Spaces:
Running
Running
File size: 15,605 Bytes
3c1db6c eff2120 3c1db6c eff2120 3c1db6c eff2120 3c1db6c eff2120 3c1db6c eff2120 3c1db6c eff2120 3c1db6c eff2120 3c1db6c eff2120 3c1db6c eff2120 3c1db6c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 | """
Optimization Engine — determines the optimal inference configuration.
Solves:
x* = argmin J(x) = α·Cost(x) + β·Tokens(x) - γ·Quality(x)
Subject to:
Capability(model) >= Complexity(query)
Quality >= threshold
Latency <= max_latency
V1: deterministic rule engine with fixed per-budget-mode weights.
V2: BayesianWeightOptimizer (Optuna) learns α,β,γ from past outcomes.
Falls back to V1 fixed weights if optuna is not installed.
"""
from __future__ import annotations
import json
import logging
from dataclasses import dataclass
from pathlib import Path
from typing import Optional
from llmopt.registry.model_registry import ModelRegistry, ModelSpec
from llmopt.estimator.complexity_estimator import ComplexityResult
logger = logging.getLogger(__name__)
DATA_DIR = Path(__file__).parent.parent.parent / "data"
BAYES_STUDY_PATH = DATA_DIR / "bayesian_study.json"
# ---------------------------------------------------------------------------
# User constraints schema
# ---------------------------------------------------------------------------
@dataclass
class UserConstraints:
"""
Caller-supplied constraints that bound the optimization search space.
"""
budget_mode: str = "balanced" # cheap | balanced | quality
max_cost_per_request: Optional[float] = None # USD hard cap
max_latency_score: Optional[float] = None # lower = faster model
quality_threshold: float = 0.60 # min acceptable quality proxy
exclude_providers: Optional[list[str]] = None # e.g. ["ollama"] for cloud-only
only_providers: Optional[list[str]] = None # e.g. ["openai"]
prefer_local: bool = False # prefer Ollama models
compression_enabled: Optional[bool] = None # None = auto-decide
def __post_init__(self):
if self.exclude_providers is None:
self.exclude_providers = []
if self.only_providers is None:
self.only_providers = []
# ---------------------------------------------------------------------------
# Optimization result schema
# ---------------------------------------------------------------------------
@dataclass
class OptimizationResult:
selected_model: str
provider: str
estimated_cost: float # USD for this request
estimated_input_tokens: int
estimated_output_tokens: int
max_tokens: int # hard cap to pass to the API
compression_enabled: bool
system_prompt_style: str # "verbose" | "concise" | "minimal"
rationale: list[str]
fallback_model: Optional[str] = None
objective_score: float = 0.0 # lower is better
def to_dict(self) -> dict:
return self.__dict__.copy()
# ---------------------------------------------------------------------------
# Objective weights per budget mode
# ---------------------------------------------------------------------------
_BUDGET_WEIGHTS = {
# α (cost) β (tokens) γ (quality)
"cheap": (0.60, 0.30, 0.10),
"balanced": (0.40, 0.20, 0.40),
"quality": (0.10, 0.10, 0.80),
}
# max_tokens caps per output-length bucket
_MAX_TOKENS_MAP = {
"short": 150,
"medium": 400,
"long": 900,
"very_long": 1800,
}
# ---------------------------------------------------------------------------
# Bayesian Weight Optimizer (V2)
# ---------------------------------------------------------------------------
class BayesianWeightOptimizer:
"""
Uses Optuna to find optimal α,β,γ weights for J(x) based on
accumulated feedback (cost vs quality trade-offs from past requests).
Falls back gracefully to fixed V1 weights if optuna is not installed.
"""
def __init__(self):
self._optuna_available = False
self._studies: dict = {} # one study per budget_mode
try:
import optuna # type: ignore
optuna.logging.set_verbosity(optuna.logging.WARNING)
self._optuna_available = True
logger.info("BayesianWeightOptimizer: Optuna available. Using Bayesian weight tuning.")
except ImportError:
logger.info("BayesianWeightOptimizer: Optuna not installed. Using V1 fixed weights.")
def get_weights(self, budget_mode: str) -> tuple[float, float, float]:
"""
Returns (α, β, γ) weights for the given budget mode.
Uses Bayesian optimization if optuna is available and we have
enough feedback history, otherwise falls back to V1 fixed weights.
"""
if not self._optuna_available:
return _BUDGET_WEIGHTS.get(budget_mode, _BUDGET_WEIGHTS["balanced"])
# Load saved trials
history = self._load_history(budget_mode)
if len(history) < 5:
# Not enough data yet — use V1 defaults but still warm up
logger.debug(f"Bayesian: Only {len(history)} trials for '{budget_mode}', using V1 defaults.")
return _BUDGET_WEIGHTS.get(budget_mode, _BUDGET_WEIGHTS["balanced"])
try:
import optuna # type: ignore
study_key = budget_mode
if study_key not in self._studies:
self._studies[study_key] = optuna.create_study(direction="minimize")
# Seed with historical trials
for trial_data in history:
self._studies[study_key].add_trial(
optuna.trial.create_trial(
params={"alpha": trial_data["alpha"], "beta": trial_data["beta"], "gamma": trial_data["gamma"]},
distributions={
"alpha": optuna.distributions.FloatDistribution(0.05, 0.90),
"beta": optuna.distributions.FloatDistribution(0.05, 0.60),
"gamma": optuna.distributions.FloatDistribution(0.05, 0.90),
},
value=trial_data["outcome"],
)
)
study = self._studies[study_key]
best = study.best_params
α = best["alpha"]
β = best["beta"]
γ = best["gamma"]
logger.debug(f"Bayesian weights for '{budget_mode}': α={α:.3f} β={β:.3f} γ={γ:.3f}")
return α, β, γ
except Exception as e:
logger.warning(f"Bayesian weight retrieval failed: {e}. Using V1 defaults.")
return _BUDGET_WEIGHTS.get(budget_mode, _BUDGET_WEIGHTS["balanced"])
def record_outcome(
self,
budget_mode: str,
alpha: float, beta: float, gamma: float,
actual_cost: float,
quality_score: float,
) -> None:
"""
Records the outcome of a request. The 'outcome' score is what
we want to minimize: actual cost weighted against quality.
Call this after receiving a response + evaluation score.
"""
# Composite outcome: high cost = bad, low quality = bad
# Normalise: assume max_cost ~$0.02, quality in [1,10] → [0,1]
cost_norm = min(actual_cost / 0.02, 1.0)
quality_norm = quality_score / 10.0
outcome = cost_norm - quality_norm # minimise this
history = self._load_history(budget_mode)
history.append({
"alpha": alpha, "beta": beta, "gamma": gamma,
"actual_cost": actual_cost,
"quality_score": quality_score,
"outcome": outcome,
})
self._save_history(budget_mode, history)
# Invalidate the in-memory study so it reloads next time
self._studies.pop(budget_mode, None)
def _load_history(self, budget_mode: str) -> list:
if not BAYES_STUDY_PATH.exists():
return []
try:
data = json.loads(BAYES_STUDY_PATH.read_text())
return data.get(budget_mode, [])
except Exception:
return []
def _save_history(self, budget_mode: str, history: list) -> None:
existing = {}
if BAYES_STUDY_PATH.exists():
try:
existing = json.loads(BAYES_STUDY_PATH.read_text())
except Exception:
pass
existing[budget_mode] = history
BAYES_STUDY_PATH.parent.mkdir(parents=True, exist_ok=True)
BAYES_STUDY_PATH.write_text(json.dumps(existing, indent=2))
# ---------------------------------------------------------------------------
# Engine
# ---------------------------------------------------------------------------
class OptimizationEngine:
"""
Core decision engine. Selects model + config that minimizes
J(x) = α·Cost + β·Tokens - γ·Quality under user constraints.
V2: Uses BayesianWeightOptimizer to learn optimal α,β,γ weights over time.
"""
def __init__(self, registry: ModelRegistry):
self.registry = registry
self.bayes = BayesianWeightOptimizer()
def optimize(
self,
complexity: ComplexityResult,
output_length_bucket: str,
constraints: Optional[UserConstraints] = None,
) -> OptimizationResult:
if constraints is None:
constraints = UserConstraints()
α, β, γ = self.bayes.get_weights(constraints.budget_mode)
logger.debug(f"Using weights α={α:.3f} β={β:.3f} γ={γ:.3f} for mode '{constraints.budget_mode}'")
# --- 1. Build candidate set ---
candidates = self.registry.capable_of(
complexity=complexity.score,
min_reasoning=complexity.required_reasoning * 0.85, # 15% headroom
min_coding=complexity.required_coding * 0.85,
min_math=complexity.required_math * 0.85,
exclude_providers=constraints.exclude_providers,
only_providers=constraints.only_providers if constraints.only_providers else None,
)
if not candidates:
# Hard fallback: use the most capable model in registry
candidates = self.registry.all_models()
candidates = [max(candidates, key=lambda m: m.capability_score)]
# --- 2. Apply hard filters ---
candidates = self._apply_hard_filters(candidates, constraints, complexity)
if not candidates:
candidates = self.registry.all_models()
# --- 3. Score candidates via objective function ---
scored = []
for model in candidates:
obj, rationale = self._objective(
model,
α, β, γ,
complexity.estimated_input_tokens,
complexity.estimated_output_tokens,
)
scored.append((obj, model, rationale))
scored.sort(key=lambda x: x[0])
best_score, best_model, best_rationale = scored[0]
fallback = scored[1][1].model_name if len(scored) > 1 else None
# --- 4. Determine configuration ---
max_tokens = _MAX_TOKENS_MAP.get(output_length_bucket, 400)
compression = self._should_compress(constraints, complexity, constraints.budget_mode)
system_prompt_style = self._system_prompt_style(constraints.budget_mode, complexity.score)
estimated_cost = best_model.cost_per_request(
complexity.estimated_input_tokens,
complexity.estimated_output_tokens,
)
return OptimizationResult(
selected_model=best_model.model_name,
provider=best_model.provider,
estimated_cost=round(estimated_cost, 6),
estimated_input_tokens=complexity.estimated_input_tokens,
estimated_output_tokens=complexity.estimated_output_tokens,
max_tokens=max_tokens,
compression_enabled=compression,
system_prompt_style=system_prompt_style,
rationale=best_rationale,
fallback_model=fallback,
objective_score=round(best_score, 4),
)
# ------------------------------------------------------------------
# Objective function J(x) = α·Cost + β·Tokens - γ·Quality
# ------------------------------------------------------------------
def _objective(
self,
model: ModelSpec,
α: float, β: float, γ: float,
input_tokens: int,
output_tokens: int,
) -> tuple[float, list[str]]:
"""
Normalized objective score. Lower is better.
Costs are normalized to [0,1] against registry max values.
"""
max_cost = self._max_cost_in_registry(input_tokens, output_tokens)
max_tokens = input_tokens + output_tokens
cost_norm = model.cost_per_request(input_tokens, output_tokens) / (max_cost + 1e-9)
token_norm = (input_tokens + output_tokens) / (max_tokens + 1e-9) # uniform here; prompt optimizer changes this
quality_norm = model.capability_score # higher = better
J = α * cost_norm + β * token_norm - γ * quality_norm
rationale = [
f"model={model.model_name}",
f"provider={model.provider}",
f"capability={model.capability_score:.3f}",
f"cost_norm={cost_norm:.4f}",
f"J={J:.4f} (α={α},β={β},γ={γ})",
]
return J, rationale
# ------------------------------------------------------------------
# Hard filters
# ------------------------------------------------------------------
def _apply_hard_filters(
self,
candidates: list[ModelSpec],
constraints: UserConstraints,
complexity: ComplexityResult,
) -> list[ModelSpec]:
filtered = []
for m in candidates:
# Cost cap
if constraints.max_cost_per_request is not None:
est = m.cost_per_request(
complexity.estimated_input_tokens,
complexity.estimated_output_tokens,
)
if est > constraints.max_cost_per_request:
continue
# Latency cap (latency_score: higher = faster)
if constraints.max_latency_score is not None:
if m.latency_score < constraints.max_latency_score:
continue
# Quality floor: capability_score >= quality_threshold
if m.capability_score < constraints.quality_threshold:
continue
filtered.append(m)
return filtered
# ------------------------------------------------------------------
# Config helpers
# ------------------------------------------------------------------
def _should_compress(
self,
constraints: UserConstraints,
complexity: ComplexityResult,
budget_mode: str,
) -> bool:
if constraints.compression_enabled is not None:
return constraints.compression_enabled
# Auto: compress for long contexts or cheap modes
return budget_mode in ("cheap", "balanced") or complexity.estimated_input_tokens > 800
def _system_prompt_style(self, budget_mode: str, complexity_score: float) -> str:
if budget_mode == "cheap":
return "minimal"
elif budget_mode == "quality" or complexity_score > 0.75:
return "verbose"
return "concise"
def _max_cost_in_registry(self, input_tokens: int, output_tokens: int) -> float:
costs = [
m.cost_per_request(input_tokens, output_tokens)
for m in self.registry.all_models()
]
return max(costs) if costs else 1.0
|