Skip to content

Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.

Prefer AVX2 over AVX512 by default on Windows (#2685)

Evo5c74473d0420
diff --git a/openviking/storage/vectordb/engine/__init__.py b/openviking/storage/vectordb/engine/__init__.py
index 564015652..29e71d3d3 100644
--- a/openviking/storage/vectordb/engine/__init__.py
+++ b/openviking/storage/vectordb/engine/__init__.py
@@ -22,6 +22,7 @@ _BACKEND_MODULES = {
}
_X86_DISPLAY_ORDER = ("x86_sse3", "x86_avx2", "x86_avx512")
_X86_PRIORITY = ("x86_avx512", "x86_avx2", "x86_sse3")
+_WINDOWS_X86_PRIORITY = ("x86_avx2", "x86_sse3")
_REQUEST_ALIASES = {
"sse3": "x86_sse3",
"avx2": "x86_avx2",
@@ -109,7 +110,8 @@ def _select_variant() -> tuple[str | None, tuple[str, ...], str | None]:
return "native", available, None
supported_x86 = _supported_x86_variants()
- for variant in _X86_PRIORITY:
+ priority = _WINDOWS_X86_PRIORITY if sys.platform == "win32" else _X86_PRIORITY
+ for variant in priority:
if variant in available and variant in supported_x86:
return variant, available, None
diff --git a/tests/misc/test_vectordb_engine_loader.py b/tests/misc/test_vectordb_engine_loader.py
index 7386727a6..79a6e6a43 100644
--- a/tests/misc/test_vectordb_engine_loader.py
+++ b/tests/misc/test_vectordb_engine_loader.py
@@ -24,7 +24,7 @@ def _install_package_stubs(monkeypatch):
def _load_engine_module(
- monkeypatch, *, machine, available_backends, cpu_variants, env_variant=None
+ monkeypatch, *, machine, available_backends, cpu_variants, env_variant=None, sys_platform=None
):
_install_package_stubs(monkeypatch)
for backend_name in available_backends:
@@ -76,6 +76,8 @@ def _load_engine_module(
monkeypatch.setattr(importlib, "import_module", fake_import_module)
monkeypatch.setattr(importlib.util, "find_spec", fake_find_spec)
+ if sys_platform is not None:
+ monkeypatch.setattr(sys, "platform", sys_platform)
spec = importlib.util.spec_from_file_location(
"openviking.storage.vectordb.engine",
@@ -141,6 +143,43 @@ def test_engine_loader_auto_selects_best_supported_x86_backend(monkeypatch):
assert module.AVAILABLE_ENGINE_VARIANTS == ("x86_sse3", "x86_avx2", "x86_avx512")
+def test_engine_loader_auto_prefers_avx2_over_avx512_on_windows(monkeypatch):
+ module = _load_engine_module(
+ monkeypatch,
+ machine="AMD64",
+ available_backends={"x86_sse3", "x86_avx2", "x86_avx512"},
+ cpu_variants={"x86_sse3", "x86_avx2", "x86_avx512"},
+ sys_platform="win32",
+ )
+
+ assert module.ENGINE_VARIANT == "x86_avx2"
+
+
+def test_engine_loader_auto_skips_avx512_on_windows(monkeypatch):
+ module = _load_engine_module(
+ monkeypatch,
+ machine="AMD64",
+ available_backends={"x86_sse3", "x86_avx512"},
+ cpu_variants={"x86_sse3", "x86_avx512"},
+ sys_platform="win32",
+ )
+
+ assert module.ENGINE_VARIANT == "x86_sse3"
+
+
+def test_engine_loader_allows_explicit_avx512_on_windows(monkeypatch):
+ module = _load_engine_module(
+ monkeypatch,
+ machine="AMD64",
+ available_backends={"x86_sse3", "x86_avx2", "x86_avx512"},
+ cpu_variants={"x86_sse3", "x86_avx2", "x86_avx512"},
+ env_variant="avx512",
+ sys_platform="win32",
+ )
+
+ assert module.ENGINE_VARIANT == "x86_avx512"
+
+
def test_engine_loader_uses_native_backend_on_non_x86(monkeypatch):
module = _load_engine_module(
monkeypatch,
@@ -178,51 +217,55 @@ def test_engine_loader_wraps_abi3_backend_with_python_api(monkeypatch):
_new_schema=lambda fields: calls.append(("new_schema", fields)) or schema_handle,
_schema_get_total_byte_length=lambda handle: calls.append(("schema_total", handle)) or 123,
_new_bytes_row=lambda handle: calls.append(("new_bytes_row", handle)) or bytes_row_handle,
- _bytes_row_serialize=lambda handle, row: calls.append(("bytes_row_serialize", handle, row))
- or b"blob",
- _bytes_row_serialize_batch=lambda handle, rows: calls.append(
- ("bytes_row_serialize_batch", handle, rows)
- )
- or [b"blob-a", b"blob-b"],
- _bytes_row_deserialize=lambda handle, payload: calls.append(
- ("bytes_row_deserialize", handle, payload)
- )
- or {"id": 42, "name": "viking"},
- _bytes_row_deserialize_field=lambda handle, payload, field_name: calls.append(
- ("bytes_row_deserialize_field", handle, payload, field_name)
- )
- or (42 if field_name == "id" else "viking"),
- _new_index_engine=lambda config: calls.append(("new_index_engine", config))
- or engine_handle,
+ _bytes_row_serialize=lambda handle, row: (
+ calls.append(("bytes_row_serialize", handle, row)) or b"blob"
+ ),
+ _bytes_row_serialize_batch=lambda handle, rows: (
+ calls.append(("bytes_row_serialize_batch", handle, rows)) or [b"blob-a", b"blob-b"]
+ ),
+ _bytes_row_deserialize=lambda handle, payload: (
+ calls.append(("bytes_row_deserialize", handle, payload)) or {"id": 42, "name": "viking"}
+ ),
+ _bytes_row_deserialize_field=lambda handle, payload, field_name: (
+ calls.append(("bytes_row_deserialize_field", handle, payload, field_name))
+ or (42 if field_name == "id" else "viking")
+ ),
+ _new_index_engine=lambda config: (
+ calls.append(("new_index_engine", config)) or engine_handle
+ ),
_index_engine_add_data=lambda handle, items: calls.append(("add_data", handle, items)) or 3,
- _index_engine_delete_data=lambda handle, items: calls.append(("delete_data", handle, items))
- or 2,
- _index_engine_search=lambda handle, req: calls.append(("search", handle, req))
- or {
- "result_num": 2,
- "labels": [101, 202],
- "scores": [0.9, 0.8],
- "extra_json": '{"ok":true}',
- },
+ _index_engine_delete_data=lambda handle, items: (
+ calls.append(("delete_data", handle, items)) or 2
+ ),
+ _index_engine_search=lambda handle, req: (
+ calls.append(("search", handle, req))
+ or {
+ "result_num": 2,
+ "labels": [101, 202],
+ "scores": [0.9, 0.8],
+ "extra_json": '{"ok":true}',
+ }
+ ),
_index_engine_dump=lambda handle, path: calls.append(("dump", handle, path)) or 11,
- _index_engine_get_state=lambda handle: calls.append(("get_state", handle))
- or {"update_timestamp": 123, "element_count": 7},
+ _index_engine_get_state=lambda handle: (
+ calls.append(("get_state", handle)) or {"update_timestamp": 123, "element_count": 7}
+ ),
_new_persist_store=lambda path: calls.append(("new_persist_store", path)) or store_handle,
_new_volatile_store=lambda: calls.append(("new_volatile_store",)) or store_handle,
_store_exec_op=lambda handle, ops: calls.append(("store_exec_op", handle, ops)) or 1,
- _store_get_data=lambda handle, keys: calls.append(("store_get_data", handle, keys))
- or [b"one", b"two"],
- _store_put_data=lambda handle, keys, values: calls.append(
- ("store_put_data", handle, keys, values)
- )
- or 0,
- _store_delete_data=lambda handle, keys: calls.append(("store_delete_data", handle, keys))
- or 0,
+ _store_get_data=lambda handle, keys: (
+ calls.append(("store_get_data", handle, keys)) or [b"one", b"two"]
+ ),
+ _store_put_data=lambda handle, keys, values: (
+ calls.append(("store_put_data", handle, keys, values)) or 0
+ ),
+ _store_delete_data=lambda handle, keys: (
+ calls.append(("store_delete_data", handle, keys)) or 0
+ ),
_store_clear_data=lambda handle: calls.append(("store_clear_data", handle)) or 0,
- _store_seek_range=lambda handle, start, end: calls.append(
- ("store_seek_range", handle, start, end)
- )
- or [("aa", b"11"), ("ab", b"22")],
+ _store_seek_range=lambda handle, start, end: (
+ calls.append(("store_seek_range", handle, start, end)) or [("aa", b"11"), ("ab", b"22")]
+ ),
_init_logging=lambda level, output, fmt: calls.append(("init_logging", level, output, fmt)),
)