FunASR v1.4.16: reliable mixed-device submodels and native Transformers guides

[LauraGPT]

FunASR v1.4.16: reliable mixed-device submodels and native Transformers guides

FunASR 1.4.16 is now available on PyPI and GitHub.

python -m pip install -U "funasr==1.4.16"

What changed

• Explicit device placement in vad_kwargs, punc_kwargs, and spk_kwargs is preserved across repeated inference calls.
• Runtime device overrides no longer mutate already-loaded models or caller-owned submodel configuration.
• Speaker inference uses the speaker model's own configuration.
• English and Chinese documentation now explains punctuation postprocessing and provides tested native Transformers adoption paths.

The device regressions use CPU-resident stand-in weights and mocked accelerator availability. They verify configuration propagation, fallback, and repeated calls; they are not CUDA, MPS, or NPU performance certification.

Verified artifacts

The signed v1.4.16 tag resolves to 904cd18681b8083de5e1039bd0ecebc4f49ede60. The GitHub Release has exactly 12 assets: the wheel, sdist, nine independently versioned llama.cpp/GGUF runtimes, and SHA256SUMS.

• Wheel: f95943f6a8111349a1b08907363b048a49e9cb9caa1e3afd8c92970d08bf450d
• Source distribution: 1e2f8e3887fbc4b164cce40f73bf1324907da6ab59d1e8bf2add827f534c69e7

The public PyPI files were downloaded again and matched the GitHub assets byte for byte. The nine native runtimes remain builds from runtime-llamacpp-v0.2.6; they do not contain the Python-only fix.

• GitHub Release
• PyPI 1.4.16
• English release guide
• 中文发布说明

If a mixed-device pipeline still moves a submodel unexpectedly, please share the model IDs/revisions, FunASR and PyTorch versions, hardware/backend, constructor kwargs, runtime overrides, and a minimal reproduction.

---

中文：FunASR v1.4.16 混合设备子模型修复与原生 Transformers 指南

FunASR 1.4.16 已发布到 PyPI 与 GitHub。本次发布保持 VAD、标点和说话人子模型的显式设备配置：重复推理和运行时 device 覆盖不再改写已经加载的模型或调用方持有的配置，说话人推理也会使用自身模型配置。

中英文文档同步补齐了标点后处理边界与原生 Transformers 接入路径。设备回归覆盖配置传播、回退和重复调用，但不代表 CUDA、MPS 或 NPU 的性能认证。

签名标签精确指向 904cd18681b8083de5e1039bd0ecebc4f49ede60。Release 恰有 12 个资产：wheel、sdist、九个独立版本的 llama.cpp/GGUF 运行时和 SHA256SUMS；从公开 PyPI 回源下载的两个 Python 制品与 GitHub 资产逐字节一致。

如果仍遇到子模型设备被意外移动，请附上模型 ID/修订、FunASR 与 PyTorch 版本、硬件/后端、构造参数、运行时覆盖参数和最小复现。