[Question] [AgentScope] Qwen2___5-7B-Instruct 比 Qwen2 -7B性能下降了不少啊? #965
Replies: 4 comments 1 reply
-
How did you test it? Which agent framework did you use? |
Beta Was this translation helpful? Give feedback.
-
请问您采用的function call是用的什么实现?qwen-agent框架提供的,还是vllm提供的? |
Beta Was this translation helpful? Give feedback.
-
qwen2.5 7B 两种方案都可以实现,qwen-agent开启了并发功能后,结果与openAI格式的一样的。
|
Beta Was this translation helpful? Give feedback.
-
据我所知,agentscope是用的非qwen官方的prompt,qwen2.5模型没有针对agentscope优化。 |
Beta Was this translation helpful? Give feedback.
-
Has this been raised before?
Description
最近用 Qwen2___5-7B-Instruct 进行了性能测试还有Function call功能测试!
发现qwen2.5 7B退步了不少啊,性能下降。Function call功能也比之前弱了。
环境:vllm6.0 部署,硬件是 一张L20 48G
Agent并发 10 用户,prompt 150汉字,输出约1000汉字。 流式输出首包平均时长0.9s。 qwen2(0.32s)
Agent并发 20 用户,prompt 150汉字,输出约1000汉字。 流式输出首包平均时长1.8s。 qwen2(0.45s)
function call 并发多函数顺序调用的时候很容易出现乱填参数的问题。
Beta Was this translation helpful? Give feedback.
All reactions