Skip to content

LLM inference in C/C++

C++ ggml

llama : allow virtual igpu devices (#26953)

Georgi Gerganov4c1a0af40d88
diff --git a/src/llama.cpp b/src/llama.cpp
index 9ff1902fc1..1609fec88d 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -257,7 +257,11 @@ static bool llama_prepare_model_devices(const llama_model_params & params, llama
}
case GGML_BACKEND_DEVICE_TYPE_IGPU:
- if (igpus.empty()) {
+ // igpus.empty() - workaround for integrated devices seen by multiple backends
+ // ref: https://github.com/ggml-org/llama.cpp/pull/23897
+ // ggml_backend_dev_backend_reg - allow devices of the same backend regardless if integrated
+ // ref: https://github.com/ggml-org/llama.cpp/pull/23897#issuecomment-5264222997
+ if (igpus.empty() || ggml_backend_dev_backend_reg(dev) == ggml_backend_dev_backend_reg(igpus.back().dev)) {
igpus.push_back({false, dev});
}
break;