Skip to content

SOTA-Class TTS at Compact Scale

Python

Document current SGLang Omni RTF

linux295650-ux05c83f39a1e9
diff --git a/README.md b/README.md
index 1730f33..daa90e5 100644
--- a/README.md
+++ b/README.md
@@ -159,6 +159,20 @@ revision instead of the latest `main` branch.
| Transformers | `4.57.1` |
| Precision | BF16 |
+### Performance
+
+Warm single-stream latency was measured on one NVIDIA H20 with BF16 weights,
+CUDA Graph, greedy decoding, and 128 generated frames. The output WAV was
+5.85-5.94 seconds long; cold start and compilation time were excluded. Lower
+RTF is better.
+
+| SGLang Omni adapter | Warm p50 latency | RTF |
+|---|---:|---:|
+| Current implementation | **0.691 s** | **0.116** |
+
+See the [SGLang Omni implementation and evaluation report](sglang_omni/OPTIMIZATION_REPORT.md)
+for the configuration, implementation details, and validation results.
+
### Install
Run these commands from the Audio8 TTS repository root. The example uses
diff --git a/README_zh.md b/README_zh.md
index 4fb26b0..7f98872 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -146,6 +146,18 @@ paged attention、动态 batching、Fast AR 固定 KV cache、参考音频编码
| Transformers | `4.57.1` |
| 精度 | BF16 |
+### 性能
+
+单流 warm latency 在单张 NVIDIA H20 上测试,使用 BF16、CUDA Graph、greedy decoding,
+并生成 128 帧。输出 WAV 时长为 5.85-5.94 秒,不包含冷启动和编译时间。RTF 越低越好。
+
+| SGLang Omni 适配器 | Warm p50 latency | RTF |
+|---|---:|---:|
+| 当前实现 | **0.691 s** | **0.116** |
+
+配置、实现细节和验证结果请参阅
+[SGLang Omni 实现与评测报告](sglang_omni/OPTIMIZATION_REPORT.md)。
+
### 安装
在 Audio8 TTS 仓库根目录执行以下命令。示例使用 Python 3.12 和