v0.2.0rc0

What's Changed

• Support FSDP2
• Support Contiguous Batching for RLHF
• readme中增加了MiniCPM的支持 by @LDLINGLINGLING in #869
• [Bug] fix dsv2 attn dispatch (softmax_scale) by @HIT-cwh in #873
• [Bug] fix openai_map_fn bugs by @HIT-cwh in #885
• support transformers >= 4.43 by @HIT-cwh in #878
• Add internlm2 5 cfgs by @HIT-cwh in #872
• [Bugs] fix qlora convert bugs by @HIT-cwh in #930
• 对Minicpm3进行了支持 by @LDLINGLINGLING in #954
• Add functionality to download models from sources other than HuggingFace by @starmountain1997 in #946
• Add Ascend NPU as a backend by @Tonyztj in #983
• Support 4 48 by @HIT-cwh in #985
• [Feature] Auto patch for different devices by @pppppM in #986
• [Fix]MLU Device Mesh by @pppppM in #987
• bump version to v0.2.0rc0 by @pppppM in #990

New Contributors

• @starmountain1997 made their first contribution in #946
• @Tonyztj made their first contribution in #983

Full Changelog: v0.1.23...v0.2.0rc0

XTuner Release V0.1.23

What's Changed

• Support InternVL 1.5/2.0 finetune by @hhaAndroid in #737
• [Bug] fix preference_collate_fn attn_mask by @HIT-cwh in #859
• bump version to 0.1.23 by @HIT-cwh in #862

Full Changelog: v0.1.22...v0.1.23

XTuner Release V0.1.22

What's Changed

• [Refactor] fix internlm2 dispatch by @HIT-cwh in #779
• Fix zero3 compatibility issue for DPO by @Johnson-Wang in #781
• [Fix] Fix map_fn in custom_dataset/sft by @fanqiNO1 in #785
• [Fix] fix configs by @HIT-cwh in #783
• [Docs] DPO and Reward Model documents by @RangiLyu in #751
• Support internlm2.5 by @HIT-cwh in #803
• [Bugs] fix dispatch bugs when model not in LOWEST_TRANSFORMERS_VERSION by @HIT-cwh in #802
• [Docs] fix benchmark table by @HIT-cwh in #801
• [Feature] support output without loss in openai_map_fn by @HIT-cwh in #816
• [Docs] fix typos in sp docs by @HIT-cwh in #821
• [Feature] Support the DatasetInfoHook of DPO training by @xu-song in #787
• [Enhance]: Fix sequence parallel memory bottleneck in DPO & ORPO by @RangiLyu in #830
• [Fix] Fix typo by @bychen7 in #795
• [Fix] fix initialization of ref_llm for full param dpo training with zero-3 by @xu-song in #778
• [Bugs] Fix attn mask by @HIT-cwh in #852
• fix lint by @HIT-cwh in #854
• [Bugs] Fix dispatch attn bug by @HIT-cwh in #829
• [Docs]: update readme and DPO en docs by @RangiLyu in #853
• Added minicpm config file to support sft、qlora、lora、dpo by @LDLINGLINGLING in #847
• fix lint by @HIT-cwh in #856
• bump version to 0.1.22 by @HIT-cwh in #855

New Contributors

• @Johnson-Wang made their first contribution in #781
• @xu-song made their first contribution in #787
• @bychen7 made their first contribution in #795
• @LDLINGLINGLING made their first contribution in #847

Full Changelog: v0.1.21...v0.1.22

XTuner Release V0.1.21

What's Changed

• [Feature] Support DPO, ORPO and Reward Model by @RangiLyu in #743
• [Bugs] fix dispatch bugs by @HIT-cwh in #775
• [Bugs] Fix HFCheckpointHook bugs when training deepseekv2 and mixtral withou… by @HIT-cwh in #774
• [Feature] Support the scenario where sp size is not divisible by attn head num by @HIT-cwh in #769
• bump version to 0.1.21 by @HIT-cwh in #776

Full Changelog: v0.1.20...v0.1.21

XTuner Release V0.1.20

What's Changed

• [Enhancement] Optimizing Memory Usage during ZeRO Checkpoint Convert by @pppppM in #582
• [Fix] ZeRO2 Checkpoint Convert Bug by @pppppM in #684
• [Feature] support auto saving tokenizer by @HIT-cwh in #696
• [Bug] fix internlm2 flash attn by @HIT-cwh in #693
• [Bug] The LoRA model will have meta-tensor during the pth_to_hf phase. by @pppppM in #697
• [Bug] fix cfg check by @HIT-cwh in #729
• [Bugs] Fix bugs caused by sequence parallel when deepspeed is not used. by @HIT-cwh in #752
• [Fix] Avoid incorrect torchrun invocation with --launcher slurm by @LZHgrla in #728
• [fix] fix save eval result failed with mutil-node pretrain by @HoBeedzc in #678
• [Improve] Support the export of various LLaVA formats with pth_to_hf by @LZHgrla in #708
• [Refactor] refactor dispatch_modules by @HIT-cwh in #731
• [Docs] Readthedocs ZH by @pppppM in #553
• [Feature] Support finetune Deepseek v2 by @HIT-cwh in #663
• bump version to 0.1.20 by @HIT-cwh in #766

New Contributors

• @HoBeedzc made their first contribution in #678

Full Changelog: v0.1.19...v0.1.20

XTuner Release V0.1.19

What's Changed

• [Fix] LLaVA-v1.5 official settings by @LZHgrla in #594
• [Feature] Release LLaVA-Llama-3-8B by @LZHgrla in #595
• [Improve] Add single-gpu configs for LLaVA-Llama-3-8B by @LZHgrla in #596
• [Docs] Add wisemodel badge by @LZHgrla in #597
• [Feature] Support load_json_file with json.load by @HIT-cwh in #610
• [Feature]Support Mircosoft Phi3 4K&128K Instruct Models by @pppppM in #603
• [Fix] set dataloader_num_workers=4 for llava training by @LZHgrla in #611
• [Fix] Do not set attn_implementation to flash_attention_2 or sdpa if users already set it in XTuner configs. by @HIT-cwh in #609
• [Release] LLaVA-Phi-3-mini by @LZHgrla in #615
• Update README.md by @eltociear in #608
• [Feature] Refine sp api by @HIT-cwh in #619
• [Feature] Add conversion scripts for LLaVA-Llama-3-8B by @LZHgrla in #618
• [Fix] Convert nan to 0 just for logging by @HIT-cwh in #625
• [Docs] Delete colab and add speed benchmark by @HIT-cwh in #617
• [Feature] Support dsz3+qlora by @HIT-cwh in #600
• [Feature] Add qwen1.5 110b cfgs by @HIT-cwh in #632
• check transformers version before dispatch by @HIT-cwh in #672
• [Fix] convert_xtuner_weights_to_hf with frozen ViT by @LZHgrla in #661
• [Fix] Fix batch-size setting of single-card LLaVA-Llama-3-8B configs by @LZHgrla in #598
• [Feature] add HFCheckpointHook to auto save hf model after the whole training phase by @HIT-cwh in #621
• Remove test info in DatasetInfoHook by @hhaAndroid in #622
• [Improve] Support safe_serialization saving by @LZHgrla in #648
• bump version to 0.1.19 by @HIT-cwh in #675

New Contributors

• @eltociear made their first contribution in #608

Full Changelog: v0.1.18...v0.1.19

XTuner Release V0.1.18

What's Changed

• set dev version by @LZHgrla in #537
• [Fix] Fix typo by @KooSung in #547
• [Feature] support mixtral varlen attn by @HIT-cwh in #564
• [Feature] Support qwen sp and varlen attn by @HIT-cwh in #565
• [Fix]Fix attention mask in default_collate_fn by @pppppM in #567
• Accept pytorch==2.2 as the bugs in triton 2.2 are fixed by @HIT-cwh in #548
• [Feature] Refine Sequence Parallel API by @HIT-cwh in #555
• [Fix] Enhance split_list to support value at the beginning by @LZHgrla in #568
• [Feature] Support cohere by @HIT-cwh in #569
• [Fix] Fix rotary_seq_len in varlen attn in qwen by @HIT-cwh in #574
• [Docs] Add sequence parallel related to readme by @HIT-cwh in #578
• [Bug] SUPPORT_FLASH1 = digit_version(torch.version) >= digit_version('2… by @HIT-cwh in #587
• [Feature] Support Llama 3 by @LZHgrla in #585
• [Docs] Add llama3 8B readme by @HIT-cwh in #588
• [Bugs] Check whether cuda is available when choose torch_dtype in sft.py by @HIT-cwh in #577
• [Bugs] fix bugs in tokenize_ftdp_datasets by @HIT-cwh in #581
• [Feature] Support qwen moe by @HIT-cwh in #579
• [Docs] Add tokenizer to sft in Case 2 by @HIT-cwh in #583
• bump version to 0.1.18 by @HIT-cwh in #590

Full Changelog: v0.1.17...v0.1.18

XTuner Release V0.1.17

What's Changed

• [Fix] Fix PyPI package by @LZHgrla in #540
• [Improve] Add LoRA fine-tuning configs for LLaVA-v1.5 by @LZHgrla in #536
• [Configs] Add sequence_parallel_size and SequenceParallelSampler to configs by @HIT-cwh in #538
• Check shape of attn_mask during attn forward by @HIT-cwh in #543
• bump version to v0.1.17 by @LZHgrla in #542

Full Changelog: v0.1.16...v0.1.17

XTuner Release V0.1.16

What's Changed

• set dev version by @LZHgrla in #487
• Fix type error when the visual encoder is not CLIP by @hhaAndroid in #496
• [Feature] Support Sequence parallel by @HIT-cwh in #456
• [Bug] Fix bugs in flash_attn1_pytorch by @HIT-cwh in #513
• [Fix] delete cat in varlen attn by @HIT-cwh in #508
• bump version to 0.1.16 by @HIT-cwh in #520
• [Improve] Add generation_kwargs for EvaluateChatHook by @LZHgrla in #501
• [Bugs] Fix bugs when training in non-distributed env by @HIT-cwh in #522
• [Fix] Support transformers>=4.38 and require transformers>=4.36.0 by @HIT-cwh in #494
• [Fix] Fix throughput hook by @HIT-cwh in #527
• Update README.md by @JianxinDong in #528
• [Fix] dispatch internlm rote by @HIT-cwh in #530
• Limit transformers != 4.38 by @HIT-cwh in #531

New Contributors

• @hhaAndroid made their first contribution in #496
• @JianxinDong made their first contribution in #528

Full Changelog: v0.1.15...v0.1.16

XTuner Release V0.1.15

What's Changed

• set dev version by @LZHgrla in #437
• [Bugs] Fix bugs when using EpochBasedRunner by @HIT-cwh in #439
• [Feature] Support processing ftdp dataset and custom dataset offline by @HIT-cwh in #410
• Update prompt_template.md by @aJupyter in #441
• [Doc] Split finetune_custom_dataset.md to 6 parts by @HIT-cwh in #445
• [Improve] Add notes for demo_data examples by @LZHgrla in #458
• [Fix] Gemma prompt_template by @LZHgrla in #454
• [Feature] Add LLaVA-InternLM2-1.8B by @LZHgrla in #449
• show more info about datasets by @amulil in #464
• [Fix] write text with encoding='utf-8' by @LZHgrla in #477
• support offline process llava data by @HIT-cwh in #448
• [Fix] msagent_react_map_fn error by @LZHgrla in #470
• [Improve] Reorg xtuner/configs/llava/ configs by @LZHgrla in #483
• limit pytorch version <= 2.1.2 as there may be some bugs in triton2… by @HIT-cwh in #452
• [Fix] fix batch sampler bs by @HIT-cwh in #468
• bump version to v0.1.15 by @LZHgrla in #486

New Contributors

• @aJupyter made their first contribution in #441

Full Changelog: v0.1.14...v0.1.15