LinkedIn reports that a Triton kernel generated by an internal kernel generation agent, auto-inserted through a torch.compile-based graph pass, cut the encoder step time of an internal recommendation model from 400ms to 40ms (10x) in its internal evaluations. Average training step time fell from 1.12s to 0.39s (3x), and 64.7% of GPU hours were saved on the end-to-end training job.
The work grew out of Liger Kernel, LinkedIn's open-source collection of optimized GPU kernels, which delivers 20% throughput improvement and 60% memory reduction across nearly 40 model architectures. It is integrated with HuggingFace Transformers, TRL, and LLaMa-Factory, and has over 7 million downloads and 100+ contributors across companies. Creating a new kernel means implementing forward and backward passes in Triton, writing correctness tests, and validating numerical accuracy, while supporting a new model means understanding its architecture and validating across multiple configurations. Each of these tasks takes hours of expert time, and manual effort does not scale with the pace of model innovation.
LinkedIn built agentic workflows, often packaged as reusable agent skills, that follow a 3-stage pipeline with human review checkpoints. In the Understand stage the agent produces a structured profile that a human reviews, in the Act stage it generates or modifies files following project conventions, and in the Verify stage it runs correctness checks and benchmarks, with hard failures blocking progress. The liger-kernel-dev agent converts a PyTorch operation into a Triton kernel and classifies it into one of three complexity tiers; for ReLU squared (PR #1171) the kernel achieved 1.9x forward speedup, 3.2x backward speedup and 37.5% memory reduction versus PyTorch. The liger-autopatch agent resolves 12 architectural decisions from a HuggingFace modeling file to add model support, and was used for Nemotron (PR #1165) and Ministral (PR #1166), both requiring only human review before merging with no manual code changes. The liger-kernel-perf agent profiles kernels, optionally runs NVIDIA NCU profiling, and generates versioned optimization variants; on the fused_add_rms_norm backward kernel (PR #1187) it delivered a 3.35x backward speedup at hidden dimension 16384 and a 59% full-pass speedup with no regression across all 40 tests.
Inside LinkedIn's training infrastructure, the team built an internal compiler-based kernel selection and replacement library that extends torch.compile. It captures the traced graph via TorchDynamo, identifies operations that can be fused, gets the best kernel from a kernel registry, and replaces the operations through a custom graph pass in torch fx that is auto-generated by agent skills. The kernel behind the 10x result computes a batched partitioned mean pooling in one GPU launch. LinkedIn says it is building further workflows, liger-debug and liger-chunked-loss, to cover the rest of the kernel engineering lifecycle.