Phase 2 - Input tooltips, autocorrect, synced run names, default names by O-J1 · Pull Request #1300 · Nerogar/OneTrainer

O-J1 · 2026-02-09T12:55:23Z

Input validation tooltips
Reworked model outputs section:
- 3 options, default, friendly and custom
- Better support for EU, Middle Eastern and Asian numerical practices/chars (safe type coercions)
- Auto-correction for basic things like floats, ints and path normalisation
Run Name is synchronised to saves, backups and tensorboard as a prefix.
Debounced workspace dir. Closes [Bug]: Always-On Tensorboard create unwanted workspace directories #1028
Fixed backup resum bug with LoHA. Closes [Bug]: Restart of Z-Image Base LoHA after stopping fails with Rank Mismatch. #1340

When autocorrect is not turned on we surface a tooltip:

…nts training if there are validation errors

…entry and dir_entry

for more information, see https://pre-commit.ci

…loat to be positive, delete dead code

…into input-validation

…nput-tooltips-and-co + minor NFKC full width char fix

…ainer into input-tooltips-and-co

O-J1 · 2026-04-26T08:44:59Z

@dxqb Would you be able to review this now?

Co-authored-by: Copilot <copilot@github.com>

dxqb · 2026-05-10T08:30:32Z

@dxqb Would you be able to review this now?

finally did
[adding this because our github bot doesn't post reviews)

Co-authored-by: Copilot <copilot@github.com>

O-J1 · 2026-05-11T08:42:40Z

Addressed review comments as best I could

dxqb · 2026-05-14T10:08:13Z

👀

O-J1 · 2026-05-31T09:48:02Z

@dxqb Merged updates and addressed some things + fixed the lokr regression.

You stated this "I think there is an easier way. Will follow up" last month. Everything else has been resolved aside from the name discussion

BitcrushedHeart · 2026-06-07T09:40:57Z

I'd like to ask that auto-correct doesn't kick in until the user clicks off the input box - currently if you are a really slow typer, or if you're remoting into your PC and adjusting the settings via a remote delay, you could be halfway through typing something and then autocorrect kicks you off. A better method considering UX is to only auto-correct after the user exits the input field.

I'd also agree with dxqb - os.mkdir is a better solution than 'Actually, can you open your file explorer, find the destination folder that you want to use, right click, click 'Create New Folder', open it, copy that path, and then copy and paste it into this textbox?'

Honestly I'm iffy in general on autocorrect at all - feels like fixing a problem that doesn't exist, I'm much more a fan of just... throw an exception.

O-J1 · 2026-06-07T09:56:41Z

Honestly I'm iffy in general on autocorrect at all - feels like fixing a problem that doesn't exist, I'm much more a fan of just... throw an exception.

I dont ever want to have to deal with users being stupid in the common pitfall fields again if I can help it+ it helps people from different cultures (arabic, jp and eu) still use their specific decimal separators too.

Ive already pared it back massively to only obvious things.

O-J1 · 2026-06-07T11:13:04Z

@dxqb @BitcrushedHeart PathIOType.OUTPUT fields now mkdir

dxqb · 2026-06-15T20:14:48Z

        if self.peft_type == PeftType.OFT_2:
            return

+        rank_key = next((k for k in state_dict if k.endswith(".lora_down.weight")), None)


I don't understand this new check. there already is a check below this new inserted code. it already finds the rank_key based on the peft type.
what does this do what the existing check doesn't also do? Please have a look at this code in context before you reply. on it's own, it might make sense, but not with the following section

dxqb · 2026-06-15T20:19:45Z

+        abs_path = os.path.abspath(trimmed)
+        output_dir = abs_path if output_is_dir else os.path.dirname(abs_path)
+        try:
+            os.makedirs(output_dir, exist_ok=True)


thanks, but validation.py shouldn't create directories as a side effect:

there is no need for this I think. all directories that need to be created, are already created elsewhere:

workspace_dir: GenericTrainer.init (modules/trainer/GenericTrainer.py:74-75) — os.makedirs(os.path.join(config.workspace_dir, "tensorboard"), exist_ok=True), which recursively creates workspace_dir too.

cache_dir: GenericTrainer.start() (modules/trainer/GenericTrainer.py:134) — os.makedirs(self.config.quantization.cache_dir, exist_ok=True) where quantization.cache_dir defaults to cache_dir + "/quantization".

final_output_dir: created by the model savers right before writing (modules/modelSaver/mixin/LoRASaverMixin.py:51,67,75, EmbeddingSaverMixin.py:32).

debug_dir: created via _save_image/_save_text in ModelSetupDebugMixin when a debug artifact is actually written.

dxqb · 2026-06-15T20:22:15Z

+            final_output_dir = parent if parent != "." else "./models"
+        else:
+            run_name = ""
+            final_output_dir = "./models"


minor but looks weird to me. why not just "models"?

dxqb · 2026-06-15T20:29:52Z

+            "Increase your total epochs to continue."
+        )
+
+    def get_last_backup_path(


I don't think we should touch the backup-finding logic at all. Why should you even have incompatible backups in the same workspace? What problem does this solve?

If you disagree, please submit it as a separate PR. It is incredible difficult to review a PR that does so many completely unrelated things.

dxqb · 2026-06-15T20:32:07Z

+            last_backup_path = self.config.get_last_backup_path(require_compatible=True)
+            assert last_backup_path is not None
+
+            if self.config.training_method == TrainingMethod.LORA:


LOHA regression? if you mean this, I think that was a user error: #1340 (comment)

dxqb · 2026-06-15T20:48:43Z

I made minor changes to be able to start it, but don't have write access to this branch. You'll find them here: O-J1#2
If you agree with the changes, click merge and it'll end up in this PR branch.

There are 2 unaddressed comments from an earlier review and I have added a few new ones.
Top-level overview,

I like the entire run name logic
I like the new "error tooltips" that tell the user what he did wrong
I think autocorrect is unnecessary and you might have only had this idea, because there were no "error tooltips" yet, so correcting automatically was the next best thing. However, if you disagree and this has limited scope now (basically, typing errors), I'll merge this.
I don't see a reason to touch the backup discovery logic at all, and don't understand what problem this solves. If you disagree, please submit this as a separate PR and let's have a closer look, but this can just subtly break too many things for a PR that should be about user interface

O-J1 and others added 11 commits February 8, 2026 03:43

Adds basic validation, undo, redo, debouncing, error display and prve…

996c462

…nts training if there are validation errors

Adds path_entry validation, prevent overwrite toggle and remove file_…

6ba5c9a

…entry and dir_entry

Merge branch 'Nerogar:master' into input-validation

adbec55

[pre-commit.ci] auto fixes from pre-commit.com hooks

9052e8e

for more information, see https://pre-commit.ci

Adds extra validation behaviour (range, required), required int and f…

cf18297

…loat to be positive, delete dead code

Merge branch 'input-validation' of https://github.com/O-J1/OneTrainer …

be0f2db

…into input-validation

Delete dead code found in TrainUI.py

0650422

Make commits require validation to pass in order to write to trainConfig

18bfe6f

Merge branch 'Nerogar:master' into input-validation

ef4a81f

Adds tooltips, autocorrection and run name functionality

12781f3

Tweak autocorrect, bump up debounce to 500ms

f1b7a4b

O-J1 added the waiting waiting for something / cannot fix currently label Feb 9, 2026

O-J1 added 4 commits February 13, 2026 00:55

Merge branch 'Nerogar:master' into input-tooltips-and-co

678df2f

Merge branch 'master' into input-tooltips-and-co

da6b032

Merge branch 'master' of https://github.com/Nerogar/OneTrainer into i…

5219964

…nput-tooltips-and-co + minor NFKC full width char fix

minor comment change

2757f77

O-J1 removed the waiting waiting for something / cannot fix currently label Feb 22, 2026

O-J1 marked this pull request as ready for review February 22, 2026 13:05

O-J1 added 2 commits February 23, 2026 00:17

Remove redundant migrations

e97bdcd

Code simplification

2959b58

maedtb reviewed Feb 25, 2026

View reviewed changes

O-J1 added 9 commits March 4, 2026 00:01

Remove treating learning rates specially

15fae60

Fix typo in config_version

2a47ba9

Allow certain float fields to be negative

9ca9806

Add st_ctime and mtime fallback for backup datetime extraction

ab4b687

Adjust constant name, safer friendly run names

832c579

Merge branch 'Nerogar:master' into input-tooltips-and-co

a6f82f1

Merge branch 'Nerogar:master' into input-tooltips-and-co

f05d54d

Address final concern validate_path issue

14a4711

Merge branch 'input-tooltips-and-co' of https://github.com/O-J1/OneTr…

3cc72eb

…ainer into input-tooltips-and-co

O-J1 added 2 commits April 7, 2026 05:43

Split model output destination

66c0aa9

Tweak tooltip dead code, modify var names, fix run name val regression

b29fee2

O-J1 marked this pull request as ready for review April 7, 2026 10:47

O-J1 and others added 2 commits April 26, 2026 19:29

Fix workspace/cache regression. Closes Nerogar#1395

d0706aa

Fix backup regression and fix loha bug

c701669

Co-authored-by: Copilot <copilot@github.com>

O-J1 requested a review from dxqb May 3, 2026 19:16

dxqb requested changes May 10, 2026

View reviewed changes

Address review comments

6c78da8

Co-authored-by: Copilot <copilot@github.com>

dxqb linked an issue May 11, 2026 that may be closed by this pull request

[Bug]: Workspace and cache directory validation are set to IO.TYPE input #1395

Open

dxqb changed the base branch from master to merge May 17, 2026 10:55

dxqb changed the base branch from merge to master May 17, 2026 10:56

O-J1 linked an issue May 31, 2026 that may be closed by this pull request

Regression: re-add non-negative UI validation for int fields (removed for LoKr) #1484

Open

O-J1 added 2 commits May 31, 2026 17:17

Merge branch 'master' into input-tooltips-and-co

42fbc73

Fixes remaining bugs I could find and newer comments.

2b35c15

This comment was marked as resolved.

Sign in to view

O-J1 added 2 commits June 7, 2026 21:08

Mkdir if PathIOType.OUTPUT

12ab2f3

Merge branch 'master' into input-tooltips-and-co

b2fa20d

dxqb requested changes Jun 15, 2026

View reviewed changes

This was referenced Jun 17, 2026

Orphaned TensorBoard subprocess can hang CLI/cloud training on training failure #1540

Open

[Bug]: Restart of Z-Image Base LoHA after stopping fails with Rank Mismatch. #1340

Closed

Disable LoRA rank check (false positive on LoHA backup/loading) #1549

Merged

Uh oh!

Conversation

O-J1 commented Feb 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

O-J1 commented Apr 26, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

dxqb commented May 10, 2026

Uh oh!

O-J1 commented May 11, 2026

Uh oh!

dxqb commented May 14, 2026

Uh oh!

O-J1 commented May 31, 2026

Uh oh!

This comment was marked as resolved.

Uh oh!

BitcrushedHeart commented Jun 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

O-J1 commented Jun 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

O-J1 commented Jun 7, 2026

Uh oh!

dxqb Jun 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

dxqb Jun 15, 2026

Choose a reason for hiding this comment

Uh oh!

dxqb Jun 15, 2026

Choose a reason for hiding this comment

Uh oh!

dxqb Jun 15, 2026

Choose a reason for hiding this comment

Uh oh!

dxqb Jun 15, 2026

Choose a reason for hiding this comment

Uh oh!

dxqb commented Jun 15, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

O-J1 commented Feb 9, 2026 •

edited

Loading

BitcrushedHeart commented Jun 7, 2026 •

edited

Loading

O-J1 commented Jun 7, 2026 •

edited

Loading

dxqb Jun 15, 2026 •

edited

Loading