Fix audio decoding: preserve long dtype for token indices

[Copilot]

Audio decoding failed with "tensors used as indices must be long, int, byte or bool tensors" because the frames tensor was converted to float dtype before detokenization.

Changes

• Removed dtype conversion in codec detokenization: Frames contain token IDs used as indices for embedding lookups and must remain as torch.long
• Preserved device placement: Frames are still moved to the codec device, just without dtype conversion

# Before (incorrect - converts indices to float)
frames_for_codec = frames.to(device=pipeline.codec_device, dtype=codec_dtype)

# After (correct - preserves long dtype for indexing)
frames_for_codec = frames.to(device=pipeline.codec_device)

The codec_dtype parameter is still used for loading the codec model itself, but not applied to the token indices passed to detokenize().

Original prompt

---

This section details on the original issue you should resolve

<issue_title>Decoding Audio Failure</issue_title>
<issue_description>After PR #14 generation succeeds at a reasonable speed, but once completed it fails with:

Deconding Audio:
Generation Failed: invalid type: 'torch.mps.FloatTensor'

Lets fix this last bug to make sure our generations are correctly decoded to audio.</issue_description>

<agent_instructions>Issue reopened with error:
Generation Failed: tensors used as indices must be long, int, byte or bool tensors</agent_instructions>

Comments on the Issue (you are @copilot in this section)

@lmangani This change was ineffective and caused a new issue:

Generation Failed: tensors used as indices must be long, int, byte or bool tensors</comment_new>

• Fixes Decoding Audio Failure #15

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.