Skip to content

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

docs: add GLM-ASR vLLM benchmark results to vllm_guide.md

游雁674ba5e53e2c
diff --git a/docs/vllm_guide.md b/docs/vllm_guide.md
index 7e0c267de..dbff38109 100644
--- a/docs/vllm_guide.md
+++ b/docs/vllm_guide.md
@@ -25,7 +25,7 @@ FunASR 集成 vLLM 高吞吐推理引擎,用于加速 LLM-based ASR 模型的
|------|-----------|------|
| **FunASRNano** | ✓ | Qwen3-0.6B LLM,支持离线和流式 |
| **LLMASR** | ✓ | Whisper + Qwen/Vicuna/LLaMA |
-| **GLMASR** | ✓ | GLM-ASR-Nano |
+| **GLMASR** | ✓ | GLM-ASR-Nano (RTFx 263, CER 12.92%) |
| **QwenAudioWarp** | ✓ | Qwen-Audio |
| Paraformer | ✗ | 非自回归模型(CIF predictor),无 LLM 解码 |
| SenseVoice | ✗ | Whisper-like encoder-decoder,非 LLM |
@@ -384,6 +384,26 @@ Client Server
| 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"])
+```
+
### WebSocket 实时服务
| 指标 | 数值 |