|
Download reports/performance.md from AXERA-TECH/Opus-MT.AXERA: direct link, hf CLI and curl.
- Browser
- Download file 1.39 kB
-
https://huggingface.co/AXERA-TECH/Opus-MT.AXERA/resolve/main/reports/performance.md
- Command line
-
hf download hf://AXERA-TECH/Opus-MT.AXERA/reports/performance.md
-
curl -L -o performance.md https://huggingface.co/AXERA-TECH/Opus-MT.AXERA/resolve/main/reports/performance.md
1.39 kB
| # 性能报告(AX650N / NPU1 / U16) | |
| ## 延迟(batch=1,已预热,Tatoeba 200 句平均) | |
| | 方向 | 路径 | encoder | decoder 单步 | 端到端 | 平均输出 token | | |
| |------|------|---------|--------------|--------|----------------| | |
| | zh→en | Python | 5.51 ms | 6.38 ms | 69.1 ms/句 | 9.4 | | |
| | zh→en | **C++** | **4.90 ms** | **5.41 ms** | **58.5 ms/句** | 9.4 | | |
| | en→zh | Python | 5.50 ms | 6.40 ms | 56.2 ms/句 | 7.4 | | |
| | en→zh | **C++** | **4.89 ms** | **5.41 ms** | **47.3 ms/句** | 7.4 | | |
| 端到端 = encoder + N × decoder step(N = 输出 token 数,含 EOS)。首句包含 warmup, | |
| SDK 初始化时自动预热 2 次。 | |
| ## 模型信息 | |
| | 模型 | MACs | ONNX FP32 | AXMODEL U16 | | |
| |------|------|-----------|-------------| | |
| | encoder | 2.516 GMACs | 209.1 MB | 87.1 MB | | |
| | decoder(单步,T=64) | 1.921 GMACs | 368.6 MB | 130.2 MB | | |
| 每方向 AXMODEL 合计约 217 MB(ONNX FP32 合计约 578 MB)。 | |
| ## 内存 | |
| - 加载单个方向的 encoder + decoder,Python 进程峰值 RSS:约 262 MB | |
| - C++ 只加载 AX Engine 运行时与 sentencepiece,内存更低(未单独统计) | |
| - 运行时不依赖 onnxruntime / torch / transformers | |
| ## 编译配置 | |
| - Pulsar2 `7.0-patch1` | |
| - `target_hardware=AX650`,`npu_mode=NPU1` | |
| - `layer_configs: DEFAULT -> U16`,权重 S8 | |
| - `highest_mix_precision=false` | |
| - 静态 shape:source 128,target 64 | |