def test_dcgm_action_run_diag_bad_validation(handle, gpuIds):
gpuIdStr = ""
for i, gpuId in enumerate(gpuIds):
if i > 0:
gpuIdStr += ","
gpuIdStr += str(gpuId)
drd = dcgm_structs.c_dcgmRunDiag_t()
drd.version = dcgm_structs.dcgmRunDiag_version
drd.validate = dcgm_structs.DCGM_POLICY_VALID_SV_LONG + 1 #use an invalid value
drd.groupId = 0 #Initializing to 0 in case the constructor above doesn't
drd.gpuList = gpuIdStr
with test_utils.assert_raises(dcgm_structs.dcgmExceptionClass(dcgm_structs.DCGM_ST_BADPARAM)):
ret = dcgm_agent.dcgmActionValidate_v2(handle, drd, dcgm_structs.dcgmRunDiag_version)
def helper_dcgm_action_run_diag_gpu_list(handle, gpuIds):
'''
Test that running the DCGM diagnostic works if you provide a GPU ID list rather
than a groupId.
'''
gpuIdStr = ""
for i, gpuId in enumerate(gpuIds):
if i > 0:
gpuIdStr += ","
gpuIdStr += str(gpuId)
drd = dcgm_structs.c_dcgmRunDiag_t()
drd.version = dcgm_structs.dcgmRunDiag_version
drd.validate = dcgm_structs.DCGM_POLICY_VALID_SV_SHORT
drd.groupId = 0 #Initializing to 0 in case the constructor above doesn't
drd.gpuList = gpuIdStr
#this will throw an exception on error
ret = dcgm_agent.dcgmActionValidate_v2(handle, drd, dcgm_structs.dcgmRunDiag_version)
def helper_check_diag_empty_group(handle, gpuIds):
handleObj = pydcgm.DcgmHandle(handle=handle)
systemObj = handleObj.GetSystem()
groupObj = systemObj.GetEmptyGroup("test1")
runDiagInfo = dcgm_structs.c_dcgmRunDiag_t()
runDiagInfo.version = dcgm_structs.dcgmRunDiag_version
runDiagInfo.groupId = groupObj.GetId()
runDiagInfo.validate = 1
with test_utils.assert_raises(
dcgm_structs.dcgmExceptionClass(
dcgm_structs.DCGM_ST_GROUP_IS_EMPTY)):
response = test_utils.action_validate_wrapper(runDiagInfo, handle)
# Now make sure everything works well with a group
groupObj.AddGpu(gpuIds[0])
response = test_utils.action_validate_wrapper(runDiagInfo, handle)
assert response, "Should have received a response now that we have a non-empty group"
def __init__(self,
gpuIds=None,
testNamesStr='',
paramsStr='',
verbose=True,
train=False,
forceTrain=False,
version=dcgm_structs.dcgmRunDiag_version):
# Make sure version is valid
if version not in DcgmDiag._versionMap:
raise ValueError("'%s' is not a valid version for dcgmRunDiag." %
version)
self.version = version
if self.version == dcgm_structs.dcgmRunDiag_version7:
self.runDiagInfo = dcgm_structs.c_dcgmRunDiag_v7()
else:
self.runDiagInfo = dcgm_structs.c_dcgmRunDiag_t()
self.numTests = 0
self.numParams = 0
self.SetVerbose(verbose)
if testNamesStr == '':
# default to a level 1 test
self.runDiagInfo.validate = 1
elif testNamesStr == '1':
self.runDiagInfo.validate = 1
elif testNamesStr == '2':
self.runDiagInfo.validate = 2
elif testNamesStr == '3':
self.runDiagInfo.validate = 3
else:
# Make sure no number other that 1-3 were submitted
if testNamesStr.isdigit():
raise ValueError("'%s' is not a valid test name." %
testNamesStr)
# Copy to the testNames portion of the object
names = testNamesStr.split(',')
if len(names) > dcgm_structs.DCGM_MAX_TEST_NAMES:
err = 'DcgmDiag cannot initialize: %d test names were specified exceeding the limit of %d.' %\
(len(names), dcgm_structs.DCGM_MAX_TEST_NAMES)
raise ValueError(err)
for testName in names:
self.AddTest(testName)
if paramsStr != '':
params = paramsStr.split(';')
if len(params) >= dcgm_structs.DCGM_MAX_TEST_PARMS:
err = 'DcgmDiag cannot initialize: %d parameters were specified, exceeding the limit of %d.' %\
(len(params), dcgm_structs.DCGM_MAX_TEST_PARMS)
raise ValueError(err)
for param in params:
self.AddParameter(param)
if train == True:
self.runDiagInfo.flags = dcgm_structs.DCGM_RUN_FLAGS_TRAIN
if forceTrain == True:
self.runDiagInfo.flags |= dcgm_structs.DCGM_RUN_FLAGS_FORCE_TRAIN
if gpuIds:
first = True
for gpu in gpuIds:
if first:
self.runDiagInfo.gpuList = str(gpu)
first = False
else:
self.runDiagInfo.gpuList = "%s,%s" % (
self.runDiagInfo.gpuList, str(gpu))
def test_dcgm_run_diagnostic_validate(handle, gpuIds):
"""
Validates structure version
"""
handleObj = pydcgm.DcgmHandle(handle=handle)
systemObj = handleObj.GetSystem()
gpuIdList = systemObj.discovery.GetAllGpuIds()
assert len(gpuIdList
) >= 0, "Not able to find devices on the node for embedded case"
groupId = dcgm_agent.dcgmGroupCreate(handle,
dcgm_structs.DCGM_GROUP_DEFAULT,
"test1")
groupInfo = dcgm_agent.dcgmGroupGetInfo(handle, groupId)
status_handle = dcgm_agent.dcgmStatusCreate()
diagLevel = dcgm_structs.DCGM_DIAG_LVL_SHORT
gpuIdStr = ""
for i, gpuId in enumerate(gpuIds):
if i > 0:
gpuIdStr += ","
gpuIdStr += str(gpuId)
drd = dcgm_structs.c_dcgmRunDiag_t()
drd.version = dcgm_structs.dcgmRunDiag_version
drd.validate = dcgm_structs.DCGM_POLICY_VALID_SV_SHORT
drd.groupId = groupId
drd.gpuList = gpuIdStr
with test_utils.assert_raises(
dcgmExceptionClass(dcgm_structs.DCGM_ST_VER_MISMATCH)):
versionTest = 0 #invalid version
ret = vtDcgmActionValidate_v2(handle, drd, versionTest)
with test_utils.assert_raises(
dcgmExceptionClass(dcgm_structs.DCGM_ST_VER_MISMATCH)):
versionTest = 50 #random number version
ret = vtDcgmActionValidate_v2(handle, drd, versionTest)
with test_utils.assert_raises(
dcgmExceptionClass(dcgm_structs.DCGM_ST_VER_MISMATCH)):
versionTest = 0 #invalid version
ret = vtDcgmActionValidate(handle, drd.groupId, drd.validate,
versionTest)
with test_utils.assert_raises(
dcgmExceptionClass(dcgm_structs.DCGM_ST_VER_MISMATCH)):
versionTest = 50 #random number version
ret = vtDcgmActionValidate(handle, drd.groupId, drd.validate,
versionTest)
with test_utils.assert_raises(
dcgmExceptionClass(dcgm_structs.DCGM_ST_VER_MISMATCH)):
versionTest = 0 #invalid version
ret = vtDcgmRunDiagnostic(handle, drd.groupId, diagLevel, versionTest)
with test_utils.assert_raises(
dcgmExceptionClass(dcgm_structs.DCGM_ST_VER_MISMATCH)):
versionTest = 50 #random number version
ret = vtDcgmRunDiagnostic(handle, drd.groupId, diagLevel, versionTest)