AMX matrix multiplication implementation of a 4-bit quantified version that is better than libxsmm when the batch_size exceeds 16.
I rewrote the AMX operation part of the 4-bit quantization code in the woq linear section, and the model used for testing is DeepSeek-R1-Distill-Qwen-32B.
I conducted performance testing and found a 20% improvement compared to the original code. The result chart is as follows.
In this repository, I only provided the location of the modified code. You just need to replace the original files and use the official script to compile it.