Repository navigation
Has anyone encountered a drop in model accuracy after migration？ #18951
|
"I recently deployed the Qwen3 model on AMD using llama.cpp and replaced it with my own trained model weight file, but the test results on AMD differ significantly from those on my NVIDIA server. The accuracy on the NVIDIA server is 80%, while the accuracy on the AMD server is only 20%. Do you think this is a hardware issue or a software impact caused by llama.cpp?" |
Replies: 1 comment
|
I believe I have identified the issue. In my project, the input stream is an array of extracted frames from a video. From the file clip.cpp in llama.cpp (lines 778 and 3332–3334), I learned that it does not support n_batch > 1. Additionally, from mtmd.cpp (lines 629–639 and 1085–1092), I discovered that image processing indeed lacks a temporal dimension, and llama.cpp does not support video input. Consequently, when I tested with video, the highest accuracy achieved was only 22.51%. |
I believe I have identified the issue. In my project, the input stream is an array of extracted frames from a video. From the file clip.cpp in llama.cpp (lines 778 and 3332–3334), I learned that it does not support n_batch > 1. Additionally, from mtmd.cpp (lines 629–639 and 1085–1092), I discovered that image processing indeed lacks a temporal dimension, and llama.cpp does not support video input. Consequently, when I tested with video, the highest accuracy achieved was only 22.51%.