v2.2.5

New Features:

1. Support for RLHF with multimodal models, including DPO, ORPO, SIMPO, and CPO
2. SWIFT infer, SWIFT deploy support using lmdeploy for inference acceleration.
3. Support the use of Megatron for performing PT and SFT on the Qwen2 series models.
4. Support the grounding task for InternVL2/Qwen-VL-Chat models

New Models:

1. mistral-nemo series, mistral-large
2. llama3.1 series

New Datasets:

1. sa1b-dense-caption, sa1b-paired-caption
2. rlaif-v
3. zhihu-kol, zhihu-kol-filtered

What's Changed

• Support api key by @Jintao-Huang in #1452
• Support mistral nemo series models by @tastelikefeet in #1454
• fix minicpm-v2.5 lora_target_modules by @Jintao-Huang in #1455
• Add two datasets by @tastelikefeet in #1459
• Update trl dependency version by @hjh0119 in #1463
• fix bugs by @Jintao-Huang in #1464
• fix yi1.5 by @Jintao-Huang in #1465
• Fix yi1.5 by @Jintao-Huang in #1467
• add activate and deactivate for part tuner by @tastelikefeet in #1470
• support llama3.1 by @Jintao-Huang in #1475
• support megatron by @Jintao-Huang in #1365
• fix megatron by @Jintao-Huang in #1476
• Support internvl2 grounding by @tastelikefeet in #1473
• update doc by @tastelikefeet in #1477
• Support alignment algorithm for vision MLLM by @hjh0119 in #1474
• fix doc by @hjh0119 in #1481
• Fix visual cpo by @hjh0119 in #1482
• support llama3.1-quant by @Jintao-Huang in #1478
• fix part tuner by @tastelikefeet in #1483
• fix import by @tastelikefeet in #1488
• Fix GLM4V by @tastelikefeet in #1490
• support mistral large by @Jintao-Huang in #1485
• fix resume_only_model & zero3 & full by @Jintao-Huang in #1497
• Fix resume_from_checkpoint & full by @Jintao-Huang in #1498
• fix part tuner by @tastelikefeet in #1495
• fix cogvlm2-video by @Jintao-Huang in #1500
• [TorchAcc] add script for qwen2 in torchacc by @Zhikaiiii in #1492
• Fix CI by @tastelikefeet in #1501
• fix vlm template by @Jintao-Huang in #1503
• fix internvl-4b by @Jintao-Huang in #1505
• support zero3 & freeze by @Jintao-Huang in #1508
• fix part mix with lora by @tastelikefeet in #1509
• fix docs by @Jintao-Huang in #1511
• Update README.md by @ArtificialZeng in #1516
• fix kto custom data by @hjh0119 in #1515
• Fix KTO doc by @hjh0119 in #1517
• Rescale image by @tastelikefeet in #1512
• fix pretrain dataset by @Jintao-Huang in #1518
• fix deepseek-vl template by @Jintao-Huang in #1521
• Support exporting of llama3.1, and awq-batch-size by @tastelikefeet in #1522
• support lmdeploy & vlm by @Jintao-Huang in #1364
• fix tf 4.43 llava by @tastelikefeet in #1525
• fix llamapro by @tastelikefeet in #1527
• fix template & docs by @Jintao-Huang in #1529
• fix lmdeploy & vlm by @Jintao-Huang in #1530
• update doc by @tastelikefeet in #1531
• fix lmdeploy & minicpm-v-2.5 by @Jintao-Huang in #1534
• fix internvl-phi3 batch infer by @hjh0119 in #1539
• Support SA1B series datasets by @tastelikefeet in #1542
• fix bug in _prepare_inputs by @guihonghao in #1543
• Support lmdeploy infer deploy by @Jintao-Huang in #1541
• add lmdeploy link by @Jintao-Huang in #1545
• support lmdeploy & app-ui by @Jintao-Huang in #1546
• fix lmdeploy bug by @Jintao-Huang in #1550
• support more models by @Jintao-Huang in #1552
• fix multi node by @tastelikefeet in #1554
• support lmdeploy awq by @Jintao-Huang in #1555
• support quant_policy by @Jintao-Huang in #1556
• fix xcomposer lora by @Jintao-Huang in #1559
• Update docs by @Jintao-Huang in #1558
• fix minicpm-v by @Jintao-Huang in #1562
• add result_dir paramerter to InferArgument & fix a minor bug by @starxhong in #1561
• fix some bugs in dpo by @hjh0119 in #1565
• Fix bugs 0801 by @Jintao-Huang in #1566
• fix dataset copy by @Jintao-Huang in #1569
• fix qwen-vl-merged lmdeploy by @Jintao-Huang in #1572

New Contributors

• @ArtificialZeng made their first contribution in #1516
• @guihonghao made their first contribution in #1543

Full Changelog: v2.2.3...v2.2.5