Agent workflows

Torch-TensorRT exposes versioned reports for preflight analysis, compilation, validation, and benchmarking. An assistant can inspect these reports without interpreting terminal output. The existing compile() return type and dryrun interface remain available.

Start with Direct Python APIs to use the Python APIs in a provisioned Torch-TensorRT environment. Core Torch-TensorRT has no agent framework or protocol dependency.

The first release targets the Dynamo ahead-of-time ExportedProgram path. Analysis lowers and partitions an isolated graph, and reports converter eligibility separately from the final TensorRT assignment. Constant folding can execute tensor operations on that copy; analysis requires the compiler’s Torch/CUDA/TensorRT environment. Eligibility does not guarantee a successful engine build or a speedup.

Validation and benchmarking use prepared inputs and an explicit stateless evaluation protocol. Persistence and orchestration policies remain the caller’s responsibility.

• Direct Python APIs
  • Analysis and compilation
  • Events and diagnostic files
  • Validate and measure compiled modules
• Evidence contracts and testing
  • Numerical semantics
  • Timing semantics
  • Report persistence
  • CPU and GPU checks