The vocoder depthformer requires 8 inputs but only 5 were provided,
causing TTS mode to crash with "input depth_slices_in is missing in feeds".
- Add depth_slices_in recurrent state (zeros at step 0, fed back from
outputs.depth_slices at steps 1-7)
- Add seqlens_k and total_seq_len for GroupQueryAttention operators
- Fix past_keys/past_values initial shape from [6,1,0,8,32] to [6,1,8,0,32]
- Add depth_slices to preferredOutputLocation to avoid GPU-CPU-GPU roundtrip
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Documents the missing depth_slices_in, seqlens_k, and total_seq_len
inputs in the vocoder feeds, root cause analysis, and step-by-step
fix for audio-model.js.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
