`IndexError: Dimension out of range` during inference with merged LoRA model

[harsh020]

Summary

Inference fails with an IndexError when using a LoRA-fine-tuned Soprano model (merged or adapter-based) via SopranoTTS.
This appears to be caused by incorrect assumptions about the shape of hidden_states inside infer_batch.

---

Environment

• OS: Linux (Colab / Kaggle)
• Python: 3.12
• torch: 2.9.0
• transformers: 4.57.6
• peft: 0.18.1
• soprano: latest from pip
• Base model: ekwek/Soprano-1.1-80M
• Fine-tuning: LoRA (PEFT), language adaptation (non-English)

---

What I did

1. Fine-tuned ekwek/Soprano-1.1-80M using LoRA for a different language.
2. Merged LoRA weights back into the base model:

   merged_model = model.merge_and_unload()
   merged_model.save_pretrained("merged")
   tokenizer.save_pretrained("merged")

3. Copied decoder.pth from the base Soprano repo into merged/.
4. Ran inference using SopranoTTS.

---

Error Trace

IndexError                                Traceback (most recent call last)
[/tmp/ipython-input-2296956188.py](https://localhost:8080/#) in <cell line: 0>()
      3 model = SopranoTTS(backend='auto', model_path='/content/merged-fixed', device='auto', cache_size_mb=100, decoder_batch_size=1)
      4 
----> 5 out = model.infer("some non enligh text")
      6 print(out)
      7 with open("audio.wav", 'wb+') as f:

1 frames
[/usr/local/lib/python3.12/dist-packages/soprano/tts.py](https://localhost:8080/#) in infer_batch(self, texts, out_dir, top_p, temperature, repetition_penalty, retries)
    181                 batch_hidden_states.append(torch.cat([
    182                     torch.zeros((1, 512, lengths[0]-lengths[i]), device=self.device),
--> 183                     hidden_states[idx+i].unsqueeze(0).transpose(1,2).to(self.device).to(torch.float32),
    184                 ], dim=2))
    185             batch_hidden_states = torch.cat(batch_hidden_states)

IndexError: Dimension out of range (expected to be in range of [-2, 1], but got 2)