|
Download docs/CPU_PORTABILITY.md from BenedictCarling/perch2-int8: direct link, hf CLI and curl.
- Browser
- Download file 982 Bytes
-
https://huggingface.co/BenedictCarling/perch2-int8/resolve/main/docs/CPU_PORTABILITY.md
- Command line
-
hf download hf://BenedictCarling/perch2-int8/docs/CPU_PORTABILITY.md
-
curl -L -o CPU_PORTABILITY.md https://huggingface.co/BenedictCarling/perch2-int8/resolve/main/docs/CPU_PORTABILITY.md
982 Bytes
| # Mac vs Linux output | |
| The same INT8 model gives slightly different embeddings on macOS arm64 and | |
| Linux x86_64 (ONNX Runtime 1.23.2), even when both get identical log-mel input. | |
| So the difference comes from the runtime's INT8 kernels, not from the frontend. | |
| On four synthetic test signals the cosine similarity between the two was | |
| between 0.991 and 1.000. Changing ONNX Runtime's graph optimisation level | |
| shifts the output a little too, even on one machine. | |
| What this means in practice: | |
| - Train and run your classifier on the same platform where you can. | |
| - The accuracy numbers in the README were measured on the original setup. Check | |
| them again on your own target before relying on them. | |
| The tests check the frontend exactly on every platform. They only compare | |
| embeddings exactly on Mac. On Linux they check that the output is valid and | |
| repeatable. | |
| The raw numbers are in `results/cpu_portability.json`. To re-run: | |
| ```bash | |
| python research/debug/compare_cpu_outputs.py | |
| ``` | |