zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-11-04 10:38:10 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

fix a couple bugs (#952)

Browse Source
This commit is contained in:
Awni Hannun
2024-04-02 12:07:41 -07:00
committed by GitHub
parent 1a87dc5ea8
commit 741eb28443
3 changed files with 9 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
mlx/io/gguf.cpp
View File

@@ -206,7 +206,7 @@ std::unordered_map<std::string, array> load_arrays(gguf_ctx* ctx) {
std::unordered_map<std::string, array> array_map;
gguf_tensor tensor;
auto check_insert = [](auto inserted) {
auto check_insert = [](const auto& inserted) {
if (!inserted.second) {
std::ostringstream msg;
msg << "[load_gguf] Duplicate parameter name " << inserted.first->second
@@ -216,6 +216,7 @@ std::unordered_map<std::string, array> load_arrays(gguf_ctx* ctx) {
};
while (gguf_get_tensor(ctx, &tensor)) {
std::string name(tensor.name, tensor.namelen);
if (tensor.type == GGUF_TYPE_Q4_0 || tensor.type == GGUF_TYPE_Q4_1 ||
tensor.type == GGUF_TYPE_Q8_0) {
gguf_load_quantized(array_map, tensor);
@@ -224,7 +225,7 @@ std::unordered_map<std::string, array> load_arrays(gguf_ctx* ctx) {
const auto& [data, dtype] = extract_tensor_data(&tensor);
array loaded_array = array(data, get_shape(tensor), dtype);
array_map.insert({name, loaded_array});
check_insert(array_map.insert({name, loaded_array}));
}
}
return array_map;

3
mlx/io/gguf_quants.cpp
View File

@@ -106,6 +106,7 @@ void gguf_load_quantized(
}
std::string name(tensor.name, tensor.namelen);
std::vector<int> shape = get_shape(tensor);
const uint64_t weights_per_block = 32;
if (shape[shape.size() - 1] % weights_per_block != 0) {
@@ -136,7 +137,7 @@ void gguf_load_quantized(
extract_q8_0_data(tensor, weights, scales, biases);
}
a.emplace(std::move(name), std::move(weights));
a.emplace(name, std::move(weights));
auto check_insert = [](const auto& inserted) {
if (!inserted.second) {