def _update_last_index_token(self, new_index_token: str) -> None:
"""
This function updates the last_index_token in the following scenarios:
1. last_complete_step >= last_index_token_step :
this means that the token isn't pointing to the latest completed step
2. number of steps available ( complete or incomplete ) - (last_completed_step+1) > window_size_limit:
we maintain a window to stop querying for older steps that have not completed.
if the total number of steps, we are querying for completion is greater than our window_size_limit
we update the last_index_token and last_complete_step by (window_size_limit // 2)
:param new_index_token:
:return:None
"""
if self.last_index_token is None:
last_index_token_step = 0
else:
last_index_token_step = IndexFileLocationUtils.parse_step_from_index_file_name(
self.last_index_token)
# Case 1: This case is not satisfied when all workers in a
# distributed training job have not written a step
if self.last_complete_step >= last_index_token_step:
prefix = IndexFileLocationUtils.get_prefix_from_index_file(
new_index_token)
# sort lexicographically and select the last worker
last_worker = sorted(list(self.worker_set))[-1]
# below converts worker_name to serialized workerName
# if it's a tf device, else no effect
last_worker_serialized = serialize_tf_device(last_worker)
self.last_index_token = IndexFileLocationUtils.get_index_key_for_step(
prefix, self.last_complete_step, last_worker_serialized)
self.logger.debug(
f"Updated last index token to:{self.last_index_token}")
# Case 2: This case is satisfied if the number of incomplete steps
# is greater than the INCOMPLETE_STEP_WAIT_WINDOW
available_step = self._global_to_mode.keys()
if (len(available_step) - (self.last_complete_step + 1) >
self._incomplete_wait_for_step_window):
prefix = IndexFileLocationUtils.get_prefix_from_index_file(
new_index_token)
last_worker = sorted(list(self.worker_set))[-1]
# below converts worker_name to serialized workerName
# if it's a tf device, else no effect
last_worker_serialized = serialize_tf_device(last_worker)
self.last_index_token = IndexFileLocationUtils.get_index_key_for_step(
prefix,
self.last_complete_step +
(self._incomplete_wait_for_step_window // 2),
last_worker_serialized,
)
self.last_complete_step = IndexFileLocationUtils.parse_step_from_index_file_name(
self.last_index_token)
self.logger.info(
f"Waiting for: {len(available_step) - (self.last_complete_step + 1)} Steps. \n"
f"INCOMPLETE_STEP_WAIT_WINDOW: {self._incomplete_wait_for_step_window}. \n"
f"Marking the last {self._incomplete_wait_for_step_window // 2} incomplete steps as complete"
f"Updating last_index_token to: {self.last_index_token}. \n"
f"Updating last_complete_step to: {self.last_complete_step}. ")
def test_index():
numpy_tensor = [
np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32),
np.array([[1.0, 2.0, 4.0], [3.0, 4.0, 5.0]], dtype=np.float32),
]
runid = "default"
logdir = "."
step = 0
worker = "worker_0"
run_dir = os.path.join(logdir, runid)
writer = FileWriter(trial_dir=run_dir,
step=step,
worker=worker,
verbose=True)
for i in (0, len(numpy_tensor) - 1):
n = "tensor" + str(i)
writer.write_tensor(tdata=numpy_tensor[i], tname=n)
writer.flush()
writer.close()
efl = TensorFileLocation(step_num=step, worker_name=worker)
eventfile = efl.get_file_location(trial_dir=run_dir)
indexfile = IndexFileLocationUtils.get_index_key_for_step(
run_dir, step, worker)
fo = open(eventfile, "rb")
with open(indexfile) as idx_file:
index_data = json.load(idx_file)
tensor_payload = index_data["tensor_payload"]
i = 0
for tensor in tensor_payload:
start_idx = int(tensor["start_idx"])
fo.seek(start_idx, 0)
length = int(tensor["length"])
line = fo.read(length)
zoo = open("test.txt", "wb")
zoo.write(line)
zoo.close()
testfile_reader = FileReader("./test.txt")
tensor_values = list(testfile_reader.read_tensors())
assert np.allclose(
tensor_values[0][2].all(),
numpy_tensor[i].all()), "indexwriter not working"
i = i + 1
fo.close()
shutil.rmtree(run_dir)
os.remove("test.txt")
def dummy_step_creator(trial_dir, global_step, mode, mode_step, worker_name):
static_step_data = (
'{"meta": {"mode": "TRAIN", "mode_step": 0, "event_file_name": ""}, '
'"tensor_payload": ['
'{"tensorname": "gradients/dummy:0", "start_idx": 0, "length": 1}'
"]}")
step = json.loads(static_step_data)
step["meta"]["mode"] = mode
step["meta"]["mode_step"] = mode_step
index_file_location = IndexFileLocationUtils.get_index_key_for_step(
trial_dir, global_step, worker_name)
Path(os.path.dirname(index_file_location)).mkdir(parents=True,
exist_ok=True)
with open(index_file_location, "w") as f:
json.dump(step, f)