Replies: 1 comment 1 reply
-
Debugging could be tedious - you can try to compare the results after each op with the reference implementation. Or open a PR and more people can have a look From a quick look, the following code incorrectly passes + // 创建两个视图张量,分别表示切分后的两部分
+ int64_t split_point = cur->ne[0] / 2;
+ struct ggml_tensor * x0 = ggml_cont(ctx, ggml_view_2d(ctx, cur, split_point, cur->ne[1], cur->nb[1], 0));
+ struct ggml_tensor * x1 = ggml_cont(ctx, ggml_view_2d(ctx, cur, split_point, cur->ne[1], cur->nb[1], split_point)); |
Beta Was this translation helpful? Give feedback.
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hello everyone, I'm currently adding support for the chatglm3-6b model on llama.cpp. I've done a quick check on the inference algorithm, but I didn't find any issues.
However, the inference results are all over the place.
How should I go about debugging this?
Are there any tricks for pinpointing the issue when inference goes wrong?
The following are my log link and code modifications:
main.log
0001-feature-add-chatglm-support.patch.txt
Beta Was this translation helpful? Give feedback.
All reactions