mishkat-whisper-tiny-ar-quran-onnx

Changes made to the original work

1. Exported to ONNX from the original PyTorch weights with optimum-cli export onnx --task automatic-speech-recognition-with-past --opset 14 (optimum 1.21.4 · transformers 4.41.2 · torch 2.2.2). No training, no fine-tuning, no change to the weights' values.
2. The tied embedding was re-tied. The exporter emits the token-embedding matrix twice — once as [vocab, d_model] for the embedding Gather and once, transposed, as [d_model, vocab] for the output projection — which alone costs ~80 MB in fp32. The duplicate is removed and the Gather reads the surviving matrix along axis=1 followed by a Transpose. Verified numerically to be the exact transpose before rewiring, and measured to give bit-identical accuracy to the untied graph on our benchmark.
3. Quantized to int8 (dynamic, per-tensor, weights only) with onnxruntime.quantization.quantize_dynamic, including the embedding Gather — which is what makes the shared matrix pay off twice.
4. generation_config.json is taken from openai/whisper-tiny (Apache-2.0), since the fine-tuned repo ships none and the exporter's default lacks the multilingual tokens.

Files

License