mirror of
https://github.com/ikawrakow/ik_llama.cpp.git
synced 2026-03-05 11:30:09 +00:00
Just in case
This commit is contained in:
@@ -422,7 +422,7 @@ struct llama_model {
|
||||
|
||||
size_t max_nodes(int n_tokens) const {
|
||||
auto n_tensors = tensors_by_name.size();
|
||||
if (split_mode == LLAMA_SPLIT_MODE_GRAPH) n_tensors *= devices.size();
|
||||
if (split_mode == LLAMA_SPLIT_MODE_GRAPH && !devices.empty()) n_tensors *= devices.size();
|
||||
if (arch == LLM_ARCH_QWEN3NEXT || arch == LLM_ARCH_QWEN35MOE) {
|
||||
return std::max<size_t>(n_tokens * 40, 32u * n_tensors);
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user