phi3 : duplicate rope factors in each layer (#7447)

* phi3 : duplicate rope factors in each layer

phi3 : set phi-3 model type as 14B

model loader : simplify the process for duplicating model tensors

llama-bench : remove default pg test

* replace bool parameters in llama_model_loader with named flags

diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp
index 8b965e1990ba5d8486a7db716b91bd2431bcc7fd..6bb1f70c3c8dc5c3e451527ed71e5a8c0c0e3e1d 100644 (file)
--- a/examples/llama-bench/llama-bench.cpp
+++ b/examples/llama-bench/llama-bench.cpp
@@ -195,7 +195,7 @@ static const cmd_params cmd_params_defaults = {
     /* model         */ {"models/7B/ggml-model-q4_0.gguf"},
     /* n_prompt      */ {512},
     /* n_gen         */ {128},
-    /* n_pg          */ {{512, 128}},
+    /* n_pg          */ {},
     /* n_batch       */ {2048},
     /* n_ubatch      */ {512},
     /* type_k        */ {GGML_TYPE_F16},

diff --git a/llama.cpp b/llama.cpp
index d8c6f29a536aa12ad5afcd23d80bca8223763e71..34137c7ade6b26203d47bf0d498594562103e78c 100644 (file)
--- a/llama.cpp
+++ b/llama.cpp
@@ -1940,6 +1940,10 @@ struct llama_layer {
     // mamba bias
     struct ggml_tensor * ssm_conv1d_b;
     struct ggml_tensor * ssm_dt_b;
+
+    // long rope factors
+    struct ggml_tensor * rope_long  = nullptr;
+    struct ggml_tensor * rope_short = nullptr;
 };
 
 struct llama_kv_cell {
@@ -2111,10 +2115,6 @@ struct llama_model {
     struct ggml_tensor * output;
     struct ggml_tensor * output_b;
 
-    // long rope factors
-    struct ggml_tensor * rope_long  = nullptr;
-    struct ggml_tensor * rope_short = nullptr;
-
     std::vector<llama_layer> layers;
 
     llama_split_mode split_mode;
@@ -3425,11 +3425,15 @@ struct llama_model_loader {
         return get_tensor_meta(get_tensor_name(i));
     }
 
-    struct ggml_tensor * create_tensor_for(struct ggml_context * ctx, const struct ggml_tensor * cur) {
+    struct ggml_tensor * create_tensor_for(struct ggml_context * ctx, const struct ggml_tensor * cur, bool duplicated) {
         struct ggml_tensor * tensor = ggml_dup_tensor(ctx, cur);
         ggml_set_name(tensor, ggml_get_name(cur));
 
-        n_created++;
+        if (duplicated) {
+            size_data += ggml_nbytes(cur);
+        } else {
+            n_created++;
+        }
 
         return tensor;
     }
@@ -3464,14 +3468,17 @@ struct llama_model_loader {
         return cur;
     }
 
-    struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector<int64_t> & ne, bool required = true) {
-        const struct ggml_tensor * cur = check_tensor_dims(name, ne, required);
+    static const int TENSOR_NOT_REQUIRED = 1;
+    static const int TENSOR_DUPLICATED   = 2;
+
+    struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector<int64_t> & ne, int flags = 0) {
+        const struct ggml_tensor * cur = check_tensor_dims(name, ne, !(flags & TENSOR_NOT_REQUIRED));
 
         if (cur == NULL) {
             return NULL;
         }
 
-        return create_tensor_for(ctx, cur);
+        return create_tensor_for(ctx, cur, flags & TENSOR_DUPLICATED);
     }
 
     struct ggml_tensor * create_tensor_as_view(struct ggml_context * ctx, struct ggml_tensor * base, const std::string & name, const std::vector<int64_t> & ne, size_t offset, bool required = true) {
@@ -4139,6 +4146,7 @@ static void llm_load_hparams(
                 switch (hparams.n_layer) {
                     case 24: model.type = e_model::MODEL_1B; break;
                     case 32: model.type = e_model::MODEL_3B; break;
+                    case 40: model.type = e_model::MODEL_14B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
@@ -4965,12 +4973,10 @@ static bool llm_load_tensors(
                     {
                         model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
                         if (model.arch != LLM_ARCH_MINICPM){
-                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false);
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                             // if output is NULL, init from the input tok embed
                             if (model.output == NULL) {
-                                model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                                ml.n_created--; // artificial tensor
-                                ml.size_data += ggml_nbytes(model.output);
+                                model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                             }
                         }
                     }
@@ -4989,10 +4995,10 @@ static bool llm_load_tensors(
                         layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
 
                         // optional bias tensors
-                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     false);
-                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, false);
-                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, false);
-                        layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd},     false);
+                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
 
@@ -5003,7 +5009,7 @@ static bool llm_load_tensors(
                         } else {
                             layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert});
 
-                            layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, false);
+                            layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, llama_model_loader::TENSOR_NOT_REQUIRED);
                             if (layer.ffn_gate_exps) {
                                 layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert});
                                 layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert});
@@ -5045,12 +5051,10 @@ static bool llm_load_tensors(
                     // output
                     {
                         model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, false);
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         // if output is NULL, init from the input tok embed
                         if (model.output == NULL) {
-                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                            ml.n_created--; // artificial tensor
-                            ml.size_data += ggml_nbytes(model.output);
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                         }
                     }
 
@@ -5073,7 +5077,7 @@ static bool llm_load_tensors(
 
                         layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert});
 
-                        layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, false);
+                        layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         if (layer.ffn_gate_exps) {
                             layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert});
                             layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert});
@@ -5175,11 +5179,9 @@ static bool llm_load_tensors(
                         model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
                         model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
 
-                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, false);
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         if (!model.output) {
-                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // needs to be on GPU
-                            ml.n_created--; // artificial tensor
-                            ml.size_data += ggml_nbytes(model.output);
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // needs to be on GPU
                         }
                     }
 
@@ -5192,8 +5194,8 @@ static bool llm_load_tensors(
                         layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
                         layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
 
-                        layer.attn_norm_2   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, false);
-                        layer.attn_norm_2_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i),   {n_embd}, false);
+                        layer.attn_norm_2   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_norm_2_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
                         layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
@@ -5211,12 +5213,10 @@ static bool llm_load_tensors(
                     {
                         model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
                         model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
-                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, false);
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         if (!model.output) {
                             // needs to be on GPU
-                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                            ml.n_created--; // artificial tensor
-                            ml.size_data += ggml_nbytes(model.output);
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                         }
 
                     }
@@ -5314,14 +5314,14 @@ static bool llm_load_tensors(
                         layer.wq   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
                         layer.bq   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i),   {n_embd});
 
-                        layer.attn_q_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, false);
-                        layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {n_embd}, false);
+                        layer.attn_q_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.wk   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
                         layer.bk   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i),   {n_embd_gqa});
 
-                        layer.attn_k_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, false);
-                        layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {n_embd}, false);
+                        layer.attn_k_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.wv   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
                         layer.bv   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i),   {n_embd_gqa});
@@ -5383,18 +5383,16 @@ static bool llm_load_tensors(
             case LLM_ARCH_MPT:
                 {
                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                    model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD,   "weight"), {n_embd, hparams.n_ctx_train}, false);
+                    model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD,   "weight"), {n_embd, hparams.n_ctx_train}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                     // output
                     {
                         model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, false);
+                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
-                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, false);
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         if (!model.output) {
-                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // needs to be on GPU
-                            ml.n_created--; // artificial tensor
-                            ml.size_data += ggml_nbytes(model.output);
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // needs to be on GPU
                         }
                     }
 
@@ -5405,31 +5403,31 @@ static bool llm_load_tensors(
                         auto & layer = model.layers[i];
 
                         layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
-                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, false);
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
-                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, false);
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
-                        layer.bo   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd}, false);
+                        layer.bo   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
-                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, false);
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
-                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, false);
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.ffn_up     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
-                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, false);
+                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
-                        layer.attn_q_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, false);
-                        layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias",   i), {n_embd}, false);
+                        layer.attn_q_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias",   i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
-                        layer.attn_k_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, false);
-                        layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias",   i), {n_embd}, false);
+                        layer.attn_k_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias",   i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         // AWQ ScaleActivation layer
-                        layer.ffn_act = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_ACT, "scales", i), {n_ff}, false);
+                        layer.ffn_act = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_ACT, "scales", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
                     }
                 } break;
             case LLM_ARCH_STABLELM:
@@ -5458,17 +5456,17 @@ static bool llm_load_tensors(
                         layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
 
                         // optional bias tensors, present in Stable LM 2 1.6B
-                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     false);
-                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, false);
-                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, false);
+                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         // optional q and k layernorms, present in StableLM 2 12B
-                        layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head}, false);
-                        layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head_kv}, false);
+                        layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head_kv}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         // optional FFN norm, not present in StableLM 2 12B which uses parallel residual
-                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, false);
-                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, false);
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
                         layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
@@ -5511,12 +5509,10 @@ static bool llm_load_tensors(
                     // output
                     {
                         model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, false);
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         // if output is NULL, init from the input tok embed
                         if (model.output == NULL) {
-                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                            ml.n_created--; // artificial tensor
-                            ml.size_data += ggml_nbytes(model.output);
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                         }
                     }
 
@@ -5614,8 +5610,8 @@ static bool llm_load_tensors(
                         layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
                         layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
 
-                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, false);
-                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, false);
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         if (layer.wqkv == nullptr) {
                             layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd});
@@ -5642,9 +5638,6 @@ static bool llm_load_tensors(
                 {
                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab });
 
-                    model.rope_long  = ml.create_tensor(ctx_input, tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight"), { n_embd_head/2 }, false);
-                    model.rope_short = ml.create_tensor(ctx_input, tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight"), { n_embd_head/2 }, false);
-
                     // output
                     {
                         model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd });
@@ -5659,13 +5652,16 @@ static bool llm_load_tensors(
 
                         layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd });
 
-                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, false);
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, llama_model_loader::TENSOR_NOT_REQUIRED);
                         layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd });
 
                         layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd });
 
                         layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd });
                         layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, 2 * n_ff });
+
+                        layer.rope_long  = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight"), { n_embd_head/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                        layer.rope_short = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight"), { n_embd_head/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
                     }
                 } break;
             case LLM_ARCH_PLAMO:
@@ -5834,9 +5830,7 @@ static bool llm_load_tensors(
 
                     // output
                     model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-                    model.output      = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}); // same as tok_embd, duplicated to allow offloading
-                    ml.n_created--; // artificial tensor
-                    ml.size_data += ggml_nbytes(model.output);
+                    model.output      = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
 
                     const int64_t n_ff          = hparams.n_ff;
                     const int64_t n_embd_head_k = hparams.n_embd_head_k;
@@ -5871,12 +5865,10 @@ static bool llm_load_tensors(
                         model.output_norm   = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
                         model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
 
-                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false);
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         // if output is NULL, init from the input tok embed
                         if (model.output == NULL) {
-                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                            ml.n_created--; // artificial tensor
-                            ml.size_data += ggml_nbytes(model.output);
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                         }
 
                     }
@@ -5927,12 +5919,10 @@ static bool llm_load_tensors(
                     {
                         model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
 
-                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false);
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         // if output is NULL, init from the input tok embed, duplicated to allow offloading
                         if (model.output == NULL) {
-                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                            ml.n_created--; // artificial tensor
-                            ml.size_data += ggml_nbytes(model.output);
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                         }
                     }
 
@@ -5993,9 +5983,7 @@ static bool llm_load_tensors(
                     {
                         model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
                         // init output from the input tok embed
-                        model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                        ml.n_created--; // artificial tensor
-                        ml.size_data += ggml_nbytes(model.output);
+                        model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                     }
 
                     for (int i = 0; i < n_layer; ++i) {
@@ -6027,12 +6015,10 @@ static bool llm_load_tensors(
 
                     // output
                     {
-                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false);
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         // if output is NULL, init from the input tok embed
                         if (model.output == NULL) {
-                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                            ml.n_created--; // artificial tensor
-                            ml.size_data += ggml_nbytes(model.output);
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                         }
                     }
 
@@ -6875,9 +6861,9 @@ struct llm_build_context {
         cb(lctx.inp_K_shift, "K_shift", -1);
         ggml_set_input(lctx.inp_K_shift);
 
-        struct ggml_tensor * rope_factors = build_rope_factors();
 
         for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * rope_factors = build_rope_factors(il);
             struct ggml_tensor * tmp =
                 // we rotate only the first n_rot dimensions
                 ggml_rope_ext_inplace(ctx0,
@@ -6991,15 +6977,15 @@ struct llm_build_context {
         return lctx.inp_pos;
     }
 
-    struct ggml_tensor * build_rope_factors() {
+    struct ggml_tensor * build_rope_factors(int il) {
         // choose long/short freq factors based on the context size
         const auto n_ctx_pre_seq = cparams.n_ctx / cparams.n_seq_max;
 
         if (n_ctx_pre_seq > hparams.n_yarn_orig_ctx) {
-            return model.rope_long;
+            return model.layers[il].rope_long;
         }
 
-        return model.rope_short;
+        return model.layers[il].rope_short;
     }
 
     struct ggml_tensor * build_inp_out_ids() {
@@ -9120,14 +9106,14 @@ 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();
 
-        // rope freq factors for 128k context
-        struct ggml_tensor * rope_factors = build_rope_factors();
-
         for (int il = 0; il < n_layer; ++il) {
             auto residual = inpL;
 
             // self-attention
             {
+                // rope freq factors for 128k context
+                struct ggml_tensor * rope_factors = build_rope_factors(il);
+
                 struct ggml_tensor* attn_norm_output = llm_build_norm(ctx0, inpL, hparams,
                     model.layers[il].attn_norm,
                     NULL,