Performance Engineer - ML Training & CUDA Kernels

Cohere; New York (NY); Hybrid; USD 150,000 - 190,000; Full time

Benefits offered by this job

Lunch stipend

Health benefits

RRSP matching

Parental leave

Enrichment benefits

Vacation

Travel budget

Home office stipend

Job summary

Cohere is a leading security‑first AI company focused on scalable language model training and deployment. You will optimize training throughput and accelerator utilization within a multidisciplinary team of engineers and researchers.

The role emphasizes low‑level kernel development (CUDA, Triton) and distributed training strategies to advance model performance across the Cohere platform.

Qualifications

• Proven strong software engineering skills.
• Proficient in Python and ML frameworks (JAX, PyTorch, XLA/MLIR).
• Experience writing GPU kernels with CUDA, Triton.
• Experience with large-scale distributed training strategies.
• Familiarity with autoregressive models like Transformers.

Responsibilities

• Design and implement high-performance software for training scalable models.
• Analyze architectural implications on throughput and quality.
• Write low-level CUDA and Triton kernels for accelerator optimization.
• Research and prototype training infrastructure and tooling.
• Collaborate with researchers to push model performance.

Skills

Software engineering

Python

JAX

PyTorch

XLA/MLIR

CUDA

Triton

Distributed training

Transformers

Job description

Cohere is a leading security‑first AI company focused on scalable language model training and deployment. You will optimize training throughput and accelerator utilization within a multidisciplinary team of engineers and researchers.

The role emphasizes low‑level kernel development (CUDA, Triton) and distributed training strategies to advance model performance across the Cohere platform.