Number of audio soft tokens for a clip — processing_gemma4.py _compute_audio_num_tokens: ceil(duration_ms / 40) capped at 750. MUST be computed from decoded samples, never from container metadata.
_compute_audio_num_tokens
Number of audio soft tokens for a clip — processing_gemma4.py
_compute_audio_num_tokens: ceil(duration_ms / 40) capped at 750. MUST be computed from decoded samples, never from container metadata.