Make Crush more "verbose"? #3596
Replies: 2 comments 1 reply
|
Crush already logs the detail you are after — it just does not put it in the TUI. It goes to a file. Turn it on permanentlyYou asked for "by default", so put it in your # crushrc
option debug true
option debug-lsp trueDrop Equivalent one-off: Watch it while Crush runsLogs land in crush logs --followThat is the setup that answers "what is it actually doing right now" while the TUI sits on Writing…. Also available: crush logs # last 1000 lines
crush logs --tail 500 # last 500Short forms If you ever drive it non-interactively
crush run -v "your prompt"That one is specific to Worth knowing given the local-model angle: with |
|
That log actually tells us something useful: debug is not on yet. Every line you pasted is level := slog.LevelInfo
if <debug> {
level = slog.LevelDebug
}With debug active you would see Confirm it in ten secondsSkip the config file and pass the flag: crush -dIf Where the file has to beA global Quickest way to sidestep path questions on Windows: put The JSON equivalent, if you prefer, in {
"options": {
"debug": true
}
}Do not keep both a What you actually get from itThis is the part that answers your question. if slog.Default().Enabled(req.Context(), slog.LevelDebug) {so once DEBUG is live you get the full HTTP exchange with llama.cpp, bodies included. That is where those 17-minute gaps become readable: whether tokens are trickling back, whether it is parked on a tool call, or whether a request never got a response. The Separately, from your logInstalling Caveat: I checked this against |

Uh oh!
There was an error while loading. Please reload this page.
Basically, I finally managed to set up Qwen 3.6 35B-A3B with a reasonable speed - but while crush runs, it just shows "Writing ..." in the logs - and I just want to see what it is actually doing, since local speeds aren't super fast. I do see activity via the monitoring, obviously, but I would like to see the details happening in Crush.
How do I, by default, make it verbose?
Thanks!
(btw, using this here: https://github.com/alekk89/llama-cpp-windows-manager - very good pairing with model auto discovery.)
All reactions