Skip to content

[ET] enabling half dtype input for quantization #11479

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 9 commits into from
Jun 14, 2025
Merged

[ET] enabling half dtype input for quantization #11479

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
7b7290d
[ET] enabling half dtype input for quantization
Jun 9, 2025
58b1237
Update on "[ET] enabling half dtype input for quantization"
Jun 9, 2025
96f1dcd
Update on "[ET] enabling half dtype input for quantization"
Jun 11, 2025
19a2fc3
Update on "[ET] enabling half dtype input for quantization"
Jun 11, 2025
849944d
Update on "[ET] enabling half dtype input for quantization"
Jun 12, 2025
0d595f7
Update on "[ET] enabling half dtype input for quantization"
Jun 12, 2025
9e31ad6
Update on "[ET] enabling half dtype input for quantization"
Jun 13, 2025
a49b0cf
Update on "[ET] enabling half dtype input for quantization"
Jun 13, 2025
316d2c0
Update on "[ET] enabling half dtype input for quantization"
Jun 13, 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
4 changes: 2 additions & 2 deletions kernels/quantized/cpu/op_quantize.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -150,7 +150,7 @@ Tensor& quantize_per_tensor_out(
break;

switch (input.scalar_type()) {
ET_FORALL_FLOAT_TYPES(CALCULATE_FLOAT_TYPE);
ET_FORALL_FLOATH_TYPES(CALCULATE_FLOAT_TYPE);
default:
ET_CHECK_MSG(
false,
Expand Down Expand Up @@ -346,7 +346,7 @@ Tensor& quantize_per_channel_out(
break;

switch (input.scalar_type()) {
ET_FORALL_FLOAT_TYPES(CALCULATE_FLOAT_TYPE);
ET_FORALL_FLOATH_TYPES(CALCULATE_FLOAT_TYPE);
default:
ET_CHECK_MSG(
false,
Expand Down
65 changes: 65 additions & 0 deletions kernels/quantized/test/op_quantize_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,6 +49,32 @@ void test_dtype() {
EXPECT_TENSOR_EQ(out, expected);
}

template <ScalarType INPUT_DTYPE>
void test_input_dtype() {
TensorFactory<INPUT_DTYPE> tf_input;

Tensor input = tf_input.full({3, 5}, 4);
double scale = 0.5;
int64_t zero_point = 108;
int64_t quant_min = 0;
int64_t quant_max = 127;

TensorFactory<ScalarType::Char> tfo;
Tensor out = tfo.zeros({3, 5});
// 4 / 0.5 + 108 = 116
Tensor expected = tfo.full({3, 5}, 116);
quantize_per_tensor_out(
input, scale, zero_point, quant_min, quant_max, ScalarType::Char, out);

EXPECT_TENSOR_EQ(out, expected);
}

TEST(OpQuantizeOutTest, AllInputDtypesSupported) {
test_input_dtype<ScalarType::Float>();
test_input_dtype<ScalarType::Half>();
test_input_dtype<ScalarType::Double>();
}

TEST(OpQuantizeOutTest, AllDtypesSupported) {
test_dtype<ScalarType::Byte>();
test_dtype<ScalarType::Char>();
Expand All @@ -58,6 +84,45 @@ TEST(OpQuantizeOutTest, AllDtypesSupported) {
test_dtype<ScalarType::Int>();
}

TEST(OpQuantizeOutTest, DoubleInputTest) {
TensorFactory<ScalarType::Double> tf_double;

// Test with a more complex value that might have precision differences
Tensor input = tf_double.full({2, 3}, 3.14159265359);
double scale = 0.01;
int64_t zero_point = -100;
int64_t quant_min = 0;
int64_t quant_max = 255;

TensorFactory<ScalarType::Byte> tfo;
Tensor out = tfo.zeros({2, 3});
// 3.14159265359 / 0.01 - 100 = 214.159265359
Tensor expected = tfo.full({2, 3}, 214);
quantize_per_tensor_out(
input, scale, zero_point, quant_min, quant_max, ScalarType::Byte, out);

EXPECT_TENSOR_EQ(out, expected);
}

TEST(OpQuantizeOutTest, HalfInputTest) {
TensorFactory<ScalarType::Half> tf_half;

Tensor input = tf_half.full({2, 3}, 2.5);
double scale = 0.5;
int64_t zero_point = 10;
int64_t quant_min = -128;
int64_t quant_max = 127;

TensorFactory<ScalarType::Char> tfo;
Tensor out = tfo.zeros({2, 3});
// 2.5 / 0.5 + 10 = 15
Tensor expected = tfo.full({2, 3}, 15);
quantize_per_tensor_out(
input, scale, zero_point, quant_min, quant_max, ScalarType::Char, out);

EXPECT_TENSOR_EQ(out, expected);
}

TEST(OpQuantizeOutTest, TensorArgOverload) {
TensorFactory<ScalarType::Float> tf_float;
TensorFactory<ScalarType::Double> tf_double;
Expand Down
Loading