optimizer.py

from typing import Dict, Any, List
import json

from utils import (
    call_llm,
    safe_json_loads,
    normalize_ir,
    validate_ir
)


# =====================================================
# FALLBACK STRATEGY
# =====================================================

def fallback_optimization(
    ir: Dict[str, Any],
    idx: int
):

    return {
        "explanation": f"Fallback optimisation strategy {idx}",
        "optimized_ir": normalize_ir(ir)
    }


# =====================================================
# OPTIMIZATION
# =====================================================

def optimize_ir(
    ir: Dict[str, Any],
    num_strategies: int = 3,
    provider: str = "openai"
) -> List[Dict[str, Any]]:

    if not validate_ir(ir):

        return [
            fallback_optimization(ir, i + 1)
            for i in range(num_strategies)
        ]

    ir_text = json.dumps(
        ir,
        indent=2,
        ensure_ascii=False
    )

    prompt = f"""
Tu es un expert en optimisation de graphes mathématiques.

Analyse cette IR.

Produis EXACTEMENT {num_strategies} stratégies
d'optimisation différentes.

Les optimisations possibles incluent :
- fusion d'opérations
- suppression de redondances
- réduction mémoire
- vectorisation
- parallélisation
- approximation numérique
- simplification algébrique
- réordonnancement

IMPORTANT :
- retourne UNIQUEMENT du JSON valide
- aucune explication hors JSON
- aucune balise markdown

FORMAT STRICT :

[
  {{
    "explanation": "description courte",
    "optimized_ir": {{
      "name": "Optimized IR",
      "strategy": "fusion",
      "nodes": [],
      "edges": []
    }}
  }}
]

IR :
{ir_text}
"""

    response = call_llm(
        prompt,
        provider=provider,
        max_tokens=3500
    )

    parsed = safe_json_loads(response)

    # -------------------------------------------------
    # VALIDATION
    # -------------------------------------------------

    valid_results = []

    if isinstance(parsed, list):

        for item in parsed:

            if not isinstance(item, dict):
                continue

            explanation = item.get(
                "explanation",
                "Optimisation automatique"
            )

            optimized_ir = item.get(
                "optimized_ir",
                ir
            )

            optimized_ir = normalize_ir(
                optimized_ir
            )

            valid_results.append({
                "explanation": explanation,
                "optimized_ir": optimized_ir
            })

    # -------------------------------------------------
    # SUCCESS
    # -------------------------------------------------

    if len(valid_results) > 0:

        return valid_results[:num_strategies]

    # -------------------------------------------------
    # FALLBACK
    # -------------------------------------------------

    return [
        fallback_optimization(ir, i + 1)
        for i in range(num_strategies)
    ]