ms-swift · Issues· 592 open
Open on GitHubLocally synced open issues (discussions stay on GitHub)
- #9051
A800通信问题。两张卡，一张rollout启动vllm，一张训练，连接不到vllm，是不是通信问题？
questionstaleUpdated Sep 18, 2026 - #10189
[Bug] Qwen3.5 MoE multimodal DPO drops output_router_logits and crashes with aux_loss=None
Updated Sep 17, 2026 - #9123
2 nodes sft Qwen3.5-35B-A3B error FlashAttnVarlenFunc.apply ValueError: basic_string::_M_create
bugstaleUpdated Sep 17, 2026 - #9593
如何重构dataset的默认getitem函数
questionUpdated Sep 16, 2026 - #10057
About qwen3.5 mtp sft, multi-modality data
questionUpdated Sep 16, 2026 - #10156
[Bug] Negative JSD loss in Megatron GKD at beta=0.5; FP32 gives a positive result on the same inputs
bugUpdated Sep 16, 2026 - #9064
gemma-4图像怎么设置最大token数量
questionstaleUpdated Sep 16, 2026 - #9357
[bug] Qwen2.5-VL vllm infer video failed on transformers v5
bugstaleUpdated Sep 16, 2026 - #10108
优化 npu_swiglu_forward：拼接投影权重而不是拼接激活以降低内存并提升 NPU 性能
questionUpdated Sep 15, 2026 - #10124
NPU环境微调qwen3-embedding的正确方式
questionUpdated Sep 15, 2026 - #9390
在Ascend 910B上swift sft训练Qwen3.5 9B时出现accelerate错误
bugstaleUpdated Sep 15, 2026 - #9561
How to train Qwen3 VL embedding by Python?
questionstaleUpdated Sep 15, 2026 - #9576
请问seq_cls分类任务lora训练如何加速推理
questionstaleUpdated Sep 15, 2026 - #10117
[RLHF] DAPO/CISPO/FIPO 的 loss 在梯度累积下被额外除以 gradient_accumulation_steps
Updated Sep 14, 2026 - #9556
How to run Qwen3.5 9B GRPO GSM8K
questionstaleUpdated Sep 14, 2026