Search results

使用全参sft后的qwen3vl-8b模型继续做grpo时（使用了padding_free），遇到RuntimeError: split_with_sizes expects split_sizes to sum exactly to 1 (input tensor's size at dimension 0), but got split_sizes=[11, 615, 615, 615, 615, 615, 615, 615, 448, 448, 448, 448, 448, 448, 448, 448]

Status: Open.

#6921 In modelscope/ms-swift;

· Dswn888 opened on Dec 6, 2025

Qwen3VL-30A3多图训练参数求助

Status: Open.

#6920 In modelscope/ms-swift;

· jiangsongtao opened on Dec 6, 2025

如何查看eval loss

Status: Open.

#6917 In modelscope/ms-swift;

· microchila opened on Dec 5, 2025

swift 3.11.0.dev0 seq_cls分类任务推理时输出结果异常，部分输出单个类别，部分输出空类别，部分输出2个类别

bugSomething isn't workingSomething isn't working

Status: Open.

#6916 In modelscope/ms-swift;

· songzihao010 opened on Dec 5, 2025

swift有没有一种方式可以读取已经进行过分词的缓存，如arrow

Status: Open.

#6915 In modelscope/ms-swift;

· yaojianchao5-droid opened on Dec 5, 2025

Template <code>glm4_1v</code> does not support padding free or packing

enhancementNew feature or requestNew feature or request

Status: Open.

#6914 In modelscope/ms-swift;

· cxfzzj opened on Dec 5, 2025

多阶段GRPO/GSPO，二阶段训练及KL不正常

Status: Open.

#6910 In modelscope/ms-swift;

· Trent888 opened on Dec 5, 2025

dynamic-reward-weighting

enhancementNew feature or requestNew feature or request

Status: Open.

#6909 In modelscope/ms-swift;

· xdaiycl opened on Dec 5, 2025

我的训练在未完成的时候中断了，是因为wandb，验证失败导致的，能在这个地方加一个容错吗

Status: Open.

#6902 In modelscope/ms-swift;

· TAOSHss opened on Dec 5, 2025

使用 ms-swift 对 Qwen2.5-VL-3B-Instruct 做 GUI-R1 风格 SFT 后，坐标点击任务推理准确率极低

Status: Open.

#6900 In modelscope/ms-swift;

· zhonganbi102-netizen opened on Dec 4, 2025

STF Qwen3系列模型的think问题

Status: Open.

#6894 In modelscope/ms-swift;

· Mrzjk opened on Dec 4, 2025

GRPO训练出现warning信息：Rollout logprobs count does not match completion tokens count

Status: Open.

#6890 In modelscope/ms-swift;

· yingzhao27 opened on Dec 4, 2025