LinkedIn Uses AI Agents to Accelerate Liger Kernel Engineering, 10x Speedup on Training Encoder

As LLMs grow larger and more complex, GPU kernel efficiency has become a critical bottleneck for both training and inference performance. Most teams rely on PyTorch, but its standard operations leave significant performance on the table. LinkedIn built Liger Kernel, an open-source collection of optimized GPU kernels delivering 20% throughput improvement and 60% memory reduction across nearly 40 model architectures. To accelerate kernel development and optimization, LinkedIn built agentic workflows for Liger Kernel engineering. These agentic workflows follow a 3-stage pipeline with human review checkpoints: Understand (agent reads source code and produces structured profile), Act (agent generates or modifies files), and Verify (agent runs correctness checks and benchmarks). LinkedIn applied the same approach internally with torch.compile integration, achieving a 10x speedup on a training encoder. The agentic workflows reduced hours of expert time per kernel task to agent-driven automation with human review checkpoints.