fix(fish_qwen3_omni): support loading quantized/converted models in sanitize()

[yoshphys]

Problem

sanitize() in fish_qwen3_omni/fish_speech.py skips any weight key that does not start with text_model.model. or audio_decoder., via else: continue. This means that when loading a model that was already converted to MLX format (e.g., produced by python -m mlx_audio.convert --quantize), all weight keys (which start with model.) are silently dropped, resulting in an empty dict.

The downstream effect:

• apply_quantization() finds no .scales keys → skips quantization entirely
• load_weights() is called with an empty dict → raises ValueError: Missing N parameters

This is the root cause of the bug reported in #578 ("Fish s2 pro Breaks when quantizing using convert").

Fix

Add a model.* passthrough case before the existing conditions so that keys already in MLX format are preserved as-is:

if key.startswith("model."):
    # Already in MLX format (e.g., previously converted/quantized model)
    remapped[key] = value
elif key.startswith("text_model.model."):
    ...

How to reproduce

python -m mlx_audio.convert \
  --hf-path mlx-community/fish-audio-s2-pro-bf16 \
  --mlx-path ./fish-audio-s2-pro-4bit \
  -q --q-bits 4 --q-group-size 64 --model-domain tts

python -c "
from mlx_audio.tts.utils import load_model
model = load_model('./fish-audio-s2-pro-4bit')  # raises ValueError before fix
"

Testing

After the fix, quantized models load and run correctly.

🤖 Generated with Claude Code

[lucasnewman]

@yoshphys Thanks! Did you test audio generation and does it sound good to your ear? Your agent can't hear very well so it needs human verification :)

[yoshphys]

@lucasnewman Thank you for paying attention on this PR. Sure. I wished I could make my agent hear the sound, but it cannot. So I had to listen to some audio outputs after modifying the code. I generated a couple of sets of audio sound with different sets of quantized parameters, 4bit and 8 bit, respectively. In the case of 4bit, I got aware that a few parts of audio with inline control had less quality, but I guess that was because of the quantization. In the case of 8bit-parameters set, there's no problem and the audio quality was good enough.

[Blaizzy]

Hey @yoshphys

Thanks for your contribution!

I would add a quant predicate to skip the codec, embedding and projections from the quantisation process or use a higher quant. This would ensure high quality even in low quants precision.

Am thinking about how we can automatically scan and suggest this with PRs such as #490 but for now has to be manual.

cc: @lucasnewman

[lucasnewman]

The conversion actually works correctly, so that part should be good. I tested your change with an 8-bit conversion and it looks good, thank you!