ad-add-fusion-transformation

To add a TensorRT-LLM fusion pass, implement the transform under transform/library/ based on observed SSA graph dumps, register it in default.yaml, and add unit tests to reliably rewrite targeted graph patterns during deployment.

You validate AutoDeploy fusion transforms by using graph dump evidence to interpret match counts before and after execution. This dump-based validation confirms the transform correctly rewrites the targeted SSA graph structure without regressions.

You should use existing kernel paths for TensorRT-LLM custom ops before falling back to Triton. This existing-kernel-first strategy validates reuse decisions to ensure reliable kernel integration during fusion transformations.

You register a fusion transformation in the TensorRT-LLM model registry by updating default.yaml and enabling model-registry rollout with guardrails. This prevents unproven fusions from deploying while ensuring metadata is preserved.

To prevent regressions when adding TensorRT-LLM fusion passes, you need transform unit tests and integration tests. These tests align with dump-based validation and match-count interpretation to ensure graph pattern stability.