Skip to content

Support Half/BFloat16 in native_group_norm (needs accuracy fix) #7846

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

Closed
swolchok wants to merge 2 commits into from
Closed

Support Half/BFloat16 in native_group_norm (needs accuracy fix) #7846

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
21f6f98
Update
swolchok Jan 22, 2025
d3ecd5b
Update
swolchok Jan 23, 2025
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: 8 additions & 7 deletions kernels/portable/cpu/op_native_group_norm.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -78,8 +78,9 @@ void group_norm(
// compute E[X] and Var[x] = E[x^2] - E[x]^2
CTYPE sum = reduce_add(x, inner_size);
CTYPE sq_sum = vec_powerf(x, inner_size);
CTYPE mean_value = sum / inner_size;
CTYPE variance = sq_sum / inner_size - mean_value * mean_value;
CTYPE mean_value = sum / static_cast<CTYPE>(inner_size);
CTYPE variance =
sq_sum / static_cast<CTYPE>(inner_size) - mean_value * mean_value;
CTYPE std = std::sqrt(variance + eps);
CTYPE rstd_value = 1.0 / std;

Expand All @@ -93,10 +94,10 @@ void group_norm(
const size_t g = i % G;
for (size_t j = 0; j < D; j++) {
const size_t ch = g * D + j;
const CTYPE scale =
rstd_value * (weight_data == nullptr ? 1.0 : weight_data[ch]);
const CTYPE beta =
-scale * mean_value + (bias_data == nullptr ? 0.0 : bias_data[ch]);
const CTYPE scale = rstd_value *
(weight_data == nullptr ? CTYPE(1.0) : weight_data[ch]);
const CTYPE beta = -scale * mean_value +
(bias_data == nullptr ? CTYPE(0.0) : bias_data[ch]);
x = input_data + (i * D + j) * HxW;
CTYPE* y = out_data + (i * D + j) * HxW;
for (size_t k = 0; k < HxW; k++) {
Expand Down Expand Up @@ -185,7 +186,7 @@ std::tuple<Tensor&, Tensor&, Tensor&> native_group_norm_out(

constexpr auto name = "native_group_norm.out";

ET_SWITCH_FLOAT_TYPES(input.scalar_type(), ctx, name, CTYPE, [&]() {
ET_SWITCH_FLOATHBF16_TYPES(input.scalar_type(), ctx, name, CTYPE, [&]() {
group_norm<CTYPE>(
input, weight, bias, N, C, HxW, group, eps, out, mean_out, rstd_out);
});
Expand Down
238 changes: 134 additions & 104 deletions kernels/test/op_native_group_norm_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,110 +20,140 @@ using exec_aten::ScalarType;
using exec_aten::Tensor;
using torch::executor::testing::TensorFactory;

::std::tuple<Tensor&, Tensor&, Tensor&> op_native_group_norm_out(
const Tensor& input,
const optional<Tensor>& weight,
const optional<Tensor>& bias,
int64_t N,
int64_t C,
int64_t HxW,
int64_t group,
double eps,
Tensor& out0,
Tensor& out1,
Tensor& out2) {
executorch::runtime::KernelRuntimeContext context{};
return torch::executor::aten::native_group_norm_outf(
context, input, weight, bias, N, C, HxW, group, eps, out0, out1, out2);
}
class OpNativeGroupNormOutTest : public OperatorTest {
protected:
::std::tuple<Tensor&, Tensor&, Tensor&> op_native_group_norm_out(
const Tensor& input,
const optional<Tensor>& weight,
const optional<Tensor>& bias,
int64_t N,
int64_t C,
int64_t HxW,
int64_t group,
double eps,
Tensor& out0,
Tensor& out1,
Tensor& out2) {
executorch::runtime::KernelRuntimeContext context{};
return torch::executor::aten::native_group_norm_outf(
context, input, weight, bias, N, C, HxW, group, eps, out0, out1, out2);
}
template <ScalarType DTYPE>
void test_dtype() {
TensorFactory<DTYPE> tf;

TEST(OpNativeGroupNormOutTest, SmokeTest) {
TensorFactory<ScalarType::Float> tfFloat;
Tensor input = tf.make(
{5, 6, 2, 2},
{-0.8125, 0.0625, -2.7500, -3.0625, -1.1250, -2.1250, -1.3125,
-4.0625, 2.8125, -2.0625, 4.2500, 3.5000, -0.3750, 1.6250,
4.3125, -1.0625, -2.8750, 3.3750, 4.9375, 4.0625, -3.0625,
-1.8750, -2.7500, -2.5625, -0.1875, -3.0000, -2.7500, 0.6875,
-3.2500, -3.1875, 1.0000, -4.6250, -0.1875, -1.7500, 4.5000,
-1.8750, -2.6875, 4.8125, -3.8125, -2.9375, -1.1875, 2.8750,
0.7500, 2.8750, 1.1250, -0.6250, -2.2500, -3.7500, 3.2500,
-0.3750, -2.0625, -4.7500, 2.0625, 3.0000, -3.1875, -4.1250,
-3.7500, 1.2500, -2.3125, 1.5625, 3.1250, 0.3125, 3.2500,
-2.7500, -3.8125, -4.2500, -4.3125, -0.5625, -0.4375, 2.9375,
-1.3750, -0.6250, -2.5625, -4.5625, 0.1250, -3.5000, -5.0000,
-1.0000, -4.6875, -0.6875, 1.1250, 1.8750, -4.5000, 4.3125,
4.5625, 0.2500, -3.6250, 4.5625, -3.5000, -2.1250, -3.6250,
-2.9375, 3.6875, 3.9375, 4.3750, 3.0625, 2.4375, 2.0625,
-2.4375, -3.9375, 3.6875, 2.7500, -0.8750, -0.9375, 2.7500,
-2.4375, -2.3750, -0.9375, -4.8750, 0.1875, 3.5000, -2.0000,
-0.2500, -2.7500, 0.3125, 1.2500, -0.5625, 0.0000, 1.8125,
1.0625});
optional<Tensor> weight =
tf.make({6}, {4.5625, -2.8750, -0.6875, 0.5625, -2.0625, -2.7500});
optional<Tensor> bias =
tf.make({6}, {-0.5000, -2.7500, 1.1875, 3.6875, 3.8125, 4.6875});
double eps = 1e-5;
Tensor out0 = tf.zeros({5, 6, 2, 2});
Tensor out1 = tf.zeros({5, 3});
Tensor out2 = tf.zeros({5, 3});
Tensor out0_expected = tf.make(
{5, 6, 2, 2},
{3.419882, 6.578348, -3.573864, -4.701888, -4.509254, -2.234663,
-4.082768, 2.172355, 0.838826, 2.270225, 0.416747, 0.636962,
3.207030, 3.687500, 4.333131, 3.041869, 5.547079, 1.649148,
0.674665, 1.220376, 7.156189, 6.168714, 6.896327, 6.740410,
3.509863, -3.022041, -2.441427, 5.542011, -0.794903, -0.886369,
-7.014627, 1.217361, 1.120617, 1.463606, 0.091652, 1.491045,
3.293219, 4.640229, 3.091168, 3.248319, 4.895990, 1.114683,
3.092597, 1.114683, 3.262238, 5.434066, 7.450763, 9.312329,
5.570122, 0.101119, -2.444796, -6.499403, -5.446074, -6.337338,
-0.454995, 0.436269, 2.228491, 0.871598, 1.838385, 0.786793,
4.362284, 3.737805, 4.390039, 3.057817, 5.814659, 6.202621,
6.258044, 2.932658, 3.366583, -0.623879, 4.475045, 3.588276,
-0.082914, -4.936279, 6.438795, -2.357929, 0.714463, -5.402106,
0.236606, -5.879963, 1.176247, 1.021916, 2.333727, 0.520341,
4.275447, 3.549392, 2.896994, 4.275447, 6.120910, 5.298480,
6.195676, 5.784461, 2.033296, 1.833920, 1.485010, 2.531738,
3.193988, 2.532378, -5.406940, -8.053379, -6.467402, -5.425139,
-1.395059, -1.325575, 0.266062, 1.622680, 1.606336, 1.230405,
2.809896, 3.893110, 4.601880, 3.425055, 4.374411, 8.283354,
3.494898, 2.029045, 6.088204, 4.915522, 1.136877, 2.700454});
Tensor out1_expected = tf.make(
{5, 3},
{-1.89843750,
1.62500000,
-0.09375000,
-1.91406250,
-0.49218744,
-0.02343750,
-0.77343756,
0.08593753,
-1.55468738,
-2.73437500,
1.07031238,
0.35937503,
0.34374997,
-0.77343750,
0.10937499});
Tensor out2_expected = tf.make(
{5, 3},
{0.79116172,
0.42708409,
0.30238494,
0.50903118,
0.31929117,
0.45128885,
0.33067191,
0.39473253,
0.42994878,
0.53187561,
0.29930803,
0.29000264,
0.38669431,
0.38038814,
0.75809801});
op_native_group_norm_out(
input, weight, bias, 5, 6, 4, 3, eps, out0, out1, out2);
if (DTYPE == ScalarType::Half || DTYPE == ScalarType::BFloat16) {
EXPECT_TENSOR_CLOSE_WITH_TOL(
out0,
out0_expected,
2e-1,
Copy link
Contributor

@manuelcandales manuelcandales Jan 23, 2025

Choose a reason for hiding this comment

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

That's 20%. This doesn't make me feel comfortable.

Copy link
Contributor Author

@swolchok swolchok Jan 23, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

group_norm is one of the ops that automatic mixed precision will autocast to float32: https://intel.github.io/intel-extension-for-pytorch/xpu/1.10.200+gpu/tutorials/features/amp.html

I think the norm ops are just particularly prone to roundoff error, but I'm certainly not a numerical analysis person.

(unresolving for posterity)

Copy link
Contributor Author

@swolchok swolchok Jan 23, 2025

Choose a reason for hiding this comment

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

Dug in a little further. Here's the PR that originally made PyTorch group_norm support Half: https://github.com/pytorch/pytorch/pull/100234/files#diff-7927db349f568afca2de9b94d74ea5c3b8cb468cb6a433d0cc1e61e65c515a36

It looks like the test is atol=rtol=5e-3. I think it's reasonable to argue that if we can't get the tolerances to be broadly similar then we have a correctness issue and thus don't actually support Half. I'll see what I can do; this one might have to wait for code sharing.

Copy link
Contributor Author

@swolchok swolchok Jan 23, 2025

Choose a reason for hiding this comment

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

test doesn't pass with atol=rtol=5e-3. Holding off on group_norm until we have code sharing.

Copy link
Contributor Author

@swolchok swolchok Oct 29, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

until we have code sharing.

this is a portable op, so code sharing is not applicable. Not sure what we should do about it going forward.

executorch::runtime::testing::internal::kDefaultAtol);
EXPECT_TENSOR_CLOSE_WITH_TOL(
out1,
out1_expected,
1e-2,
executorch::runtime::testing::internal::kDefaultAtol);
EXPECT_TENSOR_CLOSE_WITH_TOL(
out2,
out2_expected,
1e-2,
executorch::runtime::testing::internal::kDefaultAtol);
} else {
EXPECT_TENSOR_CLOSE(out0, out0_expected);
EXPECT_TENSOR_CLOSE(out1, out1_expected);
EXPECT_TENSOR_CLOSE(out2, out2_expected);
}
}
};

Tensor input = tfFloat.make(
{5, 6, 2, 2},
{-0.8125, 0.0625, -2.7500, -3.0625, -1.1250, -2.1250, -1.3125, -4.0625,
2.8125, -2.0625, 4.2500, 3.5000, -0.3750, 1.6250, 4.3125, -1.0625,
-2.8750, 3.3750, 4.9375, 4.0625, -3.0625, -1.8750, -2.7500, -2.5625,
-0.1875, -3.0000, -2.7500, 0.6875, -3.2500, -3.1875, 1.0000, -4.6250,
-0.1875, -1.7500, 4.5000, -1.8750, -2.6875, 4.8125, -3.8125, -2.9375,
-1.1875, 2.8750, 0.7500, 2.8750, 1.1250, -0.6250, -2.2500, -3.7500,
3.2500, -0.3750, -2.0625, -4.7500, 2.0625, 3.0000, -3.1875, -4.1250,
-3.7500, 1.2500, -2.3125, 1.5625, 3.1250, 0.3125, 3.2500, -2.7500,
-3.8125, -4.2500, -4.3125, -0.5625, -0.4375, 2.9375, -1.3750, -0.6250,
-2.5625, -4.5625, 0.1250, -3.5000, -5.0000, -1.0000, -4.6875, -0.6875,
1.1250, 1.8750, -4.5000, 4.3125, 4.5625, 0.2500, -3.6250, 4.5625,
-3.5000, -2.1250, -3.6250, -2.9375, 3.6875, 3.9375, 4.3750, 3.0625,
2.4375, 2.0625, -2.4375, -3.9375, 3.6875, 2.7500, -0.8750, -0.9375,
2.7500, -2.4375, -2.3750, -0.9375, -4.8750, 0.1875, 3.5000, -2.0000,
-0.2500, -2.7500, 0.3125, 1.2500, -0.5625, 0.0000, 1.8125, 1.0625});
optional<Tensor> weight =
tfFloat.make({6}, {4.5625, -2.8750, -0.6875, 0.5625, -2.0625, -2.7500});
optional<Tensor> bias =
tfFloat.make({6}, {-0.5000, -2.7500, 1.1875, 3.6875, 3.8125, 4.6875});
double eps = 1e-5;
Tensor out0 = tfFloat.zeros({5, 6, 2, 2});
Tensor out1 = tfFloat.zeros({5, 3});
Tensor out2 = tfFloat.zeros({5, 3});
Tensor out0_expected = tfFloat.make(
{5, 6, 2, 2},
{3.419882, 6.578348, -3.573864, -4.701888, -4.509254, -2.234663,
-4.082768, 2.172355, 0.838826, 2.270225, 0.416747, 0.636962,
3.207030, 3.687500, 4.333131, 3.041869, 5.547079, 1.649148,
0.674665, 1.220376, 7.156189, 6.168714, 6.896327, 6.740410,
3.509863, -3.022041, -2.441427, 5.542011, -0.794903, -0.886369,
-7.014627, 1.217361, 1.120617, 1.463606, 0.091652, 1.491045,
3.293219, 4.640229, 3.091168, 3.248319, 4.895990, 1.114683,
3.092597, 1.114683, 3.262238, 5.434066, 7.450763, 9.312329,
5.570122, 0.101119, -2.444796, -6.499403, -5.446074, -6.337338,
-0.454995, 0.436269, 2.228491, 0.871598, 1.838385, 0.786793,
4.362284, 3.737805, 4.390039, 3.057817, 5.814659, 6.202621,
6.258044, 2.932658, 3.366583, -0.623879, 4.475045, 3.588276,
-0.082914, -4.936279, 6.438795, -2.357929, 0.714463, -5.402106,
0.236606, -5.879963, 1.176247, 1.021916, 2.333727, 0.520341,
4.275447, 3.549392, 2.896994, 4.275447, 6.120910, 5.298480,
6.195676, 5.784461, 2.033296, 1.833920, 1.485010, 2.531738,
3.193988, 2.532378, -5.406940, -8.053379, -6.467402, -5.425139,
-1.395059, -1.325575, 0.266062, 1.622680, 1.606336, 1.230405,
2.809896, 3.893110, 4.601880, 3.425055, 4.374411, 8.283354,
3.494898, 2.029045, 6.088204, 4.915522, 1.136877, 2.700454});
Tensor out1_expected = tfFloat.make(
{5, 3},
{-1.89843750,
1.62500000,
-0.09375000,
-1.91406250,
-0.49218744,
-0.02343750,
-0.77343756,
0.08593753,
-1.55468738,
-2.73437500,
1.07031238,
0.35937503,
0.34374997,
-0.77343750,
0.10937499});
Tensor out2_expected = tfFloat.make(
{5, 3},
{0.79116172,
0.42708409,
0.30238494,
0.50903118,
0.31929117,
0.45128885,
0.33067191,
0.39473253,
0.42994878,
0.53187561,
0.29930803,
0.29000264,
0.38669431,
0.38038814,
0.75809801});
op_native_group_norm_out(
input, weight, bias, 5, 6, 4, 3, eps, out0, out1, out2);
EXPECT_TENSOR_CLOSE(out0, out0_expected);
EXPECT_TENSOR_CLOSE(out1, out1_expected);
EXPECT_TENSOR_CLOSE(out2, out2_expected);
TEST_F(OpNativeGroupNormOutTest, SmokeTest) {
#define TEST_ENTRY(ctype, dtype) test_dtype<ScalarType::dtype>();
ET_FORALL_FLOATHBF16_TYPES(TEST_ENTRY)
#undef TEST_ENTRY
}
Loading