From fda60ead35f76ff05008e95a4c772772dfba50db Mon Sep 17 00:00:00 2001
From: Branden Butler <bwtbutler@hotmail.com>
Date: Thu, 28 Sep 2023 12:39:34 -0500
Subject: [PATCH] Replace vector with C-style array and length in
 llama_split_layers_weighted

---
 examples/mpi/mpi.cpp | 2 +-
 llama.cpp            | 6 +++---
 llama.h              | 2 +-
 3 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/examples/mpi/mpi.cpp b/examples/mpi/mpi.cpp
index 0bf8f2f80..5dfa70f5c 100644
--- a/examples/mpi/mpi.cpp
+++ b/examples/mpi/mpi.cpp
@@ -243,7 +243,7 @@ int main(int argc, char ** argv) {
         return 0;
     }
 
-    llama_split_layers_weighted(ctx, params.mpi_layer_split);
+    llama_split_layers_weighted(ctx, params.mpi_layer_split.data(), params.mpi_layer_split.size());
 
     std::string path_session = params.path_prompt_cache;
     std::vector<llama_token> session_tokens;
diff --git a/llama.cpp b/llama.cpp
index 1f674f13f..98ffa1075 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -13087,12 +13087,12 @@ struct llama_context * llama_new_context_with_model(
     return ctx;
 }
 
-void llama_split_layers_weighted(struct llama_context * ctx, std::vector<float> device_weights) {
+void llama_split_layers_weighted(struct llama_context * ctx, float device_weights[], size_t num_weights) {
 #ifdef GGML_USE_MPI
-    if (ggml_mpi_rank(ctx->ctx_mpi) == 0 && ggml_mpi_size(ctx->ctx_mpi) != device_weights.size()) {
+    if (ggml_mpi_rank(ctx->ctx_mpi) == 0 && ggml_mpi_size(ctx->ctx_mpi) != num_weights) {
         GGML_ASSERT(false && "Must have same number of split percentages as devices");
     }
-    uint16_t** ranges = ggml_mpi_split_range(ctx->ctx_mpi, 0, ctx->model.hparams.n_layer - 1, device_weights.data());
+    uint16_t** ranges = ggml_mpi_split_range(ctx->ctx_mpi, 0, ctx->model.hparams.n_layer - 1, device_weights);
     ggml_mpi_scatter_layers(ctx->ctx_mpi, ranges);
 #endif
 }
diff --git a/llama.h b/llama.h
index 0a13b037d..7ad4c9257 100644
--- a/llama.h
+++ b/llama.h
@@ -358,7 +358,7 @@ extern "C" {
                              const char * path_model,
             struct llama_model_params     params);
 
-    LLAMA_API void llama_split_layers_weighted(struct llama_context * ctx, std::vector<float> device_weights);
+    LLAMA_API void llama_split_layers_weighted(struct llama_context * ctx, float device_weights[], size_t num_weights);
 
     LLAMA_API void llama_free_model(struct llama_model * model);