fix: remove double label-shifting in finetuning that caused speech speed-up

The HF ForCausalLMLoss already shifts labels internally, but sft_12hz.py
was also manually shifting before passing to model.talker(), causing
double-shifting and progressively faster audio with more epochs.

Additionally, the sub-talker forward_finetune used ForCausalLMLoss on
already-aligned logits/labels, adding another unwanted shift. Replaced
with direct cross_entropy.

See: https://github.com/QwenLM/Qwen3-TTS/issues/179#issuecomment-3870059313

finetuning/sft_12hz.py

@@ -93,15 +93,16 @@ def train():
                 input_embeddings = input_text_embedding + input_codec_embedding
 
                 outputs = model.talker(
-                    inputs_embeds=input_embeddings[:, :-1, :],
-                    attention_mask=attention_mask[:, :-1],
-                    labels=codec_0_labels[:, 1:],
+                    inputs_embeds=input_embeddings,
+                    attention_mask=attention_mask,
+                    labels=codec_0_labels,
                     output_hidden_states=True
                 )
 
                 hidden_states = outputs.hidden_states[0][-1]
-                talker_hidden_states = hidden_states[codec_mask[:, 1:]]
-                talker_codec_ids = codec_ids[codec_mask]
+                target_codec_mask = codec_mask[:, 1:]
+                talker_hidden_states = hidden_states[:, :-1][target_codec_mask]
+                talker_codec_ids = codec_ids[:, 1:][target_codec_mask]
 
                 sub_talker_logits, sub_talker_loss = model.talker.forward_sub_talker_finetune(talker_codec_ids, talker_hidden_states)
 

qwen_tts/core/models/modeling_qwen3_tts.py

@@ -1321,7 +1321,11 @@ class Qwen3TTSTalkerCodePredictorModelForConditionalGeneration(Qwen3TTSPreTraine
 
         loss = None
         if labels is not None:
-            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+            loss = torch.nn.functional.cross_entropy(
+                logits.reshape(-1, self.config.vocab_size),
+                labels.reshape(-1),
+                ignore_index=-100,
+            )
 
         return Qwen3TTSTalkerCodePredictorOutputWithPast(
             loss=loss,