Training Performance Engineer Cuda

Optimize frontier language-model training with CUDA and Python at Cohere, working remotely or from offices aligned with ET to CET hours. Improve throughput and accelerator utilization across advanced AI systems.As a Performance Engineer on Cohere’s Pre-Training team, you will optimize model and system performance. The work combines software engineering, machine learning, and low-level kernel development. You will identify bottlenecks, build profiling tools, and improve efficient language understanding and generation capabilities.Key Responsibilities:
Design scalable, high-performance software for model trainingAssess architectural choices affecting throughput and qualityWrite CUDA and Triton kernels for acceleratorsResearch ideas across supercompute and data infrastructureCollaborate with researchers in the fieldRequirements:
Extremely strong software engineering skillsProficiency in Python, JAX, PyTorch, and XLA/MLIRExperience writing GPU kernels using CUDA or TritonExperience with large-scale distributed training strategiesFamiliarity with autoregressive models such as TransformersBring CUDA, distributed training, and machine-learning systems expertise to Cohere’s Pre-Training performance engineering work.
#J-18808-Ljbffr