def test_client_anomaly_detection():
HID_DIM = 16
experts = {}
for i in range(4):
expert = layers.name_to_block['ffn'](HID_DIM)
experts[f'expert.{i}'] = hivemind.ExpertBackend(
name=f'expert.{i}',
expert=expert,
optimizer=torch.optim.Adam(expert.parameters()),
args_schema=(hivemind.BatchTensorDescriptor(HID_DIM), ),
outputs_schema=hivemind.BatchTensorDescriptor(HID_DIM),
max_batch_size=16,
)
experts['expert.3'].expert.ffn.weight.data[0, 0] = float('nan')
dht = hivemind.DHT(start=True)
server = hivemind.Server(dht, experts, num_connection_handlers=1)
server.start()
try:
server.ready.wait()
dmoe = hivemind.RemoteMixtureOfExperts(in_features=16,
grid_size=(3, ),
dht=dht,
k_best=3,
uid_prefix='expert.',
detect_anomalies=True)
input = torch.randn(1, 16)
input[0, 0] = float('nan')
with pytest.raises(ValueError):
dmoe(input)
input[0, 0] = 0
output = dmoe(input)
inf_loss = float('inf') * output.sum()
with pytest.raises(ValueError):
inf_loss.backward()
dmoe = hivemind.RemoteMixtureOfExperts(in_features=16,
grid_size=(4, ),
dht=dht,
k_best=4,
uid_prefix='expert.',
detect_anomalies=True)
output = dmoe(input)
assert output.isfinite().all()
finally:
server.shutdown()
def benchmark_throughput(num_experts=16, num_handlers=None, num_clients=128, num_batches_per_client=16,
expert_cls='ffn', hid_dim=1024, batch_size=2048, max_batch_size=None, backprop=True,
device=None, port=None):
assert not hasattr(torch.cuda, 'is_initialized') or not torch.cuda.is_initialized() \
or torch.device(device) == torch.device('cpu')
assert expert_cls in layers.name_to_block
port = port or find_open_port()
max_batch_size = max_batch_size or batch_size * 4
num_handlers = max(1, num_handlers or num_clients // 2)
benchmarking_failed = mp.Event()
can_start = mp.Event()
timestamps = dict(started=time.perf_counter())
try:
# start clients and await server
# Note: client processes must be launched BEFORE touching gpu, even torch.cuda.is_available can cause trouble
clients = [
mp.Process(
target=client_process, name=f'client_process-{i}',
args=(can_start, benchmarking_failed, port, num_experts, batch_size,
hid_dim, num_batches_per_client, backprop))
for i in range(num_clients)]
for client in clients:
client.daemon = True
client.start()
timestamps['launched_clients'] = timestamps['began_launching_server'] = time.perf_counter()
# start server
device = device or ('cuda' if torch.cuda.is_available() else 'cpu')
experts = {}
for i in range(num_experts):
expert = torch.jit.script(layers.name_to_block[expert_cls](hid_dim))
experts[f'expert{i}'] = hivemind.ExpertBackend(name=f'expert{i}',
expert=expert,
optimizer=torch.optim.Adam(expert.parameters()),
args_schema=(hivemind.BatchTensorDescriptor(hid_dim),),
outputs_schema=hivemind.BatchTensorDescriptor(hid_dim),
max_batch_size=max_batch_size,
)
timestamps['created_experts'] = time.perf_counter()
server = hivemind.Server(None, experts, listen_on=f"{hivemind.LOCALHOST}:{port}",
num_connection_handlers=num_handlers, device=device)
server.start()
server.ready.wait()
timestamps['server_ready'] = time.perf_counter()
can_start.set()
for client in clients:
client.join()
timestamps['clients_finished'] = time.perf_counter()
except BaseException as e:
benchmarking_failed.set()
raise e
finally:
for client in clients:
if client.is_alive():
client.terminate()
server.shutdown()
timestamps['server_shutdown_finished'] = time.perf_counter()
server.join()
sys.stdout.flush()
sys.stderr.flush()
time_between = lambda key1, key2: \
abs(timestamps[key2] - timestamps[key1]) if (key1 in timestamps and key2 in timestamps) else float('nan')
total_examples = batch_size * num_clients * num_batches_per_client
print('\n' * 3)
print("Benchmark finished, status:" + ["Success", "Failure"][benchmarking_failed.is_set()])
print(f"Server parameters: num_experts={num_experts}, num_handlers={num_handlers}, max_batch_size={max_batch_size},"
f" expert_cls={expert_cls}, hid_dim={hid_dim}, device={device}")
print(f"Client parameters: num_clients={num_clients}, num_batches_per_client={num_batches_per_client}, "
f"batch_size={batch_size}, backprop={backprop}")
print("Results: ")
print(f"\tServer startup took {time_between('began_launching_server', 'server_ready') :.3f} s. "
f"({time_between('began_launching_server', 'created_experts') :.3f} s. experts + "
f"{time_between('created_experts', 'server_ready') :.3f} s. networking)")
print(f"\tProcessed {total_examples} examples in {time_between('server_ready', 'clients_finished') :.3f}")
print(f"\tThroughput for {'forward + backward' if backprop else 'forward'} passes: "
f"{total_examples / time_between('server_ready', 'clients_finished') :.3f} samples / s.")
print(f"\tBenchmarking took {time_between('started', 'server_shutdown_finished') :.3f} s.")
if benchmarking_failed.is_set():
print("Note: benchmark code failed, timing/memory results only indicate time till failure!")
print_device_info(device)
print(flush=True)
assert not benchmarking_failed.is_set()
def create(listen_on='0.0.0.0:*', num_experts: int = None, expert_uids: str = None, expert_pattern: str = None,
expert_cls='ffn', hidden_dim=1024, optim_cls=torch.optim.Adam, num_handlers=None, max_batch_size=4096,
device=None, no_dht=False, initial_peers=(), dht_port=None, verbose=True,
compression=CompressionType.NONE, *, start: bool, **kwargs) -> Server:
"""
Instantiate a server with several identical experts. See argparse comments below for details
:param listen_on: network interface with address and (optional) port, e.g. "127.0.0.1:1337" or "[::]:80"
:param num_experts: run this many identical experts
:param expert_pattern: a string pattern or a list of expert uids, example: myprefix.[0:32].[0:256]\
means "sample random experts between myprefix.0.0 and myprefix.255.255;
:param expert_uids: spawn experts with these exact uids, overrides num_experts and expert_pattern
:param expert_cls: expert type from hivemind.server.layers, e.g. 'ffn', 'transformer', 'det_dropout' or 'nop';
:param hidden_dim: main dimension for expert_cls
:param num_handlers: server will use this many parallel processes to handle incoming requests
:param max_batch_size: total num examples in the same batch will not exceed this value
:param device: all experts will use this device in torch notation; default: cuda if available else cpu
:param optim_cls: uses this optimizer to train all experts
:param no_dht: if specified, the server will not be attached to a dht
:param initial_peers: a list of peers that will introduce this node to the dht,\
e.g. ('123.11.22.33:1337', '[fe80::abe2:db1c:be7d:5a85]:4567'), default = no peers
:param dht_port: DHT node will listen on this port, default = find open port
You can then use this node as initial peer for subsequent servers.
:param verbose: whether to print server started / finished / terminated events
:param start: if True, starts server right away and returns when server is ready for requests
"""
if verbose and len(kwargs) != 0:
print("Ignored kwargs:", kwargs)
assert expert_cls in name_to_block
# initialize dht
dht = None
if not no_dht:
logger.info(f"Bootstrapping DHT node, initial peers = {initial_peers}")
dht = hivemind.DHT(initial_peers=initial_peers, start=True,
listen_on=f"{hivemind.LOCALHOST}:{dht_port or hivemind.find_open_port()}")
if verbose:
logger.info(f"Running dht node on port {dht.port}")
# get expert uids
assert (expert_pattern is None and num_experts is None) or (expert_uids is None), \
"Please provide either expert_uids *or* num_experts and expert_pattern, but not both"
if expert_uids is None:
assert num_experts is not None, "Please specify either expert_uids or num_experts [and expert_pattern]"
expert_uids = generate_uids_from_pattern(num_experts, expert_pattern, dht=dht)
num_experts = len(expert_uids)
num_handlers = num_handlers if num_handlers is not None else num_experts * 8
optim_cls = optim_cls if optim_cls is not None else partial(torch.optim.SGD, lr=0.0)
device = device or ('cuda' if torch.cuda.is_available() else 'cpu')
sample_input = name_to_input[expert_cls](4, hidden_dim)
if isinstance(sample_input, tuple):
args_schema = tuple(hivemind.BatchTensorDescriptor.from_tensor(arg, compression) for arg in sample_input)
else:
args_schema = (hivemind.BatchTensorDescriptor.from_tensor(sample_input, compression),)
# initialize experts
experts = {}
for expert_uid in expert_uids:
expert = name_to_block[expert_cls](hidden_dim)
experts[expert_uid] = hivemind.ExpertBackend(name=expert_uid, expert=expert,
args_schema=args_schema,
outputs_schema=hivemind.BatchTensorDescriptor(
hidden_dim, compression=compression),
opt=optim_cls(expert.parameters()),
max_batch_size=max_batch_size,
)
# actually start server
server = Server(
dht, experts, listen_on=listen_on,
num_connection_handlers=num_handlers, device=device)
if start:
server.run_in_background(await_ready=True)
if verbose:
logger.info(f"Server started at {server.listen_on}")
logger.info(f"Got {len(experts)} active experts of type {expert_cls}: {list(experts.keys())}")
return server
def create(cls, listen_on='0.0.0.0:*', num_experts: int = None, expert_uids: str = None, expert_pattern: str = None,
expert_cls='ffn', hidden_dim=1024, optim_cls=torch.optim.Adam, scheduler: str = 'none',
num_warmup_steps=None, num_total_steps=None, clip_grad_norm=None, num_handlers=None, max_batch_size=4096,
device=None, no_dht=False, initial_peers=(), dht_port=None, checkpoint_dir: Optional[Path] = None,
compression=CompressionType.NONE, stats_report_interval: Optional[int] = None, *, start: bool,
**kwargs) -> Server:
"""
Instantiate a server with several identical experts. See argparse comments below for details
:param listen_on: network interface with address and (optional) port, e.g. "127.0.0.1:1337" or "[::]:80"
:param num_experts: run this many identical experts
:param expert_pattern: a string pattern or a list of expert uids, example: myprefix.[0:32].[0:256]\
means "sample random experts between myprefix.0.0 and myprefix.255.255;
:param expert_uids: spawn experts with these exact uids, overrides num_experts and expert_pattern
:param expert_cls: expert type from hivemind.server.layers, e.g. 'ffn', 'transformer', 'det_dropout' or 'nop';
:param hidden_dim: main dimension for expert_cls
:param num_handlers: server will use this many parallel processes to handle incoming requests
:param max_batch_size: total num examples in the same batch will not exceed this value
:param device: all experts will use this device in torch notation; default: cuda if available else cpu
:param optim_cls: uses this optimizer to train all experts
:param scheduler: if not `none`, the name of the expert LR scheduler
:param num_warmup_steps: the number of warmup steps for LR schedule
:param num_total_steps: the total number of steps for LR schedule
:param clip_grad_norm: maximum gradient norm used for clipping
:param no_dht: if specified, the server will not be attached to a dht
:param initial_peers: a list of peers that will introduce this node to the dht,\
e.g. ('123.11.22.33:1337', '[fe80::abe2:db1c:be7d:5a85]:4567'), default = no peers
:param dht_port: DHT node will listen on this port, default = find open port
You can then use this node as initial peer for subsequent servers.
:param checkpoint_dir: directory to save and load expert checkpoints
:param compression: if specified, use this compression to pack all inputs, outputs and gradients by all experts
hosted on this server. For a more fine-grained compression, start server in python and specify compression
for each BatchTensorProto in ExpertBackend for the respective experts.
:param start: if True, starts server right away and returns when server is ready for requests
:param stats_report_interval: interval between two reports of batch processing performance statistics
"""
if len(kwargs) != 0:
logger.info("Ignored kwargs:", kwargs)
assert expert_cls in name_to_block
if no_dht:
dht = None
else:
dht_endpoint = replace_port(listen_on, dht_port or hivemind.find_open_port())
dht = hivemind.DHT(initial_peers=initial_peers, start=True, listen_on=dht_endpoint)
logger.info(f"Running DHT node on port {dht.port}, initial peers = {initial_peers}")
assert ((expert_pattern is None and num_experts is None and expert_uids is not None) or
(num_experts is not None and expert_uids is None)), \
"Please provide either expert_uids *or* num_experts (possibly with expert_pattern), but not both"
if expert_uids is None:
if checkpoint_dir is not None:
assert is_directory(checkpoint_dir)
expert_uids = [child.name for child in checkpoint_dir.iterdir() if
(child / 'checkpoint_last.pt').exists()]
total_experts_in_checkpoint = len(expert_uids)
logger.info(f"Located {total_experts_in_checkpoint} checkpoints for experts {expert_uids}")
if total_experts_in_checkpoint > num_experts:
raise ValueError(
f"Found {total_experts_in_checkpoint} checkpoints, but num_experts is set to {num_experts}, "
f"which is smaller. Either increase num_experts or remove unneeded checkpoints.")
else:
expert_uids = []
uids_to_generate = num_experts - len(expert_uids)
if uids_to_generate > 0:
logger.info(f"Generating {uids_to_generate} expert uids from pattern {expert_pattern}")
expert_uids.extend(generate_uids_from_pattern(uids_to_generate, expert_pattern, dht))
num_experts = len(expert_uids)
num_handlers = num_handlers if num_handlers is not None else num_experts * 8
optim_cls = optim_cls if optim_cls is not None else partial(torch.optim.SGD, lr=0.0)
device = device or ('cuda' if torch.cuda.is_available() else 'cpu')
sample_input = name_to_input[expert_cls](4, hidden_dim)
if isinstance(sample_input, tuple):
args_schema = tuple(hivemind.BatchTensorDescriptor.from_tensor(arg, compression) for arg in sample_input)
else:
args_schema = (hivemind.BatchTensorDescriptor.from_tensor(sample_input, compression),)
scheduler = schedule_name_to_scheduler[scheduler]
# initialize experts
experts = {}
for expert_uid in expert_uids:
expert = name_to_block[expert_cls](hidden_dim)
experts[expert_uid] = hivemind.ExpertBackend(name=expert_uid, expert=expert,
args_schema=args_schema,
outputs_schema=hivemind.BatchTensorDescriptor(
hidden_dim, compression=compression),
optimizer=optim_cls(expert.parameters()),
scheduler=scheduler,
num_warmup_steps=num_warmup_steps,
num_total_steps=num_total_steps,
clip_grad_norm=clip_grad_norm,
max_batch_size=max_batch_size)
if checkpoint_dir is not None:
load_experts(experts, checkpoint_dir)
return cls(dht, experts, listen_on=listen_on, num_connection_handlers=num_handlers, device=device,
checkpoint_dir=checkpoint_dir, stats_report_interval=stats_report_interval, start=start)
def create(listen_on='0.0.0.0:*', num_experts: int = None, expert_uids: str = None, expert_pattern: str = None,
expert_cls='ffn', hidden_dim=1024, optim_cls=torch.optim.Adam, num_handlers=None, max_batch_size=4096,
device=None, no_dht=False, initial_peers=(), dht_port=None, checkpoint_dir: Optional[Path] = None,
load_experts=False, compression=CompressionType.NONE, *, start: bool, **kwargs) -> Server:
"""
Instantiate a server with several identical experts. See argparse comments below for details
:param listen_on: network interface with address and (optional) port, e.g. "127.0.0.1:1337" or "[::]:80"
:param num_experts: run this many identical experts
:param expert_pattern: a string pattern or a list of expert uids, example: myprefix.[0:32].[0:256]\
means "sample random experts between myprefix.0.0 and myprefix.255.255;
:param expert_uids: spawn experts with these exact uids, overrides num_experts and expert_pattern
:param expert_cls: expert type from hivemind.server.layers, e.g. 'ffn', 'transformer', 'det_dropout' or 'nop';
:param hidden_dim: main dimension for expert_cls
:param num_handlers: server will use this many parallel processes to handle incoming requests
:param max_batch_size: total num examples in the same batch will not exceed this value
:param device: all experts will use this device in torch notation; default: cuda if available else cpu
:param optim_cls: uses this optimizer to train all experts
:param no_dht: if specified, the server will not be attached to a dht
:param initial_peers: a list of peers that will introduce this node to the dht,\
e.g. ('123.11.22.33:1337', '[fe80::abe2:db1c:be7d:5a85]:4567'), default = no peers
:param dht_port: DHT node will listen on this port, default = find open port
You can then use this node as initial peer for subsequent servers.
:param checkpoint_dir: directory to save expert checkpoints
:param load_experts: whether to load expert checkpoints from checkpoint_dir
:param compression: if specified, use this compression to pack all inputs, outputs and gradients by all experts
hosted on this server. For a more fine-grained compression, start server in python and specify compression
for each BatchTensorProto in ExpertBackend for the respective experts.
:param start: if True, starts server right away and returns when server is ready for requests
"""
if len(kwargs) != 0:
logger.info("Ignored kwargs:", kwargs)
assert expert_cls in name_to_block
if no_dht:
dht = None
else:
dht_endpoint = replace_port(listen_on, dht_port or hivemind.find_open_port())
dht = hivemind.DHT(initial_peers=initial_peers, start=True, listen_on=dht_endpoint)
logger.info(f"Running DHT node on port {dht.port}, initial peers = {initial_peers}")
if load_experts:
assert dir_is_correct(checkpoint_dir)
assert expert_uids is None, "Can't both load saved experts and create new ones from given UIDs"
expert_uids = [child.name for child in checkpoint_dir.iterdir() if (child / 'checkpoint_last.pt').exists()]
if expert_uids:
logger.info(f"Located checkpoints for experts {expert_uids}, ignoring UID generation options")
else:
logger.info(f"No expert checkpoints found in {checkpoint_dir}, generating...")
assert (expert_pattern is None and num_experts is None) or (expert_uids is None) or (num_experts == 0), \
"Please provide either expert_uids *or* num_experts and expert_pattern, but not both"
# get expert uids if not loaded previously
if expert_uids is None:
assert num_experts is not None, "Please specify either expert_uids or num_experts [and expert_pattern]"
logger.info(f"Generating expert uids from pattern {expert_pattern}")
expert_uids = generate_uids_from_pattern(num_experts, expert_pattern, dht=dht)
num_experts = len(expert_uids)
num_handlers = num_handlers if num_handlers is not None else num_experts * 8
optim_cls = optim_cls if optim_cls is not None else partial(torch.optim.SGD, lr=0.0)
device = device or ('cuda' if torch.cuda.is_available() else 'cpu')
sample_input = name_to_input[expert_cls](4, hidden_dim)
if isinstance(sample_input, tuple):
args_schema = tuple(hivemind.BatchTensorDescriptor.from_tensor(arg, compression) for arg in sample_input)
else:
args_schema = (hivemind.BatchTensorDescriptor.from_tensor(sample_input, compression),)
# initialize experts
experts = {}
for expert_uid in expert_uids:
expert = name_to_block[expert_cls](hidden_dim)
experts[expert_uid] = hivemind.ExpertBackend(name=expert_uid, expert=expert,
args_schema=args_schema,
outputs_schema=hivemind.BatchTensorDescriptor(
hidden_dim, compression=compression),
opt=optim_cls(expert.parameters()),
max_batch_size=max_batch_size)
if load_experts:
load_weights(experts, checkpoint_dir)
server = Server(dht, experts, listen_on=listen_on, num_connection_handlers=num_handlers, device=device,
start=start)
return server
