Skip to content

LLM inference in C/C++

C++ ggml

graph : add precision handling for MELLUM architecture (#30584)

Mikhail Podvitskii5e03a1ff450c
diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 80be4bb573..454c214042 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -2452,7 +2452,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
}
experts = build_lora_mm_id(down_exps, cur, selected_experts, down_exps_s, slots); // [n_embd, n_expert_used, n_tokens]
- if (arch == LLM_ARCH_MISTRAL4) {
+ if (arch == LLM_ARCH_MISTRAL4 || arch == LLM_ARCH_MELLUM) {
// src1 can exceed F16 range
ggml_prec_set_src(experts, GGML_PREC_F32, 1);
}