Skip to main content

Glossary

Inference concepts

TermPlain meaningIn this project
ModelA learned function mapping inputs to outputs.Architecture and trained weights released by a model author.
TrainingCreating/updating weights from data.Out of scope; Model Connect starts from a checkpoint.
InferenceRunning fixed weights on a request.A family Task implementation drives TensorRT engines.
CheckpointSaved model/config/tokenizer/processor files.A Hugging Face ID or local snapshot consumed by one family builder.
TensorTyped multidimensional numbers.Engine inputs/outputs such as token IDs, masks, logits, pixels, audio, or features.
Token and logitsNumeric text unit and raw next-token scores.A family tokenizer maps text; a family sampler selects from logits.
Prefill and decodeInitial prompt pass and repeated next-token passes.Family-owned text runtime orchestration.
KV cacheReused attention keys/values.Avoids recomputing all prior tokens on each decode step.
OracleTrusted expected behavior.Usually an official framework/reference invoked by the family E2E test.

Deployment concepts

TermPlain meaningIn this project
TensorRT engine planCompiled GPU execution artifact.Opaque bytes stored in family-owned bundle sections.
.bundleBuild/runtime handoff container.Format-1 header plus named sections; section semantics belong to the family.
FamilyUnit of model ownership.families/<family>/ contains support, build, runtime, dependencies, and validation.
TaskUser-visible behavior.Abstract interfaces such as text generation, transcription, segmentation, embedding, and forecast.
BackendEngine implementation.trt or optional trt_rtx, selected by the bundle header.
DSOLinux shared library loaded at runtime.Core, loader, one backend, and exactly one libtrtmc_model_<family>.so.
Runtime rootExplicit DSO directory.Required by every native execution command; no fallback search exists.
PrecisionNumeric representation.A build request such as FP32, FP16, or BF16 that the family validates.
QuantizationLower-precision graph/weights such as FP8.Entirely family-owned and qualified per exact checkpoint/path.
ABIBinary compatibility contract.Runtime DSOs and TensorRT plans must match their software/hardware cohort.

Project building blocks

TermCurrent role
support.pyDependency-free exact checkpoint matching and supported/default tasks.
model.pyOne plain family build(request, writer) implementation.
BuildRequestResolved typed build inputs passed to exactly one family.
BundleWriter / BundleReaderStreaming write and bounded read mechanics; no model semantics.
Family factorytrtmc_create_family in one family DSO, returning an abstract Task.
Engine APIBackend-neutral create/bind/enqueue contract used by family runtimes.
E2E manifestExact checkpoint, build, task, topology, selection, and testcase data owned by one family.
ThresholdFamily-owned numeric override that must remain behaviorally meaningful.

The former runtime-strategy registry, provider profiles, optimized-runtime descriptor, shared config schemas, and task-strategy dispatch were removed by PR #1093. Historical pages may mention them but current APIs do not.

TensorRT-Model-Connect is not a training framework, general serving cluster, or automatic converter for arbitrary Hugging Face repositories. Unsupported checkpoints need a complete family implementation and validation.