lib/gguf/tests/quants.viv

# lib/gguf/quants.viv on the quantized tensors of tests/tiny.gguf, which hold
# random blocks (tools/gguf/make_fixture.py); tests/tiny.ref holds what
# gguf-py 0.19.0's dequantize gives for each. Paths are from the repository root.
# The Q4_1 and Q4_K blocks have minimums far larger than their scales, so
# single precision rounds q * d + m: computed in double without fround, 164
# of the 2048 Q4_1 numbers and 532 of the 2048 Q4_K numbers would differ
# from gguf-py's (counted with NumPy on 2026-10-14).
#! file lib/gguf/tests/tiny.gguf
#! file lib/gguf/tests/tiny.ref
from gguf.tensors import read_gguf, find_tensor, f32_values
from gguf.quants import tensor_values, tensor_rows

TINY = "lib/gguf/tests/tiny.gguf"
REF = "lib/gguf/tests/tiny.ref"
F32_BYTES: Int = 4

g = read_gguf(TINY)

#| What gguf-py's dequantize gives for the tensor called `name`.
def ref_values(name: Str) -> Arr {
    k = find_tensor(g, name)
    n = g.tensors[k].n_elems
    return f32_values(read_bytes(REF, k * n * F32_BYTES, n * F32_BYTES), n)
}

#| How many numbers of tensor `name` differ from gguf-py's.
def differ(name: Str) -> Int {
    x = tensor_values(TINY, g.tensors[find_tensor(g, name)])
    ref = ref_values(name)
    return int(sum(x != ref))
}

print("numbers that differ from gguf-py: q8_0", differ("t.q8_0"), "q4_0", differ("t.q4_0"), "q4_1", differ("t.q4_1"), "q4_k", differ("t.q4_k"), "q6_k", differ("t.q6_k"))
test "Q8_0, number for number as gguf-py dequantizes it" {
    assert_true(differ("t.q8_0") == 0)
    # the second row on its own
    t = g.tensors[find_tensor(g, "t.q8_0")]
    assert_allclose(tensor_rows(TINY, t, 1, 1), ref_values("t.q8_0")[t.dims[0]:2 * t.dims[0]], 0.0)
}

test "Q4_0, number for number as gguf-py dequantizes it" {
    assert_true(differ("t.q4_0") == 0)
}

test "Q4_1, number for number as gguf-py dequantizes it" {
    assert_true(differ("t.q4_1") == 0)
}

test "Q4_K, number for number as gguf-py dequantizes it" {
    assert_true(differ("t.q4_k") == 0)
}

test "Q6_K, number for number as gguf-py dequantizes it" {
    assert_true(differ("t.q6_k") == 0)
}

# SPDX-License-Identifier: MIT OR Apache-2.0
# SPDX-FileCopyrightText: 2026 The Viv authors