config: map AutoModelForTextToWaveform so pipeline() resolves
pipeline("text-to-speech", model="rumik-ai/rumik-oss-1", trust_remote_code=True) fails with:
ValueError: Unrecognized configuration class RumikOSSConfig for this kind of
AutoModel: AutoModelForTextToWaveform
That snippet is the one the Hub auto-generates from pipeline_tag, so it is the first thing most users run.
Cause. The TTS pipeline asks for AutoModelForTextToWaveform. auto_map declared only AutoConfig and AutoModelForCausalLM, so transformers fell back to its built-in config->model table, which has no rumik_oss entry.
Change. One line, pointing that Auto class at the existing RumikOSSForCausalLM. No new code, no new file. The model emits audio tokens and Mimi decoding stays the caller's job, exactly as the model card documents.
Verified on an A10G against these weights: pipeline() now loads RumikOSSForCausalLM and returns generated tokens instead of raising. The documented AutoModelForCausalLM example is unchanged.
Note. On transformers 5.x the pipeline then hits an upstream bug (BatchEncoding.to() got an unexpected keyword argument 'dtype', text_to_audio.py) that also breaks suno/bark-small. That is not fixable here.
Thanks for investigating the loading error. This mapping is not sufficient for TTS pipeline support, so please hold the merge.
RumikOSSForCausalLMreturns token IDs, whereasAutoModelForTextToWaveformpromises a waveform. In Transformers 4.57.6,TextToAudioPipeline.postprocess()simply casts these IDs to float and returns them underaudio; it never invokes Mimi. I reproduced this postprocessing behavior with a CPU-only test using the installed pipeline source, not a full checkpoint inference test.- The pipeline calls inherited
generate(), not ourgenerate_audio(). It therefore bypasses the custom audio-vocabulary restriction and learned stop predictor. Ordinary pipeline preprocessing also does not construct our speaker/text/audio prompt.
A complete fix needs a dedicated waveform adapter or custom pipeline that constructs the prompt, calls generate_audio(), removes the input prefix, recovers codec frames, and decodes through the bundled Mimi codec, returning actual audio at 24 kHz. Please retain the existing causal-LM interface unchanged. Both releases need end-to-end waveform tests on explicitly supported Transformers versions before merging. The reported Transformers 5.x error is separate from these issues; removing the loading error alone does not establish compatibility.