If this pass is removed, will the performance of some models be affected?

Some will be impact. A simple case should be linear -> add, add won't be fused into linear and cause extra computation. However, the current post-sum fusion need memory inplace and it will be conflict with memory reuse opimization and cause wrong accuracy. We'll re-enable the fusion by using post-binary fusion which doesn't actually need memory inplace but this needs more validation work.