FunASR 1.4.2: SRT output, AMD Vulkan fixes, and verified 12-asset release

[LauraGPT]

FunASR 1.4.2 is now available on PyPI and GitHub.

python -m pip install -U "funasr==1.4.2"

Highlights

• Fixed punctuation-token alignment so timestamped ASR output keeps correct sentence and subtitle boundaries.
• Fixed DDP/FSDP gradient synchronization on accumulation-window boundaries and resolved DeepSpeed/FSDP configuration initialization.
• Added native SRT output to the FunASR llama.cpp/GGUF command-line runtimes.
• Updated the v0.2.0 Vulkan runtime baseline for AMD Windows stability.
• Added validated deployment guidance for native Fun-ASR-Nano on vLLM 0.27.1 and SenseVoice through audio.cpp/native serving.

The GitHub release contains 12 verified assets: wheel, sdist, nine Linux/macOS/Windows llama.cpp runtime archives, and SHA256SUMS-v1.4.2. Windows CUDA and Linux/Windows Vulkan builds are included.

PyPI: https://pypi.org/project/funasr/1.4.2/

---

FunASR 1.4.2 已发布到 PyPI 与 GitHub。此次更新修复标点 token 与时间戳句子对齐、DDP/FSDP 梯度累积同步和 DeepSpeed/FSDP 配置初始化，并为原生 llama.cpp/GGUF 命令行增加 SRT 字幕输出。v0.2.0 运行时同步更新 AMD Windows Vulkan 稳定性，发布页提供 wheel、sdist、9 个跨平台原生运行时和统一 SHA-256 校验清单。