficiency in Python and ML frameworks such as JAX, PyTorch and XLA/MLIR.
• Experience writing GPU kernels using CUDA or Triton.
• Familiarity with large‑scale distributed training and transformer‑based autoregressive models.

Required skills

• Python
• JAX
• PyTorch
• XLA/MLIR
• CUDA
• Triton
• Distributed training
• Transformers

What we offer

• Weekly lunch stipend of $75/£75 (or local equivalent).
• Full health, dental and mental‑health benefits.
• RRSP matching, 401(k), pension scheme.
• 100 % parental‑leave top‑up for up to six months.
• Six weeks paid vacation (30 working days).
• Education and learning stipend for conferences, courses and coaching.
• Home‑office set‑up stipend and co‑working benefit for remote staff.