zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-09-19 10:48:09 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Fix rope (#1340)

Browse Source 
* add test

* fix rope

* fix test
This commit is contained in:
Awni Hannun
2024-08-20 17:37:52 -07:00
committed by GitHub
parent bb1b76d9dc
commit d40e76809f
4 changed files with 26 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
mlx/backend/metal/kernels/rope.metal
View File

@@ -22,23 +22,18 @@ void rope_single_impl(
float sintheta = metal::fast::sin(theta);
// Compute the input and output indices
uint in_index_1, in_index_2;
uint out_index_1, out_index_2;
uint index_1, index_2;
if (traditional) {
out_index_1 = 2 * pos.x + pos.y * stride;
out_index_2 = out_index_1 + 1;
in_index_1 = 2 * pos.x + pos.y * stride;
in_index_2 = in_index_1 + 1;
index_1 = 2 * pos.x + pos.y * stride;
index_2 = index_1 + 1;
} else {
out_index_1 = pos.x + pos.y * stride;
out_index_2 = out_index_1 + grid.x;
in_index_1 = pos.x + pos.y * stride;
in_index_2 = in_index_1 + grid.x;
index_1 = pos.x + pos.y * stride;
index_2 = index_1 + grid.x;
}
// Read and write the output
float x1 = static_cast<float>(in[in_index_1]);
float x2 = static_cast<float>(in[in_index_2]);
float x1 = static_cast<float>(in[index_1]);
float x2 = static_cast<float>(in[index_2]);
float rx1;
float rx2;
if (forward) {
@@ -48,8 +43,8 @@ void rope_single_impl(
rx1 = x2 * sintheta + x1 * costheta;
rx2 = x2 * costheta - x1 * sintheta;
}
out[out_index_1] = static_cast<T>(rx1);
out[out_index_2] = static_cast<T>(rx2);
out[index_1] = static_cast<T>(rx1);
out[index_2] = static_cast<T>(rx2);
}
template <typename T, bool traditional, bool forward>

2
mlx/backend/metal/rope.cpp
View File

@@ -86,7 +86,7 @@ void RoPE::eval_gpu(
MTL::Size group_dims;
MTL::Size grid_dims;
if (single) {
compute_encoder->setBytes(&out_strides[1], sizeof(size_t), 4);
compute_encoder->setBytes(out_strides, sizeof(size_t), 4);
uint32_t dim0 = dims_ / 2;
group_dims = get_block_dims(dim0, n_batch, 1);
grid_dims = MTL::Size(dim0, n_batch, 1);

2
mlx/fast.cpp
View File

@@ -340,7 +340,7 @@ array rope(
if (inputs.size() == 2 &&
(inputs[1].ndim() != 1 || inputs[1].shape(0) != dims / 2)) {
std::ostringstream msg;
msg << "[rope] freqs must be one dimensional with size " << dims
msg << "[rope] freqs must be one dimensional with size " << dims / 2
<< " but got shape " << inputs[1].shape() << ".";
throw std::invalid_argument(msg.str());
}