pruned NVIDIA's new 30B speech LLM from 67.5GB to 10GB with zero retraining, it still transcribes at 5.16% WER (teacher: 4.12%)
No snapshot captured; read the original:
https://www.reddit.com/r/LocalLLaMA/comments/1xibujk/pruned_nvidias_new_30b_speech_llm_from_675gb_to/