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): name the dtype spellings the runtime actually compares against
Aashishf9f280d31232
diff --git a/docs/docker.md b/docs/docker.mdindex 41eec446..56b9b066 100644--- a/docs/docker.md+++ b/docs/docker.md@@ -68,8 +68,8 @@ work with `docker run -e`; Compose-only settings are identified below.| Variable | Default | Purpose || --- | --- | --- || `LAYA_DEVICE` | `cpu` / `cuda` | Device selected by the base / GPU configuration |-| `LAYA_CUDA_AMP` | unset (checkpoint's `amp_dtype`) | `fp16` or `bf16` for the CUDA forward. Not cosmetic: the README's threshold section measures bf16 flipping 3 of 864 argmaxes on the parity set where fp16 flips none |-| `LAYA_CPU_AMP` | unset | `bf16` opts the CPU forward into bf16; anything else leaves it fp32 |+| `LAYA_CUDA_AMP` | unset (checkpoint's `amp_dtype`) | `fp16`/`float16` or `bf16`/`bfloat16` for the CUDA forward; anything else is ignored. Not cosmetic: the README's threshold section measures bf16 flipping 3 of 864 argmaxes on the parity set where fp16 flips none |+| `LAYA_CPU_AMP` | unset | `bf16` or `bfloat16` opts the CPU forward into bf16; anything else leaves it fp32. No fp16 spelling turns it on either: CPU autocast has no fp16 fast path that beats fp32, so bf16 is the only reduced precision core offers on this device || `LAYA_MODEL` | `auto` | Router alias: `auto`, `english`, `multilingual`, `typed-decisions` || `LAYA_MODEL_PATH` | unset | Compatible checkpoint path inside the container || `LAYA_REVISION` | unset | Hub commit, branch or tag used for every checkpoint download, or `reviewed` for the reviewed SHAs in `laya/revisions.py`; a `revision=` argument still wins |