mirror of
https://github.com/kevinveenbirkenbach/computer-playbook.git
synced 2026-11-10 20:56:39 +00:00
kevinveenbirkenbachandClaude Opus 5.5
11308259b3
Gate pass 5 flagged test_gpu_detection.py:31 in test_file_spot: the `ROLE / "vars" / "main.yml"` chain hard-codes a role-file path the utils.roles.mapping SPOT already names. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
127 lines
4.7 KiB
Python
127 lines
4.7 KiB
Python
"""Which hosts the role treats as GPU hosts.
|
|
|
|
A privileged container, such as a swarm lab node on a workstation with an
|
|
NVIDIA card, sees the host's ``/dev/nvidia0`` but carries no driver, so
|
|
``nvidia-smi`` does not exist there. The device node alone must therefore not
|
|
make a GPU host: the GPU include and the ``gpu`` lookup both have to say no
|
|
there, and a bare host in that state has to fail instead of silently running
|
|
its GPU services on the CPU.
|
|
"""
|
|
|
|
import unittest
|
|
from pathlib import Path
|
|
|
|
from jinja2 import Environment
|
|
|
|
from plugins.lookup.gpu import _device_present
|
|
from utils import PROJECT_ROOT
|
|
from utils.cache.yaml import load_yaml_any
|
|
from utils.roles.mapping import ROLE_FILE_VARS_MAIN
|
|
|
|
ROLE = Path(PROJECT_ROOT) / "roles" / "sys-svc-container"
|
|
CORE = ROLE / "tasks" / "00_core.yml"
|
|
GPU_INCLUDE = "00_core/gpu.yml"
|
|
DEVICE = "/dev/nvidia0"
|
|
NVIDIA_REGISTER = "sys_svc_container_nvidia_"
|
|
|
|
_ENV = Environment(autoescape=True)
|
|
_ENV.filters["bool"] = bool
|
|
|
|
|
|
def _role_vars() -> dict:
|
|
return load_yaml_any(str(ROLE / ROLE_FILE_VARS_MAIN))
|
|
|
|
|
|
def _render(value: str, context: dict) -> str:
|
|
return _ENV.from_string(value).render(**context)
|
|
|
|
|
|
def _all_hold(expressions, context: dict) -> bool:
|
|
guards = expressions if isinstance(expressions, list) else [expressions]
|
|
return all(bool(_ENV.compile_expression(expr)(**context)) for expr in guards)
|
|
|
|
|
|
def _walk(*, driver: bool, device: bool, in_container: bool) -> tuple[str, dict]:
|
|
"""Run the GPU detection of 00_core.yml over a fake filesystem.
|
|
|
|
Args:
|
|
driver: whether the NVIDIA driver binary exists on the host.
|
|
device: whether the host shows an NVIDIA device node.
|
|
in_container: whether the host is itself a container.
|
|
|
|
Returns:
|
|
"gpu" when the GPU include runs, "cpu" when it does not, "refused"
|
|
when an assert stops the play; and the registers the tasks left.
|
|
"""
|
|
context = dict(_role_vars())
|
|
context["DOCKER_IN_CONTAINER"] = in_container
|
|
existing = {
|
|
_render(context["SYS_SVC_CONTAINER_GPU_DRIVER"], context): driver,
|
|
DEVICE: device,
|
|
}
|
|
outcome = "cpu"
|
|
for task in load_yaml_any(str(CORE)):
|
|
stat = task.get("ansible.builtin.stat")
|
|
assertion = task.get("ansible.builtin.assert")
|
|
if assertion is not None and NVIDIA_REGISTER not in str(assertion["that"]):
|
|
assertion = None
|
|
gpu_include = task.get("include_tasks") == GPU_INCLUDE
|
|
if stat is None and assertion is None and not gpu_include:
|
|
continue
|
|
if not _all_hold(task.get("when", []), context):
|
|
continue
|
|
if stat is not None:
|
|
path = _render(stat["path"], context)
|
|
context[task["register"]] = {"stat": {"exists": existing.get(path, False)}}
|
|
elif assertion is not None and not _all_hold(assertion["that"], context):
|
|
return "refused", context
|
|
elif gpu_include:
|
|
outcome = "gpu"
|
|
return outcome, context
|
|
|
|
|
|
class TestGpuDetection(unittest.TestCase):
|
|
def test_a_container_seeing_a_driverless_device_runs_on_the_cpu(self):
|
|
outcome, registers = _walk(driver=False, device=True, in_container=True)
|
|
self.assertEqual(
|
|
outcome,
|
|
"cpu",
|
|
"a privileged container sees the host's /dev/nvidia0 without "
|
|
"nvidia-smi; the include would fail on the missing binary",
|
|
)
|
|
self.assertFalse(_device_present(registers))
|
|
|
|
def test_a_bare_host_with_a_driverless_device_is_refused(self):
|
|
outcome, _ = _walk(driver=False, device=True, in_container=False)
|
|
self.assertEqual(outcome, "refused")
|
|
|
|
def test_driver_and_device_make_a_gpu_host(self):
|
|
for in_container in (False, True):
|
|
with self.subTest(in_container=in_container):
|
|
outcome, registers = _walk(
|
|
driver=True, device=True, in_container=in_container
|
|
)
|
|
self.assertEqual(outcome, "gpu")
|
|
self.assertTrue(_device_present(registers))
|
|
|
|
def test_no_device_is_no_gpu_host(self):
|
|
for driver in (False, True):
|
|
for in_container in (False, True):
|
|
with self.subTest(driver=driver, in_container=in_container):
|
|
outcome, registers = _walk(
|
|
driver=driver, device=False, in_container=in_container
|
|
)
|
|
self.assertEqual(outcome, "cpu")
|
|
self.assertFalse(_device_present(registers))
|
|
|
|
def test_the_query_runs_the_binary_the_detection_probed(self):
|
|
self.assertTrue(
|
|
_role_vars()["SYS_SVC_CONTAINER_GPU_QUERY"].startswith(
|
|
"{{ SYS_SVC_CONTAINER_GPU_DRIVER }} "
|
|
)
|
|
)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|