support qwen3-tts-0.6B

[Evan-Wang-26]

Describe the bug

The Edge-LLM pipeline currently assumes all Qwen3-TTS models include an MTP (Multi-Token Prediction) module.

When building or running qwen3-tts-0.6B, the framework forcibly initializes the MTP module, but this 0.6B variant does not contain it.

[lix19937]

You means this 0.6B https://huggingface.co/Qwen/Qwen3-TTS-12Hz-0.6B-Base ?

[nvamberl]

Thanks for the report. I suspect a terminology mix-up — two unrelated things in this repo are both called "MTP":

1. Qwen3.5 MTP draft (speculative decoding) — This feature is not released yet up to release/0.7.0 and we plan to release that in release/0.7.1. Qwen3-TTS does not have one and the export path never builds one for qwen3_tts.
2. Qwen3-TTS CodePredictor (CP) — a small 5-layer Qwen3 decoder that predicts the residual codec tokens. The export pipeline does unconditionally build this for qwen3_tts.

On naming: The module's official name is CodePredictor. But "MTP" leaks in from upstream too — the checkpoint contains a weight literally named talker.code_predictor.small_to_mtp_projection.{weight,bias}, so Qwen themselves use "MTP" to describe the function the CP performs (multi-token prediction of the residual codebooks). Edge-LLM inherits that name from the checkpoint — that's why you'll see "MTP" in some weight files / comments while the module itself is still the CodePredictor. They're the same thing in the TTS context; "MTP" in TTS ≠ the Qwen3.5 spec-decode MTP draft.

[Evan-Wang-26]

Yes

…

---- Replied Message ---- | From | ***@***.***> | | Date | 05/15/2026 19:25 | | To | NVIDIA/TensorRT-Edge-LLM ***@***.***> | | Cc | Evan-Wang-26 ***@***.***>, Author ***@***.***> | | Subject | Re: [NVIDIA/TensorRT-Edge-LLM] support qwen3-tts-0.6B (Issue #86) | lix19937 left a comment (NVIDIA/TensorRT-Edge-LLM#86) You means this 0.6B https://huggingface.co/Qwen/Qwen3-TTS-12Hz-0.6B-Base ? — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.Message ID: ***@***.***>

[Evan-Wang-26]

Thanks a lot for your detailed clarification! I now fully understand the difference between the two "MTP" concepts in this repo, and I'll recheck my issue based on this explanation. Appreciate your help!

…

---- Replied Message ---- | From | ***@***.***> | | Date | 05/15/2026 20:36 | | To | NVIDIA/TensorRT-Edge-LLM ***@***.***> | | Cc | Evan-Wang-26 ***@***.***>, Author ***@***.***> | | Subject | Re: [NVIDIA/TensorRT-Edge-LLM] support qwen3-tts-0.6B (Issue #86) | nvamberl left a comment (NVIDIA/TensorRT-Edge-LLM#86) Thanks for the report. I suspect a terminology mix-up — two unrelated things in this repo are both called "MTP": Qwen3.5 MTP draft (speculative decoding) — mtp_draft.py, exported via --mtp. Qwen3-TTS does not have one and the export path never builds one for qwen3_tts. Qwen3-TTS CodePredictor (CP) — a small 5-layer Qwen3 decoder that predicts the residual codec tokens. The export pipeline does unconditionally build this for qwen3_tts. On naming: The module's official name is CodePredictor. But "MTP" leaks in from upstream too — the checkpoint contains a weight literally named talker.code_predictor.small_to_mtp_projection.{weight,bias}, so Qwen themselves use "MTP" to describe the function the CP performs (multi-token prediction of the residual codebooks). Edge-LLM inherits that name from the checkpoint — that's why you'll see "MTP" in some weight files / comments while the module itself is still the CodePredictor. They're the same thing in the TTS context; "MTP" in TTS ≠ the Qwen3.5 spec-decode MTP draft. — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.Message ID: ***@***.***>