laya-multilingual-mlx

Native MLX FP16 conversion of convaiinnovations/laya-multilingual for Apple silicon.

This checkpoint uses mmBERT-base, a 1024-token total context, and Laya's decision Transformer, scoring head and action head. It supports choice, ordinal score, and boolean noul questions. All model computation runs in MLX; the runtime does not require PyTorch or Transformers.

Usage

Install the dedicated runtime on an Apple silicon Mac with macOS 14+ and Python 3.11+:

python -m pip install laya-mlx

import laya_mlx as laya

agent = laya.load("aac6fef/laya-multilingual-mlx")
result = agent.predict(
    "I was billed twice. Please refund the duplicate today.",
    {
        "department": {
            "type": "choice",
            "instructions": "Which department should handle this request?",
            "criteria": ["billing", "technical", "sales"],
        },
        "refund": {
            "type": "noul",
            "instructions": "Does the customer ask for money back?",
        },
    },
)
print(result["answers"])

Use dtype="float32" for closer agreement with upstream FP32 arithmetic. The source weights themselves are FP16. Question formatting, tokenizer behavior, calibration temperatures and output schema are preserved.

This is a bidirectional decision encoder loaded with laya_mlx. The package provides the custom architecture needed to interpret the checkpoint. The repository does not include a generative language model or training implementation.

Validation

Tested locally on Apple M3 Max, 40-core GPU, 128 GB unified memory, macOS 27.2, Python 3.12.13 and MLX 0.32.2.

• FP16 agrees with upstream PyTorch MPS FP32 on the argmax of 63/63 decision distributions across 16 cases.
• Maximum calibrated probability difference: 0.0012887.
• 100 repeated calls produced finite, deterministic public outputs; measured MLX active-memory growth after clearing caches was 0 bytes.
• Every exported tensor was checked for exact equality with the corresponding source tensor cast to FP16.

The included validation.json contains numerical and stability measurements for both FP32 and FP16 arithmetic. Full performance report and raw timing samples compare MLX with the original runtime on the same machine. These checks establish port fidelity, not that every model answer is correct.

Provenance and limits

• Source checkpoint: convaiinnovations/laya-multilingual at 052592a15d198d9ad47da779604259b10b47b7aa.
• Upstream code: NandhaKishorM/laya, commit 6a5819129eb220570792e417e49723d697efd76f.
• Conversion changes parameter names for MLX and preserves FP16 weights. It does not retrain or quantize to fewer bits.
• This is an independent port. Model quality, calibration and language/task limitations remain those of the original checkpoint. Questions and options share the context budget with the input state.
• The typed-decisions checkpoint is specialized for upstream workflows; the multilingual checkpoint is the intended choice for non-English text.

Apache-2.0. Original Laya models and code are by Convai Innovations and contributors. See LICENSE, NOTICE, mlx_config.json and manifest.json for attribution and export details.