Open X9VoiD opened 1 day ago
Name | Link |
---|---|
Latest commit | d1551627b8ca846a319583ea6885ab40edf67311 |
Latest deploy log | https://app.netlify.com/sites/continuedev/deploys/673f78bbc178620008e828eb |
Binding the http requests to the IDE cancellation token was the 1st approach I used in #2935. But as @sestinj requested, and as https://github.com/continuedev/continue/pull/2935#issuecomment-2488927707 explains, the abort signal has to be controlled by the generator, so you can still reuse pending streaming if the prefix is compatible with previous completion requests, as to minimize requests. This is irrelevant to response caching, that you observed.
Description
2935 had some
AbortController
s with no way to callabort()
on it. This PR fixes that and only adds AbortSignal where it makes sense. For future proofing, AbortSignal is also passed down tocore/llm/stream.ts
functions.The abort error that shows up in debug console is also removed.
Testing
ollama serve
. No requests are left open. Cached completions still work. No regressions are detected so far with manual testing.