runtime-llamacpp-v0.2.2

modelscope/FunASRruntime-llamacpp-v0.2.2Aug 28, 2026by github-actions[bot]

AI Summary

This release focuses on improving runtime stability and error visibility by guarding backend initialization failures and flushing stderr boundaries to clearly distinguish between backend selection, initialization, and buffer resolution stages.

Key Highlights

Added guards for failed `ggml_backend_dev_init()` to prevent null pointer dereferencing.
Flushed deterministic stderr boundaries to differentiate between selection, initialization, and buffer resolution failures.
Improved Vulkan backend failure interpretation for easier debugging.
Windows Vulkan packages prepared for Radeon 780M and RX 9070 XT testing.

New Features

Backend initialization error handling improvements
Enhanced stderr logging for backend stages
Improved Vulkan backend diagnostics

Full Release Notes

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

## v0.2.2

- Guards a failed CUDA or Vulkan `ggml_backend_dev_init()` before resolving the backend's default buffer type, so initialization failure is reported instead of dereferencing a null backend.
- Flushes deterministic stderr boundaries before backend initialization, after initialization, and after buffer-type resolution. These messages distinguish selection, backend initialization, buffer resolution, and later graph/model failures.
- Built from exact commit `05be4863fc94e3df37924a849b9790484ac74a29` with all nine platform jobs successful.
- The Windows Vulkan package is ready for Radeon 780M and RX 9070 XT retesting. This release does not claim that the AMD Windows `0xC0000005` crash is fixed; a crash inside the Vulkan driver or backend initializer can still terminate before FunASR regains control.
- Android/Mali remains outside the official prebuilt and continuously validated targets.

For `--backend vulkan`, interpret the final stderr boundary as follows:

- No `initializing vulkan backend ...`: failure occurred during enumeration/selection or earlier.
- `initializing ...` without `initialized ...`: failure occurred inside `ggml_backend_dev_init()` or below it.
- `initialized ...; resolving buffer type` without `vulkan backend ready ...`: failure occurred while resolving the default buffer type.
- `vulkan backend ready ...`: initialization completed and the failure is later in model loading or graph execution.

## Quick start

Download the default quantized model with:

```bash
bash download-funasr-model.sh <sensevoice|paraformer|nano>
```

The helper requires the Hugging Face CLI (`pip install -U huggingface_hub`). Then run `llama-funasr-cli`, `llama-funasr-sensevoice`, or `llama-funasr-paraformer`.

Use the default x64 asset for maximum CPU compatibility. Use x64-avx2 on CPUs with AVX2/FMA/F16C/BMI2. The Vulkan assets require a working Vulkan driver/ICD and accelerate SenseVoiceSmall with `--backend vulkan`. The Windows CUDA asset targets CUDA architecture 86 and requires a compatible NVIDIA driver; build from source for other GPU architectures.

## SHA-256

```text
adc0e968d70a4308191a91011a3444e0a1cbefc90940c3e199b7d65ff9e59d0b funasr-llamacpp-linux-arm64.tar.gz
cb5b5679938d2001426b5ea079ba948bac5c23b19aa2fc79e7a8572d9e9516e7 funasr-llamacpp-linux-x64-avx2.tar.gz
f865659d1787a2769d4ecfba598f2a490144819945bb2397fce3e172c1a1aff9 funasr-llamacpp-linux-x64-vulkan.tar.gz
c1ad11bac292288a783c1e5eb1103c6db301b98bf861d29b5b5310de3a190404 funasr-llamacpp-linux-x64.tar.gz
cb90c64c6c251d9df9a40193037713feaee9dd602d59b470bb4735d78c00da33 funasr-llamacpp-macos-arm64.tar.gz
f1c9ba8e35c273b995877e0fd7f4080df28e40bce2c41be96d862c44e20fea53 funasr-llamacpp-windows-x64-avx2.zip
52d5ecf4220e428737f9954b148d7ee1410109a12ef77fd45f99bc1c7dd040d3 funasr-llamacpp-windows-x64-cuda.zip
86a7d5ca7c134ae2fd3c9c1b356fcff041e29baf569c1145a75c73bb5bc5ea90 funasr-llamacpp-windows-x64-vulkan.zip
19e368fe0debaf880ae5aed063a1105e6cdc2d1a57f259b065df15db45a0103a funasr-llamacpp-windows-x64.zip
```

Build evidence: https://github.com/modelscope/FunASR/actions/runs/33182316846

Docs: https://github.com/modelscope/FunASR/blob/runtime-llamacpp-v0.2.2/runtime/llama.cpp/README.md