v2.4.0

English Version

New Features:

1. Support for Liger, which accommodates models like LLaMA, Qwen, Mistral, etc., and reduces memory usage by 10% to 60%.
2. Support for custom loss function training using a registration mechanism.
3. Training now supports pushing models to ModelScope and HuggingFace.
4. Support for the freeze_vit parameter to control the behavior of full parameter training for multimodal models.

New Models:

1. Qwen2-VL series includes GPTQ/AWQ quantized models. For best practices, see here.
2. InternVL2 AWQ quantized models.

New Datasets:

1. qwen2-pro series

中文版

新特性：

1. 支持 Liger训练LLaMA、Qwen、Mistral 等模型，内存使用降低 10% 至 60%。
2. 支持使用注册机制进行自定义损失函数的训练。
3. 训练支持将模型推送至 ModelScope 和 HuggingFace。
4. 支持 freeze_vit 参数，以控制多模态模型全参数训练的行为。

新模型：

1. Qwen2-VL 系列包括 GPTQ/AWQ 量化模型，最佳实践可以查看这里。
2. InternVL2 AWQ 量化模型。

新数据集：

1. qwen2-pro 系列

What's Changed

• compat with vllm==0.5.5 by @Jintao-Huang in #1812
• Support zero2 offload by @Jintao-Huang in #1814
• fix mp+ddp & resume_from_checkpoint by @Jintao-Huang in #1815
• fix preprocess_num_proc by @Jintao-Huang in #1818
• Support liger by @tastelikefeet in #1819
• fix dora deployment by @tastelikefeet in #1821
• Support register loss func by @Jintao-Huang in #1822
• use default-lora by @Jintao-Huang in #1823
• fix minicpm-v 2.6 infer device_map by @Jintao-Huang in #1832
• Fix code by @tastelikefeet in #1824
• fix inject by @tastelikefeet in #1835
• support qwen2-pro dataset by @Jintao-Huang in #1834
• add ddp_timeout parameter by @tastelikefeet in #1836
• fix internlm-xcomposer rlhf by @hjh0119 in #1838
• Support eval_nproc by @tastelikefeet in #1843
• support qwen2-vl by @Jintao-Huang in #1842
• Add internvl2 awq models by @tastelikefeet in #1846
• Fix some datasets for streaming by @tastelikefeet in #1848
• Fix Pissa and OLoRA by @tastelikefeet in #1852
• Support qwen2 vl grounding by @tastelikefeet in #1854
• support qwen2-vl & video finetune by @Jintao-Huang in #1849
• Update new datasets by @tastelikefeet in #1855
• update qwen2-vl docs by @Jintao-Huang in #1856
• update qwen2-vl docs by @Jintao-Huang in #1858
• fix qwen2-vl docs by @Jintao-Huang in #1861
• fix requirements by @Jintao-Huang in #1864
• update docs qwen2-vl by @Jintao-Huang in #1869
• Support faster data map by @tastelikefeet in #1871
• [TorchAcc] fix serveral bugs for torchacc FSDP. by @baoleai in #1872
• Add train record by @tastelikefeet in #1873
• Fix num_proc by @Jintao-Huang in #1874
• Fix neftune doc by @tastelikefeet in #1875
• add duet by @tastelikefeet in #1877
• use model.generation_config by @Jintao-Huang in #1850
• Support freeze vit by @Jintao-Huang in #1880
• support qwen2-vl gptq awq by @Jintao-Huang in #1884
• Refactor push_to_hub by @tastelikefeet in #1883
• Fix push to hub logic by @tastelikefeet in #1888
• add vllm lmdeploy benchmark by @Jintao-Huang in #1889
• Add some warnings and fix RLHF by @tastelikefeet in #1890

Full Changelog: v2.3.2...v2.4.0