fix: replace AbortSignal.any with manual polyfill for WebKit compatibility#7552
Conversation
|
Good stuff. What MacOS version are you using? @since-2017-hub. Want to reproduce for testing. |
I am using 26(Tahoe) |
|
@nhhai please help take a look at this PR, give it a test then we are good to merge. |
|
Hi @since-2017-hub, We tried to reproduce this first to verify but none of our machine can reproduce this. If it happens on your machine, can you help verify it with this build? It's not signed so you may need to go to Security & Privacy > Open Anyway to launch it. Thank you, |
|
@nhhai |
|
@nhhai this is the PR to fix this issue. I think you might want to check the build instead. As contributors can build from source. |
|
I gave this fix a test, and there is a performance degradation issue. The token speed drop from 44 > 17 after stopping the first inference. Which means the abort signal didn't work somehow, hence the aborted stream still running in the background. We will need to look into it. |
This is a legacy issue, it seems. I will check on the main branch after it is merged. |
Describe Your Changes
AboartSignal.any()is not supported in the WebKit/JavaScriptCore runtime that Tauri uses on macOS, causing all llama.cpp and MLX model loads to immediately fail with:TypeError: AbortSignal.any is not a function-The fix creates a combined
AbortControllerthat manually forwards both timeout and user-abort signals, preserving identical behavior across all platformsFixes Issues
Self Checklist