Bundle Format

A bundle contains eight magic bytes, an unsigned little-endian 64-bit JSON header length, the UTF-8 JSON header, and concatenated named sections.

{
  "format": 1,
  "family": "gpt2",
  "task": "text_generation",
  "backend": "trt",
  "sections": {
    "runtime.json": {"offset": 0, "length": 42},
    "engine.plan": {"offset": 42, "length": 1234}
  }
}

BundleReader validates this exact shape and every section bound when it is constructed. It owns the normalized bundle path and immutable section table, then reads a requested section directly from the file. It has no write API and does not eagerly load section contents.

The core does not interpret family sections or compute section hashes. Only format 1 is supported.

Build provenance

Section names starting with trtmc. are reserved for the core; BundleWriter rejects them from families. When a build records its request with BundleWriter.record_build(request), as trtmc build and family-owned build commands do, finish adds one trtmc.provenance.json section. Readers that predate it ignore the extra section, so the header and runtime contract are unchanged.

{
  "version": 1,
  "producer": {
    "name": "tensorrt-model-connect",
    "version": "0.1.0",
    "source_revision": "<40-character Git commit or null>",
    "source_dirty": false
  },
  "checkpoint": {"hf_id": "openai-community/gpt2", "revision": "<40-character commit>"},
  "build": {"precision": "fp16", "max_sequence_length": 256, "tensor_parallel_size": 1},
  "toolchain": {"tensorrt_version": "10.16.1.11", "sm": 80, "cpu_arch": "x86_64"}
}

producer.source_revision is the checkout commit of the running builder, and source_dirty reports modified tracked files. Both are null for an installed package without its own checkout.
checkpoint identifies the primary model directory when it is a Hugging Face cache snapshot and is null otherwise. Additional checkpoints that a family combines are not listed.
build contains the shared request fields that the family request defines, such as precision, sequence length, parallelism, quantization, and FP32 layers. Family-specific inputs stay in family sections.
toolchain records the TensorRT builder version and GPU SM when the build imported TensorRT.

Every field is best effort: information that cannot be determined is recorded as null instead of failing the build. The record contains no paths, host names, or timestamps. It is unsigned metadata, not a tamper-evidence mechanism.

Build provenance​

Build provenance