Skip to content

[DeepEP Integration] Free cache after combine for forward only path. #2274

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

Draft
elfiegg wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

[DeepEP Integration] Free cache after combine for forward only path. #2274

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
7 changes: 4 additions & 3 deletions torchtitan/distributed/deepep/deepep.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -146,7 +146,7 @@ def _dispatch_backward(
if grad_recv_x is None:
return None, None, None, None, None, None, None

handle = _handle_cache.get(ctx.cache_id_int)
handle = ctx.saved_handle
assert handle is not None, f"Handle not found for cache_id={ctx.cache_id_int}"

previous_event = _create_event_if_async(True)
Expand All @@ -161,7 +161,6 @@ def _dispatch_backward(
)

_sync_stream_if_async(True, after_event)
_handle_cache.pop(ctx.cache_id_int, None)

grad_x = grad_x.to(ctx.input_dtype)
grad_topk_weights = (
Expand All @@ -176,6 +175,7 @@ def _dispatch_setup_context(ctx, inputs, output):
recv_x, recv_indices, recv_scores, num_recv, cache_id = output
ctx.cache_id_int = cache_id.item()
ctx.input_dtype = x.dtype
ctx.saved_handle = _handle_cache.get(ctx.cache_id_int)


def _combine_backward(ctx, grad_combined):
Expand Down Expand Up @@ -207,7 +207,8 @@ def _combine_backward(ctx, grad_combined):
def _combine_setup_context(ctx, inputs, output):
x, cache_id = inputs
ctx.cache_id_int = cache_id.item()
ctx.saved_handle = _handle_cache.get(ctx.cache_id_int)
# Pop from cache - safe because dispatch_setup_context already saved handle for dispatch_backward
ctx.saved_handle = _handle_cache.pop(ctx.cache_id_int, None)


torch.library.register_autograd(
Expand Down