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

Fix validation errors caused by empty buffers on larger batch sizes

Browse source
This commit is contained in:
0cc4m 2024-05-09 16:23:25 +02:00
parent d4abd657ad
commit d2088cd79d
1 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
ggml-vulkan.cpp
View file

@ -5418,7 +5418,7 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
}
static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, bool last_node){
if (ctx->disable || node->backend != GGML_BACKEND_TYPE_GPU) {
if (ctx->disable || node->backend != GGML_BACKEND_TYPE_GPU || ggml_is_empty(node)) {
return;
}
@ -6218,7 +6218,7 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen
int last_node = cgraph->n_nodes - 1;
// If the last op in the cgraph isn't backend GPU, the command buffer doesn't get closed properly
while (last_node > 0 && cgraph->nodes[last_node]->backend != GGML_BACKEND_TYPE_GPU) {
while (last_node > 0 && (cgraph->nodes[last_node]->backend != GGML_BACKEND_TYPE_GPU || ggml_is_empty(cgraph->nodes[last_node]))) {
last_node -= 1;
}