def get_experts(
self,
uids,
expiration_time: Optional[DHTExpiration] = None,
return_future: bool = False) -> List[Optional[RemoteExpert]]:
logger.warning(
"dht.get_experts is scheduled for removal in 0.9.8, please use hivemind.get_experts."
)
return hivemind.get_experts(self, uids, expiration_time, return_future)
def test_store_get_experts():
peers = [hivemind.DHT(start=True)]
for i in range(10):
neighbors_i = [
f'{LOCALHOST}:{node.port}'
for node in random.sample(peers, min(3, len(peers)))
]
peers.append(hivemind.DHT(initial_peers=neighbors_i, start=True))
first_peer = random.choice(peers)
other_peer = random.choice(peers)
expert_uids = [f"my_expert.{i}" for i in range(50)]
batch_size = 10
for batch_start in range(0, len(expert_uids), batch_size):
hivemind.declare_experts(
first_peer, expert_uids[batch_start:batch_start + batch_size],
'localhost:1234')
found = get_experts(other_peer,
random.sample(expert_uids, 5) + ['foo', 'bar'])
assert all(res is not None
for res in found[:-2]), "Could not find some existing experts"
assert all(res is None for res in found[-2:]), "Found non-existing experts"
other_expert, other_port = "my_other_expert.1337", random.randint(
1000, 9999)
hivemind.declare_experts(other_peer, [other_expert],
f'that_host:{other_port}')
first_notfound, first_found = get_experts(first_peer,
['foobar', other_expert])
assert isinstance(first_found, hivemind.RemoteExpert)
assert first_found.endpoint == f'that_host:{other_port}'
for peer in peers:
peer.shutdown()
def test_dht_single_node():
node = hivemind.DHT(start=True)
beam_search = MoEBeamSearcher(node, 'expert.', grid_size=(10, ))
assert all(
declare_experts(node, ['expert.1', 'expert.2', 'expert.3'],
f"{hivemind.LOCALHOST}:1337").values())
assert len(declare_experts(node, ["ffn.1", "ffn.2"],
endpoint="that_place")) == 4
assert len(
declare_experts(node, ['e.1.2.3', 'e.1.2.5', 'e.2.0'],
f"{hivemind.LOCALHOST}:42")) == 7
for expert in get_experts(node, ['expert.3', 'expert.2']):
assert expert.endpoint == f"{hivemind.LOCALHOST}:1337"
assert all(
declare_experts(node, ['expert.5', 'expert.2'],
f"{hivemind.LOCALHOST}:1337").values())
found_experts = beam_search.find_best_experts(
[(0., 1., 2., 3., 4., 5., 6., 7., 8.)], beam_size=2)
assert len(found_experts) == 2 and [
expert.uid for expert in found_experts
] == ['expert.5', 'expert.3']
successors = beam_search.get_active_successors(
['e.1.2.', 'e.2.', 'e.4.5.'])
assert len(successors['e.1.2.']) == 2
assert successors['e.1.2.'][3] == UidEndpoint('e.1.2.3', f'{LOCALHOST}:42')
assert successors['e.1.2.'][5] == UidEndpoint('e.1.2.5', f'{LOCALHOST}:42')
assert len(
successors['e.2.']) == 1 and successors['e.2.'][0] == UidEndpoint(
'e.2.0', f'{LOCALHOST}:42')
assert successors['e.4.5.'] == {}
initial_beam = beam_search.get_initial_beam((3, 2, 1, 0, -1, -2, -3),
beam_size=3)
assert len(initial_beam) == 3
assert initial_beam[0][:2] == (2.0, 'expert.1.')
assert initial_beam[1][:2] == (1.0, 'expert.2.')
assert initial_beam[2][:2] == (0.0, 'expert.3.')
with pytest.raises(AssertionError):
beam_search = MoEBeamSearcher(node, 'expert.1.ffn', (2, 2))
with pytest.raises(AssertionError):
beam_search.get_active_successors(['e.1.2.', 'e.2', 'e.4.5.'])
def benchmark_dht(num_peers: int, initial_peers: int, num_experts: int,
expert_batch_size: int, random_seed: int,
wait_after_request: float, wait_before_read: float,
wait_timeout: float, expiration: float):
random.seed(random_seed)
print("Creating peers...")
peers = []
for _ in trange(num_peers):
neighbors = [
f'0.0.0.0:{node.port}'
for node in random.sample(peers, min(initial_peers, len(peers)))
]
peer = hivemind.DHT(initial_peers=neighbors,
start=True,
wait_timeout=wait_timeout,
listen_on=f'0.0.0.0:*')
peers.append(peer)
store_peer, get_peer = peers[-2:]
expert_uids = list(
set(f"expert.{random.randint(0, 999)}.{random.randint(0, 999)}.{random.randint(0, 999)}"
for _ in range(num_experts)))
print(f"Sampled {len(expert_uids)} unique ids (after deduplication)")
random.shuffle(expert_uids)
print(f"Storing experts to dht in batches of {expert_batch_size}...")
successful_stores = total_stores = total_store_time = 0
benchmark_started = time.perf_counter()
endpoints = []
for start in trange(0, num_experts, expert_batch_size):
store_start = time.perf_counter()
endpoints.append(random_endpoint())
store_ok = hivemind.declare_experts(store_peer,
expert_uids[start:start +
expert_batch_size],
endpoints[-1],
expiration=expiration)
successes = store_ok.values()
total_store_time += time.perf_counter() - store_start
total_stores += len(successes)
successful_stores += sum(successes)
time.sleep(wait_after_request)
print(
f"Store success rate: {successful_stores / total_stores * 100:.1f}% ({successful_stores} / {total_stores})"
)
print(
f"Mean store time: {total_store_time / total_stores:.5}, Total: {total_store_time:.5}"
)
time.sleep(wait_before_read)
if time.perf_counter() - benchmark_started > expiration:
logger.warning(
"All keys expired before benchmark started getting them. Consider increasing expiration_time"
)
successful_gets = total_get_time = 0
for start in trange(0, len(expert_uids), expert_batch_size):
get_start = time.perf_counter()
get_result = hivemind.get_experts(
get_peer, expert_uids[start:start + expert_batch_size])
total_get_time += time.perf_counter() - get_start
for i, expert in enumerate(get_result):
if expert is not None and expert.uid == expert_uids[start + i] \
and expert.endpoint == endpoints[start // expert_batch_size]:
successful_gets += 1
if time.perf_counter() - benchmark_started > expiration:
logger.warning(
"keys expired midway during get requests. If that isn't desired, increase expiration_time param"
)
print(
f"Get success rate: {successful_gets / len(expert_uids) * 100:.1f} ({successful_gets} / {len(expert_uids)})"
)
print(
f"Mean get time: {total_get_time / len(expert_uids):.5f}, Total: {total_get_time:.5f}"
)
alive_peers = [peer.is_alive() for peer in peers]
print(f"Node survival rate: {len(alive_peers) / len(peers) * 100:.3f}%")
def generate_uids_from_pattern(num_experts: int,
expert_pattern: Optional[str],
dht: Optional[DHT] = None,
attempts_per_expert=10) -> List[str]:
"""
Sample experts from a given pattern, remove duplicates.
:param num_experts: sample this many unique expert uids
: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 dht: if specified, uses this DHT to check that expert uids are not yet occupied by other peers
:param attempts_per_expert: give up if unable to generate a new expert uid after this many attempts per uid
:note: this method is not strictly process-safe. If several servers run it concurrently, they have
a small chance of sampling duplicate expert uids.
"""
remaining_attempts = attempts_per_expert * num_experts
found_uids, attempted_uids = list(), set()
def _generate_uid():
if expert_pattern is None:
return f"expert{UID_DELIMITER}{attempts_per_expert * num_experts - remaining_attempts}"
uid = []
for block in expert_pattern.split(UID_DELIMITER):
try:
if '[' not in block and ']' not in block:
uid.append(block)
elif block.startswith('[') and block.endswith(
']') and ':' in block:
slice_start, slice_end = map(int, block[1:-1].split(':'))
uid.append(str(random.randint(slice_start, slice_end - 1)))
else:
raise ValueError(
"Block must be either fixed or a range [from:to]")
except KeyboardInterrupt:
raise
except Exception as e:
raise ValueError(
f"Expert pattern {expert_pattern} has invalid block {block}, {e}"
)
return UID_DELIMITER.join(uid)
while remaining_attempts > 0 and len(found_uids) < num_experts:
# 1. sample new expert uids at random
new_uids = []
while len(new_uids) + len(
found_uids) < num_experts and remaining_attempts > 0:
new_uid = _generate_uid()
remaining_attempts -= 1
if new_uid not in attempted_uids:
attempted_uids.add(new_uid)
new_uids.append(new_uid)
# 2. look into DHT (if given) and remove duplicates
if dht:
existing_expert_uids = {
found_expert.uid
for found_expert in hivemind.get_experts(dht, new_uids)
if found_expert is not None
}
new_uids = [
new_uid for new_uid in new_uids
if new_uid not in existing_expert_uids
]
found_uids += new_uids
if len(found_uids) != num_experts:
logger.warning(
f"Found only {len(found_uids)} out of {num_experts} free expert uids after "
f"{attempts_per_expert * num_experts} attempts")
return found_uids
