initial GLM-4.5 integration

Author: ddh0

src/llama-arch.h

@@ -66,6 +66,7 @@ enum llm_arch {
     LLM_ARCH_DEEPSEEK2,
     LLM_ARCH_CHATGLM,
     LLM_ARCH_GLM4,
+    LLM_ATCH_GLM4_MOE,
     LLM_ARCH_BITNET,
     LLM_ARCH_T5,
     LLM_ARCH_T5ENCODER,

src/llama-graph.cpp

@@ -749,8 +749,10 @@ ggml_tensor * llm_graph_context::build_ffn(
 
     if (down) {
         cur = build_lora_mm(down, cur);
-        if (arch == LLM_ARCH_GLM4) {
-            // GLM4 seems to have numerical issues with half-precision accumulators
+        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
+            // GLM4 FFNs seem to have numerical issues with half-precision accumulators
+            // -- ref: https://github.com/ggml-org/llama.cpp/pull/13101
+            // (GLM4_MOE uses some GLM4 FFNs, so we need to match it too)
             ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
         }
     }
@@ -1391,8 +1393,10 @@ ggml_tensor * llm_graph_context::build_attn(
 
     if (wo) {
         cur = build_lora_mm(wo, cur);
-        if (arch == LLM_ARCH_GLM4) {
-            // GLM4 seems to have numerical issues with half-precision accumulators
+        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
+            // GLM4 FFNs seem to have numerical issues with half-precision accumulators
+            // -- ref: https://github.com/ggml-org/llama.cpp/pull/13101
+            // (GLM4_MOE uses some GLM4 FFNs, so we need to match it too)
             ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
         }
     }

src/llama-model.cpp

@@ -111,6 +111,8 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_30B_A3B:       return "30B.A3B";
         case LLM_TYPE_235B_A22B:     return "235B.A22B";
         case LLM_TYPE_300B_A47B:     return "300B.A47B";
+        case LLM_TYPE_355B_A32B:     return "355B.A32B (GLM-4.5)";
+        case LLM_TYPE_106B_A12B:     return "106B.A12B (GLM-4.5)";
         case LLM_TYPE_E2B:           return "E2B";
         case LLM_TYPE_E4B:           return "E4B";
         default:                     return "?B";
@@ -18153,6 +18155,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_PLM:
         case LLM_ARCH_CHATGLM:
         case LLM_ARCH_GLM4:
+        case LLM_ARCH_GLM4_MOE:
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
         case LLM_ARCH_GRANITE_HYBRID:

src/llama-model.h

@@ -103,6 +103,8 @@ enum llm_type {
     LLM_TYPE_30B_A3B,
     LLM_TYPE_235B_A22B,
     LLM_TYPE_300B_A47B, // Ernie MoE big
+    LLM_TYPE_355B_A32B, // GLM-4.5
+    LLM_TYPE_106B_A12B, // GLM-4.5-Air
     LLM_TYPE_E2B,
     LLM_TYPE_E4B,
 };