@fugood/llama.node 0.3.12 → 0.3.14

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

@@ -9,6 +9,7 @@
 #include <algorithm>
 #include <cassert>
 #include <cstring>
+#include <cmath>
 #include <functional>
 #include <map>
 #include <sstream>
@@ -864,6 +865,23 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                }
             } break;
+        case LLM_ARCH_GEMMA3:
+            {
+                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);
+
+                switch (hparams.n_layer) {
+                    case 26: type = LLM_TYPE_1B; break;
+                    case 34: type = LLM_TYPE_4B; break;
+                    case 48: type = LLM_TYPE_12B; break;
+                    case 62: type = LLM_TYPE_27B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+
+                hparams.f_attention_scale = type == LLM_TYPE_27B
+                    ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0)))
+                    : 1.0f / std::sqrt(float(hparams.n_embd_head_k));
+            } break;
         case LLM_ARCH_STARCODER2:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -1275,7 +1293,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
     const bool use_mmap_buffer = true;
 
-    LLAMA_LOG_INFO("%s: loading model tensors, this can take a while... (mmap = %s)\n", __func__, use_mmap_buffer ? "true" : "false");
+    LLAMA_LOG_INFO("%s: loading model tensors, this can take a while... (mmap = %s)\n", __func__, ml.use_mmap ? "true" : "false");
 
     // build a list of buffer types for the CPU and GPU devices
     pimpl->cpu_buft_list = make_cpu_buft_list(devices);
@@ -1424,6 +1442,14 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 throw std::runtime_error(format("missing tensor info mapping for %s", tn.str().c_str()));
             }
 
+            // skip unused tensors
+            if (info.op == GGML_OP_NONE) {
+                LLAMA_LOG_WARN("model has unused tensor %s -- ignoring\n", tn.str().c_str());
+                ml.n_created++;
+
+                return nullptr;
+            }
+
             // tensors with "bias" suffix are always used with GGML_OP_ADD
             ggml_op op;
             bool bias = tn.suffix != nullptr && strcmp(tn.suffix, "bias") == 0;
@@ -2194,13 +2220,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_PHI3:
                 {
-                    const int64_t n_embd_head = n_embd / n_head;
-
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
 
                     // output
                     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
-                    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
 
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
@@ -2215,8 +2244,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, 0);
                         layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, 2 * n_ff }, 0);
 
-                        layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
-                        layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                        layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                        layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
                     }
                 } break;
             case LLM_ARCH_PHIMOE:
@@ -2443,6 +2472,35 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
                         layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
 
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+                    }
+                } break;
+            case LLM_ARCH_GEMMA3:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+                        layer.attn_k_norm    = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM,    "weight", i), {n_embd_head_k}, 0);
+                        layer.attn_q_norm    = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM,    "weight", i), {n_embd_head_k}, 0);
+
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
                         layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
                         layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
@@ -3639,6 +3697,7 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: f_clamp_kqv      = %.1e\n",   __func__, hparams.f_clamp_kqv);
         LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n",   __func__, hparams.f_max_alibi_bias);
         LLAMA_LOG_INFO("%s: f_logit_scale    = %.1e\n",   __func__, hparams.f_logit_scale);
+        LLAMA_LOG_INFO("%s: f_attn_scale     = %.1e\n",   __func__, hparams.f_attention_scale);
         LLAMA_LOG_INFO("%s: n_ff             = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_ff(il); }, hparams.n_layer).c_str());
         LLAMA_LOG_INFO("%s: n_expert         = %u\n",     __func__, hparams.n_expert);
         LLAMA_LOG_INFO("%s: n_expert_used    = %u\n",     __func__, hparams.n_expert_used);
@@ -3830,6 +3889,10 @@ int32_t llama_model_n_head(const struct llama_model * model) {
     return model->hparams.n_head();
 }
 
+int32_t llama_model_n_head_kv(const struct llama_model * model) {
+    return model->hparams.n_head_kv();
+}
+
 // deprecated
 int32_t llama_n_ctx_train(const struct llama_model * model) {
     return llama_model_n_ctx_train(model);
@@ -3908,6 +3971,7 @@ enum llama_rope_type llama_model_rope_type(const struct llama_model * model) {
         case LLM_ARCH_PHIMOE:
         case LLM_ARCH_GEMMA:
         case LLM_ARCH_GEMMA2:
+        case LLM_ARCH_GEMMA3:
         case LLM_ARCH_STARCODER2:
         case LLM_ARCH_OPENELM:
         case LLM_ARCH_GPTNEOX:

package/src/llama.cpp/src/llama-sampling.cpp

@@ -316,6 +316,13 @@ static uint32_t get_rng_seed(uint32_t seed) {
 
 // llama_sampler API
 
+struct llama_sampler * llama_sampler_init(const struct llama_sampler_i * iface, llama_sampler_context_t ctx) {
+    return new llama_sampler {
+        /* .iface = */ iface,
+        /* .ctx   = */ ctx,
+    };
+}
+
 const char * llama_sampler_name(const struct llama_sampler * smpl) {
     if (!smpl->iface) {
         return "(null)";
@@ -347,10 +354,10 @@ struct llama_sampler * llama_sampler_clone(const struct llama_sampler * smpl) {
     }
 
     if (smpl->ctx == nullptr) {
-        return new llama_sampler {
+        return llama_sampler_init(
             /* .iface = */ smpl->iface,
-            /* .ctx   = */ nullptr,
-        };
+            /* .ctx   = */ nullptr
+        );
     }
 
     GGML_ABORT("the sampler does not support cloning");
@@ -472,15 +479,15 @@ static struct llama_sampler_i llama_sampler_chain_i = {
 };
 
 struct llama_sampler * llama_sampler_chain_init(struct llama_sampler_chain_params params) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_chain_i,
         /* .ctx   = */ new llama_sampler_chain {
             /* .params      = */ params,
             /* .samplers    = */ {},
             /* .t_sample_us = */ 0,
             /* .n_sample    = */ 0,
-        },
-    };
+        }
+    );
 }
 
 void llama_sampler_chain_add(struct llama_sampler * chain, struct llama_sampler * smpl) {
@@ -546,10 +553,10 @@ static struct llama_sampler_i llama_sampler_greedy_i = {
 };
 
 struct llama_sampler * llama_sampler_init_greedy() {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_greedy_i,
-        /* .ctx   = */ nullptr,
-    };
+        /* .ctx   = */ nullptr
+    );
 }
 
 // dist
@@ -608,14 +615,14 @@ static struct llama_sampler_i llama_sampler_dist_i = {
 
 struct llama_sampler * llama_sampler_init_dist(uint32_t seed) {
     auto seed_cur = get_rng_seed(seed);
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_dist_i,
         /* .ctx   = */ new llama_sampler_dist {
             /* .seed     = */ seed,
             /* .seed_cur = */ seed_cur,
             /* .rng      = */ std::mt19937(seed_cur),
-        },
-    };
+        }
+    );
 }
 
 // softmax
@@ -638,10 +645,10 @@ static struct llama_sampler_i llama_sampler_softmax_i = {
 };
 
 struct llama_sampler * llama_sampler_init_softmax() {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_softmax_i,
-        /* .ctx   = */ nullptr,
-    };
+        /* .ctx   = */ nullptr
+    );
 }
 
 // top-k
@@ -678,12 +685,12 @@ static struct llama_sampler_i llama_sampler_top_k_i = {
 };
 
 struct llama_sampler * llama_sampler_init_top_k(int32_t k) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_top_k_i,
         /* .ctx   = */ new llama_sampler_top_k {
             /* .k = */ k,
-        },
-    };
+        }
+    );
 }
 
 // top-p
@@ -744,13 +751,13 @@ static struct llama_sampler_i llama_sampler_top_p_i = {
 };
 
 struct llama_sampler * llama_sampler_init_top_p(float p, size_t min_keep) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_top_p_i,
         /* .ctx   = */ new llama_sampler_top_p {
             /* .p        = */ p,
             /* .min_keep = */ min_keep,
-        },
-    };
+        }
+    );
 }
 
 // min-p
@@ -840,13 +847,13 @@ static struct llama_sampler_i llama_sampler_min_p_i = {
 };
 
 struct llama_sampler * llama_sampler_init_min_p(float p, size_t min_keep) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_min_p_i,
         /* .ctx   = */ new llama_sampler_min_p {
             /* .p        = */ p,
             /* .min_keep = */ min_keep,
-        },
-    };
+        }
+    );
 }
 
 // typical
@@ -939,13 +946,13 @@ static struct llama_sampler_i llama_sampler_typical_i = {
 };
 
 struct llama_sampler * llama_sampler_init_typical(float p, size_t min_keep) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_typical_i,
         /* .ctx   = */ new llama_sampler_typical {
             /* .p        = */ p,
             /* .min_keep = */ min_keep,
-        },
-    };
+        }
+    );
 }
 
 // temp
@@ -983,12 +990,12 @@ static struct llama_sampler_i llama_sampler_temp_i = {
 };
 
 struct llama_sampler * llama_sampler_init_temp(float temp) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_temp_i,
         /* .ctx   = */ new llama_sampler_temp {
             /*.temp = */ temp,
-        },
-    };
+        }
+    );
 }
 
 // temp-ext
@@ -1093,14 +1100,14 @@ static struct llama_sampler_i llama_sampler_temp_ext_i = {
 };
 
 struct llama_sampler * llama_sampler_init_temp_ext(float temp, float delta, float exponent) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_temp_ext_i,
         /* .ctx   = */ new llama_sampler_temp_ext {
             /* .temp     = */ temp,
             /* .delta    = */ delta,
             /* .exponent = */ exponent,
-        },
-    };
+        }
+    );
 }
 
 // xtc
@@ -1185,7 +1192,7 @@ static struct llama_sampler_i llama_sampler_xtc_i = {
 
 struct llama_sampler * llama_sampler_init_xtc(float p, float t, size_t min_keep, uint32_t seed) {
     auto seed_cur = get_rng_seed(seed);
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_xtc_i,
         /* .ctx   = */ new llama_sampler_xtc {
             /* .probability   = */ p,
@@ -1194,8 +1201,8 @@ struct llama_sampler * llama_sampler_init_xtc(float p, float t, size_t min_keep,
             /* .seed          = */ seed,
             /* .seed_cur      = */ seed_cur,
             /* .rng           = */ std::mt19937(seed_cur),
-        },
-    };
+        }
+    );
 }
 
 // mirostat
@@ -1292,7 +1299,7 @@ static struct llama_sampler_i llama_sampler_mirostat_i = {
 
 struct llama_sampler * llama_sampler_init_mirostat(int32_t n_vocab, uint32_t seed, float tau, float eta, int32_t m) {
     auto seed_cur = get_rng_seed(seed);
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_mirostat_i,
         /* .ctx   = */ new llama_sampler_mirostat {
             /* .n_vocab  = */ n_vocab,
@@ -1303,8 +1310,8 @@ struct llama_sampler * llama_sampler_init_mirostat(int32_t n_vocab, uint32_t see
             /* .m        = */ m,
             /* .mu       = */ 2.0f*tau,
             /* .rng      = */ std::mt19937(seed_cur),
-        },
-    };
+        }
+    );
 }
 
 // mirostat v2
@@ -1391,7 +1398,7 @@ static struct llama_sampler_i llama_sampler_mirostat_v2_i = {
 
 struct llama_sampler * llama_sampler_init_mirostat_v2(uint32_t seed, float tau, float eta) {
     auto seed_cur = get_rng_seed(seed);
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_mirostat_v2_i,
         /* .ctx   = */ new llama_sampler_mirostat_v2 {
             /* .seed     = */ seed,
@@ -1400,8 +1407,8 @@ struct llama_sampler * llama_sampler_init_mirostat_v2(uint32_t seed, float tau,
             /* .eta      = */ eta,
             /* .mu       = */ 2.0f*tau,
             /* .rng      = */ std::mt19937(seed_cur),
-        },
-    };
+        }
+    );
 }
 
 // grammar
@@ -1442,7 +1449,9 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
                      const char ** trigger_words,
                             size_t num_trigger_words,
                const llama_token * trigger_tokens,
-                            size_t num_trigger_tokens);
+                            size_t num_trigger_tokens,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns);
 
 static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
     auto * ctx = (llama_sampler_grammar *) smpl->ctx;
@@ -1450,12 +1459,14 @@ static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
         return;
     }
 
-    std::vector<const char *>  trigger_words;
-    for (auto & word : ctx->grammar->trigger_words) {
-        trigger_words.push_back(word.c_str());
+    std::vector<const char *>  trigger_patterns_c;
+    trigger_patterns_c.reserve(ctx->grammar->trigger_patterns.size());
+    for (auto & trigger_pattern : ctx->grammar->trigger_patterns) {
+        trigger_patterns_c.push_back(trigger_pattern.pattern.c_str());
     }
+
     auto * grammar_new = llama_grammar_init_impl(ctx->grammar->vocab, ctx->grammar_str.c_str(), ctx->grammar_root.c_str(),
-                                                 ctx->grammar->lazy, trigger_words.data(), trigger_words.size(),
+                                                 ctx->grammar->lazy, trigger_patterns_c.data(), trigger_patterns_c.size(),
                                                  ctx->grammar->trigger_tokens.data(), ctx->grammar->trigger_tokens.size());
 
     llama_grammar_free_impl(ctx->grammar);
@@ -1465,7 +1476,7 @@ static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
 static struct llama_sampler * llama_sampler_grammar_clone(const struct llama_sampler * smpl) {
     const auto * ctx = (const llama_sampler_grammar *) smpl->ctx;
 
-    auto * result = llama_sampler_init_grammar_impl(ctx->vocab, nullptr, nullptr, false, nullptr, 0, nullptr, 0);
+    auto * result = llama_sampler_init_grammar_impl(ctx->vocab, nullptr, nullptr, false, nullptr, 0, nullptr, 0, nullptr, 0);
 
     // copy the state
     {
@@ -1509,15 +1520,33 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
                      const char ** trigger_words,
                             size_t num_trigger_words,
                const llama_token * trigger_tokens,
-                            size_t num_trigger_tokens) {
+                            size_t num_trigger_tokens,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns) {
     auto * ctx = new llama_sampler_grammar;
 
     if (grammar_str != nullptr && grammar_str[0] != '\0') {
+        // TODO: remove trigger_words support.
+        if (trigger_words != nullptr && num_trigger_words > 0) {
+            GGML_ASSERT(trigger_patterns == nullptr && num_trigger_patterns == 0);
+            std::string trigger_pattern("[\\s\\S]*?(");
+            for (size_t i = 0; i < num_trigger_words; ++i) {
+                static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
+                if (i > 0) {
+                    trigger_pattern += "|";
+                }
+                trigger_pattern += std::regex_replace(trigger_words[i], special_chars, "\\$0");
+            }
+            trigger_pattern += ")[\\s\\S]*";
+            auto trigger_pattern_c = trigger_pattern.c_str();
+            trigger_patterns = &trigger_pattern_c;
+            num_trigger_patterns = 1;
+        }
         *ctx = {
             /* .vocab        = */ vocab,
             /* .grammar_str  = */ grammar_str,
             /* .grammar_root = */ grammar_root,
-            /* .grammar      = */ llama_grammar_init_impl(vocab, grammar_str, grammar_root, lazy, trigger_words, num_trigger_words, trigger_tokens, num_trigger_tokens),
+            /* .grammar      = */ llama_grammar_init_impl(vocab, grammar_str, grammar_root, lazy, trigger_patterns, num_trigger_patterns, trigger_tokens, num_trigger_tokens),
         };
     } else {
         *ctx = {
@@ -1528,17 +1557,17 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
         };
     }
 
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_grammar_i,
-        /* .ctx   = */ ctx,
-    };
+        /* .ctx   = */ ctx
+    );
 }
 
 struct llama_sampler * llama_sampler_init_grammar(
         const struct llama_vocab * vocab,
                       const char * grammar_str,
                       const char * grammar_root) {
-    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ false, nullptr, 0, nullptr, 0);
+    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ false, nullptr, 0, nullptr, 0, nullptr, 0);
 }
 
 struct llama_sampler * llama_sampler_init_grammar_lazy(
@@ -1549,7 +1578,18 @@ struct llama_sampler * llama_sampler_init_grammar_lazy(
                             size_t num_trigger_words,
                const llama_token * trigger_tokens,
                             size_t num_trigger_tokens) {
-    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ true, trigger_words, num_trigger_words, trigger_tokens, num_trigger_tokens);
+    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ true, trigger_words, num_trigger_words, trigger_tokens, num_trigger_tokens, nullptr, 0);
+}
+
+struct llama_sampler * llama_sampler_init_grammar_lazy_patterns(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens) {
+    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ true, nullptr, 0, trigger_tokens, num_trigger_tokens, trigger_patterns, num_trigger_patterns);
 }
 
 // penalties
@@ -1678,7 +1718,7 @@ struct llama_sampler * llama_sampler_init_penalties(
         float penalty_present) {
     penalty_last_n = std::max(penalty_last_n, 0);
 
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_penalties_i,
         /* .ctx   = */ new llama_sampler_penalties {
             /* .penalty_last_n  = */ penalty_last_n,
@@ -1687,8 +1727,75 @@ struct llama_sampler * llama_sampler_init_penalties(
             /* .penalty_present = */ penalty_present,
             /* .prev            = */ ring_buffer<llama_token>(penalty_last_n),
             /* .token_count     = */ {},
-        },
-    };
+        }
+    );
+}
+
+// top-n-sigma
+
+struct llama_sampler_top_n_sigma {
+    const float n;
+};
+
+static const char * llama_sampler_top_n_sigma_name(const struct llama_sampler * /*smpl*/) {
+    return "top-n-sigma";
+}
+
+static void llama_sampler_top_n_sigma_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    const auto * ctx = (llama_sampler_top_n_sigma *) smpl->ctx;
+
+    // find max logit and calculate mean
+    float max = cur_p->data[0].logit;
+    float logits_sum = 0;
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        if (cur_p->data[i].logit > max) {
+            max = cur_p->data[i].logit;
+        }
+        logits_sum += cur_p->data[i].logit;
+    }
+    float mean = logits_sum/cur_p->size;
+
+    // calculate standard deviation
+    float acc = 0;
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        acc += pow(cur_p->data[i].logit - mean, 2);
+    }
+    float std = sqrt(acc/cur_p->size);
+
+    //apply mask
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        if (cur_p->data[i].logit < max - (ctx->n * std)) {
+            cur_p->data[i].logit = -INFINITY;
+        }
+    }
+    llama_sampler_softmax_impl(cur_p);
+}
+
+static struct llama_sampler * llama_sampler_top_n_sigma_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_top_n_sigma *) smpl->ctx;
+    return llama_sampler_init_top_n_sigma(ctx->n);
+}
+
+static void llama_sampler_top_n_sigma_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_top_n_sigma *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_top_n_sigma_i = {
+    /* .name   = */ llama_sampler_top_n_sigma_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_top_n_sigma_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_top_n_sigma_clone,
+    /* .free   = */ llama_sampler_top_n_sigma_free,
+};
+
+struct llama_sampler * llama_sampler_init_top_n_sigma(float n) {
+    return llama_sampler_init(
+        /* .iface = */ &llama_sampler_top_n_sigma_i,
+        /* .ctx   = */ new llama_sampler_top_n_sigma {
+            /* .n = */ n,
+        }
+    );
 }
 
 // DRY
@@ -2041,7 +2148,7 @@ struct llama_sampler * llama_sampler_init_dry(const struct llama_vocab * vocab,
         }
     }
 
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_dry_i,
         /* .ctx   = */ new llama_sampler_dry {
             /* .total_context_size     = */ context_size,
@@ -2053,8 +2160,8 @@ struct llama_sampler * llama_sampler_init_dry(const struct llama_vocab * vocab,
             /* .dry_repeat_count       = */ dry_enabled ? std::vector<int>(effective_dry_penalty_last_n, 0) : std::vector<int>{},
             /* .dry_max_token_repeat   = */ {},
             /* .last_tokens            = */ dry_enabled ? ring_buffer<llama_token>(effective_dry_penalty_last_n) : ring_buffer<llama_token>(0),
-        },
-    };
+        }
+    );
 }
 
 // wrapper for test-sampling.cpp
@@ -2155,14 +2262,14 @@ struct llama_sampler * llama_sampler_init_logit_bias(
                          int32_t   n_vocab,
                          int32_t   n_logit_bias,
           const llama_logit_bias * logit_bias) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_logit_bias_i,
         /* .ctx   = */ new llama_sampler_logit_bias {
             /* .n_vocab    = */ n_vocab,
             /* .logit_bias = */ std::vector<llama_logit_bias>(logit_bias, logit_bias + n_logit_bias),
             /* .to_search  = */ {},
-        },
-    };
+        }
+    );
 }
 
 // infill
@@ -2377,14 +2484,14 @@ static struct llama_sampler_i llama_sampler_infill_i = {
 };
 
 struct llama_sampler * llama_sampler_init_infill(const struct llama_vocab * vocab) {
-    return new llama_sampler {
+    return llama_sampler_init(
         /* .iface = */ &llama_sampler_infill_i,
         /* .ctx   = */ new llama_sampler_infill {
             /* .vocab = */ vocab,
             /* .buf0  = */ std::vector<char>(512),
             /* .buf1  = */ std::vector<char>(512),
-        },
-    };
+        }
+    );
 }
 
 // utils