31 Pull requests merged by 8 people

• [grpo] fix offpolicy check

  #4852 merged Jul 7, 2025

• [grpo]Fix bug when repeatedly call inputs_to_rolloutrequest

  #4823 merged Jul 7, 2025

• [grpo] deprecated params for 3.6

  #4848 merged Jul 7, 2025

• [megatron] fix eval_iters -1

  #4847 merged Jul 7, 2025

• fix bug: grpo train error for deepseek model

  #4833 merged Jul 7, 2025

• [megatron] Fix the display issue for train_type=lora

  #4845 merged Jul 7, 2025

• update stream & fix bugs

  #4842 merged Jul 7, 2025

• [Feature] SwanLab Lark callback

  #4830 merged Jul 6, 2025

• fix multimodal padding_free prediction_step

  #4839 merged Jul 6, 2025

• [train] fix multimodal packing & padding_free

  #4838 merged Jul 6, 2025

• Support gemma3n

  #4836 merged Jul 4, 2025

• [grpo] fix apply_chat_template

  #4827 merged Jul 4, 2025

• [rollout] fix request from dict

  #4826 merged Jul 4, 2025

• [rollout] Fix non-serializable torch.dtype bug in VLLM weight sync

  #4825 merged Jul 4, 2025

• [rollout] fix external plugins

  #4822 merged Jul 4, 2025

• [GITHUB WORKFLOW]add close stale issues workflow

  #4816 merged Jul 3, 2025

• [RM] support margin & update doc

  #4817 merged Jul 3, 2025

• Support ring attention for llm sft/dpo/grpo (packing/padding_free only).

  #4814 merged Jul 3, 2025

• Refactor Web-UI

  #4687 merged Jul 3, 2025

• [train] Update split_dataset_ratio

  #4798 merged Jul 3, 2025

• [model] support GLM4.1V

  #4804 merged Jul 2, 2025

• fix template bug for qwen3 reranker

  #4795 merged Jul 2, 2025

• update custom_dataset_docs

  #4792 merged Jul 2, 2025

• update resume from checkpoint & update timeout

  #4774 merged Jul 1, 2025

• Fix media downloading from hf

  #4788 merged Jul 1, 2025

• [grpo] check eval_dataset length

  #4781 merged Jul 1, 2025

• [grpo] pass trainer state to reward funcs

  #4779 merged Jul 1, 2025

• [docs] fix grpo docs

  #4777 merged Jul 1, 2025

• [grpo] update vllm weight sync & wake up

  #4770 merged Jul 1, 2025

• update megatron shell

  #4773 merged Jun 30, 2025

• update wechat

  #4769 merged Jun 30, 2025

6 Pull requests opened by 5 people

• [Safety]Fix torch load

  #4802 opened Jul 2, 2025

• [WIP][megatron] support LoRA

  #4812 opened Jul 3, 2025

• Update template_meta.prefix bug

  #4813 opened Jul 3, 2025

• Aacedar patch 3

  #4832 opened Jul 4, 2025

• [grpo] entropy mask

  #4850 opened Jul 7, 2025

• Fix test bug

  #4851 opened Jul 7, 2025

24 Issues closed by 13 people

• GRPO训练结果异常

  #4800 closed Jul 7, 2025

• grpo + gen_rm 流程中的GenRMPlugin是否重复跑了数据

  #4846 closed Jul 7, 2025

• Padding free feature

  #4439 closed Jul 6, 2025

• 支持Gemma-3n模型

  #4759 closed Jul 5, 2025

• Zero loss in case lora qwen3-4b-reranker tuning

  #4820 closed Jul 4, 2025

• grpo can not support deepseek-6.7b-base model

  #4785 closed Jul 4, 2025

• Missing attribute when generate infer_request in VLM GRPO

  #4824 closed Jul 4, 2025

• grpo训练qwen，报通信超时错误

  #4797 closed Jul 4, 2025

• Rollout Stuck after “Core engine process 0 ready.” if use custom plugin

  #4807 closed Jul 4, 2025

• Load lora finetuned model and further finetune with GRPO

  #4821 closed Jul 4, 2025

• 使用PPO训练完毕保存时报错：zero_gather_16bit_weights_on_model_save

  #4815 closed Jul 4, 2025

• reward model的训练能否再详细整理一下文档

  #4379 closed Jul 3, 2025

• 我想使用Gemini蒸馏的特定场景带有思考过程的多轮对话数据，微调qwen3-32B，想问一下，损失计算的时候是不是思考过程只计算最后一轮的<think>\n\n</think>的内容呢？

  #4809 closed Jul 2, 2025

• Qwen3Reranker自定义数据构造bug

  #4784 closed Jul 2, 2025

• How to Input Frame Sequence Without Original Video? 如何直接输入帧图片序列?

  #4776 closed Jul 2, 2025

• megatron sft使用packing时报“cannot pickle _io.TextIOWrapper”错误

  #4778 closed Jul 2, 2025

• vllmengine推理的GuidedDecodingParams参数失效

  #4790 closed Jul 1, 2025

• 为什么我在使用GRPO微调时，我的reward始终为0

  #4789 closed Jul 1, 2025

• grpo字段解析bug

  #4783 closed Jul 1, 2025

• GRPO,Why is it that I am incoming in a video, which is recognized as a picture?

  #4772 closed Jul 1, 2025

• grpo输入格式多处描述不一致

  #4782 closed Jul 1, 2025

• 支持GRPO训练Qwen2-audio-7B-Instruct吗

  #4768 closed Jul 1, 2025

• grpo_trainer.py TypeError: must be real number, not NoneType

  #4751 closed Jun 30, 2025

• 长文本推理报错swift.llm.template.base.MaxLengthError: Current length of row(57972) is larger than the max_length(32768)

  #4767 closed Jun 30, 2025

29 Issues opened by 25 people

• qwen2.5vl是否支持4bit的kv_cache量化？

  #4849 opened Jul 7, 2025

• 使用ms-swift sft之后模型的config.json文件变了，导致我不能直接使用vllm部署模型

  #4844 opened Jul 7, 2025

• ALL_PARALLEL_STYLES argument of type 'NoneType' is not iterable

  #4843 opened Jul 7, 2025

• grpo + gen_rm padding index error

  #4841 opened Jul 7, 2025

• DDP环境下FileNotFoundError问题

  #4840 opened Jul 6, 2025

• Need to update requirements.txt

  #4837 opened Jul 5, 2025

• Trained Qwen 3 model seems to be broken.

  #4835 opened Jul 4, 2025

• Feature Request: RTX 5090 Support with ms-swift docker image with CUDA 12.8

  #4834 opened Jul 4, 2025

• [有人碰到过吗？]qwen2.5vl微调agent出现坐标点偏移问题

  #4831 opened Jul 4, 2025

• SwanLab Notification Integration

  #4829 opened Jul 4, 2025

• awq量化qwen2.5-vl-7b报错

  #4828 opened Jul 4, 2025

• Qwen2-VL-2B 预训练到后期会出现梯度爆炸，其他VLM不会出现

  #4819 opened Jul 3, 2025

• Ovis-2B 预训练报错

  #4818 opened Jul 3, 2025

• error when finetuning qwen3 in modelscope notebook.

  #4811 opened Jul 3, 2025

• 带有思考过程的多轮对话数据，微调qwen3-32B，想问一下，损失计算的时候是不是思考过程只计算最后一轮的<think>\n\n</think>的内容呢？

  #4810 opened Jul 2, 2025

• grpo微调deepseek_coder模型填充信息有误

  #4808 opened Jul 2, 2025

• Reranker training requires A LOT of VRAM

  #4805 opened Jul 2, 2025

• 设置packing_cache后，第二次训练没有从cache读取数据，又重新packing了。

  #4803 opened Jul 2, 2025

• 预训练/微调很慢

  #4801 opened Jul 2, 2025

• sft A3B模型一直卡在这里不动了

  #4799 opened Jul 2, 2025

• Logged accuracy doesn't change when training a reranker

  #4796 opened Jul 2, 2025

• qwen2.5-vl grounding GRPO

  #4794 opened Jul 2, 2025

• llava-next-110b seq_cls 微调出错，AttributeError: 'Identity' object has no attribute 'weight'

  #4793 opened Jul 2, 2025

• 关于SFT微调语料

  #4791 opened Jul 2, 2025

• ValueError: Failed to retrieve the dataset. You can avoid this issue by increasing `max_length` or modifying the `truncation_strategy`.

  #4787 opened Jul 1, 2025

• test kimi vl thinking meet error!

  #4780 opened Jul 1, 2025

• 对DeepSeek-VL2进行GRPO训练

  #4775 opened Jun 30, 2025

• GRPO training: skip_special_tokens config

  #4771 opened Jun 30, 2025

• 支持keye-vl-8b模型

  #4766 opened Jun 30, 2025

34 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• 请教下GRPO训练时出现模型多次异常触碰到Max_length的问题

  #4758 commented on Jun 30, 2025 • 0 new comments

• shape mismatch internvl3

  #4636 commented on Jun 30, 2025 • 0 new comments

• GRPO npu多机多卡显存问题

  #4748 commented on Jul 1, 2025 • 0 new comments

• Fatal Python error: none_dealloc: deallocating None

  #4353 commented on Jul 1, 2025 • 0 new comments

• 关于resume_from_checkpoint加载deepspeed

  #4765 commented on Jul 2, 2025 • 0 new comments

• agent推理时是否还不支持实际的工具调用，参考demo_agent.py

  #4764 commented on Jul 2, 2025 • 0 new comments

• 蒸馏Qwen2.5-Omni模型报错：IndexError: max(): Expected reduction dim 1 to have non-zero size.

  #4755 commented on Jul 2, 2025 • 0 new comments

• wandb，开了海外代理还一直报错（网络连接超时，network error (connectiontimeout)）

  #4152 commented on Jul 2, 2025 • 0 new comments

• GPTQ量化模型GRPO强化微调报错:AttributeError: 'GPTQLoraLinear' object has no attribute 'get_delta_weight'

  #3949 commented on Jul 2, 2025 • 0 new comments

• 请问支持 webdataset 作为 qwen2.5VL 的输入么？

  #3214 commented on Jul 2, 2025 • 0 new comments

• 使用lora的方式单机多卡微调最新的Qwen3_embedding模型会报错

  #4585 commented on Jul 2, 2025 • 0 new comments

• a question for rl

  #4735 commented on Jul 2, 2025 • 0 new comments

• 请问是否支持自定义dataloader，想要实现两种不同格式dataset的训练，每个batch仅有一类数据

  #4750 commented on Jul 3, 2025 • 0 new comments

• 关于序列并行训练

  #2837 commented on Jul 3, 2025 • 0 new comments

• SFT利用上一次加载数据的缓存

  #3762 commented on Jul 4, 2025 • 0 new comments

• 评测时，长度最大只能输出2048，不知道为啥。。。。

  #3761 commented on Jul 4, 2025 • 0 new comments

• 评测时，生成参数改了没效果，评测的配置文件还是显示默认参数

  #3758 commented on Jul 4, 2025 • 0 new comments

• 自定义评测集报错

  #3757 commented on Jul 4, 2025 • 0 new comments

• ValueError: Cannot use chat template functions because tokenizer.chat_template

  #3755 commented on Jul 4, 2025 • 0 new comments

• Support SGLang in Swift

  #3750 commented on Jul 4, 2025 • 0 new comments

• 关于rejected_response的引入

  #3748 commented on Jul 4, 2025 • 0 new comments

• qwen2.5-vl的awq量化问题

  #4762 commented on Jul 4, 2025 • 0 new comments

• 支持GME微调么

  #3019 commented on Jul 4, 2025 • 0 new comments

• MaxLengthError

  #4754 commented on Jul 4, 2025 • 0 new comments

• 使用xtuner作为序列并行的实现方式时没有调用pad_and_split_inputs对输入进行pad和split

  #4760 commented on Jul 4, 2025 • 0 new comments

• 使用qwen32b-vl训练grounding发现连基本的格式遵循都做不到

  #3746 commented on Jul 5, 2025 • 0 new comments

• more logs in wandb

  #3737 commented on Jul 5, 2025 • 0 new comments

• deepseek-r蒸馏模型funcation_calling训练没有效果

  #3733 commented on Jul 5, 2025 • 0 new comments

• async_infer无法实现异步调用的疑问

  #3717 commented on Jul 6, 2025 • 0 new comments

• ModuleNotFoundError: No module named 'torch.distributed.device_mesh'

  #4092 commented on Jul 6, 2025 • 0 new comments

• 开启断点训练后，为什么剩余时间越来越大了，epoch 和max_steps也和断点训练前接不上呢

  #3783 commented on Jul 7, 2025 • 0 new comments

• Use the feature of resume_from_checkpoint when using python code to run finetuning.

  #3774 commented on Jul 7, 2025 • 0 new comments

• Update dataset_info.json

  #3723 commented on Jul 6, 2025 • 0 new comments

• support ernie_vl

  #4763 commented on Jul 1, 2025 • 0 new comments