Skip to content

Commit

Permalink
fix layer_norm decompose dtyte bugs, polish codes (#61631)
Browse files Browse the repository at this point in the history
  • Loading branch information
lxd-cumt committed Feb 6, 2024
1 parent 9250f66 commit e5a85b6
Showing 1 changed file with 26 additions and 19 deletions.
45 changes: 26 additions & 19 deletions paddle/fluid/primitive/composite/composite.h
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,9 @@ namespace paddle {
namespace primitive {
namespace details {

// empty_shape means x.shape=[]
static std::vector<int64_t> empty_shape;

template <typename T>
Tensor mean_decomp(const Tensor& x, const IntArray& axis, bool keepdim) {
auto org_dtype = x.dtype();
Expand Down Expand Up @@ -345,62 +348,66 @@ std::tuple<Tensor, Tensor, Tensor> layer_norm_decomp(

// cast dtype to float32 if dtype =float16 or bfloat16
if (need_cast) {
x_cast = cast<T>(x_cast, phi::DataType::FLOAT32);
x_cast = cast<T>(x_cast, DataType::FLOAT32);
}

auto x_dim = common::vectorize<int64_t>(x.dims());
for (size_t i = begin_norm_axis; i < x_dim.size(); i++) {
axis.push_back(static_cast<int64_t>(i));
}
auto mean_ = mean_decomp<T>(x_cast, IntArray(axis), true);
auto mean_ = mean_decomp<T>(x_cast, axis, true);
auto difference = x_cast - mean_;
auto var_tmp1 = difference * difference;
auto variance = mean_decomp<T>(var_tmp1, IntArray(axis), true);
auto variance = mean_decomp<T>(var_tmp1, axis, true);
auto var_tmp3 = variance + epsilon;
auto rsqrt_var = elementwise_pow<T>(
var_tmp3,
full<T>(common::vectorize(var_tmp3.dims()), -0.5, var_tmp3.dtype()));
var_tmp3, full<T>(empty_shape, -0.5, var_tmp3.dtype()));
auto out = difference * rsqrt_var;

auto scale_ptr = scale.get_ptr();
auto bias_ptr = bias.get_ptr();

std::vector<int64_t> slice_shape;
for (int64_t i = begin_norm_axis; i < static_cast<int64_t>(x_dim.size());
i++) {
slice_shape.push_back(x_dim[i]);
std::vector<int64_t> slice_shape_l;
std::vector<int64_t> slice_shape_r;
for (int64_t i = 0; i < static_cast<int64_t>(x_dim.size()); i++) {
if (i < begin_norm_axis) {
slice_shape_l.push_back(x_dim[i]);
} else {
slice_shape_r.push_back(x_dim[i]);
}
}
Tensor scale_cast;
if (scale_ptr) {
if (slice_shape != scale_ptr->shape()) {
scale_cast = reshape<T>(*scale_ptr, slice_shape);
if (slice_shape_r != scale_ptr->shape()) {
scale_cast = reshape<T>(*scale_ptr, slice_shape_r);
} else {
scale_cast = *scale_ptr;
}
if (need_cast) {
scale_cast = cast<T>(scale_cast, phi::DataType::FLOAT32);
scale_cast = cast<T>(scale_cast, DataType::FLOAT32);
}
out = out * scale_cast;
}
Tensor bias_cast;
if (bias_ptr) {
if (slice_shape != bias_ptr->shape()) {
bias_cast = reshape<T>(*bias_ptr, slice_shape);
if (slice_shape_r != bias_ptr->shape()) {
bias_cast = reshape<T>(*bias_ptr, slice_shape_r);
} else {
bias_cast = *bias_ptr;
}
if (need_cast) {
bias_cast = cast<T>(bias_cast, phi::DataType::FLOAT32);
bias_cast = cast<T>(bias_cast, DataType::FLOAT32);
}
out = out + bias_cast;
}
mean_ = reshape<T>(mean_, std::vector<int64_t>({-1}));
variance = reshape<T>(variance, std::vector<int64_t>({-1}));
mean_ = reshape<T>(mean_, slice_shape_l);
variance = reshape<T>(variance, slice_shape_l);

// same as LayerNormInferMeta
// x: float32 --> out: float32, mean: float32, variance: float32
// x: float16 --> out: float16, mean: float32, variance: float32
if (need_cast) {
out = cast<T>(out, org_dtype);
mean_ = cast<T>(mean_, org_dtype);
variance = cast<T>(variance, org_dtype);
}

return std::make_tuple(out, mean_, variance);
Expand Down

0 comments on commit e5a85b6

Please sign in to comment.