Repository Issue Activity (beta)

#10610 Flash attention broken in transformers 5.6.0

#10623 Qwen3.5-4B

#10620 triton kernel chunk_scaled_dot_kkt_fwd 越界访问 DDR 内存

#10608 Qwen3-VL 多视频训练时，如果不同视频采样帧数不同，会出现 video tokens 与 features 数量不匹配

#4388 关于npu训练模型总结以及疑问

#5308 Running tokenizer on dataset 一直阻塞，然后subprocesses has abruptly died during map operation

#9704 qwen3vl 多模态video sft: index out of range

#10596 合并lora微调后模型+vllm部署与直接使用llamafactory分别加载基座模型与checkpoint建立的接口推理效果不一致

#10548 Will Gemma4 12b fine-tuning be supported?

#10617 Qwen3-8B-Thinking在sft以后chat会话中，直接输出了<think>

#10566 [Doc] Misleading `_nothink` suffix warning for reasoning-only model families (e.g. Qwen3.6)

#10614 [Feature] Add support for Qwen-AgentWorld-35B-A3B

#10611 训练中eval的一些疑问

#10609 Bug in gradio webui

#10591 Qwen3.5-122B-A10B lora SFT，仅把lora_target从all改为q_proj, v_proj，加载模型出现Cuda out of memory

#10580 训练Qwen3.5-4B时报错了

#4453 8*A800 80G lora训练qwen2-72B模型 内存占用异常

#10604 [Feature] Add support for Tencent Hunyuan Hy-MT2-1.8B/7B

#10557 [Feature] Add support for MiniCPM5-1B

#10364 Gemma4目前只支持thinking吗？

#10593 Unsloth vs. LlamaFactory - incompatible torch

#10386 Liger-Kernel is now supported on LLaMA-Factory + NPU

#10457 llamafactory使用mca训练多模态数据集loss偏高

#10581 How to set a random seed in the weibuUI interface to control randomness

#10560 使用lamafactory微调时模型config文件指定了tie_word_embedding=true，但是训练结束后还是保留了lm.head权重