fix: Qwen3-TTS project input text embedding

[FelysNeko]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Running this Qwen3-TTS example with the alternative model Qwen/Qwen3-TTS-12Hz-0.6B-Base leads to this error.

  File "/root/autodl-tmp/ms-swift/swift/model/models/qwen.py", line 1802, in tts_forward
    input_embeddings = input_text_embedding + input_codec_embedding
                       ~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~
RuntimeError: The size of tensor a (2048) must match the size of tensor b (1024) at non-singleton dimension 2

The root cause is the missing text embedding projection, and the official implementation sft_12hz.py has the same issue. There are several PRs in the Qwen3-TTS repository attempting to patch it: QwenLM/Qwen3-TTS#336 and QwenLM/Qwen3-TTS#278.

Experiment results

Both Qwen/Qwen3-TTS-12Hz-1.7B-Base and Qwen/Qwen3-TTS-12Hz-0.6B-Base run without any problems.

[gemini-code-assist]

Warning

Gemini encountered an error creating the review. You can try again by commenting /gemini review.

[Jintao-Huang]

thanks!

[Jintao-Huang]

#9721