Skip to main content

Developer Guide

The Developer Guide is for readers who need to understand or modify the source tree. It separates two questions:

  • Architecture: which component owns checkpoint resolution, TensorRT graph construction, bundle format, runtime dispatch, task execution, and evidence?
  • Extension: which files and tests form the smallest complete change for a model family, runtime strategy, optimized provider, or config schema?

Read by change type

ChangeStart hereThen follow
Understand the systemArchitecture OverviewUnits, build pipeline, runtime lifecycle, validation design
Add a modelAdd a Model FamilyModel-owned Python, C++ DSO, manifest, and E2E proof
Add native runtime behaviorAdd a Runtime StrategyUnique strategy ownership and runtime registration
Add platform specializationAdd an Optimized RuntimeExact profile, provider adapter, implementation DSO, and separate qualification
Add user-facing configurationAdd a Config SchemaMatching Python/C++ schema plus the owning consumer
Submit a contributionContributor QuickstartFocused validation and PR evidence

User-facing task instructions belong in User Guides. Progressive labs belong in Tutorials. Keep architecture and contribution mechanics here so neither path becomes a prerequisite for a normal user.

General runtime strategy

For one exact model × platform × configuration tuple, Model Connect selects one runtime owner. It can be the model family's native TensorRT runtime or an exact platform-specialized provider. Selection must not create two competing owners for the same qualified tuple or silently fall through after one provider claims it.

The native and platform-specialized paths are separate implementation boundaries. Users normally interact with the same bundle/task API; advanced developers use the provider/profile metadata and runtime lifecycle pages to understand which backend owns execution.