v1.4.4

modelscope/FunASRv1.4.4Aug 26, 2026by github-actions[bot]

AI Summary

This release focuses on major performance improvements for Realtime WebSocket concurrency and enhanced compatibility. It removes serialization locks, batches decode requests, and improves exception handling while maintaining compatibility with older PyTorch versions.

Key Highlights

Significant throughput improvements in Realtime WebSocket concurrency (up to 13.173x realtime)
Removal of the process-wide lock that previously serialized realtime sessions
Isolation of per-session VAD state while sharing encoder weights
Compatibility with PyTorch versions without torch.amp

Breaking Changes

Removes the process-wide lock that serialized realtime sessions
Runtime bindings now surface real exceptions instead of silently masking failures

New Features

Realtime WebSocket concurrency batching for compatible vLLM decode requests
VAD state isolation per session
Improved exception visibility and handling
Compatibility layer for PyTorch versions without torch.amp

Full Release Notes

## FunASR 1.4.4

Install or upgrade:

```bash
python -m pip install -U "funasr==1.4.4"
```

[PyPI package](https://pypi.org/project/funasr/1.4.4/)

### Realtime WebSocket concurrency

- Removes the process-wide lock that serialized realtime sessions.
- Batches compatible vLLM decode requests across sessions while preserving request-local outputs.
- Isolates per-session VAD state while sharing encoder weights.
- Recovers individual requests after batch failures and keeps the worker alive.

On one H100 with 12 concurrent clients, aggregate throughput improved from 8.431x to 11.566x realtime and STOP p95 fell from 19,832.9 ms to 414.4 ms. At 16 clients, throughput improved from 8.555x to 13.173x realtime.

### Compatibility and artifacts

- Keeps compatibility with PyTorch versions without `torch.amp`.
- Runtime bindings surface real exceptions instead of silently masking failures.
- The Python wheel, source distribution, prebuilt llama.cpp/GGUF runtimes, and SHA-256 manifest are attached below.

[Full changelog](https://github.com/modelscope/FunASR/compare/v1.4.3...v1.4.4)

<!-- funasr-runtime-downloads:start -->
## Runtime downloads

This Python release pairs with the current prebuilt llama.cpp / GGUF runtime release: [runtime-llamacpp-v0.2.0](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.0).

The same verified runtime assets are attached directly to this Python release so users can find the package and self-contained `llama-funasr-*` binaries in one place.

| Platform | Asset | SHA-256 |
|---|---|---|
| Linux arm64 | [funasr-llamacpp-linux-arm64.tar.gz](https://github.com/modelscope/FunASR/releases/download/v1.4.4/funasr-llamacpp-linux-arm64.tar.gz) | `c78987b2384c6aef339aea1bcd0e130070455d6394fa7ab7ca26840ead10d5da` |
| Linux x64 AVX2 | [funasr-llamacpp-linux-x64-avx2.tar.gz](https://github.com/modelscope/FunASR/releases/download/v1.4.4/funasr-llamacpp-linux-x64-avx2.tar.gz) | `02e10e9a46ea76a040c45d431efe51a3324e64f08c24d38e18c8a4d2781490cd` |
| Linux x64 Vulkan | [funasr-llamacpp-linux-x64-vulkan.tar.gz](https://github.com/modelscope/FunASR/releases/download/v1.4.4/funasr-llamacpp-linux-x64-vulkan.tar.gz) | `caf71b8c0b4c3249cebc4175e5406d3c588eb9e8966a00d571d4cc5070405385` |
| Linux x64 portable | [funasr-llamacpp-linux-x64.tar.gz](https://github.com/modelscope/FunASR/releases/download/v1.4.4/funasr-llamacpp-linux-x64.tar.gz) | `15e6407143b4fb91d90bb37f2a41c64c4d48ea0fbe6404b88a9b70269c84f240` |
| macOS arm64 | [funasr-llamacpp-macos-arm64.tar.gz](https://github.com/modelscope/FunASR/releases/download/v1.4.4/funasr-llamacpp-macos-arm64.tar.gz) | `416cbb289e31cb7575365d382155074e922fd061807a37b9ca0247dabd9bc6f9` |
| Windows x64 AVX2 | [funasr-llamacpp-windows-x64-avx2.zip](https://github.com/modelscope/FunASR/releases/download/v1.4.4/funasr-llamacpp-windows-x64-avx2.zip) | `4db0f11f603c324a63545cd7009cdd45bb45576efe282cec22796b5fd42d8ea1` |
| Windows x64 CUDA | [funasr-llamacpp-windows-x64-cuda.zip](https://github.com/modelscope/FunASR/releases/download/v1.4.4/funasr-llamacpp-windows-x64-cuda.zip) | `7f2f9ef4d7e0291b284a295ec74bbeca9ea635a7f5f42d0ad06eb780c0d6efc1` |
| Windows x64 Vulkan | [funasr-llamacpp-windows-x64-vulkan.zip](https://github.com/modelscope/FunASR/releases/download/v1.4.4/funasr-llamacpp-windows-x64-vulkan.zip) | `90b45240c6ccc9177c25490a11848de60a406e129391c8736b14521c0c28cdcb` |
| Windows x64 portable | [funasr-llamacpp-windows-x64.zip](https://github.com/modelscope/FunASR/releases/download/v1.4.4/funasr-llamacpp-windows-x64.zip) | `297c962346d7e30d7a7c2c860dfaab3ff07d01fddf15e6fc5212ca9545441a51` |

Quick start: download one asset, unpack it, then run the bundled `download-funasr-model.sh <sensevoice|paraformer|nano>` helper and one of `llama-funasr-cli`, `llama-funasr-sensevoice`, or `llama-funasr-paraformer`.

For Python users, install from PyPI:

```bash
python -m pip install -U "funasr==1.4.4"
```
<!-- funasr-runtime-downloads:end -->