def test_gpu_e2e(ray_8_cpus_gpus):
with fault_tolerance_patches():
discovery_schedule = [
(20, ['host-1:1']),
(60, ['host-1:1', 'host-4:1', 'host-5:1']),
(None, ['host-1:1', 'host-4:1', 'host-5:1', 'host-6:1']),
]
nics = list(psutil.net_if_addrs().keys())[0]
settings = ElasticRayExecutor.create_settings(min_np=4,
max_np=4,
nics={nics})
settings.discovery = SimpleTestDiscovery(discovery_schedule)
executor = ElasticRayExecutor(settings,
gpus_per_slot=1,
use_gpu=True,
override_discovery=False)
logger, training_fn = _create_training_function(iterations=100)
executor.start()
results = executor.run(training_fn)
assert len(results) == 4
events = ray.get(logger.fetch.remote())
assert sum(int("started" in e) for e in events) == 4, events
assert sum(int("finished" in e) for e in events) == 4, events
def test_min_np(ray_8_cpus):
with fault_tolerance_patches():
discovery_schedule = [
(10, ['host-1:1']),
(10, ['host-1:1', 'host-4:1', 'host-5:1']),
(None, ['host-1:1', 'host-4:1', 'host-5:1', 'host-6:1']),
]
nics = list(psutil.net_if_addrs().keys())[0]
settings = ElasticRayExecutor.create_settings(min_np=4,
max_np=4,
nics={nics})
settings.discovery = SimpleTestDiscovery(discovery_schedule)
executor = ElasticRayExecutor(settings,
cpus_per_slot=1,
override_discovery=False)
training_fn = _create_training_function(iterations=30)
executor.start()
trace = StatusCallback()
results = executor.run(training_fn, callbacks=[trace])
assert len(results) == 4
events = trace.fetch()
assert sum(int("started" in e) for e in events) == 4, events
assert sum(int("finished" in e) for e in events) == 4, events
