runtime-llamacpp-v0.2.1

QwenAudio/SenseVoiceruntime-llamacpp-v0.2.1Aug 27, 2026by LauraGPT

AI Summary

Prebuilt self-contained binaries for the FunASR llama.cpp runtime v0.2.1, supporting SenseVoice, Fun-ASR-Nano, and Paraformer with built-in FSMN-VAD. This release improves GPU selection logic to prefer discrete GPUs over integrated ones and updates Vulkan support.

Key Highlights

Fixes Vulkan selection for integrated GPUs (IGPU) and prefers discrete GPUs.
Provides nine desktop archives for Linux, macOS, and Windows including Vulkan and CUDA variants.
Windows Vulkan package is ready for Radeon 780M retesting.
Android/Mali is not an official target in this release.

New Features

Improved Vulkan device selection logic
Multi-platform binary support (Linux, macOS, Windows)
CUDA and Vulkan backend support

Full Release Notes

Prebuilt self-contained binaries for the FunASR llama.cpp / GGUF runtime, including SenseVoice, Fun-ASR-Nano, and Paraformer with built-in FSMN-VAD.

## v0.2.1

- Fixes Vulkan selection when a requested device is reported as an integrated GPU (`GGML_BACKEND_DEVICE_TYPE_IGPU`).
- Prefers a matching discrete GPU and falls back to a matching integrated GPU.
- Provides nine desktop archives for Linux, macOS, and Windows, including Vulkan and Windows CUDA variants.
- The Windows Vulkan package is ready for Radeon 780M retesting. This selector fix does not claim to resolve the separate Radeon RX 9070 XT `0xC0000005` initialization crash.
- Android/Mali is not an official prebuilt or validated target in this release.

## Quick start

```bash
pip install -U huggingface_hub
bash download-funasr-model.sh sensevoice
```

Then run `llama-funasr-sensevoice`. Use the default x64 asset for broad CPU compatibility, the x64-avx2 asset on CPUs with AVX2/FMA/F16C/BMI2, or a Vulkan/CUDA asset only with a compatible driver. The Windows CUDA asset targets architecture 86; build from source for other GPU architectures.

## Provenance

These nine binaries are byte-for-byte identical to the digest-verified assets from the canonical [`modelscope/FunASR` runtime v0.2.1](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.1). They were built from exact FunASR commit `c69e69a0b78fb9becc8b2be91db8ae57d2a41f3c` by [Actions run 32991388379](https://github.com/modelscope/FunASR/actions/runs/32991388379). GitHub's generated source archives on this mirror tag are snapshots of the SenseVoice repository; use the canonical FunASR commit above for the native runtime source.

Canonical documentation: https://www.funasr.com/en/deploy/llama-cpp.html

## SHA-256

```text
9657e519986a5db72082aab508439559987fd4e563733d11da4662f898c75c45  funasr-llamacpp-linux-arm64.tar.gz
32e459673297fc01fcda6901ca7d63cc9e64896a43a3c9edcb7a8e1ef70fcd04  funasr-llamacpp-linux-x64-avx2.tar.gz
d45d3acb77e05c4bae94d818ed4a5c66852be87b22c52723bd576377c055c6e6  funasr-llamacpp-linux-x64-vulkan.tar.gz
1bc83dd36d11b1a2b9a45faaf77a5d1014000ac641c93dfec73d24d71ec51caf  funasr-llamacpp-linux-x64.tar.gz
bc63c4d4b96f2465f1d258600668a971f4f600d661f1859b03797cefaa417167  funasr-llamacpp-macos-arm64.tar.gz
67e428a91614f8a0c1d53dc2d79cce8efc7916c030ae2ddce2cbcf667ec1c502  funasr-llamacpp-windows-x64-avx2.zip
af32e0ef352880fcfd5589e0152c44038bc81f462f42c7a6ca91a0d9879bac58  funasr-llamacpp-windows-x64-cuda.zip
5fd179c6aefc18477eb79fa6235c70d131a96eba621318bb84e901ae85035578  funasr-llamacpp-windows-x64-vulkan.zip
b8f2b8f241b57921d82d64068d9b5695629779f3db5f3205a730cb3810232bb4  funasr-llamacpp-windows-x64.zip
```