feat(granitemoe): Implement granitemoe

GraniteMoE follows the mixtral architecture (once the input_linear layers
are split into gate_exps/up_exps). The main delta is the addition of the
same four multipliers used in Granite.

Branch: GraniteMoE

Signed-off-by: Gabe Goodhart <[email protected]>

Author: gabe-l-hart

src/llama.cpp

@@ -213,6 +213,7 @@ enum llm_arch {
     LLM_ARCH_EXAONE,
     LLM_ARCH_RWKV6,
     LLM_ARCH_GRANITE,
+    LLM_ARCH_GRANITE_MOE,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -262,6 +263,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_EXAONE,          "exaone"       },
     { LLM_ARCH_RWKV6,           "rwkv6"        },
     { LLM_ARCH_GRANITE,         "granite"      },
+    { LLM_ARCH_GRANITE_MOE,     "granitemoe"   },
     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
 };
 
@@ -1431,6 +1433,23 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_GRANITE_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -2344,7 +2363,7 @@ struct llama_hparams {
     float f_max_alibi_bias = 0.0f;
     float f_logit_scale    = 0.0f;
 
-    // For Granite architecture
+    // For Granite architectures
     float f_residual_multiplier = 0.0f;
     float f_embedding_multiplier = 0.0f;
     float f_attention_multiplier = 0.0f;
@@ -5385,6 +5404,7 @@ static void llm_load_hparams(
     switch (model.arch) {
         case LLM_ARCH_LLAMA:
         case LLM_ARCH_GRANITE:
+        case LLM_ARCH_GRANITE_MOE:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
@@ -5408,8 +5428,8 @@ static void llm_load_hparams(
                         default: model.type = e_model::MODEL_UNKNOWN;
                     }
                 }
-                // Extra multipliers for Granite architecture
-                if (model.arch == LLM_ARCH_GRANITE) {
+                // Extra multipliers for Granite architectures
+                if (model.arch == LLM_ARCH_GRANITE || model.arch == LLM_ARCH_GRANITE_MOE) {
                     ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale);
                     ml.get_key(LLM_KV_RESIDUAL_MULTIPLIER, hparams.f_residual_multiplier);
                     ml.get_key(LLM_KV_EMBEDDING_MULTIPLIER, hparams.f_embedding_multiplier);
@@ -6685,7 +6705,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
         LLAMA_LOG_INFO("%s: n_ff_shexp       = %d\n",     __func__, hparams.n_ff_shexp);
     }
 
-    if (model.arch == LLM_ARCH_GRANITE) {
+    if (model.arch == LLM_ARCH_GRANITE || model.arch == LLM_ARCH_GRANITE_MOE) {
         LLAMA_LOG_INFO("%s: f_embedding_multiplier = %f\n", __func__, hparams.f_embedding_multiplier);
         LLAMA_LOG_INFO("%s: f_residual_multiplier  = %f\n", __func__, hparams.f_residual_multiplier);
         LLAMA_LOG_INFO("%s: f_attention_multiplier = %f\n", __func__, hparams.f_attention_multiplier);
@@ -6861,6 +6881,7 @@ static bool llm_load_tensors(
             case LLM_ARCH_REFACT:
             case LLM_ARCH_MINICPM:
             case LLM_ARCH_GRANITE:
+            case LLM_ARCH_GRANITE_MOE:
                 {
                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
 
@@ -15362,6 +15383,7 @@ static struct ggml_cgraph * llama_build_graph(
     switch (model.arch) {
         case LLM_ARCH_LLAMA:
         case LLM_ARCH_GRANITE:
+        case LLM_ARCH_GRANITE_MOE:
             {
                 result = llm.build_llama();
             } break;
@@ -18649,6 +18671,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_CHATGLM:
         case LLM_ARCH_GRANITE:
+        case LLM_ARCH_GRANITE_MOE:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2