Search results

建议引入Training-Free GRPO这个免参数更新、无需训练的强化学习算法

Status: Open.

#6396 In modelscope/ms-swift;

· QiZishi opened on Nov 1, 2025

On-Policy Call to collective_rpc method failed: 'Worker' object has no attribute 'pynccl_comm'

Status: Open.

#6394 In modelscope/ms-swift;

· airlsyn opened on Nov 1, 2025

CUDA OUT OF MEMORY!!!!!!!!!!!!!!!!!! WHEN GRPO..... HELP!!!

Status: Open.

#6393 In modelscope/ms-swift;

· ragelgq opened on Nov 1, 2025

<code>save_total_limit</code> does not work when <code>save_steps</code> not equals to <code>eval_steps</code>

Status: Open.

#6392 In modelscope/ms-swift;

· auto-Dog opened on Nov 1, 2025

使用自定义多模态标准messages格式数据 可以正常训练推理，评测无法跑起来

Status: Open.

#6391 In modelscope/ms-swift;

· wengchunlei opened on Nov 1, 2025

Does flash attention backend work with B200? (Qwen3-VL example super slow on B200)

Status: Open.

#6390 In modelscope/ms-swift;

· mertunsall opened on Oct 31, 2025

不同节点间rollout和grpo同时进行时的报错求解决

Status: Open.

#6389 In modelscope/ms-swift;

· ragelgq opened on Oct 31, 2025

k2模型如何多机多卡转mcore

Status: Open.

#6388 In modelscope/ms-swift;

· ray075hl opened on Oct 31, 2025

Qwen3-Next-80B-A3B-Instruct 8卡H20，lora微调报错OOM Failed to CUDA calloc async 136 bytes

Status: Open.

#6387 In modelscope/ms-swift;

· xiaotianns opened on Oct 31, 2025

[Bug] Freeze in 2-node GRPO training with rollout server

Status: Open.

#6386 In modelscope/ms-swift;

· StupidYoshiaki opened on Oct 31, 2025

vlm训练二阶段全参sft/dpo加载checkpoint会报错

Status: Open.

#6384 In modelscope/ms-swift;

· Yimi81 opened on Oct 31, 2025

megatron 模型保存报错

Status: Open.

#6381 In modelscope/ms-swift;

· RonanDou opened on Oct 31, 2025