1 Release published by 1 person

• v3.2.2

  published Mar 26, 2025

26 Pull requests merged by 3 people

• fix grpo train dataloader

  #3736 merged Apr 1, 2025

• fix qwen2_5 omni

  #3734 merged Apr 1, 2025

• support qwen2_5_vl packing

  #3694 merged Mar 31, 2025

• Fix grpo dora

  #3709 merged Mar 31, 2025

• fix qwen2_5-omni

  #3716 merged Mar 28, 2025

• fix adalora

  #3714 merged Mar 28, 2025

• fix grpo template copy

  #3708 merged Mar 28, 2025

• update warning_once

  #3706 merged Mar 28, 2025

• fix grpo vl

  #3704 merged Mar 28, 2025

• fix grpo qwen2_5_omni

  #3701 merged Mar 28, 2025

• [megatron] fix val_dataset streaming

  #3699 merged Mar 27, 2025

• fix import error

  #3700 merged Mar 27, 2025

• Grpo vl72b script

  #3692 merged Mar 27, 2025

• fix grpo rank

  #3687 merged Mar 27, 2025

• support Qwen/Qwen2.5-Omni-7B (sft/dpo/grpo)

  #3613 merged Mar 26, 2025

• fix shell

  #3675 merged Mar 26, 2025

• fix npu context

  #3664 merged Mar 25, 2025

• update readme

  #3663 merged Mar 25, 2025

• Fix evaluation of embedding

  #3661 merged Mar 25, 2025

• compat vllm0.8.1

  #3656 merged Mar 25, 2025

• fix grpo vllm tp

  #3658 merged Mar 25, 2025

• fix label_names

  #3657 merged Mar 25, 2025

• set grpo multi turn max tokens

  #3655 merged Mar 25, 2025

• update docs

  #3653 merged Mar 25, 2025

• fix grpo epsilon

  #3652 merged Mar 25, 2025

• Fix template torch_dtype

  #3651 merged Mar 25, 2025

6 Pull requests opened by 4 people

• [megatron] support max_epochs

  #3677 opened Mar 26, 2025

• 优化_replace_image_tags防止数据中不包含具体图像的<img></img> tag中断训练

  #3683 opened Mar 26, 2025

• Update readme to vllm082

  #3707 opened Mar 28, 2025

• Update dataset_info.json

  #3723 opened Mar 31, 2025

• [WIP] dapo

  #3725 opened Mar 31, 2025

• support internvl2 packing

  #3735 opened Apr 1, 2025

17 Issues closed by 11 people

• Qwen2.5_omni 部署bug

  #3727 closed Apr 1, 2025

• 请问有支持vllm v1引擎的计划吗

  #3720 closed Mar 31, 2025

• GRPO微调，gpu利用率很低

  #3693 closed Mar 28, 2025

• suspicious GRPO training temperature setting error

  #3696 closed Mar 27, 2025

• how to set wanb project and run

  #3691 closed Mar 27, 2025

• Ascend NPU下，GRPO训练时死循环

  #3650 closed Mar 27, 2025

• grpo 多机训练qwen2.5-72b，开tensor并行会timeout

  #3401 closed Mar 27, 2025

• 请问Qwen2.5-VL-7B-Instruct 跑GRPO需要多少显存？

  #3426 closed Mar 27, 2025

• How to switch on the multi-GPU for GRPO Training?

  #3473 closed Mar 27, 2025

• Qwen25VL72B GRPO Lora Bug

  #3483 closed Mar 27, 2025

• train_72b_4gpu.sh how to change to other number of GPUs?

  #3507 closed Mar 27, 2025

• 仿照train_72b_4gpu.sh开启sleep_leval进行GRPO训练，KL一直为0

  #3558 closed Mar 27, 2025

• 自定义损失函数

  #3495 closed Mar 26, 2025

• 多轮对话时，swift可以让history中assistant部分的内容不计算loss吗？有参数可以控制这个吗？

  #3679 closed Mar 26, 2025

• 为什么调用reward函数时传入的样本数量不固定？

  #3676 closed Mar 26, 2025

• qwen2.5vl npu infer

  #3670 closed Mar 26, 2025

• GRPO多节点训练，训练比较慢，脚本还能优化吗

  #3186 closed Mar 26, 2025

48 Issues opened by 41 people

• more logs in wandb

  #3737 opened Apr 1, 2025

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

  #3733 opened Apr 1, 2025

• grounding数据集格式，多类别+多box怎么写

  #3732 opened Apr 1, 2025

• Support Ulysses in Swift

  #3731 opened Apr 1, 2025

• TypeError: embedding(): argument 'indices' (position 2) must be Tensor, not NoneType

  #3730 opened Mar 31, 2025

• swift训练时 Qwenvl2.0/2.5 是否采用了smart_resize

  #3729 opened Mar 31, 2025

• GRPO 训练，数据格式解析有bug

  #3728 opened Mar 31, 2025

• npu环境GRPO训练，使用vllm时，官方脚本无法正常启动，其他脚本则可以

  #3726 opened Mar 31, 2025

• Qwen2.5-Omni-7B 部署api推理报错

  #3724 opened Mar 31, 2025

• Qwen2.5-VL-32B-Instruct-AWQ无法进行推理

  #3722 opened Mar 31, 2025

• max_pixels到底是怎么发挥作用呢？

  #3721 opened Mar 30, 2025

• It is recommended to use a dedicated device for vLLM

  #3719 opened Mar 29, 2025

• SimPO and ORPO support for VLM (Qwen2.5VL)

  #3718 opened Mar 29, 2025

• async_infer无法实现异步调用的疑问

  #3717 opened Mar 29, 2025

• qwen2.5 omni TypeError: 'NoneType' object is not iterable

  #3715 opened Mar 28, 2025

• GRPO max_grad_norm seems don't work

  #3713 opened Mar 28, 2025

• grpo中的async模式是否能够支持tensor_parallel_size>1

  #3712 opened Mar 28, 2025

• loss_scale 疑问

  #3711 opened Mar 28, 2025

• qwen2.5-7b-Instruct进行lora微调合并后推理报错

  #3710 opened Mar 28, 2025

• 微调Qwen2.5VL，ref输入只有一个字符

  #3705 opened Mar 28, 2025

• ovis 一定要flash_attn才能训练吗？

  #3703 opened Mar 28, 2025

• Hanging after tqdm starts [COLOCATE MODE]

  #3702 opened Mar 28, 2025

• No training with AWQ QLORA (Qwen 2.5 VL)

  #3698 opened Mar 27, 2025

• Deepspeed Zero++ 会出现Nan

  #3697 opened Mar 27, 2025

• multi-node grpo training hangs

  #3695 opened Mar 27, 2025

• 支持Qwen/Qwen2.5-Omni-7B的talker微调，用于微调音色、方言等

  #3690 opened Mar 27, 2025

• Cache Inference Optimization

  #3689 opened Mar 27, 2025

• cannot import name 'UnencryptedCookieSessionFactoryConfig' from 'pyramid.session' (unknown location)

  #3688 opened Mar 27, 2025

• 支持对多个 source 数据集进行 loss 输出，方便查看不同数据集的 loss

  #3686 opened Mar 27, 2025

• 关于多机分布式训练数据读取不到的问题

  #3685 opened Mar 27, 2025

• pack_to_max_length is only available with distributed training.

  #3684 opened Mar 27, 2025

• unsloth error when sft qwen2.5-vl-7b-instruct

  #3682 opened Mar 26, 2025

• deepspeed错误

  #3681 opened Mar 26, 2025

• 如何设置lm_head为可训练?全量微调

  #3680 opened Mar 26, 2025

• 微调以后效果不好怎么办？

  #3678 opened Mar 26, 2025

• 批次问题

  #3674 opened Mar 26, 2025

• ValueError: RLHF do not support sequence parallel

  #3673 opened Mar 26, 2025

• xgrammar support

  #3672 opened Mar 26, 2025

• 有没有4*V100能跑起来GRPO的训练脚本和环境配置呀？

  #3671 opened Mar 26, 2025

• KTO 训练每次保持ckt 都报错

  #3669 opened Mar 26, 2025

• 奇怪的问题：/opt/conda/envs/swift/bin/python: can't open file '/data/swift'

  #3668 opened Mar 26, 2025

• 【bug】Failed to open local file in cache

  #3667 opened Mar 25, 2025

• 使用GRPO训练llava-1.5以及qwen2-vl时，使用vllm推理，在eval时报错

  #3666 opened Mar 25, 2025

• cannot import name 'LoRA' from 'swift'

  #3665 opened Mar 25, 2025

• [Bug]: RuntimeError: setup failed!

  #3662 opened Mar 25, 2025

• gemma3使用grpo用vllm的bug

  #3660 opened Mar 25, 2025

• 微调qwen2.5vl grouding

  #3659 opened Mar 25, 2025

• weighted loss for different class 为不同数据的loss分配不同的权重

  #3654 opened Mar 25, 2025

21 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.

• 建议增加稳定版本的安装文档

  #3644 commented on Mar 25, 2025 • 0 new comments

• unsloth不支持多模态模型

  #3546 commented on Mar 25, 2025 • 0 new comments

• How to deploy a trained reward model?

  #2437 commented on Mar 25, 2025 • 0 new comments

• 是否可以支持使用r1的合成数据，蒸馏小模型的lora微调方法

  #3200 commented on Mar 26, 2025 • 0 new comments

• 使用GRPO 使用我已经训练的LLava模型加载问题

  #3195 commented on Mar 26, 2025 • 0 new comments

• SWIFT支持embedding模型的推理和部署如何实现

  #3606 commented on Mar 26, 2025 • 0 new comments

• RTX3090上运行sft-rlhf-grpo微调，报错：torch.distributed.DistBackendError: [3] is setting up NCCL communicator and retrieving ncclUniqueId from [0] via c10d key-value store by key '0', but store->get('0') got error: wait timeout after 1800000ms,

  #3612 commented on Mar 26, 2025 • 0 new comments

• 关于GRPO训练采样重复的问题

  #3450 commented on Mar 27, 2025 • 0 new comments

• 训练保存checkpoint的时候报错，但本地又有相应的文件。

  #3420 commented on Mar 27, 2025 • 0 new comments

• GRPO算法训练，后期训练时，显存暴增

  #3600 commented on Mar 27, 2025 • 0 new comments

• lora 微调 ovis2-34B loss=0.0 grad_norm=nan

  #3494 commented on Mar 27, 2025 • 0 new comments

• 4*v100环境执行lora_vllm脚本报错：Assertion `!(srcMmaLayout && dstMmaLayout && !srcMmaLayout.isAmpere()) && "mma -> mma layout conversion is only supported on Ampere"' failed.

  #3549 commented on Mar 27, 2025 • 0 new comments

• Qwen2.5-vl-72b 使用 vllm V1 引擎时，无法正常推理，提示`'AsyncLLM' object has no attribute 'engine'`

  #3139 commented on Mar 27, 2025 • 0 new comments

• Qwen25VL 72B GRPO training (lora) would hang for no reason.

  #3592 commented on Mar 28, 2025 • 0 new comments

• grpo 多机多卡训练timeout

  #3343 commented on Mar 28, 2025 • 0 new comments

• torch stack error for molmo

  #3389 commented on Mar 30, 2025 • 0 new comments

• 微调Qwen2_5_VL模型时报错：AssertionError: Input and cos/sin must have the same dtype, got torch.float32 and torch.bfloat16

  #3156 commented on Mar 31, 2025 • 0 new comments

• 请求支持健康检查

  #3474 commented on Mar 31, 2025 • 0 new comments

• 支持GME微调么

  #3019 commented on Mar 31, 2025 • 0 new comments

• qwen2.5-vl grouding task with json output。

  #3511 commented on Apr 1, 2025 • 0 new comments

• Megatron-SWIFT训练交流群

  #3604 commented on Apr 1, 2025 • 0 new comments