6 Pull requests merged by 4 people

• [assets] update windows installation

  #8042 merged May 13, 2025

• [model] add seed coder and qwen3 quant models

  #8039 merged May 13, 2025

• [data] fix kimi vl template

  #8015 merged May 11, 2025

• [infer] add video params for vllm infer

  #7992 merged May 9, 2025

• [data] Avoid repetitive tool description warp

  #8000 merged May 9, 2025

• [docs] add GraphGen

  #7974 merged May 7, 2025

3 Pull requests opened by 3 people

• Update llama3_lora_sft.yaml

  #7988 opened May 8, 2025

• feat: add smollm support

  #8050 opened May 13, 2025

• [infer] Modify vllm_infer.py to batch preprocess to avoid too much files opened error

  #8051 opened May 14, 2025

43 Issues closed by 9 people

• 运行llamafactory-cli, log的info太啰嗦, 干扰debug, 如何屏蔽?

  #8052 closed May 14, 2025

• 【报错】[rank2]: max_cache_length = past_key_values.get_max_length() [rank2]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ [rank2]: AttributeError: 'DynamicCache' object has no attribute 'get_max_length'. Did you mean: 'get_seq_length'?

  #8048 closed May 14, 2025

• 是否支持qwen moe 进行lora微调

  #8047 closed May 13, 2025

• 指定的template 不生效？

  #8044 closed May 13, 2025

• 请问是否支持qwen3 moe 进行grpo训练

  #8040 closed May 13, 2025

• SFT Loss Exhibits Step Function Behavior evey Epoch, Not Monotonic Decrease

  #8037 closed May 12, 2025

• Qwen3-14B训练过程中出现torch.distributed.elastic.multiprocessing.errors.ChildFailedError:

  #7967 closed May 12, 2025

• 使用vllm推理Kimi_VL时出现RuntimeError: Engine core initialization failed. See root cause above

  #8025 closed May 12, 2025

• 训练gemma3报错

  #7370 closed May 12, 2025

• 进行模型量化导出没有后续输出

  #8030 closed May 12, 2025

• 做文本分类训练，训练后模型返回的结果会有不是训练的内容，有什么办法控制吗？

  #8032 closed May 12, 2025

• lora微调Kimk_vl合并后，vllm推理出现RuntimeError: Engine core initialization failed. See root cause above

  #8026 closed May 12, 2025

• 每次保存模型global保存的内存占的太多，有什么办法可以减少保存模型的内存吗，没用的就不保存

  #8027 closed May 12, 2025

• qwen3原始模型或lora合并参数后的模型结果输出问题

  #8024 closed May 12, 2025

• 使用vllm推理Kimi_vl时提示ValueError: `limit_mm_per_prompt` is only supported for multimodal models.

  #8022 closed May 12, 2025

• I am facing issue while using custom system msg

  #8017 closed May 11, 2025

• 训练成功后chat加载聊天报错

  #8016 closed May 11, 2025

• docker部署跑测试有问题

  #8010 closed May 11, 2025

• Kimi-VL模型推理时UnboundLocalError: cannot access local variable 'mm_inputs' where it is not associated with a value

  #8008 closed May 11, 2025

• Qwen3 not stopping generation after lora finetuning

  #7943 closed May 11, 2025

• ValueError: Processor was not found, please check and update your model file.

  #8014 closed May 11, 2025

• qwen38B模型部署

  #8011 closed May 11, 2025

• ValueError: Processor was not found, please check and update your model file.

  #8009 closed May 11, 2025

• qwen2.5VL-7B lora微调问题咨询

  #7991 closed May 10, 2025

• How to resize_token_embeddings of Qwen2.5-Omni-7B?

  #8005 closed May 10, 2025

• llama框架所需的transformer版本与qwen所需要的transformer版本冲突问题

  #8006 closed May 10, 2025

• 如何改代码使训练的时候标签(labels)部分不分词, 问题部分分词？

  #7980 closed May 9, 2025

• flash_attn: fa2 并没有改变模型的attn_implementation参数。

  #7979 closed May 9, 2025

• mllm_demo.json用的格式是shareGPT还是alpaca?

  #7982 closed May 9, 2025

• Gemma 3 FSDP PT [ValueError: Must flatten tensors with uniform dtype but got torch.bfloat16 and torch.float32]

  #7993 closed May 9, 2025

• 关于添加 add_tokens 导致Lora参数和模型参数不匹配

  #7997 closed May 9, 2025

• lora微调时使用bf16，但是合并lora之后输出fp32，导致flashattn不可用

  #7998 closed May 9, 2025

• 安装LLaMA Factory依赖的依赖hf-xet时出错

  #7999 closed May 9, 2025

• TypeError: MMPluginMixin._get_video_sample_indices() missing 2 required positional arguments: 'video_fps' and 'video_maxlen'

  #7990 closed May 9, 2025

• 单卡4080 或者双卡4060ti 使用docker 微调损失函数无法收敛

  #8002 closed May 9, 2025

• chat 功能用hugginface框架不能加载模型

  #8001 closed May 9, 2025

• 使用deepseed训练ppo阶段时候出现报错

  #7989 closed May 9, 2025

• 请问qwen3的微调支持图像输入吗？

  #7994 closed May 9, 2025

• 微调Kimi时出现OSError: Failed to load tokenizer.

  #7984 closed May 8, 2025

• Qwen2.5-Omni-3B使用BF16训练报错

  #7970 closed May 8, 2025

• Qwen3-4B lora微调合并后用vllm推理仍然输出thinking过程

  #7978 closed May 7, 2025

• qwen3训练数据格式问题

  #7977 closed May 7, 2025

• 直接下载main的代码，不支持Qwen3

  #7972 closed May 7, 2025

23 Issues opened by 22 people

• Support for SmolVLM

  #8049 opened May 13, 2025

• size mismatch for base_model.model.lm_head.weight: copying a param with shape torch.Size([128259, 4096]) from checkpoint, the shape in current model is torch.Size([128264, 4096]).

  #8046 opened May 13, 2025

• NCCL INFO socketPollConnect: connect returned No route to host

  #8045 opened May 13, 2025

• DataLoader exited unexpectedly after one epoch

  #8043 opened May 13, 2025

• [question] Dose Muon Optimizer support deepspeed ?

  #8041 opened May 13, 2025

• qwen3:30b-a3b SFT with lora-rank as 16 is very very slow

  #8038 opened May 13, 2025

• Single-machine dual-GPU fine-tuning environment based on WSL2 Ubuntu,Although the GPU memory is sufficient, the system reports insufficient GPU memory.

  #8036 opened May 12, 2025

• invalid conversion from ‘int’ to ‘CUresult’ {aka ‘cudaError_enum’}

  #8035 opened May 12, 2025

• 训练模型出错

  #8031 opened May 12, 2025

• Discrepancy in ROUGE BLEU Scores During Training vs. Final Results

  #8029 opened May 12, 2025

• 模型保存问题

  #8028 opened May 12, 2025

• unsloth train error

  #8021 opened May 12, 2025

• [PPO] failed with reward model in PPO training for llama3-8B, huggingface_hub.errors.HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name': 'saves/llama3-8b/lora/reward'. Use `repo_type` argument if needed.

  #8020 opened May 12, 2025

• 4080 单卡微调 使用bf16 第二轮之后会的损失函数变成0

  #8019 opened May 12, 2025

• 请问有计划支持4月底最新发布的kimi-audio吗

  #8013 opened May 11, 2025

• The performance has deteriorated significantly after fine-tuning with TRL but increase while using llama-factory

  #8012 opened May 11, 2025

• [eval]在模型评测时支持自定义系统提示词

  #8007 opened May 10, 2025

• 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 opened May 9, 2025

• [FSDP LORA] No device index is set for xpu when running FSDP+LORA

  #8003 opened May 9, 2025

• transformers高于 4.49 的版本对 Ascend NPU 不兼容

  #7986 opened May 8, 2025

• cutoff_len 语义混淆

  #7985 opened May 8, 2025

• sft train code pending

  #7983 opened May 8, 2025

• How to save the tokenized data?

  #7981 opened May 7, 2025

27 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.

• Distributed improvement of Muon implementation

  #7808 commented on May 13, 2025 • 1 new comment

• support repetition_penalty and align presence_penalty with OpenAI Client

  #7958 commented on May 7, 2025 • 0 new comments

• add Sequence Parallelism (reopened #6506)

  #7338 commented on May 8, 2025 • 0 new comments

• This change adds support for Intel Gaudi HPUs.

  #7275 commented on May 13, 2025 • 0 new comments

• 使用vllm_infer一次性推理多个样本会报错

  #7513 commented on May 13, 2025 • 0 new comments

• 🚨FAQs | 常见问题🚨

  #4614 commented on May 13, 2025 • 0 new comments

• 使用huggingface加载Qwen2.5-32B-Instruct 回答乱码

  #7371 commented on May 13, 2025 • 0 new comments

• streaming 训练卡在第一个step

  #7261 commented on May 13, 2025 • 0 new comments

• 有计划支持LoRAMoE吗？

  #2749 commented on May 13, 2025 • 0 new comments

• 310P 微调报错 RuntimeError: call aclnnCast failed, detail:EZ9999: Inner Error

  #5425 commented on May 13, 2025 • 0 new comments

• Can't Finetune Qwen2.5-VL-3B-Instruct-AWQ Using LoRA and Multiple GPUs

  #7963 commented on May 12, 2025 • 0 new comments

• 单机单卡4090，lora微调deepseek-r1-distill-1.5B卡死，稳定复现，会导致整个机器卡死，nvdia-smi命令失效，

  #7661 commented on May 12, 2025 • 0 new comments

• Support for images_root field for multimodal training data

  #7814 commented on May 12, 2025 • 0 new comments

• deepseek-moe-16B预训练问题

  #7165 commented on May 12, 2025 • 0 new comments

• Why Speed per iteration slower when dataset is large

  #6410 commented on May 12, 2025 • 0 new comments

• 多模态数据集多的时候，数据加载失败

  #7266 commented on May 12, 2025 • 0 new comments

• 使用neat_packing进行sft训练，模型性能指标下降明显

  #5426 commented on May 12, 2025 • 0 new comments

• ChatGLM3-6b在阿里云上使用lora微调模型后，训练完成后，点击chat加载模型出现以西报错

  #5995 commented on May 12, 2025 • 0 new comments

• RuntimeError: CUDA driver error: invalid argument

  #7832 commented on May 11, 2025 • 0 new comments

• simpo算法实现肯定存在bug，可以稳定复现

  #7954 commented on May 10, 2025 • 0 new comments

• LLava Series (7B, 14B) freeze_vision_tower=false bug

  #6376 commented on May 10, 2025 • 0 new comments

• Tuned Model's processor can't tokenize my text input properly.

  #7608 commented on May 10, 2025 • 0 new comments

• How to implement weight decay towards the pre-trained model?

  #5603 commented on May 9, 2025 • 0 new comments

• 请问有使用bitsandbytes量化方法的环境配置分享吗

  #7861 commented on May 9, 2025 • 0 new comments

• Is it possible to support microsoft/bitnet-b1.58-2B-4T ?

  #7775 commented on May 8, 2025 • 0 new comments

• Size mismatch error in the middle of training

  #6699 commented on May 8, 2025 • 0 new comments

• 昇腾多机训练 怎么设置

  #5360 commented on May 8, 2025 • 0 new comments