Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

update readme

wenhuach21fb82da38a67c
diff --git a/README.md b/README.md
index 0295e9121..9d9749220 100644
--- a/README.md
+++ b/README.md
@@ -265,12 +265,7 @@ For wikitext2/ptb-new/c4-new ppl, we follow the code of gptq and set the seqence
</table>
-We also shows some results compared to other methods. For a fair comparison, we utilized 512 samples from Pile-10k for all methods during calibration. Due to memory constraints, we maintained the original sequence length of 512 for AWQ, while for GPTQ and our approach, a sequence length of 2048 is used. We have enalbed act-order and true-seqential in GPTQ, and the notation GPTQ* indicates that we adjusted the random seed or data preprocessing to address issues related to the non-positive definite Hessian matrix or other issues.
-![](./figs/W4G-1.png)
-![](./figs/W4G128.png)
-![](./figs/W3G128.png)
-![](./figs/W2G128.png)
-
+We provide a comparative analysis with other methods [link](./acc_data/REAME.md) in our accuracy data section. Notably, our approach has outperformed GPTQ with a score of 30/32 and AWQ with a score of 27/32 across llamv1/llamav2/mistral-7b on W4G-1, W4G128, W3G128, W2G128. And the tuning costs are comparable.
### Modes passed smoke test
LaMini-GPT-124M; QWEN1-8B; OPT-125M; Bloom-560m;falcon-7b;gpt-leo-125m;stablelm-base-alpha-3b;dolly-v2-3b;mpt-7b;gpt-j-6b;chatglm2-6b
diff --git a/acc_data/REAME.md b/acc_data/REAME.md
new file mode 100644
index 000000000..fd8df5954
--- /dev/null
+++ b/acc_data/REAME.md
@@ -0,0 +1,5 @@
+For a fair comparison, we utilized 512 samples from Pile-10k for all methods during calibration. Due to memory constraints, we maintained the original sequence length of 512 for AWQ, while for GPTQ and our approach, a sequence length of 2048 is used. We have enalbed act-order and true-seqential in GPTQ, and the notation GPTQ* indicates that we adjusted the random seed or data preprocessing to address issues related to the non-positive definite Hessian matrix or other issues.
+![](W4G-1.png)
+![](W4G128.png)
+![](W3G128.png)
+![](W2G128.png)
diff --git a/figs/W2G128.png b/acc_data/W2G128.png
similarity index 100%
rename from figs/W2G128.png
rename to acc_data/W2G128.png
diff --git a/figs/W3G128.png b/acc_data/W3G128.png
similarity index 100%
rename from figs/W3G128.png
rename to acc_data/W3G128.png
diff --git a/figs/W4G-1.png b/acc_data/W4G-1.png
similarity index 100%
rename from figs/W4G-1.png
rename to acc_data/W4G-1.png
diff --git a/figs/W4G128.png b/acc_data/W4G128.png
similarity index 100%
rename from figs/W4G128.png
rename to acc_data/W4G128.png