File size: 15,732 Bytes
5347c49 5911d9a 5347c49 5911d9a 5347c49 | 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 | """
Layer 1: Property-Based Falsifier
==================================
Attacks the extracted C++20 kernels via Hypothesis.
Every test corresponds to a named Lean 4 theorem.
Run:
pip install hypothesis pytest
cmake -S .. -B ../build && cmake --build ../build
pytest falsify_properties.py -x -v --tb=short
Pass criteria: zero failures after 100k shrunk examples.
If Hypothesis finds one counterexample, the Lean theorem was weaker than intent.
"""
import ctypes
import math
import os
import pathlib
import sys
import pytest
from hypothesis import given, settings, assume, Phase
from hypothesis import strategies as st
# ── Load the shared library ───────────────────────────────────────────────────
def _add_dll_dirs():
"""Add MinGW/Strawberry runtime dirs so ctypes can resolve DLL deps on Windows."""
for d in [
r"C:\Strawberry\c\bin",
r"C:\Program Files\mingw64\bin",
r"C:\mingw64\bin",
r"C:\msys64\mingw64\bin",
]:
if os.path.isdir(d):
try:
os.add_dll_directory(d)
except AttributeError:
pass # Python < 3.8
if os.name == "nt":
_add_dll_dirs()
def _load_lib():
repo = pathlib.Path(__file__).resolve().parent.parent
candidates = [
repo / "build" / "libsovereign_array.so",
repo / "build" / "libsovereign_array.dll",
repo / "build" / "sovereign_array.dll",
repo / "build" / "libsovereign_array.dylib",
]
for p in candidates:
if p.exists():
return ctypes.CDLL(str(p))
raise FileNotFoundError(
"Build the shared library first:\n"
" cmake -S .. -B ../build && cmake --build ../build\n"
f"Looked in: {[str(c) for c in candidates]}"
)
try:
_lib = _load_lib()
# void sovarr_softmax(const float*, float*, size_t)
_lib.sovarr_softmax.argtypes = [
ctypes.POINTER(ctypes.c_float),
ctypes.POINTER(ctypes.c_float),
ctypes.c_size_t,
]
_lib.sovarr_softmax.restype = None
# void sovarr_face_centroid(const int*, size_t, float*, size_t)
_lib.sovarr_face_centroid.argtypes = [
ctypes.POINTER(ctypes.c_int),
ctypes.c_size_t,
ctypes.POINTER(ctypes.c_float),
ctypes.c_size_t,
]
_lib.sovarr_face_centroid.restype = None
# void sovarr_nand_attention(const float*, float*, float*, float*, size_t)
_lib.sovarr_nand_attention.argtypes = [
ctypes.POINTER(ctypes.c_float),
ctypes.POINTER(ctypes.c_float),
ctypes.POINTER(ctypes.c_float),
ctypes.POINTER(ctypes.c_float),
ctypes.c_size_t,
]
_lib.sovarr_nand_attention.restype = None
# int sovarr_nand(int, int) etc.
for fn in ("sovarr_nand", "sovarr_not", "sovarr_and", "sovarr_or"):
getattr(_lib, fn).argtypes = [ctypes.c_int, ctypes.c_int]
getattr(_lib, fn).restype = ctypes.c_int
LIB_OK = True
except FileNotFoundError as e:
print(f"WARNING: {e}", file=sys.stderr)
LIB_OK = False
def softmax_c(xs: list[float]) -> list[float]:
n = len(xs)
in_arr = (ctypes.c_float * n)(*xs)
out_arr = (ctypes.c_float * n)()
_lib.sovarr_softmax(in_arr, out_arr, n)
return list(out_arr)
def face_centroid_c(support: list[int], n: int) -> list[float]:
sup_arr = (ctypes.c_int * len(support))(*support)
out_arr = (ctypes.c_float * n)()
_lib.sovarr_face_centroid(sup_arr, len(support), out_arr, n)
return list(out_arr)
def nand_attention_c(q, k, v):
n = len(q)
def fa(xs): return (ctypes.c_float * n)(*xs)
out = (ctypes.c_float * n)()
_lib.sovarr_nand_attention(fa(q), fa(k), fa(v), out, n)
return list(out)
def softmax_ref(xs: list[float]) -> list[float]:
"""Pure-Python reference implementation (exact, slow)."""
m = max(xs) # shift for numerical stability
exps = [math.exp(x - m) for x in xs]
s = sum(exps)
return [e / s for e in exps]
skip_no_lib = pytest.mark.skipif(not LIB_OK, reason="shared lib not built")
# ─────────────────────────────────────────────────────────────────────────────
# PAPER I — NAND universality
# Lean theorems: notGate_eq, andGate_eq, orGate_eq
# ─────────────────────────────────────────────────────────────────────────────
@skip_no_lib
class TestPaperI_NAND:
@given(st.integers(0, 1), st.integers(0, 1))
@settings(max_examples=4, phases=[Phase.generate]) # truth table is 4 entries
def test_nand_truth_table(self, a, b):
"""sovarr_nand(a,b) == not(a and b) [andGate_eq / notGate_eq]"""
expected = 0 if (a == 1 and b == 1) else 1
assert _lib.sovarr_nand(a, b) == expected
@given(st.integers(0, 1))
@settings(max_examples=2)
def test_not_via_nand(self, a):
"""nand(a,a) == not(a) [notGate_eq]"""
assert _lib.sovarr_not(a, a) == (0 if a == 1 else 1)
@given(st.integers(0, 1), st.integers(0, 1))
@settings(max_examples=4)
def test_and_via_nand(self, a, b):
"""nand(nand(a,b), nand(a,b)) == a and b [andGate_eq]"""
assert _lib.sovarr_and(a, b) == (1 if a == 1 and b == 1 else 0)
@given(st.integers(0, 1), st.integers(0, 1))
@settings(max_examples=4)
def test_or_via_nand(self, a, b):
"""nand(nand(a,a), nand(b,b)) == a or b [orGate_eq]"""
assert _lib.sovarr_or(a, b) == (1 if a == 1 or b == 1 else 0)
@given(st.integers(0, 1), st.integers(0, 1), st.integers(0, 1))
@settings(max_examples=8)
def test_demorgan(self, a, b, c):
"""not(a or b) == not(a) and not(b) — derived via NAND"""
lhs = _lib.sovarr_not(_lib.sovarr_or(a, b), _lib.sovarr_or(a, b))
rhs = _lib.sovarr_and(_lib.sovarr_not(a, a), _lib.sovarr_not(b, b))
assert lhs == rhs
# ─────────────────────────────────────────────────────────────────────────────
# PAPER II — Simplex / Softmax / Face Centroid
# Lean theorems: softmax_is_pmap, softmax_shift_invariant,
# faceCentroid_nonneg, faceCentroid_support, vertex_centroid_eq
# ─────────────────────────────────────────────────────────────────────────────
@skip_no_lib
class TestPaperII_Simplex:
@given(st.lists(st.floats(-20, 20, allow_nan=False, allow_infinity=False),
min_size=2, max_size=256))
@settings(max_examples=50_000, phases=[Phase.generate, Phase.shrink])
def test_softmax_sums_to_one(self, logits):
"""Σ softmax(z)_i = 1 [softmax_is_pmap, sum normalisation]"""
out = softmax_c(logits)
assert abs(sum(out) - 1.0) < 1e-5, f"sum={sum(out)} logits={logits[:4]}"
@given(st.lists(st.floats(-10, 10, allow_nan=False, allow_infinity=False),
min_size=2, max_size=256),
st.floats(-50, 50, allow_nan=False, allow_infinity=False))
@settings(max_examples=20_000, phases=[Phase.generate, Phase.shrink])
def test_softmax_shift_invariant(self, logits, c):
"""softmax(z+c) == softmax(z) [softmax_shift_invariant]"""
shifted = [x + c for x in logits]
out1 = softmax_c(logits)
out2 = softmax_c(shifted)
for i, (a, b) in enumerate(zip(out1, out2)):
assert abs(a - b) < 1e-4, f"shift broke at i={i}: {a} vs {b} (c={c})"
@given(st.lists(st.floats(-10, 10, allow_nan=False, allow_infinity=False),
min_size=2, max_size=256))
@settings(max_examples=20_000)
def test_softmax_matches_reference(self, logits):
"""sovarr_softmax matches pure-Python reference [kernel correctness]"""
c_out = softmax_c(logits)
py_out = softmax_ref(logits)
for i, (a, b) in enumerate(zip(c_out, py_out)):
assert abs(a - b) < 1e-4, f"mismatch at i={i}: C={a} ref={b}"
@given(st.integers(2, 64),
st.lists(st.integers(0, 63), min_size=1, max_size=64, unique=True))
@settings(max_examples=10_000, phases=[Phase.generate, Phase.shrink])
def test_face_centroid_sums_to_one(self, n, raw_support):
"""Σ faceCentroid(F)_i = 1 [faceCentroid = uniform on F]"""
support = [i for i in raw_support if i < n]
assume(len(support) > 0)
out = face_centroid_c(support, n)
assert abs(sum(out) - 1.0) < 1e-6, f"sum={sum(out)} F={support} n={n}"
@given(st.integers(2, 64),
st.lists(st.integers(0, 63), min_size=1, max_size=64, unique=True))
@settings(max_examples=10_000)
def test_face_centroid_support_matches(self, n, raw_support):
"""faceCentroid(F)_i ≠ 0 ↔ i ∈ F [faceCentroid_support]"""
support = sorted(set(i for i in raw_support if i < n))
assume(len(support) > 0)
out = face_centroid_c(support, n)
support_set = set(support)
for i, v in enumerate(out):
if i in support_set:
assert v > 0, f"active coord {i} is zero"
else:
assert v == 0.0, f"inactive coord {i} is nonzero: {v}"
@given(st.integers(2, 64),
st.integers(0, 63))
@settings(max_examples=5_000)
def test_vertex_centroid_is_indicator(self, n, raw_v):
"""faceCentroid({v})_i = 1 if i==v else 0 [vertex_centroid_eq]"""
v = raw_v % n
out = face_centroid_c([v], n)
assert abs(out[v] - 1.0) < 1e-7, f"vertex {v}: {out[v]} ≠ 1"
for i, val in enumerate(out):
if i != v:
assert val == 0.0, f"non-vertex {i} nonzero: {val}"
@given(st.lists(st.floats(-5, 5, allow_nan=False, allow_infinity=False),
min_size=2, max_size=64))
@settings(max_examples=10_000)
def test_softmax_all_positive(self, logits):
"""softmax(z)_i > 0 for all i [softmax maps to interior of simplex]"""
out = softmax_c(logits)
for i, v in enumerate(out):
assert v > 0, f"softmax[{i}]={v} ≤ 0 on {logits}"
# ─────────────────────────────────────────────────────────────────────────────
# PAPER III — NAND Attention
# Lean theorem: attention_is_pmap
# ─────────────────────────────────────────────────────────────────────────────
def attention_ref(q, k, v):
"""Pure-Python reference: scores_i = Σ q_i*k_j, w=softmax(scores), out_i = Σ w_i*v_j"""
n = len(q)
scores = [sum(q[i] * k[j] for j in range(n)) for i in range(n)]
w = softmax_ref(scores)
return [sum(w[i] * v[j] for j in range(n)) for i in range(n)]
@skip_no_lib
class TestPaperIII_Attention:
@given(st.integers(1, 32),
st.data())
@settings(max_examples=5_000, phases=[Phase.generate, Phase.shrink])
def test_attention_matches_reference(self, n, data):
"""sovarr_nand_attention ≡ reference attention [attention_is_pmap]"""
flt = st.floats(-5, 5, allow_nan=False, allow_infinity=False)
q = data.draw(st.lists(flt, min_size=n, max_size=n))
k = data.draw(st.lists(flt, min_size=n, max_size=n))
v = data.draw(st.lists(flt, min_size=n, max_size=n))
c_out = nand_attention_c(q, k, v)
py_out = attention_ref(q, k, v)
for i, (a, b) in enumerate(zip(c_out, py_out)):
assert abs(a - b) < 1e-3, \
f"attention divergence at i={i}: C={a:.6f} ref={b:.6f}"
@given(st.integers(1, 32), st.data())
@settings(max_examples=2_000)
def test_attention_output_size(self, n, data):
"""output has same size as input [shape preservation]"""
flt = st.floats(-3, 3, allow_nan=False, allow_infinity=False)
q = data.draw(st.lists(flt, min_size=n, max_size=n))
k = data.draw(st.lists(flt, min_size=n, max_size=n))
v = data.draw(st.lists(flt, min_size=n, max_size=n))
out = nand_attention_c(q, k, v)
assert len(out) == n
@given(st.integers(1, 16), st.data())
@settings(max_examples=2_000)
def test_attention_uniform_k_is_constant(self, n, data):
"""When k is uniform, all scores are equal → attention weights are uniform."""
flt = st.floats(-3, 3, allow_nan=False, allow_infinity=False)
q = data.draw(st.lists(flt, min_size=n, max_size=n))
k = [1.0] * n # uniform key
v = data.draw(st.lists(flt, min_size=n, max_size=n))
out = nand_attention_c(q, k, v)
# All outputs should be identical (uniform weight over v)
mean = sum(out) / n
for i, val in enumerate(out):
assert abs(val - mean) < 1e-4, \
f"uniform-k: out[{i}]={val} ≠ mean={mean}"
# ─────────────────────────────────────────────────────────────────────────────
# BROADCAST — pullback semantics
# Lean theorem: broadcast_is_pullback
# ─────────────────────────────────────────────────────────────────────────────
@skip_no_lib
class TestBroadcast:
@given(st.lists(st.floats(-100, 100, allow_nan=False, allow_infinity=False),
min_size=1, max_size=256))
@settings(max_examples=10_000)
def test_broadcast_1d_pointwise(self, xs):
"""broadcast_1d(v, w)_i = v_i + w_i [broadcast_is_pullback, 1D case]"""
n = len(xs)
# split into two halves (or use same list for both)
v_arr = (ctypes.c_float * n)(*xs)
w_arr = (ctypes.c_float * n)(*xs)
out = (ctypes.c_float * n)()
_lib.sovarr_broadcast_1d.argtypes = [
ctypes.POINTER(ctypes.c_float),
ctypes.POINTER(ctypes.c_float),
ctypes.POINTER(ctypes.c_float),
ctypes.c_size_t,
]
_lib.sovarr_broadcast_1d.restype = None
_lib.sovarr_broadcast_1d(v_arr, w_arr, out, n)
for i, (a, b, o) in enumerate(zip(xs, xs, out)):
assert abs(o - (a + b)) < 1e-4, f"broadcast[{i}]: {o} ≠ {a+b}"
# ─────────────────────────────────────────────────────────────────────────────
# Standalone: run without pytest
# ─────────────────────────────────────────────────────────────────────────────
if __name__ == "__main__":
import subprocess, sys
result = subprocess.run(
[sys.executable, "-m", "pytest", __file__, "-x", "-v", "--tb=short"],
cwd=pathlib.Path(__file__).parent,
)
sys.exit(result.returncode)
|