The talker_hidden_states not obtained correctly

Description

Error code location:

Reproduction

1. Error code location: https://github.com/QwenLM/Qwen3-TTS/blob/main/finetuning/sft_12hz.py#L108
2. Error Description： The position where codec_mask is True corresponds to the actual codec. hidden_states[codec_mask[:, :-1]] retrieves the hidden_states of the actual codec. Note that the hidden state at time t is used to predict the codec at time t+1. The above logic cannot be matched with the code: https://github.com/QwenLM/Qwen3-TTS/blob/main/qwen_tts/core/models/modeling_qwen3_tts.py#L1612 The hidden state at time t should not be concatenated with the codec embedding at time t (the hidden state at time t-1 is used to predict the codec at time t), but should be concatenated with the codec embedding at time t+1. The correct value is talker_hidden_states=hidden_states[codec_mask[:, 1:]], https://github.com/QwenLM/Qwen3-TTS/blob/main/finetuning/sft_12hz.py#L108

Logs

Environment Information

no Environment

Known Issue

• The issue hasn't been already addressed in Documentation, Issues, and Discussions.