Franok / source /tests /quantize.py
yunusemrejr's picture
Publish source, film and evaluations from 8b3b199f3d72
eb373ad verified
Raw History Blame Contribute Delete
1.94 kB
"""Verify signed nibble serialization and covariance-aware error minimization."""
import numpy as np
from training.export import quantize_grouped, refine_block, refine_pairs
def main():
rng = np.random.default_rng(83742)
group = rng.standard_normal((128, 32)).astype("float32")
group[0] = 0
group[1, 1] = 6
group[2, 2] = -7
samples = rng.standard_normal((128, 54, 32)).astype("float32")
cov = samples.transpose(0, 2, 1) @ samples / 54 + 0.25 * np.eye(32, dtype="float32")
scale, q = quantize_grouped(
group, 4, True, np.diagonal(cov, axis1=1, axis2=2), True
)
def errors(s, codes):
d = group - s * codes
return np.einsum("bi,bij,bj->b", d, cov, d)
old = errors(scale, q)
step, codes = refine_block(group, scale.copy(), q, cov, -8)
assert np.all(errors(step, codes) <= old + 1e-5)
assert (step < 0).any() and (step > 0).any()
assert codes.min() == -8 and codes.max() == 7 and np.isfinite(step).all()
biased = (codes + 8).astype("uint8")
packed = biased[:, :16] | (biased[:, 16:] << 4)
restored = np.concatenate((packed & 15, packed >> 4), 1).astype("int8") - 8
np.testing.assert_array_equal(codes, restored)
weights = group.reshape(32, 2, 64)
activations = rng.standard_normal((2, 90, 64)).astype("float32")
gram = (
activations.transpose(0, 2, 1) @ activations / 90
+ np.eye(64, dtype="float32") * 0.25
)
def pair_error(s, q):
residual = weights - (s * q).reshape(weights.shape)
return np.einsum("rgi,gij,rgj->rg", residual, gram, residual)
neighbour_scale, neighbour_codes = refine_pairs(group, step, codes, gram)
assert np.all(
pair_error(neighbour_scale, neighbour_codes) <= pair_error(step, codes) + 1e-5
)
print(
"All 16 codes round-trip. Every refined group preserved or reduced calibration error."
)
if __name__ == "__main__":
main()