modelscope/FunASR
1
0
Fork 0
mirror of https://github.com/modelscope/FunASR.git synced 2026-10-12 02:19:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

docs: merge benchmark tables into single unified table

Browse Source
This commit is contained in:
游雁
2026-05-25 20:29:12 +08:00
parent 653d787c21
commit 06558dfdd2
1 changed files with 11 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/vllm_guide.md
+11 -37
View File
@@ -365,44 +365,18 @@ Client Server
### Benchmark 结果（184 files, 11541s audio, 1947 VAD segments）
| 方法 | 耗时 | RTFx | CER |
|------|------|------|-----|
| PyTorch native | 589s | 19.6 | 8.94% |
| **Our vLLM (batch)** | **29.3s** | **393.9** | **8.91%** |
| yuekaizhang vLLM | 42.7s | 273.0 | 17.07% |
| 模型 | 方法 | 耗时 | RTFx | CER | 加速 |
|------|------|------|------|-----|------|
| Fun-ASR-Nano | PyTorch | 589s | 19.6 | 8.94% | 基准 |
| Fun-ASR-Nano | **vLLM (ours)** | **29.3s** | **393.9** | **8.91%** | **20.7x** |
| Fun-ASR-Nano | yuekaizhang vLLM | 42.7s | 273.0 | 17.07% | 14.3x |
| GLM-ASR-Nano | PyTorch | 338s | 34.4 | 12.94% | 基准 |
| GLM-ASR-Nano | **vLLM (ours)** | **43.9s** | **263.2** | **12.92%** | **7.6x** |
- **加速比**: 20.7x (vs PyTorch)
- **CER 一致性**: 8.91% vs 8.94%（差异 < 0.05%，完全对齐）
- **vs 第三方实现**: 比 yuekaizhang/Fun-ASR-vllm 快 44%，CER 优 8%
### 关键优化
| 优化项 | 效果 |
|--------|------|
| `use_low_frame_rate` token 截断 | CER 19.68% → 8.91% |
| Batch audio encode (groups of 8) | 音频编码加速 ~8x |
| vLLM batch generate (all prompts) | LLM 解码加速 ~20x |
| 去掉 `<think>` tokens | 减少无效生成步骤 |
### GLM-ASR-Nano
| 方法 | RTFx | CER | 加速 |
|------|------|-----|------|
| PyTorch native | 34.4 | 12.94% | 基准 |
| **vLLM (ours)** | **263.2** | **12.92%** | **7.6x** |
```python
from funasr.models.glm_asr.inference_vllm import GLMASRVLLMEngine
engine = GLMASRVLLMEngine.from_pretrained(
model="zai-org/GLM-ASR-Nano-2512",
hub="ms",
gpu_memory_utilization=0.4,
max_model_len=4096,
)
results = engine.generate(inputs=["audio.wav"])
print(results[0]["text"])
```
**关键结论：**
- vLLM 与 PyTorch CER 完全一致（Fun-ASR-Nano 差 0.03%，GLM-ASR 差 0.02%）
- Fun-ASR-Nano 加速 20.7x，GLM-ASR 加速 7.6x
- 比第三方 yuekaizhang/Fun-ASR-vllm 快 44%，且 CER 优 8%
### WebSocket 实时服务