Skip to content

[Common][PyTorch] Fuse scaling and unscaling of bf16 momentums into kernels #2632

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
yaox12 wants to merge 6 commits into
base: main
Choose a base branch
from
Open

[Common][PyTorch] Fuse scaling and unscaling of bf16 momentums into kernels #2632

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
4eb228a
fused scaling and unscaling of bf16 momentum
yaox12 Jan 29, 2026
38388ef
add more comments
yaox12 Jan 29, 2026
a608ec8
enable cuda graphs for bf16 momentums
yaox12 Jan 29, 2026
8931852
add tests
yaox12 Jan 29, 2026
b3b419c
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Jan 29, 2026
19ff141
update the check for store_param_remainders and capturable
yaox12 Jan 30, 2026
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
28 changes: 21 additions & 7 deletions tests/pytorch/test_fused_optimizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -407,6 +407,20 @@ def test_bf16_exp_avg_sq(self):
master_atol=2e-3,
)

@pytest.mark.skipif(not is_bf16_available(), reason="bf16 if not supported")
def test_bf16_exp_avg_and_exp_avg_sq(self):
self.gen_precision_aware_test(
use_fp8_params=False,
param_dtype=torch.bfloat16,
use_master_weights=True,
master_weight_dtype=torch.float32,
grad_dtype=torch.float32,
exp_avg_dtype=torch.bfloat16,
exp_avg_sq_dtype=torch.bfloat16,
master_rtol=2e-3,
master_atol=2e-3,
)
Comment on lines +410 to +422
Copy link
Contributor

@greptile-apps greptile-apps bot Jan 29, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Consider adding a test for capturable mode (CUDA Graphs) with BF16 momentums, since the PR description mentions "Enable CUDA Graphs for BF16 momentums" as a key feature. The current test only covers non-capturable mode.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

Comment on lines +410 to +422
Copy link
Contributor

@greptile-apps greptile-apps bot Jan 30, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Test only covers non-capturable mode. Add test for capturable mode with BF16 momentums since PR enables CUDA Graphs for this case.

Note: Check that gen_precision_aware_test supports a capturable parameter, or create a separate test method.


@pytest.mark.skipif(not is_bf16_available(), reason="bf16 if not supported")
@pytest.mark.skipif(not fp8_available, reason=reason_for_no_fp8)
def test_fp8_exp_avg_sq(self):
Expand Down Expand Up @@ -553,7 +567,7 @@ def forward(self, x):
return y


class AdamTest:
class TestAdamTest:

def setup_method(self, *, seed: int = 0) -> None:
torch.manual_seed(seed)
Expand All @@ -569,8 +583,8 @@ def setup_method(self, *, seed: int = 0) -> None:
def test_grad_scaler(self):
params_ = [p for p in self.model_.parameters() if p.requires_grad]
optimizer_ = te.optimizers.FusedAdam(params_, lr=self.lr, capturable=False)
scaler = torch.cuda.amp.GradScaler(enabled=True)
scaler_ = torch.cuda.amp.GradScaler(enabled=True)
scaler = torch.amp.GradScaler("cuda", enabled=True)
scaler_ = torch.amp.GradScaler("cuda", enabled=True)

for i in range(100):
x = torch.rand([32, 1, 28, 28]).cuda().to(memory_format=torch.channels_last)
Expand Down Expand Up @@ -620,8 +634,8 @@ def test_grad_scaler(self):
def test_grad_scaler_capturable(self):
params_ = [p for p in self.model_.parameters() if p.requires_grad]
optimizer_ = te.optimizers.FusedAdam(params_, lr=self.lr, capturable=True)
scaler = torch.cuda.amp.GradScaler(enabled=True)
scaler_ = torch.cuda.amp.GradScaler(enabled=True)
scaler = torch.amp.GradScaler("cuda", enabled=True)
scaler_ = torch.amp.GradScaler("cuda", enabled=True)

for i in range(100):
x = torch.rand([32, 1, 28, 28]).cuda().to(memory_format=torch.channels_last)
Expand Down Expand Up @@ -678,8 +692,8 @@ def test_grad_scaler_capturable_master(self):
optimizer_ = te.optimizers.FusedAdam(
params_, lr=self.lr, capturable=True, master_weights=master_weights
)
scaler = torch.cuda.amp.GradScaler(enabled=True)
scaler_ = torch.cuda.amp.GradScaler(enabled=True)
scaler = torch.amp.GradScaler("cuda", enabled=True)
scaler_ = torch.amp.GradScaler("cuda", enabled=True)

for i in range(100):
x = torch.rand([32, 1, 28, 28]).cuda().to(memory_format=torch.channels_last)
Expand Down
15 changes: 15 additions & 0 deletions transformer_engine/common/common.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -717,6 +717,21 @@ struct TypeInfo {
NVTE_ERROR("Invalid type."); \
}

#define TRANSFORMER_ENGINE_TYPE_SWITCH_FP32_BF16(dtype, type, ...) \
switch (dtype) { \
using namespace transformer_engine; \
case DType::kFloat32: { \
using type = float; \
{ __VA_ARGS__ } \
} break; \
case DType::kBFloat16: { \
using type = bf16; \
{ __VA_ARGS__ } \
} break; \
default: \
NVTE_ERROR("Invalid type, expected Float32 or BFloat16."); \
}

// Add a pack_size argument to select the packed type for FP4
#define TRANSFORMER_ENGINE_TYPE_SWITCH_FP4x2_ONLY(dtype, pack_size, type, ...) \
switch (dtype) { \
Expand Down
Loading
Loading