TensorRT-Model-Connect AGENTS.md

Agent Instructions

Repository Target

• Treat GitHub as the active repository for this project: https://github.com/NVIDIA/TensorRT-Model-Connect.git.
• Use the local github remote for fetch, push, PR, and CI operations.
• Do not push project changes to any non-GitHub legacy remote unless the user explicitly asks for that repository.

Branch And PR Flow

• The GitHub default branch is main.
• Do not push directly to GitHub main.
• Start new work from github/main on a short-lived branch.
• Push the branch to the GitHub remote and open a pull request targeting main.
• Wait for GitHub CI before merging.
• Merge with squash or rebase, matching the repository ruleset.
• Sign off every commit introduced by a pull request with DCO using git commit --signoff; ensure the Signed-off-by email matches the commit author.
• Preserve or re-add valid sign-offs when amending, rebasing, or cherry-picking commits, and never sign off on another author's behalf.
• Avoid commit messages containing Claude; the GitHub ruleset rejects them.

GitHub Pages

• Keep GitHub Pages dedicated to the documentation website.
• Do not publish CI reports to GitHub Pages unless the user explicitly changes that decision.

Model-Family Ownership And Execution Philosophy

• The model family is the unit of ownership, fault isolation, and horizontal scale. One team or agent must be able to implement, validate, change, and revert one family without changing or coordinating with another.
• Each model family owns all model-specific code: configuration and weights, mathematical topology, engine composition, runtime orchestration, and model validation.
• Duplicate model-specific code by design. Duplication keeps model-specific defects, merge conflicts, and rollbacks family-local; code similarity never justifies a cross-family abstraction.
• Shared code is limited to model-agnostic contracts and mechanics. It contains no model topology, model semantics, model orchestration, model-specific validation logic or evidence, or family-specific behavior.
• Generic family code defines only the mathematical computation graph and runtime orchestration. TensorRT owns all lowering to GPU execution, including fusion, tactic and kernel selection, scheduling, code generation, and hardware or version adaptation.
• Generic family semantics do not depend on GPU, SM, CUDA, driver, or TensorRT version. Platform-specific failures are project topology or orchestration defects, or upstream TensorRT issues.
• Platform specialization is limited to complete network offload, such as TensorRT Edge-LLM, and TVM-FFI BYOK kernel bindings. Target-specific plans, timing caches, and compiled kernels do not make the model family specialized.
• Existing family-owned GPU helper kernels outside TVM-FFI BYOK violate this architecture and are temporary migration debt, not a supported specialization.