File size: 13,414 Bytes
bfc7d04 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 | #!/usr/bin/env python3
"""
Fuse LoRA adapters from multiple team members into a unified model.
This script demonstrates how to merge multiple LoRA adapters trained on different
codebases or by different team members, enabling collective intelligence while
preserving individual specialization.
Algorithm: Weighted averaging with similarity-based adaptive weights.
"""
import os
import json
import torch
import argparse
from pathlib import Path
from typing import Dict, List, Optional, Tuple
import numpy as np
from collections import defaultdict
def load_lora_adapter(adapter_path: str, device: str = "cpu") -> Dict[str, torch.Tensor]:
"""
Load a LoRA adapter from a safetensors or pytorch bin file.
Returns: Dict of parameter name -> tensor
"""
adapter_path = Path(adapter_path)
# Try safetensors first (faster, no pickle)
safetensors_path = adapter_path / "adapter_model.safetensors"
if safetensors_path.exists():
from safetensors import safe_open
tensors = {}
with safe_open(safetensors_path, framework="pt", device=device) as f:
for key in f.keys():
tensors[key] = f.get_tensor(key)
return tensors
# Fall back to pytorch bin
pytorch_path = adapter_path / "adapter_model.bin"
if pytorch_path.exists():
tensors = torch.load(pytorch_path, map_location=device, weights_only=True)
return tensors
raise FileNotFoundError(f"No adapter found at {adapter_path}")
def compute_adapter_metadata(adapter_path: str) -> Dict[str, Any]:
"""
Load adapter metadata (training stats, performance, etc.) if available.
"""
metadata_path = Path(adapter_path) / "adapter_metadata.json"
if metadata_path.exists():
with open(metadata_path, 'r') as f:
return json.load(f)
# Default metadata
return {
"training_examples": 0,
"validation_score": 0.0,
"domains": [],
"team_member": "unknown"
}
def compute_similarity_matrix(
adapters: List[Tuple[str, Dict[str, torch.Tensor]]],
sample_keys: Optional[List[str]] = None
) -> np.ndarray:
"""
Compute pairwise similarity between adapters based on weight distributions.
Uses cosine similarity of normalized weight vectors.
"""
n = len(adapters)
similarity = np.zeros((n, n))
# Get parameter names common to all adapters
if sample_keys is None:
common_keys = set(adapters[0][1].keys())
for _, tensors in adapters[1:]:
common_keys &= set(tensors.keys())
sample_keys = list(common_keys)[:100] # Sample up to 100 parameters
# Flatten sampled parameters for each adapter
vectors = []
for _, tensors in adapters:
vec_parts = []
for key in sample_keys:
if key in tensors:
# Flatten and normalize
t = tensors[key].float().flatten()
norm = torch.norm(t).item()
if norm > 1e-8:
t = t / norm
vec_parts.append(t.numpy())
else:
# Missing parameter, use zeros
shape = tensors[sample_keys[0]].shape if sample_keys[0] in tensors else (1,)
vec_parts.append(np.zeros(shape).flatten())
vectors.append(np.concatenate(vec_parts))
# Compute cosine similarity
for i in range(n):
for j in range(n):
if i == j:
similarity[i, j] = 1.0
else:
v1, v2 = vectors[i], vectors[j]
sim = np.dot(v1, v2) / (np.linalg.norm(v1) * np.linalg.norm(v2) + 1e-8)
similarity[i, j] = sim
return similarity
def compute_adaptive_weights(
similarities: np.ndarray,
metadata: List[Dict[str, Any]],
base_config: Dict[str, float]
) -> np.ndarray:
"""
Compute fusion weights using adaptive strategy:
w_i = (performance_i * domain_overlap_i) / (sum(performance_j * domain_overlap_j) + epsilon)
With similarity-based adjustments:
- Higher weight adapters that are similar to each other get boosted
- Diverse adapters get balanced contributions
"""
n = len(metadata)
weights = np.zeros(n)
# Base weights from performance
base_weights = np.array([
meta.get("validation_score", 0.0) *
meta.get("training_examples", 1) / 1000.0 # Normalize by dataset size
for meta in metadata
])
# Domain overlap weights
domain_weights = np.zeros(n)
all_domains = defaultdict(int)
for i, meta in enumerate(metadata):
for domain in meta.get("domains", []):
all_domains[domain] += 1
for i, meta in enumerate(metadata):
overlap = 0.0
for domain in meta.get("domains", []):
# Rare domains get higher weight
overlap += 1.0 / all_domains[domain]
domain_weights[i] = overlap if overlap > 0 else 1.0
# Combine base weights with domain weights
raw_weights = base_weights * domain_weights
# Apply similarity-based smoothing
# If two adapters are very similar, distribute weight more evenly
similarity_threshold = base_config.get("similarity_threshold", 0.9)
similarity_damping = base_config.get("similarity_damping", 0.3)
for i in range(n):
for j in range(i+1, n):
if similarities[i, j] > similarity_threshold:
# Too similar, dampen differences
avg_weight = (raw_weights[i] + raw_weights[j]) / 2
raw_weights[i] = raw_weights[i] * (1 - similarity_damping) + avg_weight * similarity_damping
raw_weights[j] = raw_weights[j] * (1 - similarity_damping) + avg_weight * similarity_damping
# Normalize
total = np.sum(raw_weights)
if total > 0:
weights = raw_weights / total
else:
weights = np.ones(n) / n
return weights
def fuse_adapters(
adapter_paths: List[str],
output_path: str,
config: Optional[Dict] = None
) -> Tuple[Path, Dict]:
"""
Fuse multiple LoRA adapters into a single adapter.
Args:
adapter_paths: List of paths to adapter directories
output_path: Where to save the fused adapter
config: Fusion configuration (weights, similarity thresholds, etc.)
Returns:
Path to fused adapter, fusion metadata
"""
if config is None:
config = {
"fusion_method": "weighted_average",
"similarity_threshold": 0.9,
"similarity_damping": 0.3,
"normalize_weights": True,
"clip_diff": 2.0 # Clip weight differences to avoid extreme values
}
print(f"๐ Fusing {len(adapter_paths)} adapters...")
# Load all adapters
adapters = []
metadata_list = []
for path in adapter_paths:
print(f" Loading: {Path(path).name}")
try:
tensors = load_lora_adapter(path)
meta = compute_adapter_metadata(path)
adapters.append((path, tensors))
metadata_list.append(meta)
except Exception as e:
print(f" โ ๏ธ Skipped {path}: {e}")
if len(adapters) < 2:
raise ValueError("Need at least 2 adapters to fuse")
# Get common parameter keys
common_keys = set(adapters[0][1].keys())
for _, tensors in adapters[1:]:
common_keys &= set(tensors.keys())
print(f" Common parameters: {len(common_keys)}")
# Compute similarities
print(" Computing adapter similarities...")
# Sample parameters for similarity computation
sample_keys = list(common_keys)[:min(100, len(common_keys))]
similarities = compute_similarity_matrix(adapters, sample_keys)
# Compute adaptive weights
weights = compute_adaptive_weights(similarities, metadata_list, config)
print(" Fusion weights:")
for i, (path, _) in enumerate(adapters):
member = metadata_list[i].get("team_member", f"adapter_{i}")
print(f" {member}: {weights[i]:.3f}")
# Fuse weights
print(" Fusing weights...")
fused_tensors = {}
for key in common_keys:
# Start with zero tensor
fused = None
for idx, (_, tensors) in enumerate(adapters):
weight = weights[idx]
tensor = tensors[key].float()
if fused is None:
fused = tensor * weight
else:
fused += tensor * weight
# Apply clipping if configured
if config["clip_diff"] > 0:
# Clip extreme values relative to first adapter
reference = adapters[0][1][key].float()
max_diff = torch.max(torch.abs(fused - reference)) * config["clip_diff"]
# This is a simple heuristic - could be improved
fused = torch.clamp(fused,
reference - max_diff,
reference + max_diff)
fused_tensors[key] = fused.half() # Convert back to half precision
# Save fused adapter
output_path = Path(output_path)
output_path.mkdir(parents=True, exist_ok=True)
# Save tensors
fused_file = output_path / "adapter_model.safetensors"
try:
from safetensors import save_file
save_file(fused_tensors, str(fused_file))
except ImportError:
# Fallback to pytorch
torch.save(fused_tensors, output_path / "adapter_model.bin")
# Save metadata
fusion_metadata = {
"fusion_date": "2025-04-03", # Would use datetime.now()
"source_adapters": [
{
"path": path,
"team_member": meta.get("team_member", "unknown"),
"validation_score": meta.get("validation_score", 0.0),
"domains": meta.get("domains", []),
"weight": float(weights[i])
}
for i, (path, meta) in enumerate(zip([p for p, _ in adapters], metadata_list))
],
"fusion_config": config,
"similarity_matrix": similarities.tolist(),
"total_parameters": len(common_keys)
}
with open(output_path / "fusion_metadata.json", 'w') as f:
json.dump(fusion_metadata, f, indent=2)
print(f"\nโ
Fused adapter saved to: {output_path}")
print(f" Parameters: {len(common_keys)}")
print(f" Used samples: {sum(m.get('training_examples', 0) for m in metadata_list)}")
return output_path, fusion_metadata
def validate_fusion(
fused_adapter_path: str,
test_cases_path: Optional[str] = None,
base_model: str = "Qwen/Qwen2.5-Coder-32B"
) -> Dict[str, float]:
"""
Validate the fused adapter against test cases.
Returns: Dictionary with validation metrics
"""
print("๐ Validating fused adapter...")
# This would integrate with the evaluation framework
# For now, return mock metrics
metrics = {
"score": 0.0,
"test_cases": 0,
"passed": 0
}
if test_cases_path:
# Load and run test cases
test_cases_path = Path(test_cases_path)
if test_cases_path.exists():
# Would implement actual validation
pass
print(f" Validation complete (placeholder)")
return metrics
def main():
parser = argparse.ArgumentParser(
description="Fuse LoRA adapters from multiple team members."
)
parser.add_argument(
"--adapters",
nargs='+',
required=True,
help="Paths to adapter directories (each with adapter_model.safetensors)"
)
parser.add_argument(
"--output",
type=str,
default="fused-adapter",
help="Output directory for fused adapter"
)
parser.add_argument(
"--config",
type=str,
help="JSON config file with fusion parameters"
)
parser.add_argument(
"--validate",
action="store_true",
help="Run validation after fusion"
)
parser.add_argument(
"--base-model",
type=str,
default="Qwen/Qwen2.5-Coder-32B",
help="Base model identifier"
)
args = parser.parse_args()
# Load config if provided
config = None
if args.config:
with open(args.config, 'r') as f:
config = json.load(f)
# Fuse adapters
try:
output_path, metadata = fuse_adapters(
args.adapters,
args.output,
config or {}
)
# Validate if requested
if args.validate:
metrics = validate_fusion(str(output_path), base_model=args.base_model)
print("\n๐ Validation Metrics:")
for k, v in metrics.items():
print(f" {k}: {v}")
# Print summary
print("\n๐ Fusion Summary:")
print(f" Total adapters: {len(args.adapters)}")
print(f" Output: {output_path}")
print(f" Members:", ", ".join(
m["team_member"] for m in metadata["source_adapters"]
))
except Exception as e:
print(f"โ Fusion failed: {e}")
import traceback
traceback.print_exc()
return 1
return 0
if __name__ == "__main__":
exit(main())
|