Skip to content
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

Fix the logic of VarBase _to func #37193

Merged
merged 1 commit into from
Nov 16, 2021
Merged

Fix the logic of VarBase _to func #37193

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: 11 additions & 9 deletions python/paddle/fluid/dygraph/varbase_patch_methods.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -386,21 +386,18 @@ def transform(t, device, dtype, blocking):
device = t.place
if dtype is None:
dtype = t.dtype
if type(dtype) is str:
dtype = framework.convert_np_dtype_to_dtype_(dtype)

# 1. gpu place need to determine whether the memory is sufficient for allocation.
if t.place.is_gpu_place():
gpu_memory_available = core.gpu_memory_available()
# for gpu, minimum memory allocation unit is 256 bytes.
if type(dtype) is str:
size_dtype = core.size_of_dtype(
framework.convert_np_dtype_to_dtype_(dtype))
else:
size_dtype = core.size_of_dtype(dtype)
size_dtype = core.size_of_dtype(dtype)
# Note(weilong wu): Paddle GPU minimum memory allocation unit is 256 bytes,
# waiting_alloc_memory will compute the memory space occupied by 't'.
# Coefficient 1.2 is used to avoid OOM that may occur in this critical state when the memory is just enough.
waiting_alloc_memory = (
(t._numel() * size_dtype) / 256 + 1) * 256 * 1.2
gpu_memory_available = core.gpu_memory_available()
if gpu_memory_available < waiting_alloc_memory:
# Copy Tensor to cpu
t_used = t._copy_to(paddle.CPUPlace(), blocking)
Expand All @@ -414,12 +411,17 @@ def transform(t, device, dtype, blocking):

# 2. cast Tensor to dtype
if dtype is not None and dtype != t_used.dtype:
t_casted = t_used.cast(dtype=dtype)
with paddle.fluid.framework._dygraph_place_guard(
place=t_used.place):
t_casted = t_used.cast(dtype=dtype)
else:
t_casted = t_used

# 3. Copy casted Tensor(in CPU or GPU) to device
new_t = t_casted._copy_to(device, blocking)
if device is not None and not t_casted.place._equals(device):
new_t = t_casted._copy_to(device, blocking)
else:
new_t = t_casted

# 4. Share Tensor to origin Tensor
dst_tensor = t.value().get_tensor()
Expand Down