fix(fish_speech): use from_indices() instead of decode() for DAC decoder

[ianliuy]

Summary

Fix TypeError: DAC.decode() takes 2 positional arguments but 3 were given when running fish_speech TTS online server.

Root Cause

fish_speech_dac_decoder.py:298 calls self._codec.decode(codes_bqf, feature_lengths), but:

1. Wrong method: DAC.decode(z) expects a continuous latent tensor. The code passes discrete codebook indices (torch.long). The correct method is from_indices(indices).
2. Wrong arg count: Both decode() and from_indices() accept only 1 positional argument. The extra feature_lengths causes the TypeError.
3. Wrong return unpacking: decode() and from_indices() return a single tensor, not a (wav, lengths) tuple.

Fix

# Before
wav_batch, audio_lengths = self._codec.decode(codes_bqf, feature_lengths)

# After
wav_batch = self._codec.from_indices(codes_bqf)
audio_lengths = torch.clamp(
    feature_lengths * self._hop_length,
    max=wav_batch.shape[-1],
)

• from_indices() internally does quantizer.decode(indices) decoder(z) the correct indicesaudio path
• audio_lengths computed from feature_lengths * hop_length (mathematically exact for this architecture)
• torch.clamp(max=...) as defensive bound

Also updated _FakeCodec in tests to match the new API.

Fixes #2643

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[hsliuustc0106]

DCO check failed (ACTION_REQUIRED). Please amend your commit with git commit -s --amend to add the Signed-off-by line.

[ianliuy]

Amended the commit with git commit -s --amend to add the Signed-off-by line. DCO check should pass now. cc - @hsliuustc0106

[ianliuy]

The ReadTheDocs failure is unrelated to this PR it timed out during pip install .[docs] (15min limit). Other PRs (#2377, #2517) are hitting the same RTD timeout. This PR only touches fish_speech_dac_decoder.py and its test file no docs changes.

[ianliuy]

Hi @hsliuustc0106, gentle ping 🙏 This PR and two others are all approved by @lishunyang12 with green CI, waiting on your review as the remaining requested reviewer.

All three are small fixes:

• This PR (fix(fish_speech): use from_indices() instead of decode() for DAC decoder #2668): +8/-6, DAC decoder API fix (from_indices() instead of decode())
• fix: do not apply FP8 quant config to vision/audio encoders for pre-quantized checkpoints #2702: FP8 encoder quant guard (tested locally by lishunyang12)
• [Bugfix] Make mrope kwargs optional in HunyuanImage3 get_mrope_input_positions #2654: +3/-3, signature-only change (mrope kwargs optional)

Happy to address any feedback just wanted to batch these together to keep them on your radar whenever you have a moment. Thanks!

[lishunyang12]

LGTM. The fix is correct and well-documented in the PR description.

What was wrong:

• DAC.decode(z) expects a continuous latent tensor, but the code passed discrete codebook indices (torch.long). The correct entry point for indices is from_indices().
• decode() / from_indices() both accept only 1 positional arg, so the extra feature_lengths caused the TypeError.
• The old code unpacked a (wav, lengths) tuple, but both methods return a single tensor.

Why the fix is correct:

• from_indices(codes_bqf) is the right DAC API for discrete codebook indices -> audio waveform.
• audio_lengths = clamp(feature_lengths * hop_length, max=wav_batch.shape[-1]) is the standard way to recover sample-domain lengths from frame-domain lengths for a fixed-hop-length codec, with a defensive upper bound.
• The _FakeCodec test mock is correctly updated to match the new single-tensor return.

No concerns.

Replacing with inline comments

[ianliuy]

Hi @hsliuustc0106 and @lishunyang12, just a polite follow-up on this small DAC decoder fix.

The branch is mergeable, DCO/pre-commit/build checks are green, and the prior approval was dismissed after the force-push. Would you be willing to take another look when you have a chance?