llama : skip main_gpu validation when no devices are available (ggml-org#23405)

Dev-iL · web-flow · commit b4024af6c244 · 2026-06-17T17:30:26.000+03:00
diff --git a/src/llama.cpp b/src/llama.cpp
@@ -249,7 +249,7 @@ static bool llama_prepare_model_devices(const llama_model_params & params, llama
     }
 
     // if using single GPU mode, remove all except the main GPU
-    if (params.split_mode == LLAMA_SPLIT_MODE_NONE) {
+    if (params.split_mode == LLAMA_SPLIT_MODE_NONE && !model->devices.empty()) {
         if (params.main_gpu < 0) {
             model->devices.clear();
         } else {

Original file line number	Diff line number	Diff line change
`@@ -249,7 +249,7 @@ static bool llama_prepare_model_devices(const llama_model_params & params, llama`
`249`	`249`	`}`
`250`	`250`
`251`	`251`	`// if using single GPU mode, remove all except the main GPU`
`252`		`- if (params.split_mode == LLAMA_SPLIT_MODE_NONE) {`
	`252`	`+ if (params.split_mode == LLAMA_SPLIT_MODE_NONE && !model->devices.empty()) {`
`253`	`253`	`if (params.main_gpu < 0) {`
`254`	`254`	`model->devices.clear();`
`255`	`255`	`} else {`