Skip to content

fix: serialize calls to ggml_backend_tensor_set #841

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

Closed
wants to merge 2 commits into from
Closed

fix: serialize calls to ggml_backend_tensor_set #841

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
48bc8f6
refactor: simplify weight conversion code
wbruna Sep 18, 2025
cc6b7ba
fix: serialize calls to ggml_backend_tensor_set
wbruna Sep 18, 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
31 changes: 15 additions & 16 deletions model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2024,6 +2024,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
const size_t total_tensors_to_process = processed_tensor_storages.size();
const int64_t t_start = ggml_time_ms();
int last_n_threads = 1;
std::mutex tensor_backend_mutex;

for (size_t file_index = 0; file_index < file_paths_.size(); file_index++) {
std::string file_path = file_paths_[file_index];
Expand Down Expand Up @@ -2211,26 +2212,24 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
i64_to_i32_vec((int64_t*)read_buffer.data(), (int32_t*)read_buffer.data(), tensor_storage.nelements());
}

if (tensor_storage.type == dst_tensor->type) {
// copy to device memory
t1 = ggml_time_ms();
convert_time_ms.fetch_add(t1 - t0);
t0 = ggml_time_ms();
ggml_backend_tensor_set(dst_tensor, read_buffer.data(), 0, ggml_nbytes(dst_tensor));
t1 = ggml_time_ms();
copy_to_backend_time_ms.fetch_add(t1 - t0);
} else {
// convert first, then copy to device memory
auto* tensor_buffer = &read_buffer;

if (tensor_storage.type != dst_tensor->type) {
// convert first
convert_buffer.resize(ggml_nbytes(dst_tensor));
convert_tensor((void*)read_buffer.data(), tensor_storage.type, (void*)convert_buffer.data(), dst_tensor->type, (int)tensor_storage.nelements() / (int)tensor_storage.ne[0], (int)tensor_storage.ne[0]);
t1 = ggml_time_ms();
convert_time_ms.fetch_add(t1 - t0);
t0 = ggml_time_ms();
ggml_backend_tensor_set(dst_tensor, convert_buffer.data(), 0, ggml_nbytes(dst_tensor));
t1 = ggml_time_ms();
copy_to_backend_time_ms.fetch_add(t1 - t0);
tensor_buffer = &convert_buffer;
}

t1 = ggml_time_ms();
convert_time_ms.fetch_add(t1 - t0);

// copy to device memory
std::lock_guard<std::mutex> lock(tensor_backend_mutex);
t0 = ggml_time_ms();
ggml_backend_tensor_set(dst_tensor, tensor_buffer->data(), 0, ggml_nbytes(dst_tensor));
t1 = ggml_time_ms();
copy_to_backend_time_ms.fetch_add(t1 - t0);
}
}
if (zip != NULL) {
Expand Down
Loading