Don't allow zero as temperature in ModelSettings.qml

[cosmic-snow]

Describe your changes

A number lower than 0.000001 is set to that value instead.

Issue ticket number and link

N/A

Checklist before requesting a review

• I have performed a self-review of my code.
• If it is a core feature, I have added thorough tests.
• I have added thorough documentation for my code.
• I have tagged PR with relevant project labels. I acknowledge that a PR without labels may be dismissed.
• If this PR addresses a bug, I have provided both a screenshot/video of the original bug and the working solution.

Demo

Editing:

After edit:

Steps to Reproduce

Right now, setting temperature to zero is allowed, but it shouldn't be.

Notes

This PR is meant to be seen as a suggestion for now.

I'm setting this as draft for the following reasons:

• Maybe this should be handled in the backend instead, and maybe 0 as temperature should be allowed again.
• I think it prevents setting zero as a value, but it seems to be flaky in the UI regardless (e.g. when I Alt-Tab).
• I have not figured out how to get rid of the scientific notation after it corrects itself.

[cosmic-snow]

Second commit improves points 2 & 3 from the notes.

The improved behaviour uses .toLocaleString(), although it's probably the wrong locale. I think it should be the one from the selected translation, once those are ready.

Edit: Parsing might need some changes with different locales, too.

Edit 2: Turns out you're using QLocale::setDefault() in the translation logic and QML's Qt.locale() does the right thing with that.

[cosmic-snow]

Should be good now, apart from a minor cosmetic glitch (rare corner case).

I tested it by adding to:

gpt4all/gpt4all-chat/main.cpp

Lines 53 to 54 in 6b8e0f7

	MySettings::globalInstance()->setLanguageAndLocale();

on line 54:

QLocale::setDefault(QLocale("de_DE"));

Try to break it, hope these should not be possible anymore:

• No temperature of 0 or below.
• Only commas are allowed as decimal separator with de_DE as locale.

Although note: this is not hooked up to translation dropdown/locale changes. That would be the next thing to work on for all input fields which need to be locale-aware.

[cebtenzzre]

I am hesitant to disable functionality of built-in and custom OpenAI models, and to add llama.cpp-specific assumptions to the UI, just to avoid branching to the built-in llama_sample_token_greedy when the temperature is zero here. I think this problem can actually be properly fixed in fewer lines of code than this workaround.

[cosmic-snow]

@cebtenzzre Just to be clear, part of this is also making the UI input locale-aware (decimal separator can vary).

The primary goal here was to fix a bug: there's a division by zero, which is getting hit right now in the backend (oddly seems to only throw off debug builds, but that's my builds). Greedy sampling with temp 0 was once possible in GPT4All (sometime last year), I don't know when exactly that changed.

I am not at all opposed to removing the limit again if/once there's a fix for that in the backend -- actually, I am very much in favour of being allowed to set it to 0.

[cebtenzzre]

@​cebtenzzre Just to be clear, part of this is also making the UI input locale-aware (decimal separator can vary).

That's appreciated. If you'd like to make a PR or change this one such that all four instances of parseFloat in ModelSettings.qml are replaced with something locale-aware, that'd be great.

The primary goal here was to fix a bug: there's a division by zero, which is getting hit right now in the backend (oddly seems to only throw off debug builds, but that's my builds). Greedy sampling with temp 0 was once possible in GPT4All (sometime last year), I don't know when exactly that changed.

But only llama.cpp divides by zero, because of a missing branch. We are actively working (#2806) towards having first-class support for backends other than llamamodel.cpp, so I feel like adding more llamamodel-specific assumptions to the UI code is a step in the wrong direction. Both ollama and OpenAI support temp=0.

I am not at all opposed to removing the limit again if/once there's a fix for that in the backend -- actually, I am very much in favour of being allowed to set it to 0.

Okay, done: #2854

[cosmic-snow]

That's appreciated. If you'd like to make a PR or change this one such that all four instances of parseFloat in ModelSettings.qml are replaced with something locale-aware, that'd be great.

Yes, that was the idea after this one. Although there may be one other thing to consider: In this PR (and probably the next) I had to remove the direct connection in the QML (text: root.currentModelInfo.temperature) because of the parsing and conversion.

An alternative would be to have the model be "more closely aligned" with the UI and do the conversions in C++ instead of JS.

In any case, with merging of #2854 this PR is now obsolete.

[cebtenzzre]

Although there may be one other thing to consider: In this PR (and probably the next) I had to remove the direct connection in the QML (text: root.currentModelInfo.temperature) because of the parsing and conversion.

That shouldn't be a problem, since the binding is removed once onTemperatureChanged or onCurrentModelInfoChanged fires anyway. As long as the value is never initially blank—in which case, you may need a Component.onCompleted handler.