FunASR 1.4.5: lighter Python install and nine-platform llama.cpp runtimes / 更轻安装与九平台运行时

[LauraGPT]

FunASR 1.4.5 is now available from public PyPI and the signed GitHub Release. This release makes the default Python inference install lighter and puts the current prebuilt llama.cpp / GGUF runtimes beside the Python distributions.

Install

python -m pip install -U "funasr==1.4.5"

# Optional kaldi-native-fbank backend:
python -m pip install -U "funasr[knf]==1.4.5"

What changed

• torchaudio is no longer a hard dependency for standard inference. Existing torchaudio workflows remain supported when it is installed explicitly.
• The optional knf extra adds kaldi-native-fbank for a smaller inference dependency path.
• Ascend 910B validation processed 70.47 seconds of audio in 1.15 seconds (RTF 0.016) under the documented test contract. This single-run result is validation evidence, not a capacity guarantee.
• The Release contains exactly 12 assets: wheel, sdist, nine prebuilt runtime-llamacpp-v0.2.1 archives, and SHA256SUMS-v1.4.5.

The runtime set covers Linux arm64, Linux x64 portable/AVX2/Vulkan, macOS arm64, and Windows x64 portable/AVX2/Vulkan/CUDA. Download one archive, unpack it, then use the bundled download-funasr-model.sh <sensevoice|paraformer|nano> helper and the matching llama-funasr-* binary.

• PyPI: https://pypi.org/project/funasr/1.4.5/
• GitHub Release: https://github.com/modelscope/FunASR/releases/tag/v1.4.5
• Runtime release: https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.1
• Checksums: https://github.com/modelscope/FunASR/releases/download/v1.4.5/SHA256SUMS-v1.4.5

Public artifact SHA-256:

• wheel: e3f982e8d84c4c1b043e0536aaa7d4ae81ad2b5ae6c9f234bfd355bab23c6dd1
• sdist: a0fa52149a6dcdac9713b3d6fee07febb0c01a34bcb5f0b9c42ff27dc9ffb464

---

FunASR 1.4.5 已发布到 public PyPI 和带签名 tag 的 GitHub Release。本次发布让默认 Python 推理安装更轻，并把当前预编译 llama.cpp / GGUF 运行时与 Python 发行包集中到同一个下载页。

更新内容

• 标准推理不再强制依赖 torchaudio；显式安装后，已有 torchaudio 工作流仍可继续使用。
• 可选的 funasr[knf] extra 增加 kaldi-native-fbank，提供更精简的推理依赖路径。
• Ascend 910B 在已记录测试契约下处理 70.47 秒音频用时 1.15 秒，RTF 0.016；这是单次验证证据，不是容量承诺。
• GitHub Release 恰好包含 12 个资产：wheel、sdist、九个平台的 runtime-llamacpp-v0.2.1 归档和 SHA256SUMS-v1.4.5。

运行时覆盖 Linux arm64、Linux x64 portable/AVX2/Vulkan、macOS arm64，以及 Windows x64 portable/AVX2/Vulkan/CUDA。下载并解压对应归档后，可使用内置模型下载脚本和 llama-funasr-* 命令启动。

本次发布已从公网重新下载并核对所有归档完整性与实际 SHA-256。欢迎在反馈中附上平台、后端、模型、输入音频边界和最小复现命令。