I’m testing streamed responses from the DeepSeek API and want to understand what happens when my application cancels a request partway through. Does the server stop generating promptly, or does closing the client connection only stop the output from reaching me?

I’m also trying to work out how cancelled API generation affects usage: are tokens produced before cancellation counted, and is there any delay before generation actually stops? Real-world experience would be helpful, especially for apps where users often interrupt a response.