Skip to content

fix: _patch_cpu_offload_apply handles .to(cuda) in addition to .cuda() #59

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
cennn wants to merge 1 commit into
base: main
Choose a base branch
from
+13 −2
Open

fix: _patch_cpu_offload_apply handles .to(cuda) in addition to .cuda() #59

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
15 changes: 13 additions & 2 deletions magi_compiler/_api.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -509,15 +509,26 @@ def _cpu_apply(self, fn):
id_cpu_lambda = getattr(fn, "__qualname__", "") == "Module.cpu.<locals>.<lambda>"
is_to_lambda = getattr(fn, "__qualname__", "") == "Module.to.<locals>.convert"

# Detect .to("cuda:X") by probing the lambda on a small CPU tensor.
# .to("cpu") also produces is_to_lambda=True but should not trigger offload.
is_to_cuda = False
if is_to_lambda and not getattr(self, "_magi_offloaded_once", False):
try:
is_to_cuda = fn(torch.empty(0, device="cpu")).is_cuda
except Exception:
pass

is_moving_to_gpu = is_cuda_lambda or is_to_cuda

# after first time to call _apply(cuda), skip "Module.to" and "Module.cpu" and "Module.cuda"
if getattr(self, "_magi_offloaded_once", False):
if is_cuda_lambda or id_cpu_lambda or is_to_lambda:
return self
else:
return _orig_apply(self, fn)
else:
# first time to call _apply(cuda), move all parameters/buffers to CPU
if not is_cuda_lambda:
# first time to call _apply(cuda) or _apply(to_cuda), move all parameters/buffers to CPU
if not is_moving_to_gpu:
return _orig_apply(self, fn)

# move all parameters/buffers to CPU
Expand Down