transformers
719d8090 - fix(vibevoice-asr): use integer ceiling division for audio token count (#48864)

Commit
65 days ago
fix(vibevoice-asr): use integer ceiling division for audio token count (#48864) * fix(vibevoice-asr): use integer ceiling division for audio token count padding_mask sums are int64; dividing by hop length in float32 loses precision past 2**24 samples, so torch.ceil undercounts by one vs the processor's float64 np.ceil for long audio. Integer (n + d - 1) // d is exact for all counts. Fixes #48835 * Update src/transformers/models/vibevoice_asr/modeling_vibevoice_asr.py * Update src/transformers/models/vibevoice_asr/modular_vibevoice_asr.py --------- Co-authored-by: ege-arhan <ege@example.com> Co-authored-by: Eric Bezzam <4757445+ebezzam@users.noreply.github.com>
Author
Parents
Loading