Skip to content

[Tunix] Improve the reshard logging. Current logging only polls on the results for on device finish time. We need to do the same thing for the inputs for accurate logging. This cl splits the total wait time to waiting for input time and output ready time. #992

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
copybara-service wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[Tunix] Improve the reshard logging. Current logging only polls on the results for on device finish time. We need to do the same thing for the inputs for accurate logging. This cl splits the total wait time to waiting for input time and output ready time. #992

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 22 additions & 9 deletions tunix/rl/reshard.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,31 +29,36 @@
from flax import nnx
from tunix.rl import utils


# TODO(tsbao): move this to util
def callback_on_ready(
in_data: jaxtyping.PyTree,
x: jaxtyping.PyTree,
success: Callable[[], None],
failure: Callable[[Exception], None],
success: Callable[[float], None],
failure: Callable[[float, Exception], None],
):
"""Callback to invoke when the Jax array is ready."""
fut = futures.Future()
start_time = [-1.0]

def callback(f):
e = f.exception()
if e is None:
success()
success(start_time[0])
else:
failure(e)
failure(start_time[0], e)

fut.add_done_callback(callback)

def wait():
try:
jax.block_until_ready(in_data)
start_time[0] = time.perf_counter()
jax.block_until_ready(x)
except Exception as e: # pylint: disable=broad-exception-caught
fut.set_exception(e)
else:
fut.set_result(x)
fut.set_result((x))

threading.Thread(target=wait).start()

Expand Down Expand Up @@ -475,15 +480,23 @@ def _get_dst_sharding(x):
],
)

start = time.time()
start = time.perf_counter()

resharded_array = reshard_fn(source, dst_shardings)

callback_on_ready(
source,
resharded_array,
lambda: logging.info('Reshard finished in %.2fs', time.time() - start),
lambda e: logging.error(
'Reshard failed in %.2fs: %s', time.time() - start, e
lambda reshard_start: logging.info(
'Input ready in %.2fs. Reshard finished in %.2fs',
reshard_start - start,
time.perf_counter() - reshard_start,
),
lambda reshard_start, e: logging.error(
'Input ready in %.2fs. Reshard failed in %.2fs: %s',
reshard_start - start,
time.perf_counter() - reshard_start,
e,
),
)
return resharded_array
Expand Down