MMEngine Release V0.8.3

v0.8.3 (31/07/2023)

Highlights

• Support enabling efficient_conv_bn_eval for efficient convolution and batch normalization. See save memory on gpu for more details
• Add Llama2 finetune example
• Support multi-node distributed training with MLU backend

New Features & Enhancements

• Enable efficient_conv_bn_eval for memory saving convolution and batch normalization by @youkaichao in #1202, #1251 and #1259
• Add Llama2 example by @HAOCHENYE in #1264
• Compare the difference of two configs by @gachiemchiep in #1260
• Enable explicit error for deepspeed not installed by @Li-Qingyun in #1240
• Support skipping initialization in BaseModule by @HAOCHENYE in #1263
• Add parameter save_begin to control when to save checkpoints by @KerwinKai in #1271
• Support multi-node distributed training with MLU backend by @josh6688 in #1266
• Enhance error message thrown by Config, build function and ConfigDict.items by @HAOCHENYE in #1272, #1270 and #1088
• Add the loop_stage runtime information in message_hub by @zhouzaida in #1277
• Fix Visualizer that built vis_backends will not be used when save_dir is None by @Xinyu302 in #1275

Bug fixes

• Fix scalar check in RuntimeInfoHook by @i-aki-y in #1250
• Move data preprocessor to target device in FSDPStrategy by @HAOCHENYE in #1261

Docs

• Add ecosystem in README by @zhouzaida in #1247
• Add short explanation about registry scope by @mmeendez8 in #1114
• Add the data flow of Runner in README by @zhouzaida in #1257
• Introduce how to customize distributed training settings @zhouzaida in #1279

New Contributors

• @youkaichao made their first contribution in #1202
• @mmeendez8 made their first contribution in #1114
• @Xinyu302 made their first contribution in #1275

Full Changelog: v0.8.2...v0.8.3