SLICES 更新至 2.0.10:
在 SLICES 字符串的最前面添加了对称群编码。例如，在字符串 o v b OOO g DDO c DDO h DDO + YBO Nd Nd Si Si Ru Ru 0 2 ooo 0 2 oo- 0 3 ooo 0 3 -+o 0 3 -oo 0 3 o+o 0 5 ooo 0 5 -oo 0 4 o+o 0 4 ooo 1 3 ooo 1 3 oo+ 1 2 ooo 1 2 +-o 1 2 o-o 1 2 +oo 1 5 o-+ 1 5 oo+ 1 4 oo+ 1 4 +o+ 2 4 oo+ 2 4 o++ 2 5 -o+ 2 5 oo+ 3 4 +oo 3 4 ooo 3 5 ooo 3 5 o-o 中，o v b OOO g DDO c DDO h DDO + YBO 部分即为对称群的编码。
MatterGPT 大幅度更新: 引入了 SLICES-PLUS 论文中提出的 Crystal_system embedding 机制，采用双 block 架构，大幅提升了 MatterGPT 生成高对称结构的精准度。使用方法与上一版本一致，属于无缝升级。
MatterGPT 开启 Flash Attention 作为默认版本:
开启 Flash Attention 的 MatterGPT 版本现已成为默认版本，位于 MatterGPT 文件夹中。此版本仅支持本地安装，不支持 Docker 镜像（因为需要使用 cuda-devel 版本重新构建镜像）。相比旧版本（现已移至 MatterGPT_old 文件夹），该版本速度提升约 2 倍，训练和推理均实现 2 倍提速，性能提升非常明显。我们强烈推荐大家使用此默认版本。
未开启 Flash Attention 的旧版本已移至 MatterGPT_old 文件夹。
SLICES Update to 2.0.10:
Added symmetry group encoding at the beginning of SLICES strings. For example, in the string o v b OOO g DDO c DDO h DDO + YBO Nd Nd Si Si Ru Ru 0 2 ooo 0 2 oo- 0 3 ooo 0 3 -+o 0 3 -oo 0 3 o+o 0 5 ooo 0 5 -oo 0 4 o+o 0 4 ooo 1 3 ooo 1 3 oo+ 1 2 ooo 1 2 +-o 1 2 o-o 1 2 +oo 1 5 o-+ 1 5 oo+ 1 4 oo+ 1 4 +o+ 2 4 oo+ 2 4 o++ 2 5 -o+ 2 5 oo+ 3 4 +oo 3 4 ooo 3 5 ooo 3 5 o-o, the part o v b OOO g DDO c DDO h DDO + YBO represents the symmetry group encoding.
MatterGPT Major Update: Incorporated the Crystal_system embedding mechanism proposed in the SLICES-PLUS paper, adopting a dual-block architecture. This significantly improves the accuracy of MatterGPT in generating highly symmetric structures. Usage remains consistent with the previous version, making this a seamless upgrade.
MatterGPT with Flash Attention as Default:
The version of MatterGPT with Flash Attention enabled is now the default version, located in the MatterGPT folder. This version can only be installed locally and is not supported in Docker images (as it requires rebuilding the image with the cuda-devel version). Compared to the older version (now in MatterGPT_old), it offers approximately 2x speedup, with noticeable improvements in both training and inference. We highly recommend users adopt this default version for its significant performance boost.
The previous version without Flash Attention has been moved to the MatterGPT_old folder.