@fugood/llama.node 1.4.12 → 1.4.13

package/src/llama.cpp/src/llama-model.cpp

@@ -507,6 +507,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
     ml.get_key(LLM_KV_CONTEXT_LENGTH,          hparams.n_ctx_train);
     ml.get_key(LLM_KV_EMBEDDING_LENGTH,        hparams.n_embd);
+    ml.get_key(LLM_KV_EMBEDDING_LENGTH_OUT,    hparams.n_embd_out, false);
     ml.get_key(LLM_KV_BLOCK_COUNT,             hparams.n_layer);
     ml.get_key(LLM_KV_EXPERT_COUNT,            hparams.n_expert,        false);
     ml.get_key(LLM_KV_EXPERT_USED_COUNT,       hparams.n_expert_used,   false);
@@ -578,6 +579,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     hparams.rope_scaling_type_train = llama_rope_scaling_type_from_string(rope_scaling);
     GGML_ASSERT(hparams.rope_scaling_type_train != LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED);
 
+    // TODO: Handle SWA metadata similarly when models start implementing it
     // rope_freq_scale (inverse of the kv) is optional
     float ropescale = 0.0f;
     if (!ml.get_key(LLM_KV_ROPE_SCALING_FACTOR, ropescale, false)) {
@@ -586,10 +588,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     }
     hparams.rope_freq_scale_train = ropescale == 0.0f ? 1.0f : 1.0f/ropescale;
 
-    // by default assume that the sliding-window layers use the same scaling type as the non-sliding-window layers
-    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
-    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
-
     ml.get_key(LLM_KV_ROPE_SCALING_ATTN_FACTOR, hparams.rope_attn_factor, false);
 
     // non-transformer models do not have attention heads
@@ -677,6 +675,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.f_attn_temp_scale       = 0.1f;
                     hparams.f_attn_temp_offset      = 1.0f;
                     hparams.set_swa_pattern(4);   // pattern: 3 chunked - 1 full
+
+                    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+                    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
                 }
 
                 switch (hparams.n_expert) {
@@ -722,6 +724,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 if (hparams.n_swa > 0) {
                     hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                     hparams.set_swa_pattern(4);
+
+                    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+                    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
                 } else {
                     hparams.swa_type = LLAMA_SWA_TYPE_NONE;
                 }
@@ -1243,7 +1249,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 if (found_swa && hparams.n_swa > 0) {
                     uint32_t swa_period = 8;
                     hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
-                    hparams.rope_freq_scale_train_swa = 1.0f;
                     ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa);
                     ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
                     hparams.set_swa_pattern(swa_period);
@@ -1309,7 +1314,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 hparams.n_swa = 4096; // default value of gemma 2
                 hparams.set_swa_pattern(2);
                 hparams.attn_soft_cap = true;
+                hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
 
+                ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,          hparams.rope_freq_base_train_swa, false);
                 ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING,      hparams.f_attn_logit_softcapping, false);
@@ -1334,8 +1342,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                     hparams.set_swa_pattern(6);
 
-                    hparams.rope_freq_base_train_swa  = 10000.0f;
-                    hparams.rope_freq_scale_train_swa = 1.0f;
+                    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
                 } else {
                     hparams.swa_type = LLAMA_SWA_TYPE_NONE;
                 }
@@ -1365,10 +1372,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 hparams.set_swa_pattern(5);
 
                 hparams.n_layer_kv_from_start     = 20;
-                hparams.rope_freq_base_train_swa  = 10000.0f;
-                hparams.rope_freq_scale_train_swa = 1.0f;
                 hparams.f_attention_scale         = 1.0f;
 
+                ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,          hparams.rope_freq_base_train_swa, false);
                 ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
@@ -1384,9 +1390,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 hparams.set_swa_pattern(6);
 
                 hparams.causal_attn = false; // embeddings do not use causal attention
-                hparams.rope_freq_base_train_swa = 10000.0f;
-                hparams.rope_freq_scale_train_swa = 1.0f;
 
+                ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
                 ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type);
@@ -1525,7 +1530,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             {
                 hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                 hparams.set_swa_pattern(4);
+                hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
 
+                ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,       hparams.rope_freq_base_train_swa, false);
                 ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
                 ml.get_key(LLM_KV_LOGIT_SCALE,              hparams.f_logit_scale);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,  hparams.f_norm_eps);
@@ -1564,6 +1572,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 if (found_swa && hparams.n_swa > 0) {
                     hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                     hparams.set_swa_pattern(4);
+
+                    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                    hparams.rope_freq_scale_train_swa = 1.0; // See olmo2.cpp
+                    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
                 } else {
                     hparams.swa_type = LLAMA_SWA_TYPE_NONE;
                 }
@@ -1906,6 +1918,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                     hparams.n_swa = 4096;
                     hparams.set_swa_pattern(4);
+
+                    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+                    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
                 }
 
                 ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);
@@ -2208,6 +2224,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                 hparams.set_swa_pattern(2);
 
+                hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+                ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
+
                 switch (hparams.n_layer) {
                     case 24: type = LLM_TYPE_20B; break;
                     case 36: type = LLM_TYPE_120B; break;
@@ -2252,6 +2272,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.swa_type      = LLAMA_SWA_TYPE_STANDARD;
                     hparams.n_swa         = 4096;
                     hparams.set_swa_pattern(4, true);
+
+                    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+                    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+                    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
                 } else {
                     hparams.swa_type             = LLAMA_SWA_TYPE_NONE;
                     hparams.n_no_rope_layer_step = hparams.n_layer;
@@ -6446,6 +6470,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.shortconv.out_proj = create_tensor(tn(LLM_TENSOR_SHORTCONV_OUTPROJ, "weight", i), {n_embd, n_embd}, 0);
                         }
                     }
+
+                    // for LFM2-ColBert-350M
+                    dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.get_n_embd_out()}, TENSOR_NOT_REQUIRED);
                 } break;
             case LLM_ARCH_SMALLTHINKER:
                 {
@@ -7098,6 +7125,10 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: rope scaling     = %s\n",     __func__, rope_scaling_type.c_str());
         LLAMA_LOG_INFO("%s: freq_base_train  = %.1f\n",   __func__, hparams.rope_freq_base_train);
         LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
+        if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
+            LLAMA_LOG_INFO("%s: freq_base_swa    = %.1f\n",   __func__, hparams.rope_freq_base_train_swa);
+            LLAMA_LOG_INFO("%s: freq_scale_swa   = %g\n",     __func__, hparams.rope_freq_scale_train_swa);
+        }
         LLAMA_LOG_INFO("%s: n_ctx_orig_yarn  = %u\n",     __func__, hparams.n_ctx_orig_yarn);
         LLAMA_LOG_INFO("%s: rope_yarn_log_mul= %.4f\n",   __func__, hparams.rope_yarn_log_mul);
         LLAMA_LOG_INFO("%s: rope_finetuned   = %s\n",     __func__, hparams.rope_finetuned ? "yes" : "unknown");
@@ -7910,12 +7941,17 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
     // add on pooling layer
     llm->build_pooling(cls, cls_b, cls_out, cls_out_b);
 
+    // add backend sampling layers (if any)
+    llm->build_sampling();
+
     // if the gguf model was converted with --sentence-transformers-dense-modules
     // there will be two additional dense projection layers
     // dense linear projections are applied after pooling
     // TODO: move reranking logic here and generalize
     llm->build_dense_out(dense_2_out_layers, dense_3_out_layers);
 
+    llm->res->set_outputs();
+
     return llm->res->get_gf();
 }
 
@@ -7971,6 +8007,10 @@ int32_t llama_model_n_embd_inp(const llama_model * model) {
     return model->hparams.n_embd_inp();
 }
 
+int32_t llama_model_n_embd_out(const llama_model * model) {
+    return model->hparams.get_n_embd_out();
+}
+
 int32_t llama_model_n_layer(const llama_model * model) {
     return model->hparams.n_layer;
 }