CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-01 09:01:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama : fix input allocation logic

Browse Source
This commit is contained in:
Georgi Gerganov
2023-10-31 08:23:43 +02:00
parent a3f80013ad
commit 2926ef63b1
1 changed files with 6 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
llama.cpp
View File

@@ -4970,10 +4970,10 @@ static struct ggml_cgraph * llama_build_graph(
// allocate input tensors and set input data // allocate input tensors and set input data
// //
if (batch.token && !alloc_inp_tokens && strcmp(name, "inp_tokens") == 0) { if (!alloc_inp_tokens && strcmp(name, "inp_tokens") == 0) {
ggml_allocr_alloc(lctx.alloc, cur); ggml_allocr_alloc(lctx.alloc, cur);
if (!ggml_allocr_is_measure(lctx.alloc)) { if (!ggml_allocr_is_measure(lctx.alloc) && batch.token) {
const int64_t n_tokens = cur->ne[0]; const int64_t n_tokens = cur->ne[0];
memcpy(cur->data, batch.token, n_tokens*ggml_element_size(cur)); memcpy(cur->data, batch.token, n_tokens*ggml_element_size(cur));
@@ -4982,10 +4982,10 @@ static struct ggml_cgraph * llama_build_graph(
alloc_inp_tokens = true; alloc_inp_tokens = true;
} }
if (batch.embd && !alloc_inp_embd && strcmp(name, "inp_embd") == 0) { if (!alloc_inp_embd && strcmp(name, "inp_embd") == 0) {
ggml_allocr_alloc(lctx.alloc, cur); ggml_allocr_alloc(lctx.alloc, cur);
if (!ggml_allocr_is_measure(lctx.alloc)) { if (!ggml_allocr_is_measure(lctx.alloc) && batch.embd) {
const int64_t n_embd = cur->ne[0]; const int64_t n_embd = cur->ne[0];
const int64_t n_tokens = cur->ne[1]; const int64_t n_tokens = cur->ne[1];
@@ -4995,10 +4995,10 @@ static struct ggml_cgraph * llama_build_graph(
alloc_inp_embd = true; alloc_inp_embd = true;
} }
if (batch.pos && !alloc_inp_pos && strcmp(name, "inp_pos") == 0) { if (!alloc_inp_pos && strcmp(name, "inp_pos") == 0) {
ggml_allocr_alloc(lctx.alloc, cur); ggml_allocr_alloc(lctx.alloc, cur);
if (!ggml_allocr_is_measure(lctx.alloc)) { if (!ggml_allocr_is_measure(lctx.alloc) && batch.pos) {
const int64_t n_tokens = cur->ne[0]; const int64_t n_tokens = cur->ne[0];
int32_t * data = (int32_t *) cur->data; int32_t * data = (int32_t *) cur->data;