From: Diego Devesa <redacted>
Date: Thu, 31 Jul 2025 18:15:41 +0000 (-0700)
Subject: llama : add simple option to enable CPU for MoE weights (--cpu-moe) (#14992)
X-Git-Tag: upstream/0.0.6073~22
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=a06ed5feaec9f935fbf662035b2673167bc88460;p=pkg%2Fggml%2Fsources%2Fllama.cpp

llama : add simple option to enable CPU for MoE weights (--cpu-moe) (#14992)
---

diff --git a/common/arg.cpp b/common/arg.cpp
index 0a4a15e7..cd853119 100644
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -2380,6 +2380,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ));
+    add_opt(common_arg(
+        {"--cpu-moe"},
+        "use CPU for Mixture of Experts (MoE) weights",
+        [](common_params & params) {
+            params.tensor_buft_overrides.push_back({"\\.ffn_up_exps\\.weight$",   ggml_backend_cpu_buffer_type()});
+            params.tensor_buft_overrides.push_back({"\\.ffn_down_exps\\.weight$", ggml_backend_cpu_buffer_type()});
+            params.tensor_buft_overrides.push_back({"\\.ffn_gate_exps\\.weight$", ggml_backend_cpu_buffer_type()});
+        }
+    ).set_env("LLAMA_ARG_CPU_MOE"));
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
         "number of layers to store in VRAM",