Questions about replication experiments.

[BenchuYee]

Are the code and parameters in this repository consistent with the parameters used in the experiments described in the paper? I conducted an experiment on an A100 using the provided command "bash 10_optimize_rotation.sh meta-llama/Llama-2-7b 4 4 4" keeping the parameters unchanged in the script. However, the results I obtained differ somewhat from the results reported in the paper. This is my command "bash 10_optimize_rotation.sh meta-llama/Llama-2-7b 4 4 4", and I only change the nproc_per_node to 1. When I use the eval command "2_eval_ptq.sh $model_name 4 4 4", the ppl is about 8.49. The model I use is NousResearch/Llama-2-7b-chat-hf.

[cokeshao]

+1

for W4A4KV4, here is my result: I change nproc_per_node to 4 and run the code on 4*RTX 4090.

	meta-llama/Llama-2-7b-hf	meta-llama/Meta-Llama-3-8B
in paper	5.9	7.3
mine	6.81	8.06

[WeiMa01]

+1 for W4A4KV4, here is my result: with the default config and run the code on 8*A100, got the following result:

	meta-llama/Llama-2-7b-hf
in paper	5.9
mine	6.85

would you please share the config about the PPL 5.9 in the paper?

[testworldagain]

+1