File size: 16,541 Bytes
9e1709d | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 | """
Formula Engine: Discovers mathematical patterns in Qwen 0.5B weights
and creates compact formula representations for reconstruction.
Architecture of Qwen2.5-0.5B-Instruct:
- hidden_size: 896
- intermediate_size: 4864
- num_attention_heads: 14
- num_key_value_heads: 2
- num_hidden_layers: 24
- vocab_size: 151936
Weight layers per transformer block:
- self_attn.q_proj: (896, 896) - 14 heads
- self_attn.k_proj: (128, 896) - 2 KV heads
- self_attn.v_proj: (128, 896) - 2 KV heads
- self_attn.o_proj: (896, 896)
- mlp.gate_proj: (4864, 896)
- mlp.up_proj: (4864, 896)
- mlp.down_proj: (896, 4864)
- input_layernorm: (896,)
- post_attention_layernorm: (896,)
Plus:
- embed_tokens: (151936, 896) - shared with lm_head (tied)
Compression Strategy:
- MLP layers (gate, up, down): SVD low-rank factorization (biggest savings)
- Attention K/V: SVD (small rectangular matrices)
- Attention Q/O: Quantization to int8 + scale factors
- Norms: Store as-is (tiny - 896 floats each)
- Embeddings: SVD + quantization hybrid
"""
import torch
import numpy as np
import json
import os
from pathlib import Path
from typing import Dict, Tuple, Optional
import struct
import io
class FormulaEngine:
"""
Discovers compact formula representations for neural network weights.
A "formula" here means: instead of storing the full weight matrix W (shape m×n),
we store a factorized/compressed representation that can reconstruct W.
Methods:
1. SVD: W ≈ U_r @ diag(S_r) @ V_r^T (stores U_r, S_r, V_r)
2. Quantized: W ≈ scale * W_int8 + zero_point
3. Hybrid: SVD on large matrices + quantization on small ones
"""
def __init__(self, compression_ratio: float = 0.4, quality_threshold: float = 0.99):
"""
Args:
compression_ratio: Target ratio of compressed/original size (0.4 = 60% smaller)
quality_threshold: Minimum cosine similarity for reconstruction
"""
self.compression_ratio = compression_ratio
self.quality_threshold = quality_threshold
self.formulas = {} # name -> formula dict
self.metadata = {}
def analyze_weight(self, name: str, weight: torch.Tensor) -> dict:
"""Analyze a weight tensor to find best compression strategy."""
shape = weight.shape
numel = weight.numel()
# Basic statistics
stats = {
"name": name,
"shape": list(shape),
"numel": numel,
"bytes_fp16": numel * 2,
"mean": float(weight.float().mean()),
"std": float(weight.float().std()),
"min": float(weight.float().min()),
"max": float(weight.float().max()),
}
if len(shape) == 2:
m, n = shape
# SVD break-even: rank r saves space when r < m*n/(m+n+1)
breakeven_rank = (m * n) / (m + n + 1)
target_rank = int(breakeven_rank * self.compression_ratio)
stats["breakeven_rank"] = int(breakeven_rank)
stats["target_rank"] = target_rank
stats["svd_viable"] = target_rank > 0 and target_rank < min(m, n)
# Check if quantization would be better for this layer
# For square matrices, SVD savings are minimal
if m == n:
stats["recommended"] = "quantize"
elif max(m, n) / min(m, n) > 2:
stats["recommended"] = "svd"
else:
stats["recommended"] = "svd" if stats["svd_viable"] else "quantize"
elif len(shape) == 1:
stats["recommended"] = "store_raw" # Norms are tiny
else:
stats["recommended"] = "quantize"
return stats
def svd_compress(self, weight: torch.Tensor, rank: int) -> dict:
"""
Compress weight via truncated SVD.
W ≈ U[:, :r] @ diag(S[:r]) @ Vh[:r, :]
Storage: U (m×r) + S (r) + Vh (r×n) in float16
vs original: W (m×n) in float16
Savings when: r * (m + n + 1) < m * n
"""
W = weight.float()
U, S, Vh = torch.linalg.svd(W, full_matrices=False)
# Truncate to rank r
U_r = U[:, :rank].half()
S_r = S[:rank].half()
Vh_r = Vh[:rank, :].half()
# Calculate compression stats
original_size = W.numel() * 2 # fp16
compressed_size = (U_r.numel() + S_r.numel() + Vh_r.numel()) * 2
# Calculate reconstruction error
W_reconstructed = (U_r.float() @ torch.diag(S_r.float()) @ Vh_r.float())
mse = float(((W - W_reconstructed) ** 2).mean())
cos_sim = float(torch.nn.functional.cosine_similarity(
W.reshape(1, -1), W_reconstructed.reshape(1, -1)
))
# Energy captured (fraction of total singular value energy)
energy = float((S[:rank] ** 2).sum() / (S ** 2).sum())
return {
"type": "svd",
"rank": rank,
"original_shape": list(weight.shape),
"original_bytes": original_size,
"compressed_bytes": compressed_size,
"compression_ratio": compressed_size / original_size,
"mse": mse,
"cosine_similarity": cos_sim,
"energy_captured": energy,
"U": U_r,
"S": S_r,
"Vh": Vh_r,
}
def quantize_compress(self, weight: torch.Tensor, bits: int = 4) -> dict:
"""
Compress weight via quantization.
W ≈ scale * W_int + zero_point (per-channel)
For 4-bit: each value stored in 4 bits (2 values per byte)
"""
W = weight.float()
if len(W.shape) == 2:
# Per-channel (per-row) quantization
w_min = W.min(dim=1, keepdim=True).values
w_max = W.max(dim=1, keepdim=True).values
else:
w_min = W.min()
w_max = W.max()
# Compute scale and zero point
qmax = (2 ** bits) - 1
scale = (w_max - w_min) / qmax
scale = scale.clamp(min=1e-8)
zero_point = (-w_min / scale).round().clamp(0, qmax)
# Quantize
W_q = ((W - w_min) / scale).round().clamp(0, qmax).to(torch.uint8)
# Dequantize for error measurement
W_reconstructed = W_q.float() * scale + w_min
mse = float(((W - W_reconstructed) ** 2).mean())
cos_sim = float(torch.nn.functional.cosine_similarity(
W.reshape(1, -1), W_reconstructed.reshape(1, -1)
))
# Storage calculation
original_size = W.numel() * 2 # fp16
quant_size = W.numel() * bits / 8 # quantized weights
meta_size = (scale.numel() + zero_point.numel()) * 2 # scales in fp16
compressed_size = int(quant_size + meta_size)
return {
"type": "quantize",
"bits": bits,
"original_shape": list(weight.shape),
"original_bytes": original_size,
"compressed_bytes": compressed_size,
"compression_ratio": compressed_size / original_size,
"mse": mse,
"cosine_similarity": cos_sim,
"W_q": W_q,
"scale": scale.half(),
"zero_point": zero_point.half(),
"w_min": w_min.half(),
}
def find_best_formula(self, name: str, weight: torch.Tensor) -> dict:
"""
Find the best compression formula for a given weight tensor.
Tries multiple approaches and picks the one with best quality/size tradeoff.
"""
analysis = self.analyze_weight(name, weight)
if analysis["recommended"] == "store_raw":
# Tiny tensors (layer norms) - just store as-is
return {
"type": "raw",
"original_shape": list(weight.shape),
"original_bytes": weight.numel() * 2,
"compressed_bytes": weight.numel() * 2,
"compression_ratio": 1.0,
"mse": 0.0,
"cosine_similarity": 1.0,
"data": weight.half(),
}
results = []
if analysis["recommended"] == "svd" and len(weight.shape) == 2:
# Try multiple ranks
m, n = weight.shape
breakeven = analysis["breakeven_rank"]
for ratio in [0.3, 0.4, 0.5, 0.6]:
rank = max(1, int(breakeven * ratio))
rank = min(rank, min(m, n) - 1)
result = self.svd_compress(weight, rank)
result["approach"] = f"svd_rank{rank}"
results.append(result)
# Always try quantization
for bits in [4, 8]:
result = self.quantize_compress(weight, bits)
result["approach"] = f"quant_{bits}bit"
results.append(result)
# Pick best: highest quality that meets compression target
valid = [r for r in results if r["cosine_similarity"] >= self.quality_threshold]
if not valid:
# If nothing meets quality, pick best quality regardless
valid = sorted(results, key=lambda x: x["cosine_similarity"], reverse=True)
# Among valid, pick smallest
best = min(valid, key=lambda x: x["compressed_bytes"])
# Clean up - remove tensors for reporting
report = {k: v for k, v in best.items() if not isinstance(v, torch.Tensor)}
return best
def compress_model(self, model_name: str = "Qwen/Qwen2.5-0.5B-Instruct",
output_dir: str = "./formula_weights"):
"""
Compress entire model and save formulas.
"""
from transformers import AutoModelForCausalLM
print(f"Loading model: {model_name}")
model = AutoModelForCausalLM.from_pretrained(
model_name,
dtype=torch.float16,
)
os.makedirs(output_dir, exist_ok=True)
total_original = 0
total_compressed = 0
formula_index = {}
print("\n" + "="*80)
print("FORMULA DISCOVERY ENGINE - Analyzing all model weights")
print("="*80)
for name, param in model.named_parameters():
weight = param.data
print(f"\n{'─'*60}")
print(f"Layer: {name}")
print(f" Shape: {list(weight.shape)} | Elements: {weight.numel():,} | Size: {weight.numel()*2/1024/1024:.2f} MB")
# Find best formula
formula = self.find_best_formula(name, weight)
print(f" Formula: {formula['type']} | Compression: {formula['compression_ratio']:.3f}")
print(f" Quality: cos_sim={formula['cosine_similarity']:.6f} | MSE={formula['mse']:.2e}")
print(f" Size: {formula['original_bytes']/1024/1024:.2f} MB → {formula['compressed_bytes']/1024/1024:.2f} MB")
total_original += formula["original_bytes"]
total_compressed += formula["compressed_bytes"]
# Save formula tensors
formula_file = name.replace(".", "_") + ".pt"
save_data = {}
if formula["type"] == "svd":
save_data = {
"type": "svd",
"rank": formula["rank"],
"shape": formula["original_shape"],
"U": formula["U"],
"S": formula["S"],
"Vh": formula["Vh"],
}
elif formula["type"] == "quantize":
save_data = {
"type": "quantize",
"bits": formula["bits"],
"shape": formula["original_shape"],
"W_q": formula["W_q"],
"scale": formula["scale"],
"zero_point": formula["zero_point"],
"w_min": formula["w_min"],
}
elif formula["type"] == "raw":
save_data = {
"type": "raw",
"shape": formula["original_shape"],
"data": formula["data"],
}
torch.save(save_data, os.path.join(output_dir, formula_file))
# Index entry
formula_index[name] = {
"file": formula_file,
"type": formula["type"],
"shape": formula["original_shape"],
"compression_ratio": formula["compression_ratio"],
"cosine_similarity": formula["cosine_similarity"],
"mse": formula["mse"],
}
if formula["type"] == "svd":
formula_index[name]["rank"] = formula["rank"]
elif formula["type"] == "quantize":
formula_index[name]["bits"] = formula["bits"]
# Save index
with open(os.path.join(output_dir, "formula_index.json"), "w") as f:
json.dump(formula_index, f, indent=2)
# Summary
print("\n" + "="*80)
print("COMPRESSION SUMMARY")
print("="*80)
print(f"Original model size: {total_original/1024/1024:.1f} MB")
print(f"Compressed formula size: {total_compressed/1024/1024:.1f} MB")
print(f"Overall compression: {total_compressed/total_original:.3f} ({(1-total_compressed/total_original)*100:.1f}% smaller)")
print(f"Formula files saved to: {output_dir}/")
self.metadata = {
"model_name": model_name,
"original_size_mb": total_original / 1024 / 1024,
"compressed_size_mb": total_compressed / 1024 / 1024,
"compression_ratio": total_compressed / total_original,
"num_layers": len(formula_index),
}
with open(os.path.join(output_dir, "metadata.json"), "w") as f:
json.dump(self.metadata, f, indent=2)
return formula_index
class FormulaReconstructor:
"""
Reconstructs model weights from formula files.
This is what runs at inference time instead of loading full weights.
"""
def __init__(self, formula_dir: str):
self.formula_dir = formula_dir
with open(os.path.join(formula_dir, "formula_index.json"), "r") as f:
self.index = json.load(f)
def reconstruct_weight(self, name: str) -> torch.Tensor:
"""Reconstruct a single weight from its formula."""
info = self.index[name]
data = torch.load(
os.path.join(self.formula_dir, info["file"]),
map_location="cpu",
weights_only=True
)
if data["type"] == "svd":
# W = U @ diag(S) @ Vh
U = data["U"].float()
S = data["S"].float()
Vh = data["Vh"].float()
W = U @ torch.diag(S) @ Vh
return W.half()
elif data["type"] == "quantize":
# W = W_q * scale + w_min
W_q = data["W_q"].float()
scale = data["scale"].float()
w_min = data["w_min"].float()
W = W_q * scale + w_min
return W.half()
elif data["type"] == "raw":
return data["data"]
raise ValueError(f"Unknown formula type: {data['type']}")
def reconstruct_all(self) -> Dict[str, torch.Tensor]:
"""Reconstruct all weights."""
weights = {}
for name in self.index:
weights[name] = self.reconstruct_weight(name)
return weights
def load_into_model(self, model):
"""Load reconstructed weights into a model."""
state_dict = {}
for name in self.index:
state_dict[name] = self.reconstruct_weight(name)
model.load_state_dict(state_dict, strict=False)
return model
if __name__ == "__main__":
# Run the formula discovery engine
engine = FormulaEngine(
compression_ratio=0.4, # Target 60% compression
quality_threshold=0.995 # Minimum cosine similarity
)
formula_index = engine.compress_model(
model_name="Qwen/Qwen2.5-0.5B-Instruct",
output_dir="./formula_weights"
)
print("\n\nFormula discovery complete!")
print(f"Results saved to ./formula_weights/")
|