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
# Use with compose.yaml, and with compose.http.yaml when serving over HTTP. The
# `laya-serve` entry repeats the CUDA build args and device reservation because it is
# a separate service: overrides for `laya` never reach it, so it would serve on CPU.
services:
laya:
build:
args:
# cu130, not cu128: compose.yaml pins TORCH_VERSION 2.14.0 and the cu128 index has no
# 2.14.0 wheel (its newest is 2.11.0), so the documented CUDA path could not resolve its
# own pin (#971). cu126 also carries 2.14.0; cu130 is chosen because docker.yml already
# builds the CUDA images with it, so Compose and CI now agree.
TORCH_INDEX: "${LAYA_TORCH_INDEX:-cu130}"
environment:
LAYA_DEVICE: "${LAYA_DEVICE:-cuda}"
# The runtime reads these when it picks the autocast dtype; empty keeps the checkpoint's
# own `amp_dtype`. README's threshold section measures what fp16 vs bf16 decides differently.
LAYA_CUDA_AMP: "${LAYA_CUDA_AMP:-}"
LAYA_CPU_AMP: "${LAYA_CPU_AMP:-}"
deploy:
resources:
reservations:
devices:
- driver: nvidia
device_ids: ["${LAYA_GPU_ID:-0}"]
capabilities: [gpu]
laya-serve:
build:
context: .
args:
TORCH_INDEX: "${LAYA_TORCH_INDEX:-cu130}"
environment:
LAYA_DEVICE: "${LAYA_DEVICE:-cuda}"
# The runtime reads these when it picks the autocast dtype; empty keeps the checkpoint's
# own `amp_dtype`. README's threshold section measures what fp16 vs bf16 decides differently.
LAYA_CUDA_AMP: "${LAYA_CUDA_AMP:-}"
LAYA_CPU_AMP: "${LAYA_CPU_AMP:-}"
deploy:
resources:
reservations:
devices:
- driver: nvidia
device_ids: ["${LAYA_GPU_ID:-0}"]
capabilities: [gpu]