[diw-zw]

Overview

This PR introduces significant enhancements to the auto-benchmark search algorithms, focusing on SLA constraint evaluation improvements, and search space handling refinements.

Key Changes

1. Search Algorithm Improvements

• Removed param.go: Eliminated the old parameter expansion logic that was causing issues in TPE random phase
• Enhanced Grid Search:
  • Added proper parameter type validation and expansion
  • Support for categorical, int, and float parameters with step values
  • Rejected unsupported types (continuous params without step, pow2)
  • Improved Cartesian product computation with overflow protection

2. TPE/Optuna Integration

• Optuna Bridge Enhancement:
  • Moved optuna_bridge.py to proper location (tools/optuna/)
  • Improved distribution descriptor handling for different parameter types
  • Better support for float, int, and categorical distributions
• Search Space Handling:
  • Proper constraint and parameter type processing
  • Raw search space preservation for pow2 decoding

3. SLA Constraint Evaluation

• Relative Deviation Calculation:
  • Changed from binary pass/fail to relative deviation ratios
  • Max constraints: deviation = max(0, (actual - limit) / limit)
  • Allows Optuna sampler to perceive violation severity
• Removed EarlyStopPatience:
  • Prevents premature stopping during TPE random exploration phase
• Percentage-based Metrics:
  • Changed SLA evaluation to use percentage values instead of absolute values

4. Configuration & Parsing

• Enhanced Config Parsing:
  • Improved search space configuration handling
  • Better validation for parameter definitions
• Updated Config Types:
  • Added support for new parameter type definitions
  • Refined strategy specification structure

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[Syspretor]

/lgtm