Skip to content

fix(executorch): copy the engine in bulk instead of byte by byte #4473

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
shoumikhin wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix(executorch): copy the engine in bulk instead of byte by byte #4473

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
20 changes: 10 additions & 10 deletions py/torch_tensorrt/executorch/backend.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -282,16 +282,16 @@ def preprocess(
_validate_engine_info(engine_info)
serialized_engine = engine_info[ENGINE_IDX]
if isinstance(serialized_engine, torch.Tensor):
# Single copy out of the underlying storage. The prior
# `.numpy().tobytes()` path allocated a fresh bytes buffer
# on top of the numpy view, which for a >2 GB engine
# roughly doubled peak memory at this step. `.cpu()` and
# `.contiguous()` are no-ops when already host-side and
# contiguous (the common case for the uint8 buffer this
# backend produces).
engine_info[ENGINE_IDX] = bytes(
serialized_engine.cpu().contiguous().untyped_storage()
)
# A single copy out of the tensor's own memory. `bytes(storage)` looks
# equivalent but iterates the storage element by element in Python,
# which costs about two seconds per megabyte and turns serializing a
# multi-gigabyte set of engines into hours. `memoryview` hands the
# buffer to the copy in one shot without the intermediate numpy view
# that would double peak memory for a large engine. `.cpu()` and
# `.contiguous()` are no-ops for the host-side uint8 buffer this
# backend normally produces.
engine_bytes = serialized_engine.cpu().contiguous().view(torch.uint8)
engine_info[ENGINE_IDX] = bytes(memoryview(engine_bytes.numpy()))
elif not isinstance(serialized_engine, (bytes, bytearray)):
engine_info[ENGINE_IDX] = bytes(serialized_engine)
input_names = _reorder_input_names_for_executorch(
Expand Down
Loading