Skip to main content

Conversion Outputs and Diagnostics

Overview​

This repository is a converter plugin package. Its job is to accept a supported input model, produce a conversion-stage artifact, and hand that artifact to a later MLIA backend. That means the important outputs here are conversion results and diagnostics, not final target metrics.

What this repo contributes​

In a successful workflow, this repo contributes:

  • The expected intermediate artifact for the selected converter route.
  • Conversion-stage logs or diagnostics during the MLIA run.
  • Artifacts that downstream backends can consume for further analysis.

What success looks like​

A successful conversion result usually means:

  1. The torch.nn.Module, .pt2 model, or .pte artifact is accepted.
  2. The converter produces the expected intermediate artifact for the selected route.
  3. A downstream backend consumes that artifact.
  4. The final user-facing metrics come from the downstream backend.

What this repo does not own​

This repo does not define the final performance or target-level compatibility metrics for a run. If you are looking for values such as cycle counts, memory figures, or target-specific compatibility summaries, those belong to the backend that runs after conversion.

Useful signals to look for​

The most useful signals in this repo are usually:

  • Whether the module, .pt2 model, or .pte artifact was accepted and converted.
  • Whether the selected route produced the expected artifact shape described in conversion_flow.md.
  • Whether the produced artifact was handed to the next backend.
  • Whether operator mapping or lowering failed during conversion.
  • Whether delegation or delegate extraction failed.
  • Whether artifact writing failed.
  • Whether the failure happened during conversion or after conversion completed.

How to interpret failures​

If a run fails before target metrics appear, treat the converter output as part of the diagnostic trail. If conversion succeeds and later analysis fails, the issue is more likely to belong to the downstream backend or target flow.