feat(llama.cpp): First pass at full port of granite deviations from llama

Something is still not working right since the results are mostly terrible,
but on occasion it's producing relevant results at this point, so
_something_ is working.

Branch: GraniteLM

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

Author: gabe-l-hart

src/llama.cpp

@@ -1415,6 +1415,22 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_CHANNEL_MIX_RECEPTANCE,    "blk.%d.channel_mix_receptance" },
         },
     },
+    {
+        LLM_ARCH_GRANITE,
+        {
+            { 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,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -6844,6 +6860,7 @@ static bool llm_load_tensors(
             case LLM_ARCH_LLAMA:
             case LLM_ARCH_REFACT:
             case LLM_ARCH_MINICPM:
+            case LLM_ARCH_GRANITE:
                 {
                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
 
@@ -8738,6 +8755,11 @@ static struct ggml_tensor * llm_build_inp_embd(
         ggml_set_input(lctx.inp_embd);
     }
 
+    // For Granite architecture
+    if (hparams.f_embedding_multiplier != 0.0f) {
+        inpL = ggml_scale(ctx, inpL, hparams.f_embedding_multiplier);
+    }
+
     cb(inpL, "inp_embd", -1);
 
     return inpL;
@@ -10016,6 +10038,7 @@ struct llm_build_context {
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
 
+        const float kq_scale = hparams.f_attention_multiplier == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_multiplier;
         for (int il = 0; il < n_layer; ++il) {
             struct ggml_tensor * inpSA = inpL;
 
@@ -10068,7 +10091,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, lctx, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -10079,6 +10102,11 @@ struct llm_build_context {
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
 
+            // For Granite architecture
+            if (hparams.f_residual_multiplier) {
+                cur = ggml_scale(ctx0, cur, hparams.f_residual_multiplier);
+            }
+
             struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
             cb(ffn_inp, "ffn_inp", il);
 
@@ -10115,6 +10143,11 @@ struct llm_build_context {
                 cb(cur, "ffn_moe_out", il);
             }
 
+            // For Granite architecture
+            if (hparams.f_residual_multiplier) {
+                cur = ggml_scale(ctx0, cur, hparams.f_residual_multiplier);
+            }
+
             cur = ggml_add(ctx0, cur, ffn_inp);
             cb(cur, "ffn_out", il);
 
@@ -10134,6 +10167,12 @@ struct llm_build_context {
 
         // lm_head
         cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+
+        // For Granite architecture
+        if (hparams.f_logit_scale) {
+            cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+        }
+
         cb(cur, "result_output", -1);
 
         ggml_build_forward_expand(gf, cur);
@@ -15322,6 +15361,7 @@ static struct ggml_cgraph * llama_build_graph(
 
     switch (model.arch) {
         case LLM_ARCH_LLAMA:
+        case LLM_ARCH_GRANITE:
             {
                 result = llm.build_llama();
             } break;
@@ -18608,6 +18648,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_ARCTIC:
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_CHATGLM:
+        case LLM_ARCH_GRANITE:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2