ad-add-fusion-transformation

What problem does it solve?

This Skill helps you implement or extend TensorRT-LLM AutoDeploy fusion transformations so targeted graph patterns are rewritten correctly and reliably during deployment.

Core Features & Use Cases

• Graph-pattern-driven fusion work: Uses graph dumps and evidence to design a minimal fusion transform that actually matches the current SSA graph structure.
• Existing-kernel-first strategy: Recommends and validates reuse of existing kernels/custom ops before falling back to Triton (only when needed).
• Registration and rollout readiness: Guides you through transform implementation, default.yaml registration, and model-registry enablement with guardrails to avoid unproven fusions.
• Validation and test alignment: Emphasizes dump-based validation, match-count interpretation, and the unit/integration tests needed to prevent regressions.

Quick Start

Use the skill to implement a new AutoDeploy fusion transform under transform/library/ for a pattern confirmed by graph dumps, then register it in default.yaml and add the corresponding unit tests.