happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

vulkan: allow graph_optimize for prompt processing workloads (#17475)

This commit is contained in:
Jeff Bolz 2025-11-26 09:46:33 -06:00 committed by GitHub
parent 879d673759
commit eec1e33a9e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 0 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -13158,24 +13158,6 @@ static void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph *
return false;
};
// This function tries to reorder the graph to allow nodes to run in parallel.
// This helps with small batches, but for large batches its a slowdown, probably
// due to cache contention. So only reorder if the majority of nodes have few rows.
int num_small_nodes = 0;
int num_counted_nodes = 0;
for (int i = 0; i < graph->n_nodes; ++i) {
if (!is_empty(graph->nodes[i]) &&
graph->nodes[i]->op != GGML_OP_SET_ROWS) {
if (ggml_nrows(graph->nodes[i]) <= 8) {
num_small_nodes++;
}
num_counted_nodes++;
}
}
if (num_small_nodes < num_counted_nodes / 2) {
return;
}
std::vector<ggml_tensor *> new_order;
std::vector<bool> used(graph->n_nodes, false);
std::set<ggml_tensor *> used_node_set;