Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
"""A flash-attn / causal-conv1d build for another torch fails with "undefined symbol" and Unsloth
falls back to slower kernels; the hint says how to get matching ones back."""
import importlib.util
import pathlib
import sys
import types
import pytest
_REPO_ROOT = pathlib.Path(__file__).resolve().parents[1]
_STALE = ImportError(
"/venv/lib/python3.13/site-packages/flash_attn_2_cuda.cpython-313-x86_64-linux-gnu.so: "
"undefined symbol: _ZN3c104cuda29c10_cuda_check_implementationEiPKcS2_ib"
)
def _load_import_fixes():
# By path: importing the unsloth package needs an accelerator.
spec = importlib.util.spec_from_file_location(
"unsloth_import_fixes_stale_kernel", _REPO_ROOT / "unsloth" / "import_fixes.py"
)
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
return module
@pytest.fixture(scope = "module")
def fixes():
return _load_import_fixes()
def _host(
monkeypatch,
fixes,
*,
torch_version = "2.14.0+cu130",
cuda = "13.0",
py = (3, 13),
machine = "x86_64",
platform = "linux",
cxx11 = True,
libc = ("glibc", "2.35"),
free_threaded = False,
):
fake_torch = types.SimpleNamespace(
__version__ = torch_version,
version = types.SimpleNamespace(cuda = cuda),
_C = types.SimpleNamespace(_GLIBCXX_USE_CXX11_ABI = cxx11),
)
monkeypatch.setitem(sys.modules, "torch", fake_torch)
monkeypatch.setattr(fixes.sys, "platform", platform)
monkeypatch.setattr(fixes.sys, "version_info", (*py, 0, "final", 0))
monkeypatch.setattr(fixes.platform, "machine", lambda: machine)
monkeypatch.setattr(fixes.platform, "libc_ver", lambda: libc)
real = fixes.sysconfig.get_config_var
monkeypatch.setattr(
fixes.sysconfig,
"get_config_var",
lambda name: (1 if free_threaded else 0) if name == "Py_GIL_DISABLED" else real(name),
)
@pytest.mark.parametrize(
"package, wheel",
[
("flash_attn", "flash_attn-2.8.4+cu13torch2.14cxx11abiTRUE-cp313-cp313-linux_x86_64.whl"),
(
"causal_conv1d",
"causal_conv1d-1.7.0+cu13torch2.14cxx11abiTRUE-cp313-cp313-linux_x86_64.whl",
),
],
)
def test_a_published_cell_names_our_matching_wheel(monkeypatch, fixes, package, wheel):
_host(monkeypatch, fixes)
hint = fixes.stale_kernel_hint(package, _STALE)
assert "built for a different torch than 2.14.0+cu130" in hint
assert hint.endswith(
"pip install --no-deps --force-reinstall "
f"https://github.com/unslothai/unsloth/releases/download/prebuilt-wheels-cu13/{wheel}"
)
def test_the_wheel_names_are_the_published_ones(fixes):
# Kept in step with the release the prebuilt workflow publishes (wheel_utils.py).
source = (_REPO_ROOT / "studio" / "backend" / "utils" / "wheel_utils.py").read_text(
encoding = "utf-8"
)
for package, version in fixes._PREBUILT_KERNEL_VERSIONS.items():
assert (
f'"{package}": "{version}"' in source or f"'{package}': '{version}'" in source
), package
assert fixes._PREBUILT_KERNEL_RELEASE_URL.endswith("/prebuilt-wheels-cu13")
@pytest.mark.parametrize(
"host",
[
{"torch_version": "2.12.1+cu130"},
{"cuda": "12.8", "torch_version": "2.14.0+cu128"},
{"py": (3, 12)},
{"machine": "aarch64"},
{"platform": "win32"},
{"cxx11": False},
{"libc": ("glibc", "2.31")},
{"libc": ("", "")},
{"free_threaded": True},
{"torch_version": "2.14.0.dev20260901+cu130"},
{"torch_version": "2.13.0rc1+cu130"},
],
)
def test_anywhere_else_it_names_a_source_rebuild(monkeypatch, fixes, host):
_host(monkeypatch, fixes, **host)
hint = fixes.stale_kernel_hint("causal_conv1d", _STALE)
assert "releases/download" not in hint
assert hint.endswith(
"pip install --no-deps --no-build-isolation --no-cache-dir --force-reinstall --no-binary causal-conv1d causal-conv1d"
)
@pytest.mark.parametrize(
"host, pinned",
[
({"py": (3, 12)}, True),
({"torch_version": "2.14.0.dev20260901+cu130"}, True),
({"torch_version": "2.12.1+cu130"}, False),
],
)
def test_flash_attn_rebuilds_from_a_revision_that_compiles_on_torch_2_13(
monkeypatch, fixes, host, pinned
):
_host(monkeypatch, fixes, **host)
hint = fixes.stale_kernel_hint("flash_attn", _STALE)
if pinned:
assert hint.endswith(f'--force-reinstall "{fixes._FLASH_ATTN_TORCH213_SOURCE}"')
else:
assert hint.endswith("--no-binary flash-attn flash-attn")
def test_the_flash_attn_source_pin_is_the_one_the_prebuilt_wheels_use(fixes):
source = (_REPO_ROOT / ".github" / "scripts" / "prebuilt_wheels.py").read_text(encoding = "utf-8")
ref = fixes._FLASH_ATTN_TORCH213_SOURCE.rsplit("@", 1)[1]
assert f'"ref": "{ref}"' in source
@pytest.mark.parametrize(
"symbol",
[
"_ZN3c1013MessageLoggerC1EPKciib",
"_ZNK2at10TensorBase8data_ptrEv",
"aoti_torch_create_device_guard",
],
)
def test_every_torch_abi_symbol_seen_in_practice_is_recognised(monkeypatch, fixes, symbol):
_host(monkeypatch, fixes)
error = ImportError(f"/x/causal_conv1d_cuda.so: undefined symbol: {symbol}")
assert "causal_conv1d-1.7.0" in fixes.stale_kernel_hint("causal_conv1d", error)
def test_a_chained_cause_is_found(monkeypatch, fixes):
_host(monkeypatch, fixes)
try:
try:
raise _STALE
except ImportError as inner:
raise RuntimeError("causal_conv1d import failed") from inner
except RuntimeError as outer:
assert "causal_conv1d-1.7.0" in fixes.stale_kernel_hint("causal_conv1d", outer)
@pytest.mark.parametrize(
"error",
[
ImportError("No module named 'flash_attn_2_cuda'"),
OSError("libcudart.so.13: cannot open shared object file"),
ImportError(
"libcusparse.so.12: undefined symbol: __nvJitLinkAddData_12_8, version libnvJitLink.so.12"
),
],
)
def test_other_failures_add_nothing(monkeypatch, fixes, error):
_host(monkeypatch, fixes)
assert fixes.stale_kernel_hint("flash_attn", error) == ""
def test_both_fallback_messages_print_the_hint():
utils = (_REPO_ROOT / "unsloth" / "models" / "_utils.py").read_text(encoding = "utf-8")
fixes_src = (_REPO_ROOT / "unsloth" / "import_fixes.py").read_text(encoding = "utf-8")
assert 'if hint := stale_kernel_hint("flash_attn", error):' in utils
assert 'hint = stale_kernel_hint("causal_conv1d", error)' in fixes_src