Repository navigation
[Bug]: Frequent "Setting changed. Reload the model to continue." even after a simple "hi" prompt #54
Description
TUFF version
9.4.0
How did you install TUFF?
Release download or in-app update
Area
Chat and generation
Mac, chip and memory
Mac17,3, Apple M5, 16 GB
macOS version
26.7.1
Model
Not model-specific
Steps to reproduce
Send any prompt, even one as simple as "hi".
If it does not happen on the first one, it happens after a few.
No settings were changed in the meantime.
What you expected
Not the "setting changed" prompt, as nothing is changed by the user.
What happened instead
Frequent "Setting changed. Reload the model to continue."
could also be: "Model setting changed. Reload the model before generating." with no way to reload the model.
In this case, the "Current model" in the bottom left shows "Failed".
After unloading and (re)loading the model, Local Model still shows Failed.
Diagnostic summary (optional)
TUFF 9.4.0
macOS: 26.7.1
Mac: Mac17,3
Chip: Apple M5
Unified memory: 17179869184 bytes
Model: gemma4-heretic
Context: 16384 tokens
Sampling: temperature 0.2, top-k 64, top-p 0.95
Expert cache: 16 slots (lfu)
Prefill: on, chunk 512
Image residency: on-demand
Logs or crash report (optional)
Before submitting
- I removed credentials, private paths, personal prompts and chat content from this report.
Activity
arker2 commentedon Oct 16, 2026
Not sure if it is the same bug:
Osiriss664 commentedon Oct 16, 2026
@arker2 I think I experienced this too. Both things you describe
Osiriss664 commentedon Oct 16, 2026
@rexmhall09
I can reproduce the cause on 9.3.0 and 9.4.0 (Apple M5, 16 GB, Qwen 3.6 35B), without the GUI, with a small test that feeds the Auto planner one answer record per round, as
recordAutoPerformancedoes after each answer:The reason seems to be
InferencePlanSelector.select: any candidate with an observation under memory pressure or swap growth is dropped (for 15 minutes,AppAutoPerformanceRecords.observations). After each answerfinishTerminalRuncallsapplySettingsProfile, the new plan differs from the loaded one,hasStaleLoadedRuntimebecomes true and the "Settings changed. Reload the model to continue." banner shows, although the user changed nothing. A single pressured observation is enough, and swap growth alone counts too. Swap growth is measured system-wide (swapWrittenBytes), so any app swapping during an answer has the same effect. On a 16 GB Mac with a large model this happens often, so it repeats. If every candidate ends up dropped, the selector throwsinsufficientBudget, which could explain the "Failed" state (I did not reach that point in 12 rounds).Possible fix ideas: don't let the re-plan after an answer change the loaded settings (keep them until the next load or new chat), drop a plan only after 2-3 pressured observations or on critical pressure, and count only TUFF's own paging rather than system-wide swap.
I used an Auto analyses using OPUS 5.5 and Fable 1.1
Here I am not deep enough into this topic
Keep the loaded plan after answers; serve each model's memory setting…
rexmhall09 commentedon Oct 17, 2026
Hey, this is fixed in 9.4.1. Auto was re-planning after every answer and changing the settings of the model that was already loaded, which is what made it ask for a reload. Now it keeps the loaded settings and uses what it learned the next time you load the model. Small amounts of system-wide paging no longer count against a plan, and it can't end up with no plan at all anymore (that was the "Failed" state). Thanks for reporting, and thanks @Osiriss664 for finding the cause!