zhangyiss/gguf-tools
1
0
Fork 0
mirror of https://github.com/antirez/gguf-tools.git synced 2025-09-18 04:08:09 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Implement f16/f32 in gguf_tensor_to_float().

Browse Source
This commit is contained in:
antirez
2023-12-30 17:23:27 +01:00
parent 136e04977c
commit a4858afb4d
1 changed files with 10 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
gguflib.c
View File

@@ -511,7 +511,16 @@ float *gguf_tensor_to_float(gguf_tensor *tensor) {
gguf_get_tensor_type_features(tensor->type); gguf_get_tensor_type_features(tensor->type);
uint64_t block_size = tf->bytes_per_block; uint64_t block_size = tf->bytes_per_block;
float *f = malloc(tensor->num_weights*sizeof(float)); float *f = malloc(tensor->num_weights*sizeof(float));
if (tensor->type == GGUF_TYPE_Q8_0) { if (tensor->type == GGUF_TYPE_F32) {
memcpy(f, tensor->weights_data, tensor->num_weights*sizeof(float));
} else if (tensor->type == GGUF_TYPE_F16) {
uint64_t i = 0; // i-th weight to dequantize.
uint16_t *w16 = (uint16_t*) tensor->weights_data;
while(i < tensor->num_weights) {
f[i] = from_half(w16[i]);
i++;
}
} else if (tensor->type == GGUF_TYPE_Q8_0) {
/* Very simple layout: |16 bit delta|32 x 8bit weights| /* Very simple layout: |16 bit delta|32 x 8bit weights|
* Each weight is delta * quantized_weight[0..31] */ * Each weight is delta * quantized_weight[0..31] */
int8_t *block = (int8_t*)tensor->weights_data; int8_t *block = (int8_t*)tensor->weights_data;