Skip to content

v3.0.1

Latest

Choose a tag to compare

@Huanshere Huanshere released this 28 Feb 19:31
· 4 commits to main since this release

Changelog - Version 3.0.1
This update adds RTX 50 series (Blackwell) GPU support, upgrades core dependencies, and fixes several deprecation warnings and stability issues.

🚀 New Features

  • RTX 50 series (Blackwell) GPU support: installer now uses nvidia-smi to detect driver CUDA version and automatically selects cu129 wheels with sm_100 kernels for Blackwell GPUs.
  • Upgraded to PyTorch 2.8.0 / whisperX 3.8.1 with improved CUDA wheel auto-selection (cu129 / cu128 / cu126).

🐛 Bug Fixes

  • Fixed ffmpeg fallback to WAV encoding when libmp3lame is unavailable (#531).
  • Fixed JSON encoding for Chinese characters in prompts.
  • Fixed term extraction prompt to handle proper noun translation correctly.
  • Fixed Streamlit audio deadlock caused by librosa.load() in non-main thread.

🔧 Improvements

  • Replaced deprecated pynvml package with official nvidia-ml-py.
  • Replaced deprecated Streamlit use_container_width API with new width parameter.
  • Resolved dependency conflicts for PyTorch / torchaudio upgrade.
  • Updated installation guide and Docker docs with cu129 support, model recommendations, and pitfall warnings.

更新日志 - 版本 3.0.1
本次更新新增了 RTX 50 系列(Blackwell)显卡支持,升级核心依赖,并修复了多个弃用警告和稳定性问题。

🚀 新功能

  • RTX 50 系列(Blackwell)显卡支持:安装脚本现通过 nvidia-smi 检测驱动 CUDA 版本,自动为 Blackwell 显卡选择包含 sm_100 内核的 cu129 轮子。
  • 升级至 PyTorch 2.8.0 / whisperX 3.8.1,改进 CUDA 轮子自动选择(cu129 / cu128 / cu126)。

🐛 问题修复

  • 修复 ffmpeg 缺少 libmp3lame 时回退到 WAV 编码的问题(#531)。
  • 修复提示词中中文字符的 JSON 编码问题。
  • 修复术语提取 prompt 中专有名词翻译的处理。
  • 修复 Streamlit 中 librosa.load() 在非主线程导致的音频死锁。

🔧 改进

  • 将已弃用的 pynvml 替换为官方 nvidia-ml-py
  • 将已弃用的 Streamlit use_container_width API 替换为新的 width 参数。
  • 解决 PyTorch / torchaudio 升级的依赖冲突。
  • 更新安装指南和 Docker 文档:cu129 支持说明、模型推荐、踩坑提示。