Commit History
-
whisper-medium int8 beamsearch with cross-QK collection, for word-level timestamps
Verified signatureExported with onnxruntime.transformers convert_to_onnx --collect_cross_qk --use_gpu --provider cuda --use_decoder_masked_mha --precision int8, then topologically repaired. Adds the cross_qk_layer_head input and the cross_qk output. CUDA execution provider only.f06d078
-
ONNX export: fused WhisperBeamSearch, int8 and fp32
Verified signature7071260 -
initial commit
Verified signature9ac3521