Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.
Python calibrationclassificationdecision-modelhuggingfacejevmodernbertmultilingualnlppythonpytorchroutingtyped-decisionszero-shot
docs(docker): clarify conditional ModelScope digest verification
gqchend717361eebc7
diff --git a/docker/prefetch_modelscope.py b/docker/prefetch_modelscope.pyindex 95ca680c..b986542f 100644--- a/docker/prefetch_modelscope.py+++ b/docker/prefetch_modelscope.py@@ -9,13 +9,13 @@ way the Hub would have, and every code path that names a checkpoint keeps its idrevision handling: `Agent(...)`, the `Router` laya-serve builds, `laya.cli`, the integrations.Only the standard library is used, because an image build has no reason to install the-ModelScope SDK for a handful of HTTP calls. Every file is streamed with resume and its size and SHA-256-are checked against the values the repository API reports, so a truncated or substituted-download fails the build instead of shipping inside a layer that looks healthy.+ModelScope SDK for a handful of HTTP calls. Every file is streamed with resume and its size is+checked against the value the repository API reports. SHA-256 is checked only when the API+publishes a digest; without one, a same-size substitution cannot be detected.Usage inside a Dockerfile RUN:- python prefetch_modelscope.py --models convaiinnovations/laya:multilingual \+ python prefetch_modelscope.py --model convaiinnovations/laya:multilingual \--revision master --cache-dir "$HF_HOME/hub"Each spec is `repo[:subfolder]`; the subfolder selects one checkpoint from the bundleddiff --git a/docs/docker.md b/docs/docker.mdindex 3aafb4b8..458664e6 100644--- a/docs/docker.md+++ b/docs/docker.md@@ -204,10 +204,9 @@ snapshot. Nothing else changes: `Agent`, the `Router` that `laya-serve` builds,`laya.cli` and the integrations keep their repo ids and resolve them to the bakedsnapshot, so a container built this way needs no network at all. Each file's sizeis checked against what the repository reports before the snapshot is published,-and its SHA-256 as well when the repository publishes one, so a truncated or-substituted download fails the build instead of shipping inside a layer that-looks healthy. A repository that publishes no digest leaves the check to size-alone.+and its SHA-256 as well when the repository publishes one. A size or digest+mismatch fails the build. A repository that publishes no digest leaves the+check to size alone, which cannot detect a same-size substitution.| Variable | Default | Purpose || --- | --- | --- |@@ -256,7 +255,8 @@ repository-wide key there is. Hub-side pins do not describe a mirror snapshot:artifact hashes, so neither applies here, and no digest pin exists for a mirrorsnapshot either. The build already refuses a download that does not match themirror's own reported size -- and its digest, when the mirror publishes one ---which is the equivalent guarantee for this path. And the weights come from+but this checks consistency with mirror metadata, not an independently pinned+digest. The weights come fromwhichever mirror account the argument names, which is its own supply-chaindecision for the deployment to make.