Skip to content

Add litertlm_manifest.json - #6288

Merged
madskristensen merged 2 commits into
SchemaStore:masterfrom
john-rocky:litertlm-manifest
Sep 1, 2026
Merged

Add litertlm_manifest.json#6288
madskristensen merged 2 commits into
SchemaStore:masterfrom
john-rocky:litertlm-manifest

Conversation

@john-rocky

@john-rocky john-rocky commented Sep 1, 2026

Copy link
Copy Markdown
Contributor

litertlm_manifest.json sits at the root of a Hugging Face model repo that ships .litertlm bundles for the LiteRT-LM on-device runtime. It tells a client which file fits which device: the backends each file is verified on, what it requires, and how fast it measured, with conditions and provenance. I maintain the spec, the generator, and the reference readers, so the CODEOWNERS entry points at me for follow-ups.

Uptake

  • 27 model repos on the Hub ship one today (litert-community/*, mlboydaisuke/*), all produced by the same generator.
  • flutter_gemma_litertlm 1.6.0+ reads it for manifest-driven installs (LitertlmManifestResolver).

Schema notes

  • Draft-07. Every object the spec declares open keeps additionalProperties: true, so a 0.1.x manifest that adds optional fields still validates.
  • Backend and platform enums are closed: the reference readers type them as closed unions, and the resolver refuses a backend outside the list.
  • Speed fields accept a number or a string, matching the generator (a "lo-hi" range string is used when runs spread).
  • All 27 published manifests validate against this schema. Two of them are the positive tests, plus a minimal and a full 0.1.2 sample. Ten negative tests each fail on exactly one constraint.

Links

node ./cli.js check, npm run prettier, and codespell pass locally. Thanks for maintaining the catalog.

Deployment manifest for Hugging Face model repos that ship .litertlm
bundles for the LiteRT-LM runtime. Draft-07, mirrors the schema the
generator validates against (manifest_schema 0.1 line); positive tests
include two manifests as published on the Hub plus a minimal and a
full 0.1.1 sample; ten negative tests each fail on one constraint.
@github-actions github-actions Bot added the CI ".github/**/*" folder is updated (auto-generated by labeler action) label Sep 1, 2026
@github-actions

github-actions Bot commented Sep 1, 2026

Copy link
Copy Markdown
Contributor

Thanks for the PR!

This section of the codebase is owned by @madskristensen and @hyperupcall - if they write a comment saying "LGTM" then it will be merged.

john-rocky added a commit to john-rocky/hf-to-litertlm that referenced this pull request Sep 1, 2026
…strips every evidence key; schema submitted to SchemaStore

Additive on the 0.1 line: measured rows may carry load_s (engine load
time under the row's cache condition) and peak_memory_mb (peak resident
memory during the run) — both were already shipping in the wild.
--public now strips every `evidence` key under a variant, not only
measured rows. derive_capabilities warns and omits channels[].is_reasoning
on a builder whose proto predates Channel.is_reasoning_channel instead of
crashing. quality[] is documented as present-but-unvalidated.

The TypeScript reader types the two new optional fields. The Qwen3
example is refreshed to the manifest as published today (both reader
suites pass: TS 19, Dart 18). Published-repo counts updated to 27.

A draft-07 JSON Schema for this format is submitted to SchemaStore
(github.com/SchemaStore/schemastore/pull/6288) so editors validate
litertlm_manifest.json by file name.
@madskristensen

Copy link
Copy Markdown
Contributor

Thanks!

@madskristensen
madskristensen merged commit 488f797 into SchemaStore:master Sep 1, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CI ".github/**/*" folder is updated (auto-generated by labeler action)

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants