Training Performance Engineer — AI Systems & CUDA

Cohere; United States; Remote; USD 175,000 - 325,000; Full time

Benefits offered by this job

Weekly lunch stipend

Health and dental benefits

RRSP matching / 401K / Pension

Parental leave up to 6 months

Enrichment benefits + learning stipend

6 weeks paid vacation

Travel budget to other offices

Home office stipend $500

Job summary

Cohere is seeking a Performance Engineer for the Pre-Training team to optimize training performance of its language models. You will focus on throughput, accelerator utilization, and cutting-edge tooling in a globally distributed setting.

The role blends software engineering with ML kernel design, requiring expertise in GPUs, Triton/CUDA and distributed training. Remote-friendly with time-zone alignment across ET to CET.

Qualifications

• Extremely strong software engineering skills.
• Proficiency in Python and ML frameworks such as JAX, PyTorch and XLA/MLIR.
• Experience writing kernels for GPUs using CUDA, Triton, etc.
• Experience using large-scale distributed training strategies.
• Familiarity with autoregressive sequence models, such as Transformers.

Responsibilities

• Design and write high-performant and scalable software for training.
• Understand architectural modifications and their effects on throughput and quality.
• Write low-level CUDA, Triton kernels to maximize accelerator performance.
• Research, implement, and experiment with ideas on supercompute and data infrastructure.
• Collaborate with researchers to advance state-of-the-art language models.

Skills

Python

ML frameworks (JAX, PyTorch, XLA/MLIR)

CUDA kernels

Triton

Distributed training

Transformers

Kernel design

Tools

CUDA

Job description

Cohere is seeking a Performance Engineer for the Pre-Training team to optimize training performance of its language models. You will focus on throughput, accelerator utilization, and cutting-edge tooling in a globally distributed setting.

The role blends software engineering with ML kernel design, requiring expertise in GPUs, Triton/CUDA and distributed training. Remote-friendly with time-zone alignment across ET to CET.