File size: 13,847 Bytes
fed2333 0ac64e3 fed2333 0de2901 fed2333 0de2901 fed2333 0ac64e3 fed2333 0de2901 fed2333 | 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 | """
Benchmark Runner: Orchestrates base vs Cortex-enhanced model comparison.
Usage:
runner = BenchmarkRunner(model_name="HuggingFaceTB/SmolLM2-135M")
results = runner.run_comparison(tasks=["hellaswag", "piqa"], n=50)
runner.print_results(results)
"""
import sys
import os
import time
import json
import torch
from typing import Dict, List, Optional, Any
from transformers import AutoModelForCausalLM, AutoTokenizer
# Add parent dir so cortex can be imported
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from benchmark.scoring import log_likelihood_score, accuracy_from_loglikelihoods
from benchmark.tasks import TASK_REGISTRY, BenchmarkTask
from benchmark.memory_tasks import PasskeyRetrieval, MultiHopMemory
from cortex.torch_device import resolve_torch_device
class BenchmarkRunner:
"""
Runs a full comparison between base model and Cortex-enhanced model.
Workflow:
1. Load base model, run all tasks → base results
2. Inject Cortex modules via CortexSurgeon → enhanced model
3. Run all tasks again → cortex results
4. Compare and report
"""
def __init__(
self,
model_name: str = "HuggingFaceTB/SmolLM2-135M",
device: str = "auto",
dtype: str = "float32",
cortex_weights: Optional[str] = None,
):
self.model_name = model_name
self.cortex_weights = cortex_weights
if device == "auto":
self.device = resolve_torch_device("auto")
else:
self.device = device
self.dtype = getattr(torch, dtype)
print(f"Loading model: {model_name} on {self.device} ({dtype})")
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
self.model = AutoModelForCausalLM.from_pretrained(
model_name,
dtype=self.dtype,
device_map=self.device,
)
if self.tokenizer.pad_token is None:
self.tokenizer.pad_token = self.tokenizer.eos_token
self.model.eval()
print(f"Model loaded: {self.model.config.hidden_size}d, {self.model.config.num_hidden_layers}L")
def _run_loglikelihood_task(
self,
task: BenchmarkTask,
model,
n: Optional[int] = None,
) -> Dict:
"""Run a log-likelihood scoring task."""
print(f" Loading examples for {task.name}...")
examples = task.load_examples(n=n)
print(f" Scoring {len(examples)} examples...")
scores_and_golds = []
for i, ex in enumerate(examples):
if (i + 1) % 10 == 0:
print(f" [{i+1}/{len(examples)}]")
scores = log_likelihood_score(
model, self.tokenizer,
ex["context"], ex["continuations"],
device=self.device,
)
scores_and_golds.append((scores, ex["gold_idx"]))
return accuracy_from_loglikelihoods(scores_and_golds)
def _run_memory_tasks(
self,
model,
n_passkey: int = 5,
passkey_lengths: Optional[List[int]] = None,
n_multihop: Optional[int] = None,
) -> Dict:
"""Run memory-specific benchmarks."""
results = {}
# Passkey retrieval
print(" Running passkey retrieval...")
passkey = PasskeyRetrieval(context_lengths=passkey_lengths or [128, 256, 512])
results["passkey_retrieval"] = passkey.run(
model, self.tokenizer,
n_per_length=n_passkey, device=self.device,
)
# Multi-hop memory
print(" Running multi-hop memory...")
multihop = MultiHopMemory()
results["multi_hop_memory"] = multihop.run(
model, self.tokenizer,
n=n_multihop, device=self.device,
)
return results
def inject_cortex(self) -> Dict:
"""
Inject all Cortex modules into the model.
Returns dict with module info.
"""
from cortex import (
CortexSurgeon, MemoryBank, HallucinationGate,
PauseAndThink, BacktrackHead, SteeringVector, AdaptiveDepth,
)
surgeon = CortexSurgeon(self.model)
hidden_dim = surgeon.hidden_dim
num_layers = surgeon.num_layers
# Find valid num_heads for cross-attention
num_heads = 8
while hidden_dim % num_heads != 0 and num_heads > 1:
num_heads -= 1
middle_layers = list(range(num_layers // 3, 2 * num_layers // 3))
deep_layers = list(range(2 * num_layers // 3, num_layers))
surgeon.add_module("memory", MemoryBank(
hidden_dim=hidden_dim, num_slots=32, num_heads=num_heads,
target_layers=middle_layers,
))
surgeon.add_module("halluc_gate", HallucinationGate(
hidden_dim=hidden_dim, bottleneck_dim=32,
target_layers=deep_layers,
))
surgeon.add_module("pause_think", PauseAndThink(
hidden_dim=hidden_dim, num_think_tokens=4,
target_layers=middle_layers,
))
surgeon.add_module("backtrack", BacktrackHead(
hidden_dim=hidden_dim, confidence_bottleneck=32,
num_layers=num_layers, target_layers="all",
))
surgeon.add_module("steering", SteeringVector(
hidden_dim=hidden_dim, num_directions=2,
direction_names=["truthfulness", "helpfulness"],
target_layers=middle_layers,
))
surgeon.add_module("adaptive_depth", AdaptiveDepth(
hidden_dim=hidden_dim, target_layers="all",
))
surgeon.operate(freeze_base=True)
if self.cortex_weights:
surgeon.load_cortex_modules(self.cortex_weights)
print(f" Loaded Cortex weights: {self.cortex_weights}")
report = surgeon.get_parameter_report()
total_cortex = sum(info["trainable"] for info in report.values())
total_model = sum(p.numel() for p in self.model.parameters())
self._surgeon = surgeon
return {
"total_cortex_params": total_cortex,
"total_model_params": total_model,
"overhead_pct": total_cortex / total_model * 100,
"per_module": report,
}
def remove_cortex(self):
"""Remove Cortex modules and restore base model."""
if hasattr(self, "_surgeon"):
self._surgeon.remove_all()
del self._surgeon
def run_comparison(
self,
tasks: Optional[List[str]] = None,
n: int = 50,
include_memory: bool = True,
n_passkey: int = 5,
passkey_lengths: Optional[List[int]] = None,
) -> Dict:
"""
Run full comparison: base model vs Cortex-enhanced.
Args:
tasks: List of task names from TASK_REGISTRY. None = all.
n: Number of examples per task.
include_memory: Whether to run memory benchmarks.
n_passkey: Number of passkey examples per context length.
passkey_lengths: Context lengths for passkey test.
Returns:
Dict with base_results, cortex_results, and comparison.
"""
if tasks is None:
tasks = ["hellaswag", "piqa", "arc-easy", "winogrande"]
results = {
"model": self.model_name,
"device": self.device,
"dtype": str(self.dtype),
"n_per_task": n,
"tasks": tasks,
"base": {},
"cortex": {},
"comparison": {},
}
# ===== BASE MODEL =====
print("\n" + "=" * 60)
print("PHASE 1: BASE MODEL EVALUATION")
print("=" * 60)
for task_name in tasks:
print(f"\n[BASE] Running {task_name}...")
t0 = time.time()
task_cls = TASK_REGISTRY[task_name]
task = task_cls() if callable(task_cls) else task_cls
result = self._run_loglikelihood_task(task, self.model, n=n)
result["time_seconds"] = time.time() - t0
results["base"][task_name] = result
print(f" {task_name}: {result['accuracy']:.4f} ({result['correct']}/{result['total']}) "
f"[{result['time_seconds']:.1f}s]")
if include_memory:
print(f"\n[BASE] Running memory benchmarks...")
t0 = time.time()
mem_results = self._run_memory_tasks(
self.model, n_passkey=n_passkey,
passkey_lengths=passkey_lengths,
)
mem_results["time_seconds"] = time.time() - t0
results["base"]["memory"] = mem_results
pk = mem_results["passkey_retrieval"]["overall"]
mh = mem_results["multi_hop_memory"]
print(f" passkey: {pk['accuracy']:.4f} ({pk['correct']}/{pk['total']})")
print(f" multi_hop: {mh['accuracy']:.4f} ({mh['correct']}/{mh['total']})")
# ===== CORTEX-ENHANCED MODEL =====
print("\n" + "=" * 60)
print("PHASE 2: CORTEX-ENHANCED MODEL EVALUATION")
print("=" * 60)
print("\nInjecting Cortex modules...")
module_info = self.inject_cortex()
print(f" Cortex params: {module_info['total_cortex_params']:,} "
f"({module_info['overhead_pct']:.2f}% overhead)")
results["cortex_info"] = module_info
for task_name in tasks:
print(f"\n[CORTEX] Running {task_name}...")
t0 = time.time()
task_cls = TASK_REGISTRY[task_name]
task = task_cls() if callable(task_cls) else task_cls
result = self._run_loglikelihood_task(task, self.model, n=n)
result["time_seconds"] = time.time() - t0
results["cortex"][task_name] = result
print(f" {task_name}: {result['accuracy']:.4f} ({result['correct']}/{result['total']}) "
f"[{result['time_seconds']:.1f}s]")
if include_memory:
print(f"\n[CORTEX] Running memory benchmarks...")
t0 = time.time()
mem_results = self._run_memory_tasks(
self.model, n_passkey=n_passkey,
passkey_lengths=passkey_lengths,
)
mem_results["time_seconds"] = time.time() - t0
results["cortex"]["memory"] = mem_results
pk = mem_results["passkey_retrieval"]["overall"]
mh = mem_results["multi_hop_memory"]
print(f" passkey: {pk['accuracy']:.4f} ({pk['correct']}/{pk['total']})")
print(f" multi_hop: {mh['accuracy']:.4f} ({mh['correct']}/{mh['total']})")
# ===== COMPARISON =====
print("\n" + "=" * 60)
print("COMPARISON: BASE vs CORTEX")
print("=" * 60)
for task_name in tasks:
base_acc = results["base"][task_name]["accuracy"]
cortex_acc = results["cortex"][task_name]["accuracy"]
delta = cortex_acc - base_acc
symbol = "↑" if delta > 0 else "↓" if delta < 0 else "="
results["comparison"][task_name] = {
"base": base_acc,
"cortex": cortex_acc,
"delta": delta,
}
print(f" {task_name:20s} base={base_acc:.4f} cortex={cortex_acc:.4f} "
f"Δ={delta:+.4f} {symbol}")
if include_memory:
base_pk = results["base"]["memory"]["passkey_retrieval"]["overall"]["accuracy"]
cortex_pk = results["cortex"]["memory"]["passkey_retrieval"]["overall"]["accuracy"]
base_mh = results["base"]["memory"]["multi_hop_memory"]["accuracy"]
cortex_mh = results["cortex"]["memory"]["multi_hop_memory"]["accuracy"]
results["comparison"]["passkey"] = {
"base": base_pk, "cortex": cortex_pk, "delta": cortex_pk - base_pk,
}
results["comparison"]["multi_hop"] = {
"base": base_mh, "cortex": cortex_mh, "delta": cortex_mh - base_mh,
}
print(f" {'passkey':20s} base={base_pk:.4f} cortex={cortex_pk:.4f} "
f"Δ={cortex_pk - base_pk:+.4f}")
print(f" {'multi_hop':20s} base={base_mh:.4f} cortex={cortex_mh:.4f} "
f"Δ={cortex_mh - base_mh:+.4f}")
# Remove cortex modules to restore base model
self.remove_cortex()
return results
@staticmethod
def print_summary(results: Dict):
"""Print a formatted summary of benchmark results."""
print("\n" + "=" * 70)
print(f"BENCHMARK SUMMARY: {results['model']}")
print(f"n={results['n_per_task']} per task, device={results['device']}")
print("=" * 70)
print(f"\n{'Task':22s} {'Base':>8s} {'Cortex':>8s} {'Delta':>8s}")
print("-" * 50)
for task_name, comp in results["comparison"].items():
delta_str = f"{comp['delta']:+.4f}"
symbol = " ↑" if comp["delta"] > 0.001 else " ↓" if comp["delta"] < -0.001 else " "
print(f"{task_name:22s} {comp['base']:8.4f} {comp['cortex']:8.4f} {delta_str:>8s}{symbol}")
if "cortex_info" in results:
info = results["cortex_info"]
print(f"\nCortex overhead: {info['total_cortex_params']:,} params "
f"({info['overhead_pct']:.2f}%)")
print("=" * 70)
|