extract-kernel-definitions

To extract GPU kernel schemas from SGLang models, run the extraction command with the target model name. This automates scanning the model implementation to detect kernel definitions like MLA, MoE, GQA, RMSNorm, and GEMM for FlashInfer trace.

Kernel definition deduplication prevents duplicate kernel schemas by consolidating shared definitions and tagging model usage. You need it when adding new SGLang models to maintain consistent, reusable FlashInfer trace definitions without redundancy.

Yes, you can generate Definition JSON files for MLA, MoE, GQA, RMSNorm, and GEMM kernels. The extraction process scans SGLang models and outputs JSON-based definitions suitable for flashinfer_trace and downstream testing.

Yes, the kernel definition extraction relies on the existing project structure and FlashInfer tests to validate references. This ensures the extracted kernel schemas are accurate and the generated definitions remain reusable across models.

The best way to handle duplicate kernel definitions is to run the extraction tool and then review the generated schemas. The tool automatically consolidates shared kernels and tags model usage to prevent redundant FlashInfer trace definitions.

FlashInfer trace definitions may not be reusable if deduplication was skipped during extraction. Running the extraction process ensures shared kernels are consolidated and tagged by model usage, maintaining consistent reusable definitions.