Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-27 06:39:25 +01:00
Code Issues Packages Projects Releases Wiki Activity

llama : reduce useless copies when saving session (#8916)

Browse Source 
* llama : avoid useless copies in dummy session writer

* llama : avoid double tensor copy when saving session to buffer
This commit is contained in:
compilade 2024-08-08 23:54:00 -04:00 committed by GitHub
parent 3a14e00366
commit 345a686d82
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 28 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

39
src/llama.cpp
View File

@ -17343,6 +17343,7 @@ bool llama_save_session_file(struct llama_context * ctx, const char * path_sessi
// TODO: replace all non-fatal assertions with returned errors or exceptions
struct llama_data_write {
virtual void write(const void * src, size_t size) = 0;
virtual void write_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) = 0;
virtual size_t get_size_written() = 0;
virtual ~llama_data_write() = default;
@ -17465,9 +17466,8 @@ struct llama_data_write {
// Read each range of cells of k_size length each into tmp_buf and write out
for (const auto & range : cell_ranges) {
const size_t range_size = range.second - range.first;
tmp_buf.resize(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);
write(tmp_buf.data(), tmp_buf.size());
const size_t buf_size = range_size * k_size_row;
write_tensor_data(kv_self.k_l[il], range.first * k_size_row, buf_size);
}
}
@ -17486,9 +17486,8 @@ struct llama_data_write {
// Read each range of cells of v_size length each into tmp_buf and write out
for (const auto & range : cell_ranges) {
const size_t range_size = range.second - range.first;
tmp_buf.resize(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);
write(tmp_buf.data(), tmp_buf.size());
const size_t buf_size = range_size * v_size_row;
write_tensor_data(kv_self.v_l[il], range.first * v_size_row, buf_size);
}
}
} else {
@ -17514,9 +17513,8 @@ struct llama_data_write {
for (const auto & range : cell_ranges) {
const size_t range_size = range.second - range.first;
const size_t src_offset = (range.first + j * kv_size) * v_size_el;
tmp_buf.resize(range_size * v_size_el);
ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), src_offset, tmp_buf.size());
write(tmp_buf.data(), tmp_buf.size());
const size_t buf_size = range_size * v_size_el;
write_tensor_data(kv_self.v_l[il], src_offset, buf_size);
}
}
}
@ -17875,12 +17873,14 @@ struct llama_data_write_dummy : llama_data_write {
llama_data_write_dummy() {}
// TODO: avoid unnecessary calls to ggml_backend_tensor_get in a dummy context
void write(const void * /* src */, size_t size) override {
size_written += size;
}
void write_tensor_data(const struct ggml_tensor * /* tensor */, size_t /* offset */, size_t size) override {
size_written += size;
}
size_t get_size_written() override {
return size_written;
}
@ -17903,6 +17903,16 @@ struct llama_data_write_buffer : llama_data_write {
buf_size -= size;
}
void write_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) override {
if (size > buf_size) {
throw std::runtime_error("unexpectedly reached end of buffer");
}
ggml_backend_tensor_get(tensor, ptr, offset, size);
ptr += size;
size_written += size;
buf_size -= size;
}
size_t get_size_written() override {
return size_written;
}
@ -17938,6 +17948,7 @@ struct llama_data_read_buffer : llama_data_read {
struct llama_data_write_file : llama_data_write {
llama_file * file;
size_t size_written = 0;
std::vector<uint8_t> temp_buffer;
llama_data_write_file(llama_file * f) : file(f) {}
@ -17946,6 +17957,12 @@ struct llama_data_write_file : llama_data_write {
size_written += size;
}
void write_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) override {
temp_buffer.resize(size);
ggml_backend_tensor_get(tensor, temp_buffer.data(), offset, size);
write(temp_buffer.data(), temp_buffer.size());
}
size_t get_size_written() override {
return size_written;
}