Fix quantization param overrides in task grid

[Sbboss]

Summary

• Fix quantization parameter overrides from task parameters (issue Bug: Quantization Configuration in task creation doesn't override default quant parameters #2).

Changes

• Promote quantization fields from parameters into __quant__* for grid expansion.
• Add tests to verify user-specified kv-cache dtype overrides defaults.

Testing

• python tests/test_quant_param_override.py

Checklist

• I have added/updated tests where needed.
• I have run the relevant tests.
• I have updated documentation if required.

Notes

• Related issue: Bug: Quantization Configuration in task creation doesn't override default quant parameters #2

[k-l-lambda]

Code Review

The core fix looks correct and tests are valuable. A few suggestions before merging:

1. Precedence Rule Clarification

When both __quant__* and alias keys (e.g., kv-cache-dtype) are provided simultaneously, the current behavior is that the alias wins. Please:

• Document this precedence rule explicitly
• Add a test case for this scenario

2. Test Robustness

test_prepare_runtime_parameters_respects_override uses grid[0] which assumes a specific ordering. Consider:

• Iterating through all grid entries, or
• Using set-based assertions to avoid flakiness

3. Code Style

• The test file uses tabs for indentation while the project appears to use spaces
• Consider using standard imports instead of sys.path manipulation

4. PR Template Issue

The .github/PULL_REQUEST_TEMPLATE.md contains this PR's actual content rather than a generic template with placeholders. Should be a separate commit or removed.

---

Review assisted by GPT-5.2

[Sbboss]

Thanks for the review.

I have made the pertinent changes, please check once.

[Sbboss]

Hey @k-l-lambda ,
Reminder