- Change vllm_inference.py app name from example-vllm-inference to vllm-inference
- Replace hardcoded _VLLM_BASE_URL with _modal.Function.from_name() lookup
- Make llm a lazy singleton via _get_llm() to support dynamic URL resolution
Rewrite transcribe_generator.py to match Cohere quickstart logic
- Use CohereAsrForConditionalGeneration instead of AutoModelForSpeechSeq2Seq
- Use device_map="auto" instead of manual .to("cuda")
- Use load_audio from transformers.audio_utils instead of soundfile+librosa
- Add language="en" to processor call
- Use model.generate(**inputs, max_new_tokens=256)
- Use processor.decode() instead of processor.batch_decode()
- Remove monkey-patch for transformers 5.x compat
- Remove unused soundfile, librosa, sentencepiece dependencies
Fix Cohere transcription: correct model class, load_audio, device_map, inputs pattern
