zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-07-30 06:31:21 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Donation bug (#933)

Browse Source 
* donation

* buf

* fix bug in softmax

* comment

* remove print
This commit is contained in:
Awni Hannun 2024-03-30 10:08:54 -07:00 committed by GitHub
parent f48bc496c7
commit 8915901966
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 31 additions and 33 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

54
mlx/backend/metal/normalization.cpp
View File

@ -19,7 +19,7 @@ void RMSNorm::eval_gpu(
// Make sure that the last dimension is contiguous // Make sure that the last dimension is contiguous
std::vector<array> copies; std::vector<array> copies;
auto check_input = [&copies, &s](const array& x) { auto check_input = [&copies, &s](const array& x) -> const array& {
bool no_copy = x.strides()[x.ndim() - 1] == 1; bool no_copy = x.strides()[x.ndim() - 1] == 1;
if (x.ndim() > 1) { if (x.ndim() > 1) {
auto s = x.strides()[x.ndim() - 2]; auto s = x.strides()[x.ndim() - 2];
@ -28,10 +28,9 @@ void RMSNorm::eval_gpu(
if (no_copy) { if (no_copy) {
return x; return x;
} else { } else {
array x_copy(x.shape(), x.dtype(), nullptr, {}); copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
copy_gpu(x, x_copy, CopyType::General, s); copy_gpu(x, copies.back(), CopyType::General, s);
copies.push_back(x_copy); return copies.back();
return x_copy;
} }
}; };
const array& x = check_input(inputs[0]); const array& x = check_input(inputs[0]);
@ -106,15 +105,13 @@ void RMSNormVJP::eval_gpu(
// is contiguous (no broadcasts or holes) and that the input strides are the // is contiguous (no broadcasts or holes) and that the input strides are the
// same as the cotangent strides but for now this is simpler. // same as the cotangent strides but for now this is simpler.
std::vector<array> copies; std::vector<array> copies;
auto check_input = [&copies, &s](const array& x) { auto check_input = [&copies, &s](const array& x) -> const array& {
if (x.flags().row_contiguous) { if (x.flags().row_contiguous) {
return x; return x;
} }
copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
array x_copy(x.shape(), x.dtype(), nullptr, {}); copy_gpu(x, copies.back(), CopyType::General, s);
copy_gpu(x, x_copy, CopyType::General, s); return copies.back();
copies.push_back(x_copy);
return x_copy;
}; };
const array& x = check_input(inputs[0]); const array& x = check_input(inputs[0]);
const array& w = inputs[1]; const array& w = inputs[1];
@ -149,8 +146,11 @@ void RMSNormVJP::eval_gpu(
gw_temp.set_data(allocator::malloc_or_wait(gw_temp.nbytes())); gw_temp.set_data(allocator::malloc_or_wait(gw_temp.nbytes()));
} }
copies.push_back(gw_temp); copies.push_back(gw_temp);
array zero(0, gw.dtype()); {
copy_gpu(zero, gw, CopyType::Scalar, s); array zero(0, gw.dtype());
copy_gpu(zero, gw, CopyType::Scalar, s);
copies.push_back(std::move(zero));
}
const int simd_size = 32; const int simd_size = 32;
const int n_reads = RMS_N_READS; const int n_reads = RMS_N_READS;
@ -212,7 +212,7 @@ void LayerNorm::eval_gpu(
// Make sure that the last dimension is contiguous // Make sure that the last dimension is contiguous
std::vector<array> copies; std::vector<array> copies;
auto check_input = [&copies, &s](const array& x) { auto check_input = [&copies, &s](const array& x) -> const array& {
bool no_copy = x.strides()[x.ndim() - 1] == 1; bool no_copy = x.strides()[x.ndim() - 1] == 1;
if (x.ndim() > 1) { if (x.ndim() > 1) {
auto s = x.strides()[x.ndim() - 2]; auto s = x.strides()[x.ndim() - 2];
@ -221,10 +221,9 @@ void LayerNorm::eval_gpu(
if (no_copy) { if (no_copy) {
return x; return x;
} else { } else {
array x_copy(x.shape(), x.dtype(), nullptr, {}); copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
copy_gpu(x, x_copy, CopyType::General, s); copy_gpu(x, copies.back(), CopyType::General, s);
copies.push_back(x_copy); return copies.back();
return x_copy;
} }
}; };
const array& x = check_input(inputs[0]); const array& x = check_input(inputs[0]);
@ -300,15 +299,13 @@ void LayerNormVJP::eval_gpu(
// is contiguous (no broadcasts or holes) and that the input strides are the // is contiguous (no broadcasts or holes) and that the input strides are the
// same as the cotangent strides but for now this is simpler. // same as the cotangent strides but for now this is simpler.
std::vector<array> copies; std::vector<array> copies;
auto check_input = [&copies, &s](const array& x) { auto check_input = [&copies, &s](const array& x) -> const array& {
if (x.flags().row_contiguous) { if (x.flags().row_contiguous) {
return x; return x;
} }
copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
array x_copy(x.shape(), x.dtype(), nullptr, {}); copy_gpu(x, copies.back(), CopyType::General, s);
copy_gpu(x, x_copy, CopyType::General, s); return copies.back();
copies.push_back(x_copy);
return x_copy;
}; };
const array& x = check_input(inputs[0]); const array& x = check_input(inputs[0]);
const array& w = inputs[1]; const array& w = inputs[1];
@ -345,9 +342,12 @@ void LayerNormVJP::eval_gpu(
gw_temp.set_data(allocator::malloc_or_wait(gw_temp.nbytes())); gw_temp.set_data(allocator::malloc_or_wait(gw_temp.nbytes()));
} }
copies.push_back(gw_temp); copies.push_back(gw_temp);
array zero(0, gw.dtype()); {
copy_gpu(zero, gw, CopyType::Scalar, s); array zero(0, gw.dtype());
copy_gpu(zero, gb, CopyType::Scalar, s); copy_gpu(zero, gw, CopyType::Scalar, s);
copy_gpu(zero, gb, CopyType::Scalar, s);
copies.push_back(std::move(zero));
}
// Finish with the gradient for b in case we had a b // Finish with the gradient for b in case we had a b
auto compute_encoder = d.get_command_encoder(s.index); auto compute_encoder = d.get_command_encoder(s.index);

10
mlx/backend/metal/softmax.cpp
View File

@ -21,7 +21,7 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
// Make sure that the last dimension is contiguous // Make sure that the last dimension is contiguous
std::vector<array> copies; std::vector<array> copies;
auto check_input = [&copies, &s](const array& x) { auto check_input = [&copies, &s](const array& x) -> const array& {
bool no_copy = x.strides()[x.ndim() - 1] == 1; bool no_copy = x.strides()[x.ndim() - 1] == 1;
if (x.ndim() > 1) { if (x.ndim() > 1) {
auto s = x.strides()[x.ndim() - 2]; auto s = x.strides()[x.ndim() - 2];
@ -30,10 +30,9 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
if (no_copy) { if (no_copy) {
return x; return x;
} else { } else {
array x_copy(x.shape(), x.dtype(), nullptr, {}); copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
copy_gpu(x, x_copy, CopyType::General, s); copy_gpu(x, copies.back(), CopyType::General, s);
copies.push_back(x_copy); return copies.back();
return x_copy;
} }
}; };
const array& in = check_input(inputs[0]); const array& in = check_input(inputs[0]);
@ -81,7 +80,6 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
compute_encoder->setComputePipelineState(kernel); compute_encoder->setComputePipelineState(kernel);
set_array_buffer( set_array_buffer(
compute_encoder, in.data_shared_ptr() == nullptr ? out : in, 0); compute_encoder, in.data_shared_ptr() == nullptr ? out : in, 0);
set_array_buffer(compute_encoder, in, 0);
set_array_buffer(compute_encoder, out, 1); set_array_buffer(compute_encoder, out, 1);
compute_encoder->setBytes(&axis_size, sizeof(int), 2); compute_encoder->setBytes(&axis_size, sizeof(int), 2);
compute_encoder->setThreadgroupMemoryLength(simd_size * in.itemsize(), 0); compute_encoder->setThreadgroupMemoryLength(simd_size * in.itemsize(), 0);