Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

Clarify GPU quickstart PyTorch setup

zhifu gao18cc7dc3998e
diff --git a/README.md b/README.md
index 761342995..f8749deab 100644
--- a/README.md
+++ b/README.md
@@ -33,10 +33,25 @@
No local setup? Open the [Colab quickstart](./examples/colab/) to transcribe a public sample or upload your own audio in a browser.
```bash
+# CPU-only installs can use the default PyPI wheels.
pip install torch torchaudio
pip install funasr
```
+For GPU quickstarts, install the PyTorch and torchaudio wheels that match your
+NVIDIA driver from [pytorch.org](https://pytorch.org/get-started/locally/)
+before installing FunASR. After installation, confirm the GPU is visible:
+
+```bash
+python - <<'PY'
+import torch
+print(torch.cuda.is_available())
+PY
+```
+
+Only use `device="cuda"` when this prints `True`; otherwise use `device="cpu"`
+or reinstall PyTorch with the correct CUDA wheel.
+
**Flagship model — Fun-ASR-Nano** (LLM-ASR for Chinese, English, and Japanese, plus Chinese dialect groups and regional accents; needs a GPU):
```python
diff --git a/README_zh.md b/README_zh.md
index d77c13048..197497bcd 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -33,9 +33,24 @@
不想先配置本地环境?可以打开 [Colab 快速体验](./examples/colab/README_zh.md) 在浏览器里转写公开样例或上传自己的音频。
```bash
+pip install torch torchaudio
pip install funasr
```
+如果要运行 GPU quickstart,请先按 [pytorch.org](https://pytorch.org/get-started/locally/)
+选择与你的 NVIDIA driver 匹配的 PyTorch / torchaudio CUDA wheel,再安装 FunASR。
+安装后先确认 GPU 可见:
+
+```bash
+python - <<'PY'
+import torch
+print(torch.cuda.is_available())
+PY
+```
+
+只有这里输出 `True` 时才使用 `device="cuda"`;否则请先使用
+`device="cpu"`,或重新安装匹配 CUDA 的 PyTorch wheel。
+
```python
from funasr import AutoModel
from funasr.utils.postprocess_utils import rich_transcription_postprocess
diff --git a/docs/community_growth_20k.md b/docs/community_growth_20k.md
index 4520d1862..2e9e228a9 100644
--- a/docs/community_growth_20k.md
+++ b/docs/community_growth_20k.md
@@ -341,7 +341,7 @@ Paste the Markdown output into launch notes, weekly community updates, or releas
### Week 1: Repository conversion
-- [ ] Verify README quick start on clean CPU and GPU environments.
+- [ ] Verify README quick start on clean CPU and GPU environments. On 2026-07-23, clean venv checks covered README Quick Start SenseVoiceSmall CPU and README Quick Start Fun-ASR-Nano GPU setup paths with `torch 2.13.0+cu130`, `torchaudio 2.11.0+cu130`, and `funasr 1.3.26`. The clean CPU venv ran `AutoModel(model="iic/SenseVoiceSmall", vad_model="fsmn-vad", spk_model="cam++", device="cpu")` and produced `欢迎大家来体验达摩院推出的语音识别模型`. The clean GPU venv ran `AutoModel(model="FunAudioLLM/Fun-ASR-Nano-2512", device="cuda")` and produced the same sample text, but `torch.cuda.is_available()` was `False` on an H100 host after the bare `pip install torch torchaudio` path, so README and README_zh now tell GPU users to install the PyTorch / torchaudio CUDA wheels from pytorch.org and verify `torch.cuda.is_available()` before using `device="cuda"`. Do not mark this complete until a clean GPU venv reports CUDA visible. Logs are kept under `/cpfs_speech/user/zhifu.gzf/.cache/funasr-ops/readme-quickstart-cpu-20260723.log` and `/cpfs_speech/user/zhifu.gzf/.cache/funasr-ops/readme-quickstart-gpu-20260723.log`.
- [x] Verify high-traffic README, docs, runtime, and model_zoo links. On 2026-07-23, PRs #3360 through #3363 refreshed localized README Paraformer links, runtime tutorial download commands, runtime deploy-script fallback URLs, package/help metadata, model_zoo Hugging Face links, and verified core ModelScope entries; regression guards now cover these public entry points.
- [x] Add or refresh CONTRIBUTING, PR template, and issue templates. On 2026-07-23, the four core repositories now have `CONTRIBUTING.md`, issue templates, and PR templates: `modelscope/FunClip#188`, `modelscope/FunClip#189`, `FunAudioLLM/Fun-ASR#151`, and `FunAudioLLM/SenseVoice#327` closed the remaining gaps after the existing FunASR, Fun-ASR, and SenseVoice templates were audited.
- [x] Add a short FAQ for install/deployment failures. On 2026-07-23, `modelscope/FunASR#3366` added [`docs/troubleshooting.md`](./troubleshooting.md) and [`docs/troubleshooting_zh.md`](./troubleshooting_zh.md), linked both from README and README_zh, and covered `torch` / `torchaudio`, ModelScope versus Hugging Face downloads, `funasr-server` `/v1/audio/transcriptions`, `WebSocket` realtime checks, `llama.cpp` / `GGUF`, and `Deployment Help` issue details.
diff --git a/tests/test_growth_plan_doc.py b/tests/test_growth_plan_doc.py
index 69b20aff3..59c516578 100644
--- a/tests/test_growth_plan_doc.py
+++ b/tests/test_growth_plan_doc.py
@@ -42,6 +42,32 @@ def test_growth_plan_records_install_deployment_faq_completion():
assert "[ ] Add a short FAQ for install/deployment failures" not in text
+def test_growth_plan_records_readme_quick_start_cpu_gpu_validation():
+ text = PLAN.read_text()
+
+ required_markers = [
+ "[ ] Verify README quick start on clean CPU and GPU environments",
+ "clean CPU venv",
+ "clean GPU venv",
+ "README Quick Start SenseVoiceSmall CPU",
+ "README Quick Start Fun-ASR-Nano GPU",
+ "`torch 2.13.0+cu130`",
+ "`torchaudio 2.11.0+cu130`",
+ "`funasr 1.3.26`",
+ "`AutoModel(model=\"iic/SenseVoiceSmall\"",
+ "`AutoModel(model=\"FunAudioLLM/Fun-ASR-Nano-2512\"",
+ "`torch.cuda.is_available()` was `False`",
+ "install the PyTorch / torchaudio CUDA wheels from pytorch.org",
+ "verify `torch.cuda.is_available()` before using `device=\"cuda\"`",
+ "Do not mark this complete until a clean GPU venv reports CUDA visible",
+ "欢迎大家来体验达摩院推出的语音识别模型",
+ "readme-quickstart-cpu-20260723.log",
+ "readme-quickstart-gpu-20260723.log",
+ ]
+ for marker in required_markers:
+ assert marker in text
+
+
def test_growth_plan_records_github_metadata_completion():
text = PLAN.read_text()