5 Pull requests merged by 3 people

• [model] add qwen3 2507 models

  #8750 merged Jul 25, 2025

• [model] add glm4moe

  #8689 merged Jul 25, 2025

• [assets] update readme

  #8739 merged Jul 24, 2025

• [assets] update wechat

  #8731 merged Jul 23, 2025

• [model] support granite4

  #8680 merged Jul 21, 2025

2 Pull requests opened by 2 people

• [assets] update usage of apptainer in readme

  #8722 opened Jul 23, 2025

• Add Channel loss

  #8736 opened Jul 24, 2025

34 Issues closed by 8 people

• 导出模型 qwen2.5vl 3b导出bug

  #8751 closed Jul 25, 2025

• 支持昆仑芯P800进行训练么

  #8742 closed Jul 25, 2025

• [Help]精度验证不一致

  #8744 closed Jul 25, 2025

• 导出模型时报ChatGLM model is not supported yet.

  #8741 closed Jul 25, 2025

• Gemma3n finetuning with audio inputs does not work

  #8693 closed Jul 25, 2025

• 文档错误，按照文档构建数据集运行被完全忽略（from和value）

  #8738 closed Jul 25, 2025

• 请问各位佬，LLaMA-Factory的PPO训练器目前支不支持从checkpoint恢复训练

  #8715 closed Jul 24, 2025

• 关于量化export_quantization_dataset的问题

  #8735 closed Jul 24, 2025

• 训练qwen2.5vl7b bug ValueError: The number of images does not match the number of <image> tokens in

  #8734 closed Jul 24, 2025

• 微调glm4报错

  #8732 closed Jul 24, 2025

• UI-TARS-7B-SFT经过训练后推理极慢

  #8728 closed Jul 23, 2025

• 训练Qwen2.5VL的坐标漂移问题

  #8727 closed Jul 23, 2025

• 内存需求参考

  #8726 closed Jul 23, 2025

• `accelerate` 启动多机多卡没有生效，每个节点各训各的。

  #8729 closed Jul 23, 2025

• 在帮助文档中，[--pref_loss {sigmoid,hinge,ipo,kto_pair,orpo,simpo}]，但是，报错ValueError: Unknown loss type: kto_pair. Should be one of ['sigmoid', 'hinge', 'ipo', 'bco_pair', 'sppo_hard', 'nca_pair', 'robust', 'exo_pair']，还没更新吗？

  #8725 closed Jul 23, 2025

• 使用llamafactory-cli api部署模型时，如何设置api-key

  #8723 closed Jul 23, 2025

• GRPO training support

  #8717 closed Jul 23, 2025

• 使用vllm推理和在llama-factory chat上输出不一致

  #8720 closed Jul 23, 2025

• RuntimeError: CUDA driver error: invalid argument

  #8703 closed Jul 23, 2025

• qwen2.5-7b-vl使用0.9.3微调并合并后，使用vllm0.8.4拉起合并后的模型，无微调效果

  #8716 closed Jul 22, 2025

• Processor was not found [mistralai--Mistral-Small-3.2-24B-Instruct-2506] for finetune lora

  #8709 closed Jul 22, 2025

• GRADIO_SERVER_PORT 设置后 webui.py源码用修改吗

  #8711 closed Jul 22, 2025

• Failed to run LlamaFactory

  #8710 closed Jul 22, 2025

• 如何实现Qwen 2.5-VL的Full-parameter on ViT, LoRA on LLM?

  #8704 closed Jul 22, 2025

• 单机多卡微调训练异常：deepspeed 模式 ，且最新tags 0.9.3版本：Unknown command: --local_rank=0/1

  #8706 closed Jul 22, 2025

• webui默认暴露端口是7860，如何改为别的端口

  #8707 closed Jul 22, 2025

• 未支持模型的LoRA微调

  #8700 closed Jul 21, 2025

• Correct format for DPO Qwen2.5-VL training example

  #8692 closed Jul 21, 2025

• 都是lora训练。dpo 的create_new_adapter为false。dpo的lora参数需要和sft一致吗，比如lora_rank、lora_target这些

  #8696 closed Jul 21, 2025

• 感谢你的工作，请问下如何在您的代码基础上，添加其他PEFT方法的微调，如 VERA FourierFT HRA 等。

  #8697 closed Jul 21, 2025

• tokenizer越来越慢

  #8691 closed Jul 20, 2025

• SFT了奖励模型后，用PPO训练模型报错没有value_head.bin文件

  #8688 closed Jul 20, 2025

• llama-factory利用Ray进行sft3机2卡运行报错

  #8641 closed Jul 20, 2025

• ValueError: Template intern_vl does not exist.

  #8687 closed Jul 20, 2025

13 Issues opened by 13 people

• 如题：训练过程中如何实现train和eval的任务不一样，比如我训练用的是qa数据，阶段性eval用的是其他任务，

  #8754 opened Jul 26, 2025

• 在使用多图像数据微调kimi-vl时训练卡死

  #8748 opened Jul 25, 2025

• 修改compute_loss实现自定义权重，但是获取的logits为none

  #8746 opened Jul 25, 2025

• 用2卡910B微调完成Qwen3-32B后chat报错

  #8745 opened Jul 25, 2025

• validation (webui) should support vllm and disabling thinking mode

  #8743 opened Jul 25, 2025

• Gemma3n SFT trained model failing to load

  #8740 opened Jul 24, 2025

• 微调qwen-vl的时在Running tokenizer on dataset出现image file is truncated (15 bytes not processed)错误

  #8733 opened Jul 24, 2025

• pointwise reward model

  #8713 opened Jul 22, 2025

• webui显示错误

  #8712 opened Jul 22, 2025

• ppo训练时出现了RuntimeError("shape '[0, 4, -1]' is invalid for input of size 1280") 报错

  #8708 opened Jul 22, 2025

• Qwen3-1.7B lora微调报错：Expected a string, got None，停在特定百分比或是直接0%

  #8705 opened Jul 22, 2025

• 训练成功后无法进行问答的问题

  #8702 opened Jul 21, 2025

• v100微调qwen-2.5vl报错

  #8699 opened Jul 21, 2025

19 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.

• NCCL INFO socketPollConnect: connect returned No route to host

  #8045 commented on Jul 19, 2025 • 0 new comments

• NCCL broken pipe

  #6849 commented on Jul 19, 2025 • 0 new comments

• 多模态qwen-2.5vl训练卡死

  #8665 commented on Jul 19, 2025 • 0 new comments

• 8卡h200，deepspeed-zero2，qwen32b全量微调出现oom现象

  #8642 commented on Jul 22, 2025 • 0 new comments

• api形式的reward_model出入参文档

  #8371 commented on Jul 22, 2025 • 0 new comments

• 请问DPO训练的时候有什么注意事项吗？我训练出来效果很差。

  #5484 commented on Jul 22, 2025 • 0 new comments

• Template should not be truncated

  #1575 commented on Jul 22, 2025 • 0 new comments

• Why does the loss suddenly drop to zero and the grad_norm stabilize at 1 at specific training steps when fine-tuning Llama3.2-vision-11B?

  #8004 commented on Jul 22, 2025 • 0 new comments

• StopIteration Issue in streaming mode

  #8560 commented on Jul 22, 2025 • 0 new comments

• Feature Request: Separate Learning Rates for Vision Encoder and Language Backbone in VLM Tuning

  #5831 commented on Jul 23, 2025 • 0 new comments

• 在Qwen3上进行新语言的CPT和sft后, 模型无法正常停止，会一直重复回答。同样的数据与参数在qwen2.5则正常

  #8365 commented on Jul 23, 2025 • 0 new comments

• 多机多卡训练时卡住

  #7300 commented on Jul 23, 2025 • 0 new comments

• RuntimeError: CUDA error: operation not supported

  #8085 commented on Jul 23, 2025 • 0 new comments

• Qwen-Omni在混合模态数据上dpo训练时，训练卡住

  #8151 commented on Jul 24, 2025 • 0 new comments

• OOM and slow tokenization after upgrade LLaMA-Factory

  #8600 commented on Jul 24, 2025 • 0 new comments

• 在GPU 5090上使用docker安装llama factory 出现问题

  #8290 commented on Jul 24, 2025 • 0 new comments

• h20 SFT Qwen2.5_VL_7B Loss为0，grad固定值不变

  #8563 commented on Jul 25, 2025 • 0 new comments

• Watchdog caught collective operation timeout: WorkNCCL(SeqNum=1580, OpType=ALLREDUCE, NumelIn=466119168, NumelOut=466119168, Timeout(ms)=600000) ran for 600004 milliseconds before timing out.

  #7396 commented on Jul 25, 2025 • 0 new comments

• fix some features

  #8556 commented on Jul 22, 2025 • 0 new comments