Fix CSM TextToAudioPipeline missing <bos> token

[jiqing-feng]

What does this PR do?

CsmProcessor defaults add_special_tokens=False (designed for apply_chat_template, which includes <bos> in the Jinja template). When the pipeline calls preprocessor(text) directly for raw text input, <bos> (128000) and <eos> (128001) are missing from the tokenized sequence. Without these tokens the model receives malformed input it was never trained on, making generation unstable — certain seed/sampling parameter combinations cause the model to emit all-zero codebook frames, which are treated as EOS (codebook_eos_token_id=0), resulting in an empty audio tensor that crashes Mimi's Conv1d decoder.

Fix: set add_special_tokens=True for CSM in pipeline preprocess.

Reproduction

from transformers import pipeline, set_seed

pipe = pipeline("text-to-speech", model="sesame/csm-1b")
set_seed(777)
output = pipe("Hello, my dog is cooler than you!", forward_params={"do_sample": True, "temperature": 0.7, "top_k": 50, "top_p": 0.95})
# RuntimeError: Calculated padded input size per channel: (0). Kernel size: (1).
# Kernel size can't be greater than actual input size

error:

......
  File "/home/jiqing/transformers/src/transformers/models/csm/generation_csm.py", line 478, in generate
    codec_decode_output = self.codec_model.decode(audio_codes_batch.transpose(0, 1).unsqueeze(0))                                                              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^                                      File "/home/jiqing/transformers/src/transformers/models/mimi/modeling_mimi.py", line 1666, in decode
    audio_values, decoder_past_key_values = self._decode_frame(                                                                                                                  ^^^^^^^^^^^^^^^^^^^
  File "/home/jiqing/transformers/src/transformers/models/mimi/modeling_mimi.py", line 1619, in _decode_frame
    embeddings = self.quantizer.decode(codes)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/jiqing/transformers/src/transformers/models/mimi/modeling_mimi.py", line 1344, in decode
    quantized_out = self.semantic_residual_vector_quantizer.decode(codes[:, : self.num_semantic_quantizers])
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/jiqing/transformers/src/transformers/models/mimi/modeling_mimi.py", line 1292, in decode
    quantized_out = self.output_proj(quantized_out)
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1778, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1789, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/conv.py", line 385, in forward
    return self._conv_forward(input, self.weight, self.bias)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/conv.py", line 380, in _conv_forward
    return F.conv1d(
           ^^^^^^^^^
RuntimeError: Calculated padded input size per channel: (0). Kernel size: (1). Kernel size can't be greater than actual input size

Hi @Rocketknight1 . Would you please review this PR? Thanks!

[Rocketknight1]

Yes, the fix makes sense, thank you!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.