Qwen3-TTS CoreML converter: 1.7B Base + 1024 context parameterization?

[davidqyc]

Hi @Alex-Wengg — I'm adapting the public Qwen3-TTS CoreML conversion work in this repo for Qwen/Qwen3-TTS-12Hz-1.7B-Base.

I found the current converter under models/tts/qwen3/coreml/convert/, especially convert_argmax_code_decoder.py. This is much closer to what we need than starting from scratch.

The current CodeDecoder converter appears to have two 0.6B/256-specific assumptions:

• MAX_SEQ_LEN = 256;
• the test/input embedding tensor is fixed at hidden size 1024, while 1.7B Base uses hidden size 2048.

Before we patch it blindly, could you clarify:

1. Have you already tested or internally converted Qwen3-TTS 1.7B (Base or CustomVoice) with this converter family?
2. Is changing MAX_SEQ_LEN from 256 to 1024 intended to be a mechanical regeneration of cache/mask/attention shapes, or are there other sequence-length-dependent graph constants/steps we need to change?
3. For 1.7B, is the intended adaptation simply to derive all hidden/KV dimensions from talker.config (including the input embedding shape), or are there known 1.7B-specific conversion issues?
4. Is the current explicit-KV export + later MLState/stateful conversion still the recommended path, or is there a newer stateful exporter we should use?
5. If you have a newer 1.7B/longer-context branch or script that is not yet on main, a pointer would be extremely helpful.

We only need the auditable recipe/constraints, not prebuilt weights. Thanks!