### What this PR does / why we need it? We proposed the FastPatch method, which optimized patch embedding (Conv3D) for Qwen2VL. ### Does this PR introduce _any_ user-facing change? No. ### How was this patch tested? We've tested it on benchmark, it meets our satisfaction and is better than original patch_embed layer. --------- Signed-off-by: baifanxxx <baifanxxx@gmail.com> Signed-off-by: zouyida <zouyida@huawei.com> Co-authored-by: zouyida <zouyida@huawei.com>