vllm-project/vllm

[Feature]: Improve DCP error messages

Open

#28,407 创建于 2025年11月10日

在 GitHub 查看
 (4 评论) (2 反应) (0 负责人)Python (80,034 star) (16,816 fork)batch import
feature requestgood first issuestale

描述

🚀 The feature, motivation and pitch

Currently if a backend doesn't support DCP we get the following error message

AssertionError: DCP requires attention impls to return the softmax lse for decode, but the impl FlashInferImpl does not return the softmax lse for decode.

It would good to suggest to the user to try an alternative backend using VLLM_ATTENTION_BACKEND

Alternatives

No response

Additional context

No response

Before submitting a new issue...

  • Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

贡献者指南

技术栈
python
领域
backenddeveloper experience
议题类型
feature
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
2
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
under 1 hour
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
fresh
清晰度议题是否清楚说明期望改动、验收标准和下一步。
clear
前置要求
Pythonvllm framework familiarity
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
85
研究方向
Locate the file where the AssertionError about DCP attention is raised (likely in vllm/attention/backends/ or vllm/attention/selector.py). Modify the error message to include a suggestion to set the VLLM ATTENTION BACKEND environment variable to an alternative backend (e.g., 'FLASH ATTN' or 'FLASHINFER'). Ensure the change follows the project's error message conventions. Optionally, add a test to verify the new message appears when expected.