llama: fix magic number of 999 for GPU layers (#18266)

* llama: fix magic number of 999 for GPU layers

* use strings for -ngl, -ngld

* enacapsulate n_gpu_layers, split_mode

src/llama-context.cpp

@@ -294,8 +294,8 @@ llama_context::llama_context(
         // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
         bool pipeline_parallel =
             model.n_devices() > 1 &&
-            model.params.n_gpu_layers > (int) model.hparams.n_layer &&
-            model.params.split_mode == LLAMA_SPLIT_MODE_LAYER &&
+            model.n_gpu_layers() > model.hparams.n_layer &&
+            model.split_mode() == LLAMA_SPLIT_MODE_LAYER &&
             cparams.offload_kqv &&
             !model.has_tensor_overrides();
 
@@ -1570,7 +1570,7 @@ llm_graph_cb llama_context::graph_get_cb() const {
 
         // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
         // FIXME: fix in ggml_backend_sched
-        const bool full_offload = model.params.n_gpu_layers > (int) model.hparams.n_layer;
+        const bool full_offload = model.n_gpu_layers() > model.hparams.n_layer;
         if (ubatch.n_tokens < 32 || full_offload) {
             if (il != -1 && strcmp(name, "norm") == 0) {
                 const auto & dev_layer = model.dev_layer(il);

src/llama-model.cpp

@@ -2378,11 +2378,11 @@ void llama_model::load_vocab(llama_model_loader & ml) {
 
 bool llama_model::load_tensors(llama_model_loader & ml) {
     const auto & split_mode   = params.split_mode;
-    const auto & n_gpu_layers = params.n_gpu_layers;
     const auto & use_mlock    = params.use_mlock;
     const auto & tensor_split = params.tensor_split;
 
-    const int n_layer = hparams.n_layer;
+    const int n_layer      = hparams.n_layer;
+    const int n_gpu_layers = this->n_gpu_layers();
 
     const bool use_mmap_buffer = true;
 
@@ -6884,6 +6884,14 @@ size_t llama_model::n_devices() const {
     return devices.size();
 }
 
+uint32_t llama_model::n_gpu_layers() const {
+    return params.n_gpu_layers >= 0 ? params.n_gpu_layers : hparams.n_layer + 1;
+}
+
+llama_split_mode llama_model::split_mode() const {
+    return params.split_mode;
+}
+
 std::map<ggml_backend_buffer_type_t, size_t> llama_model::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
     for (const auto & [ctx, bufs] : pimpl->ctxs_bufs) {
@@ -7794,7 +7802,7 @@ llama_model_params llama_model_default_params() {
     llama_model_params result = {
         /*.devices                     =*/ nullptr,
         /*.tensor_buft_overrides       =*/ nullptr,
-        /*.n_gpu_layers                =*/ 999,
+        /*.n_gpu_layers                =*/ -1,
         /*.split_mode                  =*/ LLAMA_SPLIT_MODE_LAYER,
         /*.main_gpu                    =*/ 0,
         /*.tensor_split                =*/ nullptr,

src/llama-model.h

@@ -466,8 +466,6 @@ struct llama_model {
     struct ggml_tensor * dense_2_out_layers = nullptr;
     struct ggml_tensor * dense_3_out_layers = nullptr;
 
-    llama_model_params params;
-
     // gguf metadata
     std::unordered_map<std::string, std::string> gguf_kv;
 
@@ -498,6 +496,9 @@ struct llama_model {
     size_t n_tensors() const;
     size_t n_devices() const;
 
+    uint32_t n_gpu_layers() const;
+    llama_split_mode split_mode() const;
+
     std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const;
 
     // total number of parameters in the model
@@ -526,6 +527,8 @@ struct llama_model {
     ggml_cgraph * build_graph(const llm_graph_params & params) const;
 
 private:
+    llama_model_params params;
+
     struct impl;
     std::unique_ptr<impl> pimpl;
 };