Skip to content

Design: ingest Triton model repositories (sonic-models) - #14

Draft
ngpaladi wants to merge 1 commit into
mainfrom
triton-repo-ingestion
Draft

Design: ingest Triton model repositories (sonic-models)#14
ngpaladi wants to merge 1 commit into
mainfrom
triton-repo-ingestion

Conversation

@ngpaladi

Copy link
Copy Markdown
Owner

Draft design PR — proposal, not implementation — for serving an existing Triton model repository, concretely the fastmachinelearning/sonic-models layout, with minimal hand-editing.

Doc: design/triton-model-repository-ingestion.md

What it covers

  • Mappings grounded in real sonic-models configs: Triton platform → nereid backend, and config.pbtxt fields → the core Contract (incl. TYPE_ dtype normalization and batch/dims semantics).
  • Version-directory resolution — nereid serves one artifact per model; Triton keeps numbered 1/, 2/ dirs under version_policy.
  • The real gaps present in the repo, each to fail fast with an actionable message: tensorflow_graphdef (nereid's TF is SavedModel-only), Triton-style python (model.py / TritonPythonModel), ensemble, and TYPE_STRING/TYPE_BYTES.
  • Three options — offline import/convert, native runtime ingestion, and a recommended hybrid (native config.pbtxtContract reader + version-aware loading, with an import escape hatch) — plus a phased plan and open questions.

Direct (supported) set

onnxruntime_onnx → onnx · pytorch_libtorch → torch · tensorflow_savedmodel → tensorflow.

The doc is placed under design/ (outside docs/), so it isn't part of the published docs site. Open questions at the end are the main things to weigh in on — model selection, device/concurrency source, native-vs-import, and gap-handling granularity.

🤖 Generated with Claude Code

Add a design doc for serving an existing Triton model repository (the
sonic-models layout) with minimal hand-editing. Covers the platform ->
backend and config.pbtxt -> Contract mappings, version-directory
resolution, the concrete gaps present in sonic-models (tensorflow_graphdef,
Triton-style python, ensemble, string dtype), three implementation
options with a recommended hybrid, a phased plan, and open questions.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant