Skip to content

add bf16 for some ops in static mode #51582

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 8 commits into from
Apr 9, 2023
Merged

add bf16 for some ops in static mode #51582

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
6974f54
add bf16 for some ops in static mode
shaojiewang Mar 13, 2023
ec49464
1 split bf16 grad list in clip py from fp16 list, 2 add bf16 scale va…
shaojiewang Mar 16, 2023
d4d57ee
formatting
shaojiewang Mar 16, 2023
d613412
check if bf16-fp16 amp mode is used in grad clip
shaojiewang Mar 16, 2023
7b3bc3e
merge remote develop branch
shaojiewang Mar 16, 2023
ecf5535
merge remote develop branch
shaojiewang Apr 5, 2023
1c5c423
formatting
shaojiewang Apr 5, 2023
80a80ad
fix clip type bug
shaojiewang Apr 7, 2023
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
1 change: 1 addition & 0 deletions paddle/phi/kernels/gpu/matmul_grad_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,6 +55,7 @@ PD_REGISTER_KERNEL(matmul_with_flatten_grad,
phi::MatmulWithFlattenGradKernel,
float,
double,
phi::dtype::bfloat16,
phi::dtype::float16) {}

PD_REGISTER_KERNEL(matmul_with_flatten_double_grad,
Expand Down
1 change: 1 addition & 0 deletions paddle/phi/kernels/gpu/matmul_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,4 +36,5 @@ PD_REGISTER_KERNEL(matmul_with_flatten,
phi::MatmulWithFlattenKernel,
float,
double,
phi::dtype::bfloat16,
phi::dtype::float16) {}
3 changes: 2 additions & 1 deletion python/paddle/fluid/layers/layer_function_generator.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -278,7 +278,7 @@ def func(x, name=None):

if op_type not in ["abs", "exp", "square"]:
check_variable_and_dtype(
x, 'x', ['float16', 'float32', 'float64'], op_type
x, 'x', ['float16', 'float32', 'float64', 'uint16'], op_type
)
else:
# abs exp square ops support dtype(int32, int64, float16, float32, float64)
Expand All @@ -293,6 +293,7 @@ def func(x, name=None):
'float64',
'complex64',
'complex128',
'uint16',
],
op_type,
)
Expand Down
12 changes: 9 additions & 3 deletions python/paddle/static/amp/amp_nn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,7 +49,7 @@ def check_finite_and_unscale(x, scale, name=None, float_status=None):
check_variable_and_dtype(
e,
"x",
['float16', 'float32', 'float64'],
['float16', 'float32', 'float64', 'uint16'],
'check_finite_and_unscale',
)

Expand Down Expand Up @@ -133,9 +133,15 @@ def update_loss_scaling(
check_type(x, 'x', (tuple, list), 'update_loss_scaling')
for e in x:
check_variable_and_dtype(
e, "x", ['float16', 'float32', 'float64'], 'update_loss_scaling'
e,
"x",
['float16', 'float32', 'float64', 'uint16'],
'update_loss_scaling',
)
if e.dtype == core.VarDesc.VarType.FP16:
if (
e.dtype == core.VarDesc.VarType.FP16
or e.dtype == core.VarDesc.VarType.BF16
):
assert (
prev_loss_scaling.dtype == core.VarDesc.VarType.FP32
), "The dtype of prev_loss_scaling should be float32 when the dtype of x is float16."
Expand Down
1 change: 1 addition & 0 deletions python/paddle/tensor/layer_function_generator.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -294,6 +294,7 @@ def func(x, name=None):
'float64',
'complex64',
'complex128',
'uint16',
],
op_type,
)
Expand Down