Skip to content

QA: Make make run-mic easier to read and to target - #50652

Open
tarekziade wants to merge 11 commits into
mainfrom
mic-env-differs-warning
Open

tarekziade wants to merge 11 commits into
mainfrom
mic-env-differs-warning

Conversation

@tarekziade

@tarekziade tarekziade commented Nov 12, 2026 •

Copy link
Copy Markdown
Collaborator

🤗 Transformers CI

make run-mic output was hard to act on locally, improved it.

Also improved how to select models, not an exact directory and you can use + to select severals.
example: qwen+

last, pins to the newest baselines, that were recorded with torch 2.15

After the summary, list failing contracts whose package versions differ
from the ones their baseline was recorded with. They still count as
failures; the warning flags numeric deltas that may be version drift
rather than regressions.
Show a dict mismatch from the runner as one line per differing key instead
of two full dicts, list each mismatched package on its own line, and keep
only the target names in the end-of-run environment warning.
A heartbeat now redraws in place like the progress bar, so it no longer
stays above the first result, and blank lines separate the results from
the setup lines and the summary.
MODEL may now name a model the way it is spelled elsewhere (Wav2Vec2,
audio-flamingo-3): case and punctuation are ignored against directory
names, model types and config class names. A typo is not guessed; close
directory names are suggested instead. Before running, the equivalent
`tmic check` line is printed so it can be copied into a shell.
MODEL='qwen*' selects the contracts of every matching models/ directory.
The unknown-model error now explains both names and patterns.
A target with no reviewed baseline for this platform passes after it
loads and runs, without any output comparison. Name those targets after
the summary so a pass is not read as a checked result.
An unquoted * is expanded (or rejected) by the shell; + is not. MODEL=qwen+
now selects every qwen* model directory; * still works when quoted. Document
model names and patterns in the testing guide.
The MIC heartbeat printed "... 16s in" lines into the output while a
contract loaded. The checker window's spinner title now carries a running
timer (for every checker), so MIC prints nothing while it waits.
checkers.py sets CHECKERS_COLOR=1 for its children when it shows them in
a terminal (and NO_COLOR is unset); check_mic then colors failing result
lines and the failed/errored summary counts red. CI and piped output stay
plain. The checker window resets the color after truncating a line.
@tarekziade
tarekziade requested a review from ydshieh November 12, 2026 10:50
@HuggingFaceDocBuilderDev

Copy link
Copy Markdown

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

Pages changed in this PR:

@ydshieh ydshieh left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You are a better checkers than me, so I don't what else I could check but simply ✅

Thanks!

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants