Fix MLflowTracker.store_init_configuration mutating the caller's config dict

[ATOM00blue]

What does this PR do?

MLflowTracker.store_init_configuration drops config values that exceed MLflow's
MAX_PARAM_VAL_LENGTH, but it did so by deleting them from the input dict in place:

for name, value in list(values.items()):
    if len(str(value)) > mlflow.utils.validation.MAX_PARAM_VAL_LENGTH:
        logger.warning_once(...)
        del values[name]   # mutates the caller's dict

Accelerator.init_trackers passes the same config object to every tracker:

if config is not None:
    for tracker in self.trackers:
        tracker.store_init_configuration(config)

So when MLflow drops an over-long value, that key is permanently removed from the
user's config dict and is therefore missing from every other tracker (TensorBoard,
WandB, etc.) that runs after MLflow, as well as from any later use of the dict by the
user. This is silent, order-dependent data loss.

Fix

Build a filtered list of in-range items instead of mutating the input, which is the
same approach already used in MLflowTracker.log (it builds a separate metrics dict).
The warning and the dropping of over-long values from what gets logged to MLflow are
both preserved; only the in-place mutation of the caller's dict is removed.

Reproduction (before this PR)

import tempfile, mlflow
from accelerate.tracking import MLflowTracker

tracker = MLflowTracker(experiment_name="repro", logging_dir=tempfile.mkdtemp())
tracker.start()
config = {"learning_rate": 0.001, "long_note": "x" * (mlflow.utils.validation.MAX_PARAM_VAL_LENGTH + 1)}
tracker.store_init_configuration(config)
print(config)  # -> {'learning_rate': 0.001}  ("long_note" was deleted from the caller's dict)

After this PR, config still contains long_note; only the value sent to MLflow is dropped.

Before submitting

• Did you write any new necessary tests?

A regression test (MLflowTrackingTest::test_store_init_configuration_does_not_mutate_input)
is added; it fails on main and passes with this change.

Who can review?

@BenjaminBossan @SunMarc

---

Developed with AI coding assistance; reviewed and submitted by the author.

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Adds regression coverage and adjusts MLflow init-config logging to avoid mutating caller-provided configuration dictionaries.

Changes:

• Update MLflowTracker.store_init_configuration() to filter over-long values without deleting keys from the input dict.
• Add a regression test asserting the caller’s dict is not mutated while ensuring MLflow still drops over-long params.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
tests/test_tracking.py	Adds regression test for non-mutating init configuration logging to MLflow.
src/accelerate/tracking.py	Avoids in-place deletion from values when filtering MLflow params.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.