Closed tianylijun closed 10 hours ago
Hi @tianylijun, For the Q1, tasks supported here inherit from this popular repo EleutherAI/lm-evaluation-harness. Usually, these tasks cover what you need. As to Q2, calibration has no need to record the generated token. The data distribution is recorded during inference by each layer. If you have any other questions, please let me know.
version: tag V3.0 pytorch: examples/3.x_api/pytorch/nlp/huggingface_models/language-modeling/quantization/smooth_quant/run_clm_no_trainer.py
LLM smoothquant, how to add a customer evaluate func,current seems only given task can be used, how to add a customer evaluate for new dataset ?
run_fn for calibration called by autotun, only do one round forward for a promot, generated token not used, generated token not contribute to fmap distribution?