Skip to content

dynamic_shape support balanced vpp #72386

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

Merged
merged 1 commit into from
Apr 21, 2025
Merged

dynamic_shape support balanced vpp #72386

Show file tree
Hide file tree
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
10 changes: 10 additions & 0 deletions python/paddle/distributed/fleet/meta_parallel/pipeline_parallel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2760,6 +2760,9 @@ def backward_async_comm(

# reset dynamic meta counter
if self._dynamic_shape:
assert self._p2p_helper._dynamic_cnt == len(
self._p2p_helper._send_recv_meta_list
), "p2p dynamic_cnt should equal to send_recv_meta_list"
self._p2p_helper._dynamic_cnt = 0

return train_loss
Expand Down Expand Up @@ -3359,6 +3362,13 @@ def forward_backward_pipeline(
backward_send_recv_buffer_queue.empty()
), "send_recv buffer should be empty"

# reset dynamic meta counter
if self._dynamic_shape:
assert self._p2p_helper._dynamic_cnt == len(
self._p2p_helper._send_recv_meta_list
), "p2p dynamic_cnt should equal to send_recv_meta_list"
self._p2p_helper._dynamic_cnt = 0

self._flush_records()
self._sync_overlap_grads()

Expand Down
21 changes: 10 additions & 11 deletions python/paddle/distributed/fleet/meta_parallel/pp_utils/p2p_communication.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -763,6 +763,7 @@ def recv_forward(self, pp_first_stage, sync_recv=True, batch_p2p_comm=True):
sync_recv=sync_recv,
send_recv_meta=self._send_recv_meta,
batch_p2p_comm=batch_p2p_comm,
dynamic_shape=self._dynamic_shape,
)
if self._dynamic_shape:
self._dynamic_cnt += 1
Expand Down Expand Up @@ -807,9 +808,6 @@ def recv_backward(
if _timers is not None:
_timers("recv_backward").stop()

if self._dynamic_shape and need_increase_cnt:
self._dynamic_cnt += 1

return output_tensor_grad

def send_forward(
Expand All @@ -823,10 +821,6 @@ def send_forward(
if _timers is not None:
_timers("send_forward").start()

assert (
not self._dynamic_shape
), "p2p_helper.send_forward function doesn't support dynamic_shape now"

if not pp_last_stage:
self._send_meta(output_tensor, skip_check_meta=skip_check_meta)
_p2p_helper(
Expand All @@ -836,7 +830,11 @@ def send_forward(
recv_next=False,
send_recv_meta=self._send_recv_meta,
batch_p2p_comm=batch_p2p_comm,
dynamic_shape=self._dynamic_shape,
)
if self._dynamic_shape:
self._dynamic_cnt += 1

if _timers is not None:
_timers("send_forward").stop()

Expand All @@ -847,19 +845,20 @@ def send_backward(
if _timers is not None:
_timers("send_backward").start()

assert (
not self._dynamic_shape
), "p2p_helper.send_backward function doesn't support dynamic_shape now"

if not pp_first_stage:
if self._dynamic_shape:
self._send_meta(input_tensor_grad, reverse=True)
_p2p_helper(
tensor_send_next=None,
tensor_send_prev=input_tensor_grad,
recv_prev=False,
recv_next=False,
send_recv_meta=self._send_recv_meta,
batch_p2p_comm=batch_p2p_comm,
dynamic_shape=self._dynamic_shape,
)
if self._dynamic_shape:
self._dynamic_cnt += 1
if _timers is not None:
_timers("send_backward").stop()

Expand Down
Loading