model: centralize hidden_act -> llm_ffn_op_type mapping

Per CISC review: the GGUF '<arch>.hidden_activation' string -> enum
mapping should be centralized in llama-model.cpp, mirroring how
rope_scaling_type is handled, rather than open-coded inside each
arch-specific loader.

Adds llm_ffn_op_type_from_string() in llama-model.{h,cpp} and rewires
modern-bert's hparams loader to use it. The default fallback stays
LLM_FFN_GEGLU, matching the prior behavior.

Author: hansolosan

src/llama-model.cpp

@@ -813,6 +813,28 @@ static llama_rope_scaling_type llama_rope_scaling_type_from_string(const std::st
     return LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
 }
 
+// Maps the GGUF `<arch>.hidden_activation` string to the FFN op type used by the
+// graph builders. Only gated activations that map cleanly to llm_ffn_op_type are
+// listed; unrecognized values fall back to GeGLU, which matches the historical
+// default for ModernBert-style architectures.
+static const std::map<std::string, llm_ffn_op_type> LLM_FFN_OP_TYPES_FROM_STRING = {
+    { "gelu",   LLM_FFN_GEGLU  },
+    { "geglu",  LLM_FFN_GEGLU  },
+    { "silu",   LLM_FFN_SWIGLU },
+    { "swish",  LLM_FFN_SWIGLU },
+    { "swiglu", LLM_FFN_SWIGLU },
+    { "relu",   LLM_FFN_RELU   },
+    { "reglu",  LLM_FFN_REGLU  },
+};
+
+llm_ffn_op_type llm_ffn_op_type_from_string(const std::string & name, llm_ffn_op_type fallback) {
+    const auto it = LLM_FFN_OP_TYPES_FROM_STRING.find(name);
+    if (it != LLM_FFN_OP_TYPES_FROM_STRING.end()) {
+        return it->second;
+    }
+    return fallback;
+}
+
 // CPU: ACCEL -> GPU host -> CPU extra -> CPU
 static buft_list_t make_cpu_buft_list(const std::vector<llama_device> & devices, bool use_extra_bufts, bool no_host) {
     buft_list_t buft_list;

src/llama-model.h

@@ -144,6 +144,10 @@ enum llm_type {
 
 std::string llama_rope_scaling_type_name(llama_rope_scaling_type rope_scaling_type);
 
+// Map a GGUF activation-name string to llm_ffn_op_type. Returns `fallback` if
+// the string is empty or not recognized.
+llm_ffn_op_type llm_ffn_op_type_from_string(const std::string & name, llm_ffn_op_type fallback);
+
 struct llama_layer_posnet {
     // resnet
     struct ggml_tensor * norm1   = nullptr;

src/models/modern-bert.cpp

@@ -16,10 +16,9 @@ void llama_model_modern_bert::load_arch_hparams(llama_model_loader & ml) {
 
     // Some ModernBert derivatives (e.g. IBM Granite Embedding 97m R2) use
     // SiLU/SwiGLU in the FFN instead of the default GELU/GeGLU.
-    std::string hidden_act = "gelu";
-    ml.get_key(LLM_KV_HIDDEN_ACT, hidden_act, false);
-    if (hidden_act == "silu" || hidden_act == "swish") {
-        hparams.llm_ffn_op = LLM_FFN_SWIGLU;
+    std::string hidden_act;
+    if (ml.get_key(LLM_KV_HIDDEN_ACT, hidden_act, false)) {
+        hparams.llm_ffn_op = llm_ffn_op_type_from_string(hidden_act, LLM_FFN_GEGLU);
     }
 
     switch (hparams.n_layer) {