Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-12 05:17:21 +01:00
Code Issues Packages Projects Releases Wiki Activity

llama : avoid useless copies in dummy session writer

Browse Source
This commit is contained in:
Francis Couture-Harpin 2024-08-07 15:42:11 -04:00
parent 15fa07a5c5
commit dca7ad8627
1 changed files with 21 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

30
src/llama.cpp
View File

@ -17346,6 +17346,14 @@ struct llama_data_write {
virtual size_t get_size_written() = 0; virtual size_t get_size_written() = 0;
virtual ~llama_data_write() = default; virtual ~llama_data_write() = default;
std::vector<uint8_t> temp_buffer;
virtual void * get_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) {
temp_buffer.resize(size);
ggml_backend_tensor_get(tensor, temp_buffer.data(), offset, size);
return temp_buffer.data();
}
void write_string(const std::string & str) { void write_string(const std::string & str) {
uint32_t str_size = str.size(); uint32_t str_size = str.size();
@ -17465,9 +17473,9 @@ struct llama_data_write {
// Read each range of cells of k_size length each into tmp_buf and write out // Read each range of cells of k_size length each into tmp_buf and write out
for (const auto & range : cell_ranges) { for (const auto & range : cell_ranges) {
const size_t range_size = range.second - range.first; const size_t range_size = range.second - range.first;
tmp_buf.resize(range_size * k_size_row); const size_t buf_size = range_size * k_size_row;
ggml_backend_tensor_get(kv_self.k_l[il], tmp_buf.data(), range.first * k_size_row, range_size * k_size_row); const void * data = get_tensor_data(kv_self.k_l[il], range.first * k_size_row, buf_size);
write(tmp_buf.data(), tmp_buf.size()); write(data, buf_size);
} }
} }
@ -17486,9 +17494,9 @@ struct llama_data_write {
// Read each range of cells of v_size length each into tmp_buf and write out // Read each range of cells of v_size length each into tmp_buf and write out
for (const auto & range : cell_ranges) { for (const auto & range : cell_ranges) {
const size_t range_size = range.second - range.first; const size_t range_size = range.second - range.first;
tmp_buf.resize(range_size * v_size_row); const size_t buf_size = range_size * v_size_row;
ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), range.first * v_size_row, range_size * v_size_row); const void * data = get_tensor_data(kv_self.v_l[il], range.first * v_size_row, buf_size);
write(tmp_buf.data(), tmp_buf.size()); write(data, buf_size);
} }
} }
} else { } else {
@ -17514,9 +17522,9 @@ struct llama_data_write {
for (const auto & range : cell_ranges) { for (const auto & range : cell_ranges) {
const size_t range_size = range.second - range.first; const size_t range_size = range.second - range.first;
const size_t src_offset = (range.first + j * kv_size) * v_size_el; const size_t src_offset = (range.first + j * kv_size) * v_size_el;
tmp_buf.resize(range_size * v_size_el); const size_t buf_size = range_size * v_size_el;
ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), src_offset, tmp_buf.size()); const void * data = get_tensor_data(kv_self.v_l[il], src_offset, buf_size);
write(tmp_buf.data(), tmp_buf.size()); write(data, buf_size);
} }
} }
} }
@ -17881,6 +17889,10 @@ struct llama_data_write_dummy : llama_data_write {
size_written += size; size_written += size;
} }
void * get_tensor_data(const struct ggml_tensor * /* tensor */, size_t /* offset */, size_t /* size */) override {
return nullptr;
}
size_t get_size_written() override { size_t get_size_written() override {
return size_written; return size_written;
} }