4 Releases published by 1 person

• v3.5.0

  published Jun 8, 2025

• v3.5.1 Patch release v3.5.1

  published Jun 13, 2025

• v3.5.2 Patch release v3.5.2

  published Jun 20, 2025

• v3.5.3 Patch release v3.5.3

  published Jun 27, 2025

121 Pull requests merged by 11 people

• [grpo]Tool rl: add reward func for ToolRL

  #4694 merged Jun 27, 2025

• compat transformers==4.52 (vlm)

  #4738 merged Jun 26, 2025

• [grpo] check liger & sp

  #4734 merged Jun 26, 2025

• [grpo] fix max_step for dataloader when applying sequence parallel

  #4731 merged Jun 26, 2025

• [quant] Support fp8

  #4729 merged Jun 26, 2025

• support Kimi-VL-A3B-Thinking-2506 & Kimi-Dev-72B

  #4719 merged Jun 25, 2025

• [doc] simplify environment variables & update best practices documentation

  #4715 merged Jun 25, 2025

• [grpo] fix colocate seed

  #4712 merged Jun 25, 2025

• [megatron] support rednote-hilab/dots.llm1.inst

  #4707 merged Jun 25, 2025

• [megatron] support DeepseekV2ForCausalLM and DeepseekV3ForCausalLM

  #4659 merged Jun 25, 2025

• fix links

  #4690 merged Jun 24, 2025

• [feat] support fine-tuning of reranker models

  #4671 merged Jun 24, 2025

• [grpo] fix grpo pt

  #4683 merged Jun 24, 2025

• [rollout] fix dp args

  #4678 merged Jun 23, 2025

• [doc] fix doc

  #4675 merged Jun 23, 2025

• [doc] fix image link

  #4674 merged Jun 23, 2025

• docs: correct typo "resonse" to "response"

  #4672 merged Jun 23, 2025

• [channel loss]support packing & padding free

  #4666 merged Jun 23, 2025

• [docs] update docs

  #4665 merged Jun 23, 2025

• [dataset] fix grounding_dataset

  #4664 merged Jun 23, 2025

• [grpo] refactor multi turn & support async engine & refactor grpo docs

  #4380 merged Jun 23, 2025

• [template] optimize remove_unused_columns

  #4661 merged Jun 22, 2025

• [gkd] support use_logits_to_keep/padding_free/packing & update gkd shell

  #4658 merged Jun 21, 2025

• [docs] update gkd

  #4657 merged Jun 20, 2025

• compat megatron-core 0.11

  #4655 merged Jun 20, 2025

• [megatron] fix eval data_collator

  #4654 merged Jun 20, 2025

• fix device_map & ddp rank0

  #4650 merged Jun 20, 2025

• fix packing & load_from_cache_file

  #4649 merged Jun 20, 2025

• [model] fix model_meta

  #4647 merged Jun 20, 2025

• [template] optimize get_length

  #4641 merged Jun 20, 2025

• [docs] update qwen3 best_practice

  #4300 merged Jun 19, 2025

• update docs readme

  #4639 merged Jun 19, 2025

• update docs & shell

  #4637 merged Jun 19, 2025

• [infer/deploy/eval/app] support sglang engine

  #3810 merged Jun 19, 2025

• [doc] LaTeX rendering

  #4629 merged Jun 18, 2025

• [rollout] swift rollout add template

  #4626 merged Jun 17, 2025

• [loss_scale] support last_round_with_ignore_empty_think for rag

  #4623 merged Jun 17, 2025

• fix max_epochs tp

  #4624 merged Jun 17, 2025

• [ppo] fix ppo

  #4622 merged Jun 17, 2025

• [docs] remove Qwen3-32B-Base

  #4621 merged Jun 17, 2025

• [gkd] support gkd_trainer

  #4587 merged Jun 17, 2025

• Fix minimax & fix agent_template

  #4618 merged Jun 17, 2025

• [megatron] fix megatron pp max_epochs

  #4608 merged Jun 16, 2025

• Update FAQ

  #4612 merged Jun 16, 2025

• [model] support minimax

  #4610 merged Jun 16, 2025

• [megatron] compat megatron-core main branch

  #4606 merged Jun 15, 2025

• [mirror] update swift mirror

  #4601 merged Jun 14, 2025

• Fix UI llm_train

  #4592 merged Jun 13, 2025

• fix gc_kwargs

  #4591 merged Jun 13, 2025

• [grpo] restore num_generations check

  #4590 merged Jun 13, 2025

• [megatron] support more rope_scaling & support deepseek-r1-qwen3-8b/internlm3/mimo-7b

  #4576 merged Jun 12, 2025

• [grpo] model weight synchronization before first turn rollout with async generation

  #4584 merged Jun 12, 2025

• [grpo] remove data collator to top-level to avoid pickle error in spawn mode

  #4582 merged Jun 12, 2025

• [megatron] Fix megatron all_reduce warning

  #4568 merged Jun 12, 2025

• [model] fix ovis gradient_checkpointing vit no_grad

  #4571 merged Jun 12, 2025

• fix args.json

  #4566 merged Jun 11, 2025

• [Bug]Fix ulysses train steps, embedding negative sample length

  #4565 merged Jun 11, 2025

• [dataset] fix toolbench (local)

  #4563 merged Jun 11, 2025

• [grpo] fix the pickle data collator

  #4562 merged Jun 11, 2025

• [grpo] support offloading reference model

  #4554 merged Jun 11, 2025

• support dots1

  #4560 merged Jun 11, 2025

• [megatron] support DPO

  #4193 merged Jun 11, 2025

• [megatron/dpo] fix megatron packing_cache & update DPOTrainer

  #4556 merged Jun 11, 2025

• fix qwen3 embedding saving

  #4548 merged Jun 10, 2025

• fix: handle INFONCE_HARD_NEGATIVES as integer if provided

  #4545 merged Jun 10, 2025

• support cognitivecomputations/DeepSeek-R1-0528-AWQ

  #4537 merged Jun 9, 2025

• fix LoraModel

  #4536 merged Jun 9, 2025

• [grpo] update doc about move_model_batches

  #4523 merged Jun 8, 2025

• fix emb script and docs

  #4521 merged Jun 8, 2025

• [qwen2.5-omni] Fix omni get_template

  #4518 merged Jun 7, 2025

• [qwen2.5-omni] Fix omni save checkpoint

  #4517 merged Jun 7, 2025

• [megatron] fix pp4

  #4516 merged Jun 7, 2025

• [dataset] fix dpo emoji dataset

  #4514 merged Jun 7, 2025

• Support minicpm4

  #4508 merged Jun 6, 2025

• [train] Fix vlm use_logits_to_keep

  #4506 merged Jun 6, 2025

• [Dataset]add stsb positive subset

  #4502 merged Jun 6, 2025

• [loss] fix vlm channel loss

  #4497 merged Jun 6, 2025

• fix sft eval

  #4494 merged Jun 5, 2025

• [grpo] update grpo check

  #4493 merged Jun 5, 2025

• [grpo] support None reward & multi-task doc & more profiling

  #4459 merged Jun 5, 2025

• [grpo] support move_model_batches for external mode

  #4453 merged Jun 5, 2025

• Fix multi modal bugs in ulysses

  #4484 merged Jun 5, 2025

• [infer] fix infer stream print

  #4485 merged Jun 5, 2025

• [vlm] fix llm_lora vlm_full

  #4482 merged Jun 5, 2025

• [grpo] fix infer url

  #4480 merged Jun 5, 2025

• [pt/sft] Feature channel loss

  #4405 merged Jun 4, 2025

• [megatron] fix val_dataset

  #4478 merged Jun 4, 2025

• fix

  #4475 merged Jun 4, 2025

• [seq_parallel] fix sp compute_acc

  #4474 merged Jun 4, 2025

• Support qwen3 embedding

  #4357 merged Jun 4, 2025

• [eval] fix eval dependence

  #4472 merged Jun 4, 2025

• Fix omni grpo

  #4469 merged Jun 4, 2025

• Fix create checkpoint symlink & grpo omni

  #4468 merged Jun 4, 2025

• [grpo] fix base url

  #4463 merged Jun 4, 2025

• [train] Fix qwen2.5-vl use_cache

  #4458 merged Jun 3, 2025

• [seq_parallel] fix sp compute_acc

  #4456 merged Jun 3, 2025

• [dpo] support dpo padding_free & dpo compat trl==0.18

  #4394 merged Jun 3, 2025

• [grpo] fix hang in colocate lora settings

  #4451 merged Jun 3, 2025

• [grpo] Two-Sided Clipping for GRPO Trainer

  #4450 merged Jun 3, 2025

• [grpo] support vllm_server_base_url for vLLMClient

  #4449 merged Jun 3, 2025

• [template] fix vlm padding_free

  #4444 merged Jun 2, 2025

• fix qwen2_5_vl awq

  #4436 merged Jun 1, 2025

• [megatron] support megatron num_train_epochs

  #4432 merged Jun 1, 2025

• fix emb docs

  #4434 merged May 31, 2025

• fix model_meta

  #4431 merged May 31, 2025

• [model] Support MiMo-VL

  #4429 merged May 30, 2025

• [dataset] add ms_logger_context

  #4428 merged May 30, 2025

• [dataset] fix self-cognition & load_from_cache_file

  #4426 merged May 30, 2025

• fix transformers 4.52 device_map ddp

  #4424 merged May 30, 2025

• Fix cmdline parsing error on Windows system

  #4422 merged May 30, 2025

• support DeepSeek-R1-0528-Qwen3-8B

  #4417 merged May 29, 2025

• [pt/sft] support use_logits_to_keep & support DeepSeek-R1-0528

  #4409 merged May 29, 2025

• fix ulysses

  #4404 merged May 29, 2025

• [dataset] Fix multinode packing

  #4402 merged May 29, 2025

• Support ulysses padding-free grpo

  #4377 merged May 28, 2025

• [deploy] fix client timeout

  #4399 merged May 28, 2025

• update requirements

  #4397 merged May 28, 2025

• Standardize think templates

  #4395 merged May 28, 2025

• [megatron] fix split_dataset_ratio

  #4391 merged May 28, 2025

• [grpo] QwenLong-L1 reward model plugin

  #4385 merged May 28, 2025

• [megatron] fix save timeout & pp4 hang

  #4381 merged May 28, 2025

6 Pull requests opened by 5 people

• swift-megatron qwen3-235b-a22b

  #4401 opened May 29, 2025

• fix: add SO_REUSEADDR to find_free_port to handle TIME_WAIT state

  #4573 opened Jun 12, 2025

• solve the default 'template_backend' bug in llm.tempalte.base.Templte._encode

  #4669 opened Jun 23, 2025

• Refactor Web-UI

  #4687 opened Jun 24, 2025

• [megatron] support fp8

  #4730 opened Jun 26, 2025

• [model] support Tencent-Hunyuan/Hunyuan-A13B-Instruct

  #4745 opened Jun 27, 2025

294 Issues closed by 47 people

• Megatron不支持GRPO训练

  #4744 closed Jun 27, 2025

• DPO的full微调后Qwen3-4B模型不再输出think

  #4701 closed Jun 27, 2025

• GRPO怎么自定义format reward

  #4667 closed Jun 26, 2025

• [grpo] loading BERT model in reward

  #4580 closed Jun 26, 2025

• GRPO训练中Loss和grad_norm一直为0

  #4570 closed Jun 26, 2025

• GRPO什么时候支持多机megatron训练

  #4558 closed Jun 26, 2025

• GRPO训练reward的std始终为0

  #4512 closed Jun 26, 2025

• 多机训练使用--vllm_mode server 会卡死无法运行

  #4532 closed Jun 26, 2025

• GRPO Qwen3 32B training torch issue

  #4491 closed Jun 26, 2025

• qwen3强化训练，grpo训练结束后，爆通信错误

  #4170 closed Jun 26, 2025

• The expanded size of the tensor (8) must match the existing size (5) at non-singleton dimension 0.

  #4056 closed Jun 26, 2025

• 训练结束报错/data/chatglm/retrieval_agent_new/ms_swift_train/ms-swift/swift/cli/rlhf.py FAILED

  #4302 closed Jun 26, 2025

• dapo时在UserWarning: None of the inputs have requires_grad=True. Gradients will be None一直卡住，直至timeout

  #4050 closed Jun 26, 2025

• 用grpo训练qwen2.5-7b-instruct出现!!!!

  #4060 closed Jun 26, 2025

• 训练正常 eval时报assert error

  #4081 closed Jun 26, 2025

• Batch size in GRPO.

  #4341 closed Jun 26, 2025

• grpo训练奖励函数注册失败

  #4351 closed Jun 26, 2025

• GRPO数据传递失败

  #4362 closed Jun 26, 2025

• Qwen-Omni 全量微调grpo报错ValueError: `max_new_tokens` must be greater than 0, but is -16384

  #4392 closed Jun 26, 2025

• GRPO微调多模态训练报错

  #4470 closed Jun 26, 2025

• 双卡A6000使用GRPO微调Qwen2.5-VL-3B会OOM吗？

  #4477 closed Jun 26, 2025

• 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 closed Jun 26, 2025

• Any plans to support megatron for GRPO training?

  #3760 closed Jun 26, 2025

• LLava 跑GRPO 无法跑通

  #3928 closed Jun 26, 2025

• QWQ：GRPO训练无法跑通，报错”RuntimeError: ACL stream synchronize failed, error code:107020“

  #3932 closed Jun 26, 2025

• While training GRPO, I noticed that my model crashes. Its loss is 0, its grad_norm and kl are both Nan, and it completes as “!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!””

  #3930 closed Jun 26, 2025

• GRPO训练中间一部分后报错

  #3771 closed Jun 26, 2025

• grpo训练卡住，一直显示一下问题。

  #3794 closed Jun 26, 2025

• GRPO训练报错

  #3769 closed Jun 26, 2025

• Various traceback error during GRPO training

  #3836 closed Jun 26, 2025

• 贡献一个dockerfile吧，这个测试了 多模态的grpo训练 可以基本可以复现示例里面的结果

  #3812 closed Jun 26, 2025

• GRPO 算法如果设置 reward_model 而不是--reward_funcs ，reward模型和 model都加载到一张卡里去了

  #3843 closed Jun 26, 2025

• Meet GPU OutOfMemory in GRPO training

  #3848 closed Jun 26, 2025

• grpo训练32b模型OOM

  #3871 closed Jun 26, 2025

• GRPO 训练100 steps后性能骤降，请问是什么原因

  #3876 closed Jun 26, 2025

• Bug! Checkpoint resume failure - deepspeed different DP size. Is there a quick checkpoint converter anywere?

  #3989 closed Jun 26, 2025

• Bug! Help! MS-SWIFT GRPO + LoRA training hung/stuck after training 1 step from full merged model merged from lora adapter

  #3990 closed Jun 26, 2025

• if sleep_level > 0, gradient_accumulation_steps will be forced to 1

  #3943 closed Jun 26, 2025

• The GRPO training process hangs for multi-node training.

  #3934 closed Jun 26, 2025

• NPU环境训练速度问题

  #3331 closed Jun 26, 2025

• 求一个能8卡A100使用GRPO跑通Qwen2.5 72B模型的脚本

  #3416 closed Jun 26, 2025

• GRPO 训练时使用2个节点并且设置--num_infer_workers 2 时会报错

  #3393 closed Jun 26, 2025

• 基于qwenvl-7b-instruct训练grpo，eval过程会oom

  #3541 closed Jun 26, 2025

• 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 closed Jun 26, 2025

• 单机多卡跑grpo，多个step后会报错

  #3576 closed Jun 26, 2025

• Loss goes to 0, Gibberish Outputs

  #3582 closed Jun 26, 2025

• 日志怎么添加训练数据中的字段

  #3591 closed Jun 26, 2025

• 多机多卡GRPO assert self.cpu_group is not None

  #3583 closed Jun 26, 2025

• 设置NPROC_PER_NODE后会直接报错 failed (exitcode: -11) local_rank: 1

  #3611 closed Jun 26, 2025

• GRPO算法训练，后期训练时，显存暴增

  #3600 closed Jun 26, 2025

• grpo 固定seed，结果依旧不可复现

  #3607 closed Jun 26, 2025

• gemma3使用grpo用vllm的bug

  #3660 closed Jun 26, 2025

• 【bug】Failed to open local file in cache

  #3667 closed Jun 26, 2025

• [Bug]: RuntimeError: setup failed!

  #3662 closed Jun 26, 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 closed Jun 26, 2025

• 有没有4*V100能跑起来GRPO的训练脚本和环境配置呀？

  #3671 closed Jun 26, 2025

• ValueError: RLHF do not support sequence parallel

  #3673 closed Jun 26, 2025

• Hanging after tqdm starts [COLOCATE MODE]

  #3702 closed Jun 26, 2025

• GRPO max_grad_norm seems don't work

  #3713 closed Jun 26, 2025

• It is recommended to use a dedicated device for vLLM

  #3719 closed Jun 26, 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 closed Jun 26, 2025

• GRPO 训练，数据格式解析有bug

  #3728 closed Jun 26, 2025

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

  #3730 closed Jun 26, 2025

• Support Ulysses in Swift

  #3731 closed Jun 26, 2025

• GRPO tutorial bug: world_size (8) is not equal to tensor_model_parallel_size (4) x pipeline_model_parallel_size (1)

  #3739 closed Jun 26, 2025

• 多模态qwen2.5-vl-3B,grpo实验报错

  #3398 closed Jun 26, 2025

• grpo微调deepseek v2，训练过程中到eval阶段，就会卡住，然后就会停止训练

  #3528 closed Jun 26, 2025

• 请问如何在grpo中配置自定义的数据集路径，并进行数据格式转换？

  #3525 closed Jun 26, 2025

• 2workers_async_iterations2_vllm help

  #3522 closed Jun 26, 2025

• Bug in GRPO best practices document!

  #3501 closed Jun 26, 2025

• unhashable type: 'list'

  #3490 closed Jun 26, 2025

• 请求支持GRPO训练中，vllm推理后端支持多张卡🙏 Request for support for using multiple cards in the vLLM inference backend during GRPO training

  #3477 closed Jun 26, 2025

• 使用GRPO进行Qwen2.5-vl-7B-Instruct训练，报错：无法多卡训练，只能加载1张卡并oom

  #3404 closed Jun 26, 2025

• GRPO训练功能建议

  #3415 closed Jun 26, 2025

• GRPO 训练loss和reward异常

  #3372 closed Jun 26, 2025

• grpo 多机多卡训练timeout

  #3343 closed Jun 26, 2025

• GRPO训练LLAVA CUDA Error

  #3264 closed Jun 26, 2025

• GRPO LLava 训练报错，无法多卡训练，1卡可以

  #3228 closed Jun 26, 2025

• GRPO 4卡A100训练BUG

  #3223 closed Jun 26, 2025

• 如何对deepseek r1做sft和grpo微调

  #3211 closed Jun 26, 2025

• 使用GRPO 使用我已经训练的LLava模型加载问题

  #3195 closed Jun 26, 2025

• GRPO deepspeed lmdeploy训练InternVL2d5 报错

  #3151 closed Jun 26, 2025

• Using Unsloth in conjunction with GRPO to train a model for OOM

  #3183 closed Jun 26, 2025

• grpo训练如何设置vllm_device使用多张卡

  #3098 closed Jun 26, 2025

• Does ms-swift support tensor(model)-parallel GRPO training?

  #3068 closed Jun 26, 2025

• ValueError: Image features and image tokens do not match: tokens: 5589, features 5805

  #2460 closed Jun 26, 2025

• grad_norm nan

  #2280 closed Jun 26, 2025

• 期望RLHF能支持序列并行（sequence_parallel）

  #1958 closed Jun 26, 2025

• GRPO训练的old_per_token_logps计算是不是有bug

  #4727 closed Jun 26, 2025

• rerank 数据加载错误

  #4728 closed Jun 26, 2025

• Issue with Multi-GPU Training

  #4718 closed Jun 26, 2025

• Qwen3 Full Sft设置predict_with_generate=true报错keyerror"messages"，为false时可以正常训练结束

  #4695 closed Jun 26, 2025

• 支持 moonshotai/Kimi-VL-A3B-Thinking-2506

  #4708 closed Jun 25, 2025

• grpo训练qwen2.5-vl报错

  #4364 closed Jun 25, 2025

• 全量微调grpo 相同数量的样本ms-swift效果比unsloth效果差很多

  #4393 closed Jun 25, 2025

• GRPO OOM USE resume_from_checkpoint

  #4406 closed Jun 25, 2025

• GRPO训练报错：AssertionError: Forward context is not set. Please use `set_forward_context` to set the forward context.

  #4418 closed Jun 25, 2025

• 支持的DeepSeek-R1训练是指671B的模型吗还是蒸馏的模型？

  #3132 closed Jun 25, 2025

• seq_cls训练时候开启flash_attn指标大幅度低于不开flash_attn

  #4384 closed Jun 25, 2025

• 多回归任务，推理问题

  #4705 closed Jun 25, 2025

• 请问使用zero2/zero3导致max_steps相差八倍的原因是什么？

  #4616 closed Jun 23, 2025

• 请求增加对Qwen3-8B的自我认知训练的NoteBook文件

  #4034 closed Jun 23, 2025

• raise KeyError(f"Column {key} not in the dataset. Current columns in the dataset: {columns}") [rank1]: KeyError: 'Column length not in the dataset. Current columns in the dataset: []'

  #4058 closed Jun 23, 2025

• InternVL3-9B LoRA微调数据集预处理速度缓慢问题（大约7h）

  #4076 closed Jun 23, 2025

• 单坐标点定位物体位置

  #4292 closed Jun 23, 2025

• data_load

  #4288 closed Jun 23, 2025

• Seq CLS Infer 问题咨询

  #4325 closed Jun 23, 2025

• UI-TARS冻结参数推理无法均匀分配显存导致超出显存

  #4359 closed Jun 23, 2025

• 微调Qwen3在默认脚本上加上zero2/3会OOM

  #4371 closed Jun 23, 2025

• VLLM Engine Batch 推理咨询

  #4386 closed Jun 23, 2025

• swift infer这个些命令如何转为python命令运行的，内部原理

  #4555 closed Jun 23, 2025

• Failing to preprocess hf dataset

  #4564 closed Jun 23, 2025

• How-to use on Apple Mac?

  #4572 closed Jun 23, 2025

• Multimodal finetune llava1.6-mistral bug: RuntimeError: Tensors must have same number of dimensions

  #4578 closed Jun 23, 2025

• 关于ms-swift 3.x的template和2.x的不同

  #4602 closed Jun 23, 2025

• ovis2 微调失败，loss计算时报ValueError: Expected input batch_size (1384) to match target batch_size (16384)

  #4611 closed Jun 23, 2025

• 关于pip install -e '.[all]' 的安装、evalscope的安装的咨询

  #4605 closed Jun 23, 2025

• loss_scale hermes not work

  #4607 closed Jun 23, 2025

• 华为910B lora qwen2.5vl报错：AssertionError: Torch not compiled with CUDA enabled

  #4619 closed Jun 23, 2025

• 满血版R1/Qwen3-235B-30A HF参数转megatron OOM

  #4648 closed Jun 23, 2025

• 10分钟改变大模型自我认知教程报错'Qwen2_5VLTemplate' object has no attribute 'model'

  #4662 closed Jun 23, 2025

• DPO训练到 100 步时，遇到 StopIteration ERROR during training 问题

  #4644 closed Jun 23, 2025

• 多卡多进程使用orpo卡死，触发watchdog caught collective operation timeout.

  #3564 closed Jun 20, 2025

• 无

  #4481 closed Jun 20, 2025

• 无

  #4504 closed Jun 20, 2025

• 我希望在训练reward model的时候添加一个分类损失应该怎么做

  #4640 closed Jun 20, 2025

• 我如何基于一个qwen2.5vl创建一个新的reward model结构并进行训练

  #4635 closed Jun 19, 2025

• RuntimeError: shape '[-1, 151936]' is invalid for input of size 6266880

  #4318 closed Jun 19, 2025

• Finetune Qwen3-235B-a22b with LoRA end with error: AttributeError: 'LigerQwen3MoeSwiGLUMLP' object has no attribute 'down_proj'

  #4531 closed Jun 19, 2025

• When to support SGLang

  #3510 closed Jun 19, 2025

• RuntimeError on NPU: a leaf Variable that requires grad is being used in an in-place operation.

  #4613 closed Jun 19, 2025

• [INFO:swift.trainers.rlhf_trainer.vllm_client] Server is not up yet.

  #4525 closed Jun 18, 2025

• 实现Qwen3技术报告中的on policy distillation

  #4533 closed Jun 18, 2025

• Liger kernel not working with Qwen2.5 VL

  #4543 closed Jun 18, 2025

• qwen3-1..7B GRPO 训练几个轮次后爆显存

  #4388 closed Jun 18, 2025

• qwenvl2.5 lora training

  #4615 closed Jun 18, 2025

• agent template issue with no argument

  #4600 closed Jun 17, 2025

• 多轮对话数据的训练，但只训练最后一轮的assistant回答

  #4596 closed Jun 16, 2025

• Agent训练qwen2.5-base eos token 训练推理不一致问题

  #4498 closed Jun 15, 2025

• GRPO lora

  #4593 closed Jun 14, 2025

• 8*A100执行GRPO完整流程，使用vllm爆显存，不使用vllm可以正常训练但缓慢

  #4594 closed Jun 14, 2025

• 修改num_generations参数，报错ValueError: range() arg 3 must not be zero

  #4589 closed Jun 13, 2025

• 最新代码分支超长文本训练报错

  #4583 closed Jun 13, 2025

• lora训练Can't pickle local object 'DeepSpeedEngine._create_module_forward_post_hook.<locals>._module_forward_post_hook'

  #4586 closed Jun 13, 2025

• megatron-swift支持DeepSeek-R1-0528-Qwen3-8B

  #4438 closed Jun 12, 2025

• Internvl2.5-4B GRPO训练视频数据时报错

  #4579 closed Jun 12, 2025

• 自我认知demo mac上运行报错

  #4575 closed Jun 12, 2025

• GRPO训练过程中loss和grad_norm都为0，提示没有label_names

  #4547 closed Jun 12, 2025

• transformers4.52: if v not in ALL_PARALLEL_STYLES: TypeError: argument of type 'NoneType' is not iterable

  #4577 closed Jun 12, 2025

• grpo训练卡住

  #4549 closed Jun 12, 2025

• Why does applying sequence parallelism reduce the step count?

  #4553 closed Jun 11, 2025

• InfoNCE数据处理阶段报错

  #4546 closed Jun 11, 2025

• 使用ToolBench数据集出错

  #3947 closed Jun 11, 2025

• CoundownORM是什么？

  #4528 closed Jun 11, 2025

• 新版本ms-swift训练GRPO时报错 AttributeError: Can't pickle local object 'GRPOTrainer.__init__.<locals>.<lambda>'

  #4557 closed Jun 11, 2025

• null

  #3208 closed Jun 11, 2025

• Qwen2.5-vl lora GRPO 微调后怎么用 hg 推理呢

  #3187 closed Jun 11, 2025

• Qwen3 Full Sft后export hf失败

  #4550 closed Jun 11, 2025

• Intern3VL进行GRPO训练时报错：KeyError: 'input_ids'

  #4519 closed Jun 10, 2025

• 载入模型一个比较奇怪的事情

  #4539 closed Jun 10, 2025

• peft热补丁是否加载的判断函数存在问题

  #4534 closed Jun 9, 2025

• dpo train qwen2.5-7b

  #4526 closed Jun 9, 2025

• DPO Sequence_parallel_size == 8 Error NotImplementedError

  #4420 closed Jun 8, 2025

• 断点续训lora模型，模型加载不上

  #4507 closed Jun 6, 2025

• GRPO训练时可以根据不同的数据集定义不同的奖励函数吗？

  #4488 closed Jun 6, 2025

• Seq CLS 任务 PT engine token 量咨询

  #4490 closed Jun 6, 2025

• Ovis2 finetune: ValueError: Expected input batch_size (35) to match target batch_size (2025)

  #4495 closed Jun 6, 2025

• 关于Qwen2.5-Omni训练时Freeze_vit参数的疑问

  #4489 closed Jun 6, 2025

• 自定义数据集中的格式问题

  #4500 closed Jun 6, 2025

• AssertionError: Data does not have channel field.

  #4487 closed Jun 6, 2025

• How to get DPO chosen_rewards, rejected_rewards for inference?

  #4441 closed Jun 6, 2025

• GRPO with vLLM error when reading images "Input should be a valid string"

  #4447 closed Jun 6, 2025

• 8卡H20 GRPO 3B模型OOM

  #4486 closed Jun 5, 2025

• GRPO OOM: Qwen 72B (6xH100 Training, 4xH100 External vLLM) - Master GPU Low Load & Crash

  #4440 closed Jun 5, 2025

• qwen2.5-vl-72b, vllm_server_host方式运行，CUDA out of memory

  #4023 closed Jun 5, 2025

• MiMo-VL-7B-RL微调grounding任务，推理结果的中标示包围框的special token为乱码

  #4483 closed Jun 5, 2025

• QwenVL2.5 单机多卡，无法开启packing和sequence_parallel_size

  #4277 closed Jun 5, 2025

• OSError: [Errno 39] Directory not empty

  #3618 closed Jun 5, 2025

• 二分类任务大模型微调，token_acc表示什么

  #3639 closed Jun 5, 2025

• KTO 训练每次保持ckt 都报错

  #3669 closed Jun 5, 2025

• ovis 一定要flash_attn才能训练吗？

  #3703 closed Jun 5, 2025

• global_step由于磁盘空间原因没有保存完整，如何接着训练

  #3740 closed Jun 5, 2025

• [WARNING:swift] No training was carried out, which may be due to the dataset being too small or incorrect usage of resume_from_checkpoint.

  #3863 closed Jun 5, 2025

• DPO训练log打印日志:logits/chosen和logits/rejected完全一样

  #3880 closed Jun 5, 2025

• `generation_config` default values have been modified to match model-specific defaults: {'temperature': 0.1}. If this is not desired, please set these values explicitly.

  #3892 closed Jun 5, 2025

• 对于自定义的数据集，如何计算其token数量？

  #3899 closed Jun 5, 2025

• 关于系统提示词

  #4020 closed Jun 5, 2025

• Qwen3 Function Calling Fine Tuning

  #4479 closed Jun 4, 2025

• About the support for Kimi-Audio

  #4008 closed Jun 4, 2025

• Qwen2.5-omni-7b 单机多卡，无法开启sequence_parallel_size

  #4281 closed Jun 4, 2025

• fsdp_qlora 8*4090 save model的时候报错了

  #3543 closed Jun 4, 2025

• 2张3090训练70B模型的脚本为啥是7B的模型

  #3929 closed Jun 4, 2025

• TypeError in rollout_server.sh: EngineArgs.__init__() got unexpected keyword 'worker_extension_cls'

  #4202 closed Jun 4, 2025

• qwen2-vl 的 pretrain 是否支持

  #2222 closed Jun 4, 2025

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

  #3715 closed Jun 4, 2025

• Fine-tuning Qwen2.5-Omni-7B with additional new layers on the audio tower

  #4070 closed Jun 4, 2025

• inference error with vllm 0.8.5

  #4063 closed Jun 4, 2025

• 对于一个已经完成sft之后的任务，如果我想加入新的知识但不想掉点，我应该选择ms-swift实现的强化微调和GRPO哪个来完成呢？

  #4107 closed Jun 4, 2025

• qwen omni注册的问题

  #4110 closed Jun 4, 2025

• grpo use speical token

  #4162 closed Jun 4, 2025

• grpo Async mode nccl timeout

  #4306 closed Jun 4, 2025

• Exception: Current loss scale already at minimum - cannot decrease scale anymore. Exiting run.

  #4374 closed Jun 4, 2025

• DPO是否支持多模态的微调？数据集格式有大佬知道吗？

  #4169 closed Jun 4, 2025

• 请问是否支持qwen3 moe 进行grpo+lora训练

  #4190 closed Jun 4, 2025

• OSError: [Errno 16] Device or resource busy: '/mnt/hdfs/harunava/home/byte_magellan_va/chenzhiyang/product/output/dpo/logging.jsonl'.

  #4266 closed Jun 4, 2025

• support Qwen 3 and Qwen 3 MoE

  #3922 closed Jun 4, 2025

• ImportError: cannot import name 'Qwen2_5OmniModel from 'transformers'

  #3897 closed Jun 4, 2025

• [Bug]: Wrong context length for Qwen 2.5 7B-Instruct?

  #3907 closed Jun 4, 2025

• dapo时在UserWarning: None of the inputs have requires_grad=True. Gradients will be None一直卡住，直至timeout

  #4049 closed Jun 4, 2025

• 更新swift后GRPO训练Qwen-Omini报错

  #4455 closed Jun 4, 2025

• 如何微调自有的COT数据集

  #3535 closed Jun 4, 2025

• OSError: [Errno 28] No space left on device

  #3551 closed Jun 4, 2025

• deepseek_r1_distill无法使用liger

  #3555 closed Jun 4, 2025

• 8*H20全量微调qwen32b爆显存

  #3560 closed Jun 4, 2025

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

  #3688 closed Jun 4, 2025

• No training with AWQ QLORA (Qwen 2.5 VL)

  #3698 closed Jun 4, 2025

• making llm_max_batch_size and mllm_max_batch_size configurable

  #4077 closed Jun 4, 2025

• function call 微调报错 TypeError: string indices must be integers, not 'str'

  #4082 closed Jun 4, 2025

• 在NPU上SFTQwen3遇到Default process group has not been initialized, please make sure to call init_process_group.

  #4086 closed Jun 4, 2025

• 为啥现做RLHF 不支持sequence_parallel

  #4089 closed Jun 4, 2025

• max_pixels到底是怎么发挥作用呢？

  #3721 closed Jun 4, 2025

• Qwen2.5-VL-32B-Instruct-AWQ无法进行推理

  #3722 closed Jun 4, 2025

• Qwen2.5-Omni-7B 部署api推理报错

  #3724 closed Jun 4, 2025

• vl类模型微调是否支持图片url

  #3738 closed Jun 4, 2025

• Qwen2.5-VL训练train memory逐步升高

  #3741 closed Jun 4, 2025

• [Question] ImportError when running mPLUG-Owl3-7B with Swift

  #3753 closed Jun 4, 2025

• Qwen2VL微调到某个step就报错

  #3924 closed Jun 4, 2025

• UserWarning: None of the inputs have requires_grad=True

  #3935 closed Jun 4, 2025

• File "/home/work/.local/torch/nn/modules/module.py", line 1928, in __getattr__ [rank0]: raise AttributeError( [rank0]: AttributeError: 'Qwen2ForCausalLM' object has no attribute 'zero_gather_16bit_weights_on_model_save'

  #3280 closed Jun 4, 2025

• PPO报错：AttributeError: 'Qwen2ForCausalLM' object has no attribute 'zero_gather_16bit_weights_on_model_save'

  #3937 closed Jun 4, 2025

• 关于 DeepSpeed Config 的问题

  #3953 closed Jun 4, 2025

• swift app

  #3958 closed Jun 4, 2025

• deepspeed报错

  #3991 closed Jun 4, 2025

• infer速度远远小于train期间infer val数据集的速度

  #3992 closed Jun 4, 2025

• qwen2.5-vl-72b多卡推理卡住

  #4021 closed Jun 4, 2025

• 3.4.0版本的swift会过滤数据集，是什么因素导致？

  #4026 closed Jun 4, 2025

• About the smart_resize of qwen2.5-vl

  #4027 closed Jun 4, 2025

• qwen2_5VL missing apply_liger_kernel

  #4028 closed Jun 4, 2025

• 不支持bf16报错

  #4036 closed Jun 4, 2025

• GRPO NPU多卡训练报错

  #4421 closed Jun 4, 2025

• Will the future model expand llama3.2-vision-instruct？

  #4460 closed Jun 4, 2025

• grpo时，kl系数为时不时出现nan

  #4465 closed Jun 4, 2025

• qwen3如何不使用思维链微调

  #4038 closed Jun 4, 2025

• 能否支持一下多卡并行推理

  #1444 closed Jun 4, 2025

• Dataset stucked when using --dataloader_num_workers 1 and --streaming true

  #1968 closed Jun 4, 2025

• NPU+deepspeed zero3环境下--max_grad_norm失效

  #4198 closed Jun 4, 2025

• 全量微调，模型参数 lm_head 没有存下来

  #4205 closed Jun 4, 2025

• multiple GPU trainning sft error

  #4219 closed Jun 4, 2025

• PtEngine在多模态推理时，输出被强制截断，导致输出不全，看了下没找到哪里设置长度的，使用的是test_vision.py这个脚本

  #4305 closed Jun 4, 2025

• 微调数据格式音频标签问题

  #4310 closed Jun 4, 2025

• When making inferences in [swift infer], how can I specify the use of a custom chat_template?

  #4312 closed Jun 4, 2025

• qwen2.5 vl 进行全参，即full模型训练没有权重文件保存。

  #4331 closed Jun 4, 2025

• 微调后的权重args中norm_bbox参数一直为null

  #4333 closed Jun 4, 2025

• swift pt

  #4339 closed Jun 4, 2025

• Packing过程中报错

  #4340 closed Jun 4, 2025

• NPU训练gme模型并行dataset map报错

  #4425 closed Jun 4, 2025

• Qwen3微调数据集格式

  #4346 closed Jun 4, 2025

• pip install -e '.[eval]' 时报错

  #4347 closed Jun 4, 2025

• ValueError: Attempting to unscale FP16 gradients.

  #4352 closed Jun 4, 2025

• swift sft下面无法使用log_completions

  #4356 closed Jun 4, 2025

• 执行 pip install -e .[all]，absl-py 误判 Python 版本，导致安装失败

  #4466 closed Jun 4, 2025

• transformer_engine 安装失败

  #4051 closed Jun 4, 2025

• eval包安装resolution-too-deep

  #4349 closed Jun 4, 2025

• Exporting to the mcore model will cause errors when using multiple devices.

  #4350 closed Jun 4, 2025

• qwen3 4b模型保存的ckpt中tokenizer的chat template没有了，多了一个chat_template.jinja文件

  #4358 closed Jun 4, 2025

• Megatron rlhf 支持

  #4370 closed Jun 4, 2025

• qwen3训练时报错 KeyError: 'ignore_empty_think'

  #4382 closed Jun 4, 2025

• megatron train_iters and max_epochs

  #4396 closed Jun 4, 2025

• swift3.1.0.dev0在推理Florence2时错误

  #3133 closed Jun 4, 2025

• 在新版本（3.4）中，如果nproc_per_node小于CUDA_VISIBLE_DEVICES的数量时无法运行，老版本（3.2）可以

  #4019 closed Jun 4, 2025

• 使用 AutoModelForSequenceClassification 训练 seq_cls 任务时出错

  #3927 closed Jun 4, 2025

• GRPO external 模式训练报错

  #4467 closed Jun 4, 2025

• Total training steps 计算有误

  #4366 closed Jun 4, 2025

• Feature request：预训练数据集packing处理逻辑问题

  #2352 closed Jun 4, 2025

• Megatron-SWIFT 的依赖 apex 难装，频繁遇到问题

  #4414 closed Jun 3, 2025

• qwen2.5-vl-7b sft，开启packing后内存持续增长，开启streaming后显存占用极高

  #4423 closed Jun 3, 2025

• 安装ms-swift[all]时resolution-too-deep

  #4437 closed Jun 3, 2025

• RuntimeError: indices should be either on cpu or on the same device as the indexed tensor模型并行问题

  #4442 closed Jun 3, 2025

• Can I run full parameter fine-tuning for vision + LoRA for text decoder?

  #4446 closed Jun 3, 2025

• MP(device_map) + DDP 无法启动，inferred_max_memory为nonetype

  #4452 closed Jun 3, 2025

• embedding task 支持注入自定义 loss 或者支持 MatryoshkaLoss

  #4430 closed Jun 3, 2025

• target_modules

  #4427 closed Jun 2, 2025

• 是否计划支持MiMo-VL模型的微调？

  #4435 closed Jun 1, 2025

• 希望支持图片列表形式的videos训练

  #4290 closed May 30, 2025

• 序列分类模型训练疑问

  #4332 closed May 30, 2025

• windows下面web-ui转换的cmd命令无法执行

  #4403 closed May 30, 2025

• 关于MS-LongWrite中的损失函数loss-ce

  #4218 closed May 29, 2025

• 自定义模型get_function被跳过

  #4243 closed May 29, 2025

• 使用GRPO训练Qwen2.5-Omni的时候启用use_liger_kernel参数报错

  #4398 closed May 29, 2025

• Inconsistent `<think>` Encoding Behavior in Qwen3 Multi-Turn GRPO

  #4390 closed May 28, 2025

• grpo vllm似乎在无效重复加载

  #4368 closed May 28, 2025

• GRPO Training Generates Identical Completions for Same Prompts in Async Mode

  #4383 closed May 28, 2025

• 自定义metric不生效

  #4372 closed May 28, 2025

• KeyError: 'eval_epoch'

  #4323 closed May 27, 2025

135 Issues opened by 110 people

• 基于本地加载数据集进行多卡并行训练，停在Init COMPLETE... 无法进入train阶段

  #4743 opened Jun 27, 2025

• 输入多图的编号问题

  #4742 opened Jun 27, 2025

• ms swift如何加入early stop

  #4741 opened Jun 27, 2025

• [WARNING:swift] Please install the package: pip install "decord" -U

  #4740 opened Jun 27, 2025

• Qwen2.5-omni GRPO训练出现内存OOM

  #4739 opened Jun 27, 2025

• 微调DeepSeek模型报错：AssertionError: noaux_tc not supported for training

  #4737 opened Jun 26, 2025

• Does the packing feature block attention score between different samples?

  #4736 opened Jun 26, 2025

• a question for rl

  #4735 opened Jun 26, 2025

• Please open Security Advisories for vulnerability reporting

  #4733 opened Jun 26, 2025

• 在学习全部轮次的SFT训练中，中间轮次结束符号不能被学习，导致训练后的模型无法停止

  #4732 opened Jun 26, 2025

• swift推理精度差异

  #4726 opened Jun 26, 2025

• 使用lora 训练qwen2.5vl3b之后，lora未合并，使用deploy部署，使用pt, 跟vllm 结果不一致

  #4725 opened Jun 26, 2025

• GKD代码加载模型卡死

  #4724 opened Jun 26, 2025

• Swift代码库进行lora checkpoint的continue sft，加载模型和checkpoint后可训练参数为0%

  #4723 opened Jun 26, 2025

• qwen2.5vl lora sft关于freeze_vit

  #4722 opened Jun 26, 2025

• [rank4]: AssertionError: Expected multimodal embeddings to be a list/tuple of 2D tensors, or a single 3D tensor, but got <class 'NoneType'> instead.

  #4721 opened Jun 25, 2025

• qwen3 embedding 微调在评估阶段报错：'NoneType' object has no attribute 'get'

  #4720 opened Jun 25, 2025

• 添加python示例代码

  #4717 opened Jun 25, 2025

• 如何传入自定义的causal_attention_mask

  #4716 opened Jun 25, 2025

• local_repo_path参数，在python脚本里如何添加

  #4714 opened Jun 25, 2025

• hf格式模型文件转megatron报错: CUDA error: operation not supported

  #4713 opened Jun 25, 2025

• lora 微调 Ovis2-34B loss=0.0

  #4711 opened Jun 25, 2025

• Deepspeed zero3 多 GPU 训练没法设置 batch_size 为1

  #4710 opened Jun 25, 2025

• [Bug]: [WARNING:swift] Please install the package: pip install "decord" -U

  #4709 opened Jun 25, 2025

• 多回归任务 输出问题

  #4706 opened Jun 25, 2025

• 序列分类任务，能否多卡训练？

  #4704 opened Jun 25, 2025

• Swift rollout卡住

  #4703 opened Jun 25, 2025

• NuminaMath-TIR数据集上训GRPO不work

  #4702 opened Jun 25, 2025

• 使用msswift框架，基于QwQ-32B模型，微调自制的function-call数据集，效果很差，不知道原因

  #4700 opened Jun 25, 2025

• 自定义数据集包含了'messages'、'rejected_response'、'label'、'images'、'videos'、'audios'、'tools'和'objects'之外的key，该如何写template？

  #4699 opened Jun 24, 2025

• lora微调qwen3 embedding模型弹出警告find_unused_parameters

  #4698 opened Jun 24, 2025

• Qwen2-VL merge lora报错

  #4697 opened Jun 24, 2025

• 求问 Qwen 235B A22 训练成本和 Qwen 32B dense 对比

  #4696 opened Jun 24, 2025

• 用CLI推理时，有办法能在推理结果中保存输入的dataset中的额外参数嘛？

  #4693 opened Jun 24, 2025

• VLLM Engine 咨询

  #4692 opened Jun 24, 2025

• 多机加载大数据集时，会多台机子先后串行加载

  #4691 opened Jun 24, 2025

• UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown

  #4689 opened Jun 24, 2025

• 我想要给PPO设置两个reward model和两个value model，通过两者的value和reward加权计算loss损失，应该怎么做？

  #4688 opened Jun 24, 2025

• 请问是否支持QWenVL等多模态模型的增量预训练？

  #4686 opened Jun 24, 2025

• 改变 IMAGE_FACTOR 是不是意味着视觉部分需要重新训练？

  #4685 opened Jun 24, 2025

• 如何关闭自动模型并行呢？

  #4684 opened Jun 24, 2025

• Help: Multi Turn SFT

  #4681 opened Jun 24, 2025

• mllm模型训练，一个epoch训练完任务卡住，gpu利用率100%，无法save checkpoint

  #4680 opened Jun 23, 2025

• GRPO训练失败，模型似乎学习困难

  #4679 opened Jun 23, 2025

• 奖励函数一直震荡不上升，似乎学不到东西

  #4677 opened Jun 23, 2025

• SFT训练一个回归任务后，推理使用vllm加速，模型load会报错，有办法解决吗

  #4676 opened Jun 23, 2025

• 持续输出Trainer.tokenizer is now deprecated. You should use Trainer.processing_class instead.，但是不报错，请问是什么原因？

  #4673 opened Jun 23, 2025

• 关于 rlhf 数据的preprocess

  #4670 opened Jun 23, 2025

• 微调 MiniCPM-o-2_6 报错 assert media_type in {'image', 'video'}

  #4668 opened Jun 23, 2025

• Applying sequence parallelism causes the training to finish early, even though it hasn't reached the specified max_steps

  #4663 opened Jun 23, 2025

• Any way to run evaluation before training starts?

  #4660 opened Jun 22, 2025

• 'weight' must be 2-D

  #4656 opened Jun 20, 2025

• grpo 多任务训练 奖励函数设置返回None 这样的话，如果想要查看单个任务的reward曲线，在tensorboard中会出现nan的情况

  #4653 opened Jun 20, 2025

• 为什么没有loss

  #4652 opened Jun 20, 2025

• 训练Omni的时候会卡住不动

  #4651 opened Jun 20, 2025

• 请问下swift中集成的lora-ga是否支持多卡训练呢

  #4646 opened Jun 20, 2025

• 更新以后我应该如何获得history呢

  #4645 opened Jun 20, 2025

• multi-gpu GRPO training with sequence parallelism terminated unexpectedly during execution, with the following warning: destroy_process_group() was not called before program exit, which may lead to resource leaks.

  #4643 opened Jun 19, 2025

• 如何新增一个vlm从而做embedding任务？

  #4642 opened Jun 19, 2025

• 想问下embedding的训练如何加入system or instructions？

  #4638 opened Jun 19, 2025

• shape mismatch internvl3

  #4636 opened Jun 19, 2025

• Qwen2.5-vl预训练过程中loss突然激增

  #4634 opened Jun 19, 2025

• 训练日志停止且GPU利用率异常

  #4633 opened Jun 19, 2025

• qwen2.5vl定位训练同样的参数环境多次训练结果波动非常大

  #4632 opened Jun 18, 2025

• 请问支持Agent的RL训练吗？

  #4631 opened Jun 18, 2025

• 新增megatron sft中的freeze_parameters_regex参数支持

  #4630 opened Jun 18, 2025

• QWEN3-32B LORA GRPO 无报错卡住

  #4628 opened Jun 18, 2025

• swift infer 设置了temperature，top_p 但是每次生成都是同样的结果

  #4627 opened Jun 18, 2025

• 环境变量设置了NPROC_PER_NODE=2，一台机器2张卡，为什么在推理时还是发生了MP而没有发生DP

  #4625 opened Jun 17, 2025

• Large Language Diffusion Moldes支持

  #4620 opened Jun 17, 2025

• Qwen3 Embedding训练抛出多线程的错误

  #4617 opened Jun 17, 2025

• qwen2.5-vl grounding任务里同时有分类，是否支持？

  #4614 opened Jun 16, 2025

• lora微调后merge完模型进行lmdeploy推理用时比Qwen2.5-VL-7B-Instruct多一倍，原因为何？

  #4609 opened Jun 16, 2025

• Any example on training llama on function calling dataset?

  #4604 opened Jun 15, 2025

• qwen2.5-7B GRPO训练时卡住，未显示任何报错

  #4603 opened Jun 15, 2025

• qwen3-32B全参数ppo训练一步报错

  #4599 opened Jun 13, 2025

• GRPO 训练到100步，评估保存的位置报错，还未进行评估

  #4598 opened Jun 13, 2025

• 关于多模态目标检测多轮对话数据集的训练

  #4597 opened Jun 13, 2025

• 采用swift infer 测试qwen2.5-omni模型结果，与官方测试方法结果不一致

  #4595 opened Jun 13, 2025

• Infonce loss hard negatives type error

  #4588 opened Jun 12, 2025

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

  #4585 opened Jun 12, 2025

• 训练qwen2.5vl时，开启序列并行+packing，loss会掉到0

  #4581 opened Jun 12, 2025

• GRPO的时候怎么保存最后一步的checkpoints

  #4574 opened Jun 12, 2025

• cachedqwen2tokenizer does not exist

  #4569 opened Jun 12, 2025

• 使用--device_map auto 报错

  #4567 opened Jun 11, 2025

• 🍭[Roadmap] ms-swift3.6

  #4561 opened Jun 11, 2025

• Megatron-SWIFT 是否支持Qwen2.5 VL模型呀

  #4559 opened Jun 11, 2025

• 如何进行多轮对话的训练

  #4552 opened Jun 10, 2025

• GRPO使用自定义预处理器加载的多模态数据集时卡死无法训练

  #4551 opened Jun 10, 2025

• 训练中评测示例中为什么使用一个中文qa数据集去训练但是用一个数学类数据集去评测？

  #4544 opened Jun 10, 2025

• vllm不支持微调的qwen2.5-omni模型

  #4542 opened Jun 10, 2025

• math_verify 解决表达式能力有限？

  #4541 opened Jun 10, 2025

• 微调qwen3-235B-A22B-AWQ

  #4540 opened Jun 10, 2025

• 怎么保存性能最好的几个checkpoint

  #4538 opened Jun 10, 2025

• KTO训练数据构造

  #4535 opened Jun 9, 2025

• 我想知道我的最终自定义数据集最终长什么样子应该如何操作?

  #4530 opened Jun 9, 2025

• 目标检测自定义数据集咨询

  #4529 opened Jun 9, 2025

• 多模态预训练性能问题

  #4527 opened Jun 9, 2025

• Any plan to add DeepEyes implementation?

  #4524 opened Jun 8, 2025

• LISA显存占用

  #4522 opened Jun 8, 2025

• 实现优势样本回放(SSR)机制

  #4520 opened Jun 8, 2025

• Is there a simple way to add online augmentation to the data ?

  #4515 opened Jun 7, 2025

• lora merge 的 swift export 是不是已经过时了

  #4513 opened Jun 7, 2025

• export qwen2.5-vl-3b 的lora模型存在问题

  #4511 opened Jun 6, 2025

• huggingface上下载的数据集在finetuning的时候还需要重新下载

  #4510 opened Jun 6, 2025

• best_model_checkpoint 为null

  #4509 opened Jun 6, 2025

• [Question] Does Megatron-SWIFT restore the streaming-dataset offset when resuming from a checkpoint?

  #4505 opened Jun 6, 2025

• finetuning qwen2.5vl+qwen3-8b acc=0

  #4503 opened Jun 6, 2025

• RL训练能否支持ProRL

  #4501 opened Jun 6, 2025

• GRPO训练Qwen2.5Omni异常

  #4499 opened Jun 6, 2025

• dpo多机单卡训练tensorboard日志路径问题

  #4496 opened Jun 6, 2025

• Qwen2.5-omni vllm 推理异常

  #4492 opened Jun 5, 2025

• TypeError: Accelerator.unwrap_model() got an unexpected keyword argument 'keep_torch_compile'

  #4473 opened Jun 4, 2025

• Qwen2.5-VL-3B 在 2卡a100中推理会爆显存

  #4471 opened Jun 4, 2025

• swift deploy 命令中，--max_model_len 80000没生效

  #4464 opened Jun 4, 2025

• 多节点megatron训练，packing_share报错

  #4462 opened Jun 4, 2025

• No matching distribution found for math_verify==0.5.2

  #4461 opened Jun 3, 2025

• _dynamic_preprocess will fail on small image like 448x364

  #4457 opened Jun 3, 2025

• Unsloth Lora fine_tunning doesn't work

  #4454 opened Jun 3, 2025

• Multi-node slurm training?

  #4448 opened Jun 3, 2025

• 支持读取云存储中的多模态数据集

  #4445 opened Jun 2, 2025

• 是否可以设置每个数据集的repeat_time

  #4443 opened Jun 2, 2025

• Padding free feature

  #4439 opened Jun 2, 2025

• 使用vllm批量推理时卡间通信报错

  #4433 opened May 31, 2025

• 支持reinfoce++

  #4419 opened May 29, 2025

• 对InternVL3-38B进行GRPO训练，lora和全参两种方式分别需要多少显存？

  #4416 opened May 29, 2025

• 支持多轮DPO吗

  #4415 opened May 29, 2025

• 多模态GRPO避免视觉信息重复读取

  #4413 opened May 29, 2025

• Swift支持在NPU启动序列并行吗

  #4412 opened May 29, 2025

• Error occurred when saving checkpoints during Qwen3 multi-GPU SFT

  #4411 opened May 29, 2025

• 训练完reward model，如何用rm预测单个样本的分数呢？

  #4410 opened May 29, 2025

• Saving best ckpt leads to an empty symlink.

  #4408 opened May 29, 2025

• 通过swift部署的模型随着推理次数增加，显存不断增加

  #4407 opened May 29, 2025

• Unsloth problem with LoRA

  #4400 opened May 28, 2025

• AssertionError: `output/v31-20250527-181405/checkpoint-9900` is not an adapter, please try using `--model` to pass it.

  #4389 opened May 28, 2025

54 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.

• SWIFT支持embedding模型的推理和部署如何实现

  #3606 commented on May 27, 2025 • 0 new comments

• LLava-next 1.6 KTO Error

  #3806 commented on May 28, 2025 • 0 new comments

• lora 微调的模型使用--resume_from_checkpoint参数，继续训练报显存不足；不使用--resume_from_checkpoint参数可以正常训练

  #2505 commented on May 29, 2025 • 0 new comments

• cannot import name 'LoRA' from 'swift'

  #3665 commented on May 29, 2025 • 0 new comments

• 华为910B上使用device_map auto 参数merge lora会出现卡死

  #4188 commented on May 30, 2025 • 0 new comments

• 训练中途突然报错 NCCL watchdog thread terminated with exception

  #1817 commented on May 30, 2025 • 0 new comments

• 训练的时候总提示： RuntimeError: CUDA driver error: invalid argument

  #4103 commented on May 30, 2025 • 0 new comments

• qwen2.5vl 多图微调 模型退化

  #3617 commented on May 30, 2025 • 0 new comments

• 多卡微调报错该如何解决？

  #2776 commented on Jun 2, 2025 • 0 new comments

• internvl2_5微调出现几个问题，希望能帮一下忙

  #3319 commented on Jun 2, 2025 • 0 new comments

• 支持Qwen3 MoE的Megatron LoRA训练

  #4126 commented on Jun 3, 2025 • 0 new comments

• Qwen2-audio-instruct用lora微调后inference，出现tensor维度不对应的问题

  #4128 commented on Jun 4, 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 Jun 4, 2025 • 0 new comments

• 请教怎么使用swift infer

  #3936 commented on Jun 4, 2025 • 0 new comments

• 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 commented on Jun 4, 2025 • 0 new comments

• 无法单服务器多卡训练

  #4334 commented on Jun 4, 2025 • 0 new comments

• VLLM 运行GLM-4-32B-0414的推理需要的配置？

  #3919 commented on Jun 4, 2025 • 0 new comments

• 强化微调脚本出现AttributeError: 'NoneType' object has no attribute 'infer'

  #4201 commented on Jun 4, 2025 • 0 new comments

• AttributeError: 'Tab' object has no attribute 'constructor_args'

  #2985 commented on Jun 4, 2025 • 0 new comments

• ms-swift 3.0.0.dev0 version is incompatible with gradio 5.9.1 version

  #2782 commented on Jun 4, 2025 • 0 new comments

• 请问有支持步骤奖励的rl方法么

  #4024 commented on Jun 4, 2025 • 0 new comments

• RuntimeError: Event loop stopped before Future completed causes 502 errors; background logs continue running

  #3839 commented on Jun 5, 2025 • 0 new comments

• swift微调模型有没有建议的图像数据增强方式如动态分辨率，随机旋转等

  #3174 commented on Jun 5, 2025 • 0 new comments

• 有无懂哥说说internvl3_8B微调完后怎么做awq量化呀

  #4115 commented on Jun 6, 2025 • 0 new comments

• qwen3 14b zero3_offload oom

  #4177 commented on Jun 6, 2025 • 0 new comments

• LISA训练要么OOM，要么使用deepseed就报错

  #2035 commented on Jun 6, 2025 • 0 new comments

• Qwen2.5vl 7b全参数训练显存异常

  #3504 commented on Jun 9, 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 Jun 9, 2025 • 0 new comments

• NPU训练qwen2.5-vl报错

  #3408 commented on Jun 9, 2025 • 0 new comments

• DPO微调报错，老是出现Storage size calculation overflowed with sizes。

  #2538 commented on Jun 10, 2025 • 0 new comments

• ms-swift使用vllm作为后端推理qwen2.5-omni-7b时报错

  #4210 commented on Jun 10, 2025 • 0 new comments

• 奇怪的out of memory报错

  #3964 commented on Jun 11, 2025 • 0 new comments

• 纯文本数据+添加 special token，全参数微调训练Qwen2VL，模型不收敛

  #3804 commented on Jun 13, 2025 • 0 new comments

• grpo每次eval结束后就卡住，然后超时训练中断

  #4355 commented on Jun 14, 2025 • 0 new comments

• 端口监听错误

  #3988 commented on Jun 15, 2025 • 0 new comments

• Streaming + Packing + resume_from_checkpoint时出现报错

  #4083 commented on Jun 15, 2025 • 0 new comments

• gme 7b在H20上lora微调时出现 Out Of Memory

  #4361 commented on Jun 16, 2025 • 0 new comments

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

  #3690 commented on Jun 17, 2025 • 0 new comments

• lora微调占用显存**逐渐增大**直到**爆炸**

  #2364 commented on Jun 17, 2025 • 0 new comments

• grounding数据集格式，多类别+多box怎么写

  #3732 commented on Jun 18, 2025 • 0 new comments

• DPO微调多模态qwen2.5-7B，在图片处理时报错，Caught ValueError in DataLoader worker process 0与cannot reshape array of size 1843200 into shape (1,2,3,17,2,14,22,2,14)

  #4181 commented on Jun 18, 2025 • 0 new comments

• 训练过程中卡死，进程处于睡眠状态，GPU利用率为0

  #3290 commented on Jun 19, 2025 • 0 new comments

• 怎样为Qwen2.5-VL的视觉和文本设置不同的 lora rank？

  #4223 commented on Jun 20, 2025 • 0 new comments

• 请问会支持qwen2.5中hermes的function call的训练方式吗？

  #3523 commented on Jun 20, 2025 • 0 new comments

• pretrain报错进度异常问题

  #2692 commented on Jun 20, 2025 • 0 new comments

• Fatal Python error: none_dealloc: deallocating None

  #4353 commented on Jun 23, 2025 • 0 new comments

• 能否支持MiniCPM-o 2.6 audio模态训练

  #2961 commented on Jun 23, 2025 • 0 new comments

• 可以在moe的模型训练中 增加专家并行的参数吗

  #1631 commented on Jun 24, 2025 • 0 new comments

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

  #3494 commented on Jun 25, 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 Jun 25, 2025 • 0 new comments

• 🚀 Best Practices for Training Qwen3/Qwen3-MoE

  #4030 commented on Jun 25, 2025 • 0 new comments

• 支持GME微调么

  #3019 commented on Jun 25, 2025 • 0 new comments

• 训练后的RM模型，支持推理引擎sglang/vllm部署

  #3610 commented on Jun 26, 2025 • 0 new comments

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

  #4092 commented on Jun 27, 2025 • 0 new comments