use no_alloc instead of fixing f32 fallback

common/common.cpp

@@ -1434,6 +1434,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
 
     mparams.progress_callback           = params.load_progress_callback;
     mparams.progress_callback_user_data = params.load_progress_callback_user_data;
+    mparams.no_alloc                    = params.no_alloc;
 
     return mparams;
 }

common/common.h

@@ -679,6 +679,7 @@ struct common_params {
     // return false from callback to abort model loading or true to continue
     llama_progress_callback load_progress_callback = NULL;
     void *                  load_progress_callback_user_data = NULL;
+    bool no_alloc = false; // Don't allocate model buffers
 };
 
 // call once at the start of a program if it uses libcommon

src/llama-model-loader.cpp

@@ -1215,10 +1215,6 @@ struct ggml_tensor * llama_model_loader::create_tensor(
         const int64_t tid = gguf_find_tensor(metadata, tn.str().c_str());
         if (tid != -1) {
             type = gguf_get_tensor_type(metadata, tid);
-        } else if (flags & TENSOR_NOT_REQUIRED) {
-            // If the tensor is not found and not required, return nullptr to allow
-            // the caller to fall back
-            return nullptr;
         }
 
         // for tensors that are not required some of the dimensions can be invalid:

tests/export-graph-ops.cpp

@@ -167,6 +167,7 @@ int main(int argc, char ** argv) {
 
         llama_model_params model_params = llama_model_default_params();
         model_params.devices = params.devices.data();
+        model_params.no_alloc = true;
 
         model.reset(llama_model_init_from_user(gguf_ctx.get(), set_tensor_data, nullptr, model_params));