Skip to content

fix: accept avx2 and vnni outputs of int8 models in tests #804

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
joein merged 3 commits into from
Oct 12, 2026
Merged

fix: accept avx2 and vnni outputs of int8 models in tests #804

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1387e6c
fix: accept avx2 and vnni outputs of int8 models in tests
joein Oct 12, 2026
7806a6a
tests: add a temporary ci job for the int8 models on several runners
joein Oct 12, 2026
37332b1
remove tmp ci
joein Oct 12, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Binary file added tests/misc/logo.png
View file
Open in desktop
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
27 changes: 19 additions & 8 deletions tests/test_image_onnx_embeddings.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,7 +5,6 @@

import numpy as np
import pytest
import requests
from PIL import Image

from fastembed import ImageEmbedding
Expand All @@ -29,8 +28,15 @@
"nomic-ai/nomic-embed-vision-v1.5": np.array(
[0.0048, -0.0254, 0.0067, -0.0296, -0.0435, -0.0123, 0.0024, -0.0361, -0.0703, -0.0186]
),
"nomic-ai/nomic-embed-vision-v1.5-Q": np.array(
[-0.0011, -0.0477, 0.0024, -0.049, -0.0458, -0.0314, 0.017, -0.0383, -0.0537, -0.021]
# onnxruntime's avx2 int8 kernel saturates on this model's full-range int8 weights, while vnni
# and amx compute them exactly, ci runners come with either
"nomic-ai/nomic-embed-vision-v1.5-Q": (
np.array( # x86-64 avx2
[-0.0006, -0.0446, 0.0069, -0.0553, -0.0459, -0.0261, 0.0058, -0.0463, -0.0606, -0.019]
),
np.array( # x86-64 vnni
[0.0020, -0.0376, 0.0094, -0.0402, -0.0457, -0.0216, 0.0100, -0.0378, -0.0650, -0.0205]
),
),
"google/siglip2-base-patch16-224": np.array(
[-0.02095927, -0.0075177, -0.00144479, -0.0080948, 0.05031789]
Expand Down Expand Up @@ -82,21 +88,26 @@ def test_embedding(model_cache, model_name: str) -> None:
dim = model_desc.dim

with model_cache(model_desc.model) as model:
# quantized models scale activations over the whole batch, so the first embedding
# depends on every image here, keep them local files
images = [
TEST_MISC_DIR / "image.jpeg",
str(TEST_MISC_DIR / "small_image.jpeg"),
Image.open((TEST_MISC_DIR / "small_image.jpeg")),
Image.open(BytesIO(requests.get("https://qdrant.tech/img/logo.png").content)),
Image.open(BytesIO((TEST_MISC_DIR / "logo.png").read_bytes())),
]
embeddings = list(model.embed(images))
embeddings = np.stack(embeddings, axis=0)
assert embeddings.shape == (len(images), dim)

canonical_vector = CANONICAL_VECTOR_VALUES[model_desc.model]

assert np.allclose(
embeddings[0, : canonical_vector.shape[0]], canonical_vector, atol=1e-3
), model_desc.model
canonical_variants = (
canonical_vector if isinstance(canonical_vector, tuple) else (canonical_vector,)
)
prefix = embeddings[0, : canonical_variants[0].shape[0]]
assert any(
np.allclose(prefix, expected, atol=1e-3) for expected in canonical_variants
), f"{model_desc.model}: {prefix}"

assert np.allclose(embeddings[1], embeddings[2]), model_desc.model

Expand Down
5 changes: 3 additions & 2 deletions tests/test_text_cross_encoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,8 +13,9 @@
"BAAI/bge-reranker-base": np.array([6.15733337, -3.65939403]),
"BAAI/bge-reranker-v2-m3": np.array([8.78182220, -5.46485329]),
"BAAI/bge-reranker-v2-m3-int8": (
np.array([8.59202957, -5.39362288]), # ONNX Runtime 1.27.0
np.array([8.51154804, -5.56898642]), # ONNX Runtime 1.30.0
np.array([8.59202957, -5.39362288]), # arm64, ONNX Runtime 1.27.0
np.array([8.51154804, -5.56898642]), # arm64, ONNX Runtime 1.30.0
np.array([8.85104847, -5.73784494]), # x86-64 avx2 and vnni, ONNX Runtime 1.30.0
),
"jinaai/jina-reranker-v1-tiny-en": np.array([2.5911, 0.1122]),
"jinaai/jina-reranker-v1-turbo-en": np.array([1.8295, -2.8908]),
Expand Down
17 changes: 12 additions & 5 deletions tests/test_text_onnx_embeddings.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,8 +56,11 @@
"nomic-ai/nomic-embed-text-v1.5": np.array(
[-0.15407836, -0.03053198, -3.9138033, 0.1910364, 0.13224715]
),
"nomic-ai/nomic-embed-text-v1.5-Q": np.array(
[0.0802303, 0.3700881, -4.3053818, 0.4431803, -0.271572]
# onnxruntime's avx2 int8 kernel saturates on this model's full-range int8 weights, while vnni
# and amx compute them exactly, ci runners come with either
"nomic-ai/nomic-embed-text-v1.5-Q": (
np.array([0.0802303, 0.3700881, -4.3053818, 0.4431803, -0.271572]), # x86-64 avx2
np.array([0.12422204, 0.14779393, -3.8240886, 0.47303754, -0.14925367]), # x86-64 vnni
),
"thenlper/gte-large": np.array(
[-0.00986551, -0.00018734, 0.00605892, -0.03289612, -0.0387564],
Expand Down Expand Up @@ -188,9 +191,13 @@ def test_embedding(model_cache, model_name: str) -> None:
assert embeddings.shape == (2, dim)

canonical_vector = CANONICAL_VECTOR_VALUES[model_desc.model]
assert np.allclose(
embeddings[0, : canonical_vector.shape[0]], canonical_vector, atol=1e-3
), model_desc.model
canonical_variants = (
canonical_vector if isinstance(canonical_vector, tuple) else (canonical_vector,)
)
prefix = embeddings[0, : canonical_variants[0].shape[0]]
assert any(
np.allclose(prefix, expected, atol=1e-3) for expected in canonical_variants
), f"{model_desc.model}: {prefix}"


def test_query_embedding(model_cache) -> None:
Expand Down
Loading