MMEngine Release V0.9.0

v0.9.0 (10/10/2023)

Highlights

• Support training with ColossalAI. Refer to the Training Large Models for more detailed usages.
• Support gradient checkpointing. Refer to the Save Memory on GPU for more details.
• Supports multiple visualization backends, including NeptuneVisBackend, DVCLiveVisBackend and AimVisBackend. Refer to Visualization Backends for more details.

New Features & Enhancements

• Add a text translation example by @Desjajja in #1283
• Add NeptuneVisBackend by @wangerlie in #1311
• Add ColossalAI strategy by @HAOCHENYE in #1299
• Add collect_results support for Ascend NPU by @xuuyangg in #1309
• Unify the parameter style of DeepSpeedStrategy by @LZHgrla in #1320
• Add progressbar rich by @Dominic23331 in #1157
• Support using other file handlers by @KevinNuNu in #1188
• Refine error message by @zhouzaida in #1338
• Implement gradient checkpointing by @zeyuanyin in #1319
• Add DVCLiveVisBackend by @RangeKing in #1336
• Add AimVisBackend by @RangeKing in #1347
• Support bitsandbytes by @okotaku in #1357
• Support Adafactor Optimizer by @okotaku in #1361
• Add unit tests for autocast with Ascend device by @6Vvv in #1363
• Support metainfo of dataset can be a generic dict-like Mapping by @hiyyg in #1378
• Support for installing minimal runtime dependencies by @HAOCHENYE in #1362
• Support setting the number of iterations in Runner for each epoch by @ShuRaymond in #1292
• Support using gradient checkpointing in FSDP by @HAOCHENYE in #1382

Docs

• Add README for examples by @zhouzaida in #1295
• Add a new ecosystem in README by @zhouzaida in #1296
• Fix typo by @zhouzaida in #1298
• Add an image for Neptune by @zhouzaida in #1312
• Fix docs of ColossalAI by @HAOCHENYE in #1315
• Update QRCode by @crazysteeaam in #1328
• Add activation checkpointing usage by @zhouzaida in #1341
• Fix typo by @zhouzaida in #1348
• Update the usage of bitsandbytes in Chinese documents by @zhouzaida in #1359
• Fix doc typo our_dir in LoggerHook by @wangg12 in #1373
• Add the contributing doc in pr template by @zhouzaida in #1380
• Update the version info by @fanqiNO1 in #1383
• Fix typo by @fanqiNO1 in #1385

Bug Fixes

• Ignore examples in CI by @zhouzaida in #1297
• Fix multi-card issue in PyTorch v2.1 on Ascend by @LRJKD in #1321
• Fix get optimizer_cls by @HAOCHENYE in #1324
• Fix ndarray metainfo check in ConcatDataset by @NrealLzx in #1333
• Adapt to PyTorch v2.1 on Ascend by @LRJKD in #1332
• Fix the type check of tasks in progress bar by @zhouzaida in #1340
• The keyword mode appears nested multiple times in the log by @huaibovip in #1305
• Fix pydantic version to fix mlflow unit tests by @zhouzaida in #1351
• Fix get class attribute from a string by @HAOCHENYE in #1345
• Delete yapf verify by @okotaku in #1365
• Ensure from_cfg of Runner have the same defaults values as its init by @YinAoXiong in #1368
• Fix docs building error caused by deepspeed by @HAOCHENYE in #1379

New Contributors

• @Desjajja made their first contribution in #1283
• @wangerlie made their first contribution in #1311
• @xuuyangg made their first contribution in #1309
• @LRJKD made their first contribution in #1321
• @crazysteeaam made their first contribution in #1328
• @Dominic23331 made their first contribution in #1157
• @NrealLzx made their first contribution in #1333
• @zeyuanyin made their first contribution in #1319
• @huaibovip made their first contribution in #1305
• @6Vvv made their first contribution in #1363
• @YinAoXiong made their first contribution in #1368
• @wangg12 made their first contribution in #1373
• @hiyyg made their first contribution in #1378
• @ShuRaymond made their first contribution in #1292
• @fanqiNO1 made their first contribution in #1383

Full Changelog: v0.8.4...v0.9.0