lm_eval dependency should be removed, as it's causing import errors for torchao.quantization.GPTQ when lm_eval is not installed. Approach: Move InputRecorder and similar classes to utils and guard it within _lm_eval_available