Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-02-07 00:43:14 +01:00
Code Issues Packages Projects Releases Wiki Activity

format fix

Browse Source
This commit is contained in:
Howard Su 2023-05-21 22:31:19 +08:00
parent 006d5707e8
commit 80f1faac87
2 changed files with 20 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
ggml.c
View File

@ -856,10 +856,11 @@ void quantize_upgrade(enum ggml_type type, void* data, size_t * size, bool shuff
new_blk_buf.d = GGML_FP32_TO_FP16(blk[i].d); new_blk_buf.d = GGML_FP32_TO_FP16(blk[i].d);
if (shuffle) if (shuffle) {
quantize_shuffle_block(blk[i].qs, new_blk_buf.qs, qk/4); quantize_shuffle_block(blk[i].qs, new_blk_buf.qs, qk/4);
else } else {
memcpy(new_blk_buf.qs, blk[i].qs, qk / 2); memcpy(new_blk_buf.qs, blk[i].qs, qk / 2);
}
memcpy(&new_blk[i], &new_blk_buf, sizeof(block_q4_0)); memcpy(&new_blk[i], &new_blk_buf, sizeof(block_q4_0));
} }
@ -875,10 +876,11 @@ void quantize_upgrade(enum ggml_type type, void* data, size_t * size, bool shuff
new_blk_buf.d = GGML_FP32_TO_FP16(blk[i].d); new_blk_buf.d = GGML_FP32_TO_FP16(blk[i].d);
new_blk_buf.m = GGML_FP32_TO_FP16(blk[i].m); new_blk_buf.m = GGML_FP32_TO_FP16(blk[i].m);
if (shuffle) if (shuffle) {
quantize_shuffle_block(blk[i].qs, new_blk_buf.qs, qk/4); quantize_shuffle_block(blk[i].qs, new_blk_buf.qs, qk/4);
else } else {
memcpy(new_blk_buf.qs, blk[i].qs, qk / 2); memcpy(new_blk_buf.qs, blk[i].qs, qk / 2);
}
memcpy(&new_blk[i], &new_blk_buf, sizeof(block_q4_1)); memcpy(&new_blk[i], &new_blk_buf, sizeof(block_q4_1));
} }
} else if (type == GGML_TYPE_Q5_0) { } else if (type == GGML_TYPE_Q5_0) {
@ -889,10 +891,11 @@ void quantize_upgrade(enum ggml_type type, void* data, size_t * size, bool shuff
block_q5_0 new_blk; block_q5_0 new_blk;
for (size_t i = 0; i < nb ; i++) { for (size_t i = 0; i < nb ; i++) {
if (shuffle) if (shuffle) {
quantize_shuffle_block(blk[i].qs, new_blk.qs, qk/4); quantize_shuffle_block(blk[i].qs, new_blk.qs, qk/4);
else } else {
memcpy(new_blk.qs, blk[i].qs, qk / 2); memcpy(new_blk.qs, blk[i].qs, qk / 2);
}
memcpy(blk[i].qs, new_blk.qs, sizeof(new_blk.qs)); memcpy(blk[i].qs, new_blk.qs, sizeof(new_blk.qs));
} }
} else if (type == GGML_TYPE_Q5_1) { } else if (type == GGML_TYPE_Q5_1) {
@ -903,10 +906,11 @@ void quantize_upgrade(enum ggml_type type, void* data, size_t * size, bool shuff
block_q5_1 new_blk; block_q5_1 new_blk;
for (size_t i = 0; i < nb ; i++) { for (size_t i = 0; i < nb ; i++) {
if (shuffle) if (shuffle) {
quantize_shuffle_block(blk[i].qs, new_blk.qs, qk/4); quantize_shuffle_block(blk[i].qs, new_blk.qs, qk/4);
else } else {
memcpy(new_blk.qs, blk[i].qs, qk / 2); memcpy(new_blk.qs, blk[i].qs, qk / 2);
}
memcpy(&blk[i], &new_blk, sizeof(new_blk)); memcpy(&blk[i], &new_blk, sizeof(new_blk));
} }
} else if (type == GGML_TYPE_Q8_0) { } else if (type == GGML_TYPE_Q8_0) {

12
llama.cpp
View File

@ -327,6 +327,8 @@ static size_t llama_calc_tensor_size_prev3(const std::vector<uint32_t> & ne, enu
case GGML_TYPE_Q8_0: case GGML_TYPE_Q8_0:
size += 2; size += 2;
break; break;
default:
break;
} }
for (uint32_t dim : ne) { for (uint32_t dim : ne) {
@ -343,11 +345,12 @@ struct llama_load_tensor_shard {
size_t file_off; size_t file_off;
void calc_size(llama_file_version file_version) { void calc_size(llama_file_version file_version) {
if (file_version == LLAMA_FILE_VERSION_GGJT_V3) if (file_version == LLAMA_FILE_VERSION_GGJT_V3) {
size = llama_calc_tensor_size(ne, type); size = llama_calc_tensor_size(ne, type);
else } else {
size = llama_calc_tensor_size_prev3(ne, type); size = llama_calc_tensor_size_prev3(ne, type);
} }
}
}; };
enum llama_split_type { enum llama_split_type {
@ -426,11 +429,12 @@ struct llama_load_tensor {
} }
void calc_size(llama_file_version file_version) { void calc_size(llama_file_version file_version) {
if (file_version == LLAMA_FILE_VERSION_GGJT_V3) if (file_version == LLAMA_FILE_VERSION_GGJT_V3) {
size = llama_calc_tensor_size(ne, type); size = llama_calc_tensor_size(ne, type);
else } else {
size = llama_calc_tensor_size_prev3(ne, type); size = llama_calc_tensor_size_prev3(ne, type);
} }
}
}; };
struct llama_load_tensors_map { struct llama_load_tensors_map {