Cohere
Member of Technical Staff, Training Performance Engineer
$250K–$535K
London +4 · Hybrid · Staff · Full-time · Posted Feb 20, 2025
- Location
- London
New York
Paris
Toronto
Montreal
About the job
Optimizes language-model training performance by improving throughput and accelerator utilization through software, kernel, profiling, and systems work.
What you'll do
- Design scalable, high-performance training software and assess how architectural choices affect throughput and quality.
- Develop low-level CUDA and Triton kernels and identify and remove performance bottlenecks.
- Research, implement, and test ideas using supercomputing and data infrastructure.
What they ask for
- Extremely strong software engineering skills.
- Proficiency in Python and ML frameworks such as JAX, Pytorch, and XLA/MLIR.
- Experience writing GPU kernels using CUDA, Triton, or similar.
- Experience with large-scale distributed training strategies and familiarity with autoregressive sequence models such as Transformers.
Nice to have
- Paper at a top-tier venue, such as NeurIPS, ICML, ICLR, AIStats, MLSys, JMLR, AAAI, Nature, COLING, ACL, or EMNLP.
Pay
CA$250K – CA$535K (total pay)
Tools
Python, JAX, Pytorch, XLA/MLIR, CUDA, triton, Transformers
AI focus
Optimizes the performance of foundation language-model training, including throughput and accelerator utilization.
Model-written summary · the original posting is authoritative