kimi-k3 : fixes for loading and running the full-size model

Graph node budget, LLM_TYPE for the 93-layer config, four hparams promoted to
required, and an E8M0 NaN check during the MXFP4 repack. Details in the PR
description.

Author: shimmyshimmer

src/llama-context.cpp

@@ -2334,6 +2334,10 @@ void llama_context::output_reorder() {
 //
 
 uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
+    if (model.arch == LLM_ARCH_KIMI_K3) {
+        // the n_tokens*40 budget below is exhausted at ubatch 3840
+        return std::max<uint32_t>(n_tokens * 160, 64u * model.n_tensors());
+    }
     if (model.arch == LLM_ARCH_QWEN3NEXT ||
         model.arch == LLM_ARCH_KIMI_LINEAR ||
         model.arch == LLM_ARCH_QWEN35 ||

src/llama-model.cpp

@@ -832,6 +832,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_397B_A17B:     return "397B.A17B";
         case LLM_TYPE_685B_A37B:     return "685B.A37B";
         case LLM_TYPE_744B_A40B:     return "744B.A40B";
+        case LLM_TYPE_2_8T_A50B:     return "2.8T.A50B";
         case LLM_TYPE_E2B:           return "E2B";
         case LLM_TYPE_E4B:           return "E4B";
         default:                     return "?B";

src/llama-model.h

@@ -142,6 +142,7 @@ enum llm_type {
     LLM_TYPE_397B_A17B, // Qwen3.5
     LLM_TYPE_685B_A37B, // DeepSeek V3.2
     LLM_TYPE_744B_A40B, // GLM-5
+    LLM_TYPE_2_8T_A50B, // Kimi-K3
     LLM_TYPE_E2B,
     LLM_TYPE_E4B,
 };

src/models/kimi-k3.cpp

@@ -47,7 +47,13 @@ void llama_model_kimi_k3::load_arch_hparams(llama_model_loader & ml) {
     // K3 always renormalizes the top-k sigmoid weights (moe_renormalize = true)
     hparams.expert_weights_norm = true;
 
-    type = LLM_TYPE_UNKNOWN;
+    GGML_ASSERT(hparams.attn_res_block_size > 0 && "Kimi-K3 requires attn_res_block_size");
+    GGML_ASSERT(hparams.moe_latent_size    > 0 && "Kimi-K3 requires moe_latent_size");
+
+    switch (hparams.n_layer()) {
+        case 93: type = LLM_TYPE_2_8T_A50B; break; // Kimi-K3
+        default: type = LLM_TYPE_UNKNOWN;
+    }
 }
 
 void llama_model_kimi_k3::load_arch_tensors(llama_model_loader &) {