graph : avoid huge warm-up graphs for MoE models (#14753)

* graph : avoid huge warm-up graphs for MoE models

ggml-ci

* cont : bump max nodes to 8x model tensors

diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 1af19caa39dab17a98aa9e5145b2d521e5bea7bb..6eb344736de6f398e768f2788d198df6eff24c98 100644 (file)
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -1312,7 +1312,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
 //
 
 uint32_t llama_context::graph_max_nodes() const {
-    return std::max<uint32_t>(65536u, 5u*model.n_tensors());
+    return std::max<uint32_t>(1024u, 8u*model.n_tensors());
 }
 
 llm_graph_result * llama_context::get_gf_res_reserve() const {

diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 7ea7fd6156e672e5ed8eeb039021713074277432..7cac3b98fa9cafffff7bb6909238c52fbe12f6da 100644 (file)
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -906,8 +906,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     }
 
     // aggregate experts
+    // note: here we explicitly use hparams.n_expert_used instead of n_expert_used
+    //       to avoid potentially a large number of add nodes during warmup
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/14753
     ggml_tensor * moe_out = nullptr;
-    for (int i = 0; i < n_expert_used; ++i) {
+    for (uint32_t i = 0; i < hparams.n_expert_used; ++i) {
         ggml_tensor * cur_expert = ggml_view_2d(ctx0, experts, n_embd, n_tokens,
                 experts->nb[2], i*experts->nb[1]);
 
@@ -918,7 +921,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         }
     }
 
-    if (n_expert_used == 1) {
+    if (hparams.n_expert_used == 1) {
         // avoid returning a non-contiguous tensor
         moe_out = ggml_cont(ctx0, moe_out);
     }