Skip to content

Commit 9574f41

Browse files
ggerganovggerganov
committed
llama : no need to pass full file loader to the file saver
just gguf_ctx
1 parent 5c85332 commit 9574f41

File tree

1 file changed

+24
-24
lines changed

1 file changed

+24
-24
lines changed

‎gguf-llama.cpp

Lines changed: 24 additions & 24 deletions
Original file line numberDiff line numberDiff line change
@@ -697,12 +697,12 @@ struct gguf_file_saver {
697697
// we need to calculate the delta in number of bytes written with a counter as a struct member.
698698

699699
gguf_file file;
700-
gguf_file_loader * fl;
700+
gguf_context * ctx; // loaded gguf context (used to re-write the KV section (good enough for now))
701701
size_t info_offset;
702702
size_t tensor_offset = 0;
703703

704-
gguf_file_saver(const char * fname, gguf_file_loader * fl)
705-
: file(fname, "wb"), fl(fl) {
704+
gguf_file_saver(const char * fname, gguf_context * ctx)
705+
: file(fname, "wb"), ctx(ctx) {
706706
fprintf(stderr, "llama.cpp: saving model to %s\n", fname);
707707
write_header();
708708
write_kv();
@@ -711,15 +711,15 @@ struct gguf_file_saver {
711711
void write_header() {
712712
file.write_i32(GGUF_MAGIC);
713713
file.write_i32(GGUF_VERSION);
714-
file.write_i32(gguf_get_n_tensors(fl->gguf_ctx));
715-
file.write_i32(gguf_get_n_kv (fl->gguf_ctx));
714+
file.write_i32(gguf_get_n_tensors(ctx));
715+
file.write_i32(gguf_get_n_kv (ctx));
716716
}
717717

718718
void write_kv_arr_str(const std::string & key, enum gguf_type type, int i, int n_arr) {
719719
std::vector<std::string> data(n_arr);
720720

721721
for (int j = 0; j < n_arr; ++j) {
722-
std::string val = gguf_get_arr_str(fl->gguf_ctx, i, j);
722+
std::string val = gguf_get_arr_str(ctx, i, j);
723723
data[j] = val;
724724
}
725725

@@ -730,7 +730,7 @@ struct gguf_file_saver {
730730
std::vector<float> data(n_arr);
731731

732732
for (int j = 0; j < n_arr; ++j) {
733-
float val = gguf_get_arr_f32(fl->gguf_ctx, i, j);
733+
float val = gguf_get_arr_f32(ctx, i, j);
734734
data[j] = val;
735735
}
736736

@@ -739,28 +739,28 @@ struct gguf_file_saver {
739739

740740
// re-write the key-value section from the loaded file
741741
void write_kv() {
742-
const int32_t n_kv = gguf_get_n_kv(fl->gguf_ctx);
742+
const int32_t n_kv = gguf_get_n_kv(ctx);
743743
for (int i = 0; i < n_kv; ++i) {
744-
const char * key = gguf_get_key(fl->gguf_ctx, i);
744+
const char * key = gguf_get_key(ctx, i);
745745
if (strcmp(key, "general.quantization_version") == 0) {
746746
file.write_val<uint32_t>("general.quantization_version", GGUF_TYPE_UINT32, GGML_QNT_VERSION);
747747
} else {
748-
const gguf_type vtype = gguf_get_kv_type(fl->gguf_ctx, i);
748+
const gguf_type vtype = gguf_get_kv_type(ctx, i);
749749

750750
switch (vtype) {
751-
case GGUF_TYPE_BOOL: file.write_val<bool> (key, GGUF_TYPE_BOOL, gguf_get_val_bool(fl->gguf_ctx, i)); break;
752-
case GGUF_TYPE_FLOAT32: file.write_val<float> (key, GGUF_TYPE_FLOAT32, gguf_get_val_f32 (fl->gguf_ctx, i)); break;
753-
case GGUF_TYPE_INT16: file.write_val<int16_t> (key, GGUF_TYPE_INT16, gguf_get_val_i16 (fl->gguf_ctx, i)); break;
754-
case GGUF_TYPE_INT32: file.write_val<int32_t> (key, GGUF_TYPE_INT32, gguf_get_val_i32 (fl->gguf_ctx, i)); break;
755-
case GGUF_TYPE_INT8: file.write_val<int8_t> (key, GGUF_TYPE_INT8, gguf_get_val_i8 (fl->gguf_ctx, i)); break;
756-
case GGUF_TYPE_STRING: file.write_str (key, GGUF_TYPE_STRING, gguf_get_val_str (fl->gguf_ctx, i)); break;
757-
case GGUF_TYPE_UINT16: file.write_val<uint16_t>(key, GGUF_TYPE_UINT16, gguf_get_val_u16 (fl->gguf_ctx, i)); break;
758-
case GGUF_TYPE_UINT32: file.write_val<uint32_t>(key, GGUF_TYPE_UINT32, gguf_get_val_u32 (fl->gguf_ctx, i)); break;
759-
case GGUF_TYPE_UINT8: file.write_val<uint8_t> (key, GGUF_TYPE_UINT8, gguf_get_val_u8 (fl->gguf_ctx, i)); break;
751+
case GGUF_TYPE_BOOL: file.write_val<bool> (key, GGUF_TYPE_BOOL, gguf_get_val_bool(ctx, i)); break;
752+
case GGUF_TYPE_FLOAT32: file.write_val<float> (key, GGUF_TYPE_FLOAT32, gguf_get_val_f32 (ctx, i)); break;
753+
case GGUF_TYPE_INT16: file.write_val<int16_t> (key, GGUF_TYPE_INT16, gguf_get_val_i16 (ctx, i)); break;
754+
case GGUF_TYPE_INT32: file.write_val<int32_t> (key, GGUF_TYPE_INT32, gguf_get_val_i32 (ctx, i)); break;
755+
case GGUF_TYPE_INT8: file.write_val<int8_t> (key, GGUF_TYPE_INT8, gguf_get_val_i8 (ctx, i)); break;
756+
case GGUF_TYPE_STRING: file.write_str (key, GGUF_TYPE_STRING, gguf_get_val_str (ctx, i)); break;
757+
case GGUF_TYPE_UINT16: file.write_val<uint16_t>(key, GGUF_TYPE_UINT16, gguf_get_val_u16 (ctx, i)); break;
758+
case GGUF_TYPE_UINT32: file.write_val<uint32_t>(key, GGUF_TYPE_UINT32, gguf_get_val_u32 (ctx, i)); break;
759+
case GGUF_TYPE_UINT8: file.write_val<uint8_t> (key, GGUF_TYPE_UINT8, gguf_get_val_u8 (ctx, i)); break;
760760
case GGUF_TYPE_ARRAY:
761761
{
762-
const gguf_type arr_type = gguf_get_arr_type(fl->gguf_ctx, i);
763-
const int n_arr = gguf_get_arr_n (fl->gguf_ctx, i);
762+
const gguf_type arr_type = gguf_get_arr_type(ctx, i);
763+
const int n_arr = gguf_get_arr_n (ctx, i);
764764
if (arr_type == GGUF_TYPE_FLOAT32) {
765765
write_kv_arr_f32(key, arr_type, i, n_arr);
766766
} else if (arr_type == GGUF_TYPE_STRING) {
@@ -777,9 +777,9 @@ struct gguf_file_saver {
777777

778778
info_offset = file.tell();
779779

780-
GGML_ASSERT(gguf_get_data_offset(fl->gguf_ctx) >= info_offset);
780+
GGML_ASSERT(gguf_get_data_offset(ctx) >= info_offset);
781781

782-
size_t count = gguf_get_data_offset(fl->gguf_ctx) - info_offset;
782+
size_t count = gguf_get_data_offset(ctx) - info_offset;
783783
file.write_zeros(count);
784784
file.seek(info_offset, SEEK_SET);
785785
GGML_ASSERT(info_offset == file.tell());
@@ -3220,7 +3220,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
32203220
}
32213221

32223222
std::unique_ptr<llama_model_loader> model_loader(new llama_model_loader(fname_inp, /*use_mmap*/ false));
3223-
gguf_file_saver file_saver(fname_out.c_str(), model_loader->file_loader.get());
3223+
gguf_file_saver file_saver(fname_out.c_str(), model_loader->file_loader->gguf_ctx);
32243224

32253225
#ifdef GGML_USE_K_QUANTS
32263226
int n_attention_wv = 0;

0 commit comments

Comments
 (0)