vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : switch to floating-point token positions

Browse source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-02-23 12:18:30 +02:00
parent 15499eb942
commit fc775366f1
No known key found for this signature in database
GPG key ID: BF970631944C16B7
14 changed files with 68 additions and 61 deletions
Download patch file Download diff file Expand all files Collapse all files

4
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -291,7 +291,7 @@ static struct ggml_tensor * llama_build_train_graphs(
};
// KQ_pos - contains the positions
struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, N);
ggml_set_input(KQ_pos);
// rope has so much parameters that we make a custom function for it
@ -419,7 +419,7 @@ static struct ggml_tensor * llama_build_train_graphs(
ggml_gallocr_alloc_graph(alloc, gb);
if (!measure_only) {
int * data = (int *) KQ_pos->data;
float * data = (float *) KQ_pos->data;
for (int i = 0; i < N; ++i) {
data[i] = n_past + i;
}