Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

laya-Ascend

Fork of Laya that runs its choice, score and yes/no decision checkpoints on Huawei Ascend 910B NPUs as well as CPU.

ServicebetaApache-2.0
View on GitHub ↗

Overview

laya-Ascend is a fork of the Laya open decision engine that adds Huawei Ascend NPU inference alongside CPU. Laya evaluates choice, score and noul questions over a state in a single forward pass, with three checkpoints (English ModernBERT-large, a 322M multilingual mmBERT model for 100+ languages, and a typed-decisions model) and a router that picks one per request; it does not generate text.

The fork adds an Ascend setup guide covering environment, package installation, model download and CPU or NPU inference, plus Snake and Tetris game examples. Per the listing, a four-question request takes 37 to 47 ms median on a 910B, 33.8 to 70.9 times faster than the same request on a single container CPU thread, and the fork replaces the transformer encoder's fused operation with an output-equivalent SDPA decision head to avoid torch_npu falling back to CPU. It is aimed at teams with Ascend hardware, notably in China, who want local Jev-style decisions.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich department should handle this request?billing · technical · sales · other
ScoreHow urgent is this request?(inferred)—
Yes / noAny yes/no question about the state—

How it works

  1. Set up the Ascend environment per README-Ascend.md
  2. Load Laya checkpoints on the NPU
  3. Send state and typed questions through the router
  4. Receive probabilities in one forward pass

Strengths

  • Runs Laya on Ascend NPUs
  • Large speedup over single-thread CPU per the listing
  • Avoids a CPU fallback in torch_npu

Caveats

  • Fork of upstream Laya; may lag updates
  • Requires Ascend hardware and toolchain
  • Speed figures from the listing, not independently verified

Getting started

See README-Ascend.md for environment setup, then pip install laya and run CPU or NPU inference.

At a glance

Type
Service
Status
beta
Language
Python
Runtime
PyTorch on Ascend NPU (torch_npu) or CPU
Has code
yes
License
Apache-2.0

Use cases and integrations

How curated lists describe it

  • Local runtime: Ascend NPU fork of the Laya checkpoints that answers the same `Choice`, `Score` and `Noul` questions on Huawei 910B hardware — 37–47 ms median for a four-question request, 33.8x–70.9x faster than the same request on a single container CPU thread, with an output-equivalent SDPA decision head that avoids `torch_npu`'s CPU fallback on `aten::_transformer_encoder_layer_fwd`.

    yibie/awesome-jev

Frequently asked questions

What does this fork add?

Inference on Huawei Ascend 910B NPUs, with a decision head that avoids a CPU fallback.

How fast is it?

The listing reports 37 to 47 ms median for a four-question request.

Is it the same model as Laya?

Yes, it runs the same Laya checkpoints.

Related projects

  • decidealot

    Docker service exposing local Laya, Von, a CLM head or hosted Jev through one TypeSafe-compatible HTTP and MCP endpoint.

    MCP serverbetaPython
  • edgejev

    Converts open Jev reproductions (Laya, Kev, NanoJev, PlayJev) to INT8 ONNX for offline CPU decisions at about 15.6 ms per question.

    ServicebetaPython
  • go-system-one

    Native Go runtime serving Gemma 4 12B as a Jev-like decision model with hand-tuned PTX kernels and no CGo or CUDA toolkit.

    ProjectbetaGo
  • gojev

    Go harness for System One models: hosted Jev, open Kev in-process via llama.cpp, and Cloudflare Clef, with typed helpers and a CLI.

    ServicebetaGo
  • hearim

    Go gateway exposing the Jev /v1/systemone contract over ordinary LLM backends by reading single-token label logprobs.

    ServicebetaGo
  • intern-decision-mlx

    Runs Intern-Decision-0.8B on Apple Silicon behind a local /v1/systemone endpoint that accepts screenshots for agent decisions.

    ServicebetaPython