fix: LightGBM 4.0+ compatibility for early_stopping_rounds=None

[Olcmyk]

Description

This PR fixes a compatibility issue with LightGBM 4.0+ where passing early_stopping_rounds=None causes a TypeError.

Changes:

• Only create early_stopping callback when rounds is not None
• LightGBM 4.0+ requires stopping_rounds to be an integer, not None
• Apply fix to gbdt.py, ddgda.py, and highfreq_gdbt_model.py
• Follow the pattern already used in double_ensemble.py

Affected files:

• qlib/contrib/model/gbdt.py: Core LGBModel fix
• qlib/contrib/rolling/ddgda.py: Remove explicit None assignment
• qlib/contrib/model/highfreq_gdbt_model.py: Add robustness check

Motivation and Context

Fixes #2226

Problem: Starting from LightGBM 4.0, the lgb.early_stopping() function requires stopping_rounds to be an integer and no longer accepts None. This breaks qlib's code that attempts to disable early stopping by passing None.

Error:

TypeError: early_stopping_round should be an integer. Got 'NoneType'

Root cause:

• qlib/contrib/model/gbdt.py:71-73 directly passes early_stopping_rounds to lgb.early_stopping() without checking for None
• qlib/contrib/rolling/ddgda.py:244 explicitly sets early_stopping_rounds: None

Solution: Only create the early_stopping callback when the value is not None, following the pattern already correctly implemented in double_ensemble.py.

How Has This Been Tested?

• If you are adding a new feature, test on your own test scripts.

Testing environment:

• LightGBM: 4.6.0
• Python: 3.8
• OS: Linux (Ubuntu)

Test method:

1. Applied fix from PR fix(unpickler): allow Alpha158/Alpha360 handlers and the standard dataset chain #2213 (required to reach this code path)
2. Ran DDG-DA example:

   cd examples/benchmarks_dynamic/DDG-DA
   rm -rf mlruns
   python workflow.py run

3. Verified that the TypeError: early_stopping_round should be an integer no longer occurs
4. Confirmed training completes successfully (154/154 tasks)

Note: The standard pytest qlib/tests/test_all_pipeline.py does not cover this specific scenario as it requires:

• LightGBM 4.0+
• Explicit early_stopping_rounds=None usage
• The DDG-DA workflow

Screenshots of Test Results (if appropriate):

Before fix:

TypeError: early_stopping_round should be an integer. Got 'NoneType'
  File "qlib/contrib/model/gbdt.py", line 71, in fit
    early_stopping_callback = lgb.early_stopping(...)

After fix:

train tasks: 100%|████████████████████████████| 154/154 [05:23<00:00,  2.10s/it]
calc: 100%|█████████████████████████████████████| 154/154 [00:01<00:00, 101.63it/s]

Training completes successfully without TypeError.

Types of changes

• Fix bugs
• Add new feature
• Update documentation

Additional Notes

Known issue discovered during testing:

After fixing this LightGBM bug, the DDG-DA workflow encounters another bug:

TypeError: unhashable type: 'slice'
  File "qlib/contrib/meta/data_selection/dataset.py", line 101, in setup

This is a separate issue in the DDG-DA code itself (not related to LightGBM) where a slice object is being used as a dictionary key. This should be tracked and fixed in a separate issue/PR.

References

• LightGBM 4.0 Release: https://github.com/microsoft/LightGBM/releases/tag/v4.0.0
• Related: UnpicklingError: Forbidden class: qlib.contrib.data.handler.Alpha158 #2130 (UnpicklingError that currently masks this bug)
• Related: PR fix(unpickler): allow Alpha158/Alpha360 handlers and the standard dataset chain #2213 (Required to reproduce this bug)