cui-llama.rn 1.1.2 → 1.1.4

package/cpp/llama-sampling.cpp

@@ -1,12 +1,56 @@
 #include "llama-sampling.h"
 
+#include "llama-vocab.h"
+#include "llama-grammar.h"
+
+#include <cassert>
 #include <algorithm>
 #include <cstring>
 #include <ctime>
 #include <cfloat>
 #include <numeric>
+#include <random>
 #include <unordered_map>
 
+static int llama_sample_dist(llama_token_data_array * cur_p, std::mt19937 & rng, std::vector<float> & probs) {
+#if 1
+    probs.resize(cur_p->size);
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        probs[i] = cur_p->data[i].p;
+    }
+
+    std::discrete_distribution<size_t> dist(probs.begin(), probs.end());
+#else
+    // avoid the copy with a custom iterator
+    #pragma GCC diagnostic push
+    #pragma GCC diagnostic ignored "-Wunused-local-typedefs"
+
+    struct probs_iterator {
+        typedef std::input_iterator_tag iterator_category;
+        typedef float value_type;
+        typedef float * pointer;
+        typedef float & reference;
+        typedef size_t difference_type;
+
+        const llama_token_data_array * data;
+        size_t i;
+
+        bool operator==(const probs_iterator & other) const { return data + i == other.data + other.i; }
+        bool operator!=(const probs_iterator & other) const { return data + i != other.data + other.i; }
+        float operator*() const { return data->data[i].p; }
+        probs_iterator & operator++() { ++i; return *this; }
+        probs_iterator operator++(int) { probs_iterator tmp = *this; ++i; return tmp; }
+    };
+    #pragma GCC diagnostic pop
+
+    std::discrete_distribution<size_t> dist(probs_iterator{cur_p, 0}, probs_iterator{cur_p, cur_p->size});
+
+    LM_GGML_UNUSED(probs);
+#endif
+
+    return dist(rng);
+}
+
 static void llama_log_softmax(float * array, size_t size) {
     float max_l = *std::max_element(array, array + size);
     float sum = 0.f;
@@ -21,65 +65,50 @@ static void llama_log_softmax(float * array, size_t size) {
     }
 }
 
-void llama_set_rng_seed_impl(struct llama_sampling * smpl, uint32_t seed) {
-    if (seed == LLAMA_DEFAULT_SEED) {
-        seed = time(NULL);
-    }
-
-    smpl->rng.seed(seed);
-}
-
-void llama_sample_softmax_impl(struct llama_sampling * smpl, llama_token_data_array * candidates) {
-    LM_GGML_ASSERT(candidates->size > 0);
-
-    const int64_t t_start_sample_us = lm_ggml_time_us();
+static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) {
+    LM_GGML_ASSERT(cur_p->size > 0);
 
     // Sort the logits in descending order
-    if (!candidates->sorted) {
-        std::sort(candidates->data, candidates->data + candidates->size, [](const llama_token_data & a, const llama_token_data & b) {
+    if (!cur_p->sorted) {
+        std::sort(cur_p->data, cur_p->data + cur_p->size, [](const llama_token_data & a, const llama_token_data & b) {
             return a.logit > b.logit;
         });
-        candidates->sorted = true;
+        cur_p->sorted = true;
     }
 
-    float max_l = candidates->data[0].logit;
+    float max_l = cur_p->data[0].logit;
     float cum_sum = 0.0f;
-    for (size_t i = 0; i < candidates->size; ++i) {
-        float p = expf(candidates->data[i].logit - max_l);
-        candidates->data[i].p = p;
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        float p = expf(cur_p->data[i].logit - max_l);
+        cur_p->data[i].p = p;
         cum_sum += p;
     }
-    for (size_t i = 0; i < candidates->size; ++i) {
-        candidates->data[i].p /= cum_sum;
-    }
 
-    if (smpl) {
-        smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        cur_p->data[i].p /= cum_sum;
     }
 }
 
-void llama_sample_top_k_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, int32_t k, size_t min_keep) {
+static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k) {
     // TODO: move bucket sort to separate function so that top_p/tail_free/typical/softmax first is equally fast
-    // if (k >= (int32_t)candidates->size) {
+    // if (k >= (int32_t)cur_p->size) {
     //     return;
     // }
 
-    const int64_t t_start_sample_us = lm_ggml_time_us();
-
     if (k <= 0) {
-        k = candidates->size;
+        k = cur_p->size;
     }
 
-    k = std::max(k, (int) min_keep);
-    k = std::min(k, (int) candidates->size);
+    k = std::min(k, (int) cur_p->size);
 
     // Sort scores in descending order
-    if (!candidates->sorted) {
+    if (!cur_p->sorted) {
         auto comp = [](const llama_token_data & a, const llama_token_data & b) {
             return a.logit > b.logit;
         };
         if (k <= 128) {
-            std::partial_sort(candidates->data, candidates->data + k, candidates->data + candidates->size, comp);
+            std::partial_sort(cur_p->data, cur_p->data + k, cur_p->data + cur_p->size, comp);
         } else {
             constexpr int   nbuckets     = 128;
             constexpr float bucket_low   = -10.0f;
@@ -87,11 +116,11 @@ void llama_sample_top_k_impl(struct llama_sampling * smpl, llama_token_data_arra
             constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low);
             constexpr float bucket_inter = -bucket_low * bucket_scale;
 
-            std::vector<int> bucket_idx(candidates->size);
+            std::vector<int> bucket_idx(cur_p->size);
             std::vector<int> histo(nbuckets, 0);
 
-            for (int i = 0; i < (int)candidates->size; ++i) {
-                const float val = candidates->data[i].logit;
+            for (int i = 0; i < (int)cur_p->size; ++i) {
+                const float val = cur_p->data[i].logit;
                 int ib = int(bucket_scale * val + bucket_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
                 ib = std::max(0, std::min(nbuckets-1, ib));
                 bucket_idx[i] = ib;
@@ -101,20 +130,22 @@ void llama_sample_top_k_impl(struct llama_sampling * smpl, llama_token_data_arra
             int ib = nbuckets - 1;
             for ( ; ib >= 0; --ib) {
                 nhave += histo[ib];
-                if (nhave >= k) break;
+                if (nhave >= k) {
+                    break;
+                }
             }
             std::vector<llama_token_data> tmp_tokens(nhave);
-            auto ptr = tmp_tokens.data();
+            auto * ptr = tmp_tokens.data();
             std::vector<llama_token_data*> bucket_ptrs;
             bucket_ptrs.reserve(nbuckets - ib);
             for (int j = nbuckets - 1; j >= ib; --j) {
                 bucket_ptrs.push_back(ptr);
                 ptr += histo[j];
             }
-            for (int i = 0; i < (int)candidates->size; ++i) {
+            for (int i = 0; i < (int)cur_p->size; ++i) {
                 int j = bucket_idx[i];
                 if (j >= ib) {
-                    *bucket_ptrs[nbuckets-1-j]++ = candidates->data[i];
+                    *bucket_ptrs[nbuckets-1-j]++ = cur_p->data[i];
                 }
             }
 
@@ -127,69 +158,554 @@ void llama_sample_top_k_impl(struct llama_sampling * smpl, llama_token_data_arra
             }
             std::partial_sort(ptr, ptr + k - ndone, ptr + histo[ib], comp);
 
-            std::memcpy(candidates->data, tmp_tokens.data(), k*sizeof(llama_token_data));
+            std::memcpy(cur_p->data, tmp_tokens.data(), k*sizeof(llama_token_data));
 
         }
-        candidates->sorted = true;
+        cur_p->sorted = true;
     }
-    candidates->size = k;
+    cur_p->size = k;
+}
 
-    if (smpl) {
-        smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
+// llama_sampler API
+
+const char * llama_sampler_name(const struct llama_sampler * smpl) {
+    if (!smpl->iface) {
+        return "(null)";
     }
+
+    return smpl->iface->name(smpl);
 }
 
-void llama_sample_top_p_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float p, size_t min_keep) {
-    if (p >= 1.0f) {
+void llama_sampler_accept(struct llama_sampler * smpl, llama_token token) {
+    if (smpl->iface->accept) {
+        smpl->iface->accept(smpl, token);
+    }
+}
+
+void llama_sampler_apply(struct llama_sampler * smpl, struct llama_token_data_array * cur_p) {
+    LM_GGML_ASSERT(smpl->iface->apply);
+    smpl->iface->apply(smpl, cur_p);
+}
+
+void llama_sampler_reset(struct llama_sampler * smpl) {
+    if (smpl->iface->reset) {
+        smpl->iface->reset(smpl);
+    }
+}
+
+struct llama_sampler * llama_sampler_clone(const struct llama_sampler * smpl) {
+    if (smpl->iface->clone) {
+        return smpl->iface->clone(smpl);
+    }
+
+    if (smpl->ctx == nullptr) {
+        return new llama_sampler {
+            /* .iface = */ smpl->iface,
+            /* .ctx   = */ nullptr,
+        };
+    }
+
+    LM_GGML_ABORT("the sampler does not support cloning");
+}
+
+void llama_sampler_free(struct llama_sampler * smpl) {
+    if (smpl == nullptr) {
         return;
     }
 
-    llama_sample_softmax_impl(smpl, candidates);
+    if (smpl->iface->free) {
+        smpl->iface->free(smpl);
+    }
 
-    const int64_t t_start_sample_us = lm_ggml_time_us();
+    delete smpl;
+}
+
+llama_token llama_sampler_sample(struct llama_sampler * smpl, struct llama_context * ctx, int32_t idx) {
+    const auto * logits = llama_get_logits_ith(ctx, idx);
+
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+
+    // TODO: do not allocate each time
+    std::vector<llama_token_data> cur(n_vocab);
+    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+        cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
+    }
+
+    llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
+
+    llama_sampler_apply(smpl, &cur_p);
+
+    return cur_p.data[cur_p.selected].id;
+}
+
+// sampler chain
+
+static struct llama_sampler_i llama_sampler_chain_i = {
+    /* .name   = */ [](const struct llama_sampler * /*smpl*/) { return "chain"; },
+    /* .accept = */ [](struct llama_sampler * smpl, llama_token token) {
+        auto * chain = (llama_sampler_chain *) smpl->ctx;
+
+        time_meas tm(chain->t_sample_us, chain->params.no_perf);
+
+        for (auto * smpl : chain->samplers) {
+            llama_sampler_accept(smpl, token);
+        }
+
+        chain->n_sample++;
+    },
+    /* .apply  = */ [](struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+        auto * chain = (llama_sampler_chain *) smpl->ctx;
+
+        time_meas tm(chain->t_sample_us, chain->params.no_perf);
+
+        for (auto * smpl : chain->samplers) {
+            llama_sampler_apply(smpl, cur_p);
+        }
+    },
+    /* .reset  = */ [](struct llama_sampler * smpl) {
+        auto * chain = (llama_sampler_chain *) smpl->ctx;
+
+        for (auto * smpl : chain->samplers) {
+            llama_sampler_reset(smpl);
+        }
+
+        chain->t_sample_us = 0;
+        chain->n_sample    = 0;
+    },
+    /* .clone  = */ [](const struct llama_sampler * smpl) {
+        const auto * chain_src = (const llama_sampler_chain *) smpl->ctx;
+
+        auto * result = llama_sampler_chain_init(chain_src->params);
+
+        for (auto * smpl : chain_src->samplers) {
+            llama_sampler_chain_add(result, llama_sampler_clone(smpl));
+        }
+
+        return result;
+    },
+    /* .free   = */ [](struct llama_sampler * smpl) {
+        auto * chain = (llama_sampler_chain *) smpl->ctx;
+
+        for (auto * smpl : chain->samplers) {
+            llama_sampler_free(smpl);
+        }
+
+        delete chain;
+    },
+};
+
+struct llama_sampler * llama_sampler_chain_init(struct llama_sampler_chain_params params) {
+    return new llama_sampler {
+        /* .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) {
+    auto * p = (llama_sampler_chain *) chain->ctx;
+    p->samplers.push_back(smpl);
+}
+
+llama_sampler_timings llama_sampler_chain_timings(struct llama_sampler * chain) {
+    auto * p = (llama_sampler_chain *) chain->ctx;
+    struct llama_sampler_timings result = {
+        p -> t_sample_us,
+        p -> n_sample
+    };
+    return result;
+}
+
+struct llama_sampler * llama_sampler_chain_get(const struct llama_sampler * chain, int32_t i) {
+    const auto * p = (const llama_sampler_chain *) chain->ctx;
+
+    if (i < 0 || i >= (int32_t) p->samplers.size()) {
+        return nullptr;
+    }
+
+    return p->samplers[i];
+}
+
+int llama_sampler_chain_n(const struct llama_sampler * chain) {
+    const auto * p = (const llama_sampler_chain *) chain->ctx;
+
+    return p->samplers.size();
+}
+
+//
+// samplers
+//
+
+// greedy
+
+static const char * llama_sampler_greedy_name(const struct llama_sampler * /*smpl*/) {
+    return "greedy";
+}
+
+static void llama_sampler_greedy_apply(struct llama_sampler * /*smpl*/, llama_token_data_array * cur_p) {
+    cur_p->selected = 0;
+    for (size_t i = 1; i < cur_p->size; ++i) {
+        if (cur_p->data[i].logit > cur_p->data[cur_p->selected].logit) {
+            cur_p->selected = i;
+        }
+    }
+}
+
+static struct llama_sampler_i llama_sampler_greedy_i = {
+    /* .name   = */ llama_sampler_greedy_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_greedy_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ nullptr,
+    /* .free   = */ nullptr,
+};
+
+struct llama_sampler * llama_sampler_init_greedy() {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_greedy_i,
+        /* .ctx   = */ nullptr,
+    };
+}
+
+// dist
+
+struct llama_sampler_dist {
+    const uint32_t seed;
+
+    std::mt19937 rng;
+
+    std::vector<float> probs; // work array
+};
+
+static const char * llama_sampler_dist_name(const struct llama_sampler * /*smpl*/) {
+    return "dist";
+}
+
+static void llama_sampler_dist_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_dist *) smpl->ctx;
+    cur_p->selected = llama_sample_dist(cur_p, ctx->rng, ctx->probs);
+}
+
+static struct llama_sampler * llama_sampler_dist_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_dist *) smpl->ctx;
+    auto * result = llama_sampler_init_dist(ctx->seed);
+
+    // copy the state
+    {
+        auto * result_ctx = (llama_sampler_dist *) result->ctx;
+
+        result_ctx->rng = ctx->rng;
+    }
+
+    return result;
+}
+
+static void llama_sampler_dist_reset(struct llama_sampler * smpl) {
+    auto * ctx = (llama_sampler_dist *) smpl->ctx;
+    ctx->rng = std::mt19937(ctx->seed);
+}
+
+static void llama_sampler_dist_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_dist *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_dist_i = {
+    /* .name   = */ llama_sampler_dist_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_dist_apply,
+    /* .reset  = */ llama_sampler_dist_reset,
+    /* .clone  = */ llama_sampler_dist_clone,
+    /* .free   = */ llama_sampler_dist_free,
+};
+
+struct llama_sampler * llama_sampler_init_dist(uint32_t seed) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_dist_i,
+        /* .ctx   = */ new llama_sampler_dist {
+            /* .seed = */ seed,
+            /* .rng  = */ std::mt19937(seed),
+            /* .probs = */ {},
+        },
+    };
+}
+
+// softmax
+
+static const char * llama_sampler_softmax_name(const struct llama_sampler * /*smpl*/) {
+    return "softmax";
+}
+
+static void llama_sampler_softmax_apply(struct llama_sampler * /*smpl*/, llama_token_data_array * cur_p) {
+    llama_sampler_softmax_impl(cur_p);
+}
+
+static struct llama_sampler_i llama_sampler_softmax_i = {
+    /* .name   = */ llama_sampler_softmax_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_softmax_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ nullptr,
+    /* .free   = */ nullptr,
+};
+
+struct llama_sampler * llama_sampler_init_softmax() {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_softmax_i,
+        /* .ctx   = */ nullptr,
+    };
+}
+
+// top-k
+
+struct llama_sampler_top_k {
+    const int32_t k;
+};
+
+static const char * llama_sampler_top_k_name(const struct llama_sampler * /*smpl*/) {
+    return "top-k";
+}
+
+static void llama_sampler_top_k_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    const auto * ctx = (llama_sampler_top_k *) smpl->ctx;
+    llama_sampler_top_k_impl(cur_p, ctx->k);
+}
+
+static struct llama_sampler * llama_sampler_top_k_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_top_k *) smpl->ctx;
+    return llama_sampler_init_top_k(ctx->k);
+}
+
+static void llama_sampler_top_k_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_top_k *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_top_k_i = {
+    /* .name   = */ llama_sampler_top_k_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_top_k_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_top_k_clone,
+    /* .free   = */ llama_sampler_top_k_free,
+};
+
+struct llama_sampler * llama_sampler_init_top_k(int32_t k) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_top_k_i,
+        /* .ctx   = */ new llama_sampler_top_k {
+            /* .k = */ k,
+        },
+    };
+}
+
+// top-p
+
+struct llama_sampler_top_p {
+    const float  p;
+    const size_t min_keep;
+};
+
+static const char * llama_sampler_top_p_name(const struct llama_sampler * /*smpl*/) {
+    return "top-p";
+}
+
+static void llama_sampler_top_p_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    const auto * ctx = (llama_sampler_top_p *) smpl->ctx;
+
+    if (ctx->p >= 1.0f) {
+        return;
+    }
+
+    llama_sampler_softmax_impl(cur_p);
 
     // Compute the cumulative probabilities
     float cum_sum = 0.0f;
-    size_t last_idx = candidates->size;
+    size_t last_idx = cur_p->size;
 
-    for (size_t i = 0; i < candidates->size; ++i) {
-        cum_sum += candidates->data[i].p;
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        cum_sum += cur_p->data[i].p;
 
         // Check if the running sum is at least p or if we have kept at least min_keep tokens
         // we set the last index to i+1 to indicate that the current iterate should be included in the set
-        if (cum_sum >= p && i + 1 >= min_keep) {
+        if (cum_sum >= ctx->p && i + 1 >= ctx->min_keep) {
             last_idx = i + 1;
             break;
         }
     }
 
     // Resize the output vector to keep only the top-p tokens
-    candidates->size = last_idx;
+    cur_p->size = last_idx;
+}
+
+static struct llama_sampler * llama_sampler_top_p_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_top_p *) smpl->ctx;
+    return llama_sampler_init_top_p(ctx->p, ctx->min_keep);
+}
+
+static void llama_sampler_top_p_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_top_p *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_top_p_i = {
+    /* .name   = */ llama_sampler_top_p_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_top_p_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_top_p_clone,
+    /* .free   = */ llama_sampler_top_p_free,
+};
+
+struct llama_sampler * llama_sampler_init_top_p(float p, size_t min_keep) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_top_p_i,
+        /* .ctx   = */ new llama_sampler_top_p {
+            /* .p        = */ p,
+            /* .min_keep = */ min_keep,
+        },
+    };
+}
+
+// min-p
+
+struct llama_sampler_min_p {
+    const float  p;
+    const size_t min_keep;
+};
+
+static const char * llama_sampler_min_p_name(const struct llama_sampler * /*smpl*/) {
+    return "min-p";
+}
+
+static void llama_sampler_min_p_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    const auto * ctx = (llama_sampler_min_p *) smpl->ctx;
+
+    if (ctx->p <= 0.0f || !cur_p->size) {
+        return;
+    }
+
+    bool min_p_applied = false;
+
+    // if the cur_p aren't sorted, try the unsorted implementation first
+    if (!cur_p->sorted) {
+        std::vector<llama_token_data> filtered_tokens;
+
+        float max_logit = -FLT_MAX;
+        for (size_t i = 0; i < cur_p->size; ++i) {
+            max_logit = std::max(max_logit, cur_p->data[i].logit);
+        }
+        const float min_logit = max_logit + logf(ctx->p); // min logit for p_i >= p * p_max
+
+        for (size_t i = 0; i < cur_p->size; ++i) {
+            if (cur_p->data[i].logit >= min_logit) {
+                filtered_tokens.push_back(cur_p->data[i]);
+            }
+        }
+
+        // if we have enough values the operation was a success
+        if (filtered_tokens.size() >= ctx->min_keep) {
+            memcpy(cur_p->data, filtered_tokens.data(), filtered_tokens.size()*sizeof(llama_token_data));
+            cur_p->size = filtered_tokens.size();
+            min_p_applied = true;
+        }
+    }
+
+    // if the cur_p are sorted or the unsorted implementation failed, use this implementation
+    if (!min_p_applied) {
+        // Sort the logits in descending order
+        if (!cur_p->sorted) {
+            std::sort(cur_p->data, cur_p->data + cur_p->size, [](const llama_token_data & a, const llama_token_data & b) {
+                return a.logit > b.logit;
+            });
+            cur_p->sorted = true;
+        }
+
+        const float min_logit = cur_p->data[0].logit + logf(ctx->p); // min logit for p_i >= p * p_max
+        size_t i = 1; // first token always matches
+
+        for (; i < cur_p->size; ++i) {
+            if (cur_p->data[i].logit < min_logit && i >= ctx->min_keep) {
+                break; // prob too small
+            }
+        }
 
-    if (smpl) {
-        smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
+        // Resize the output vector to keep only the matching tokens
+        cur_p->size = i;
     }
 }
 
-void llama_sample_xtc_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float xtc_threshold, float xtc_probability, size_t min_keep,  std::mt19937 & rng) {
-    if(xtc_threshold <= 0.0f || !candidates-> size) {
+static struct llama_sampler * llama_sampler_min_p_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_min_p *) smpl->ctx;
+    return llama_sampler_init_min_p(ctx->p, ctx->min_keep);
+}
+
+static void llama_sampler_min_p_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_min_p *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_min_p_i = {
+    /* .name   = */ llama_sampler_min_p_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_min_p_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_min_p_clone,
+    /* .free   = */ llama_sampler_min_p_free,
+};
+
+struct llama_sampler * llama_sampler_init_min_p(float p, size_t min_keep) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_min_p_i,
+        /* .ctx   = */ new llama_sampler_min_p {
+            /* .p        = */ p,
+            /* .min_keep = */ min_keep,
+        },
+    };
+}
+
+// xtc
+
+struct llama_sampler_xtc {
+    const uint32_t seed;
+    std::mt19937 rng;
+    const float xtc_p;
+    const float xtc_t;
+    const size_t min_keep;
+};
+
+static const char * llama_sampler_xtc_name(const struct llama_sampler * /* smpl */) {
+    return "xtc";
+}
+
+static void llama_sampler_xtc_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_xtc *) smpl->ctx;
+
+    size_t min_keep = ctx -> min_keep;
+    std::mt19937 rng = ctx -> rng;
+
+    float xtc_threshold = ctx -> xtc_t;
+    float xtc_probability = ctx -> xtc_p;
+   
+
+    if(xtc_threshold <= 0.0f || !cur_p-> size) {
         return;
     }
     
     bool xtc_applied = false;
     const int64_t t_start_sample_us = lm_ggml_time_us();
-    llama_sample_softmax(nullptr, candidates);
+    llama_sampler_softmax_impl(cur_p);
     
     // unsorted iteration
-    if (!candidates->sorted) {
+    if (!cur_p->sorted) {
         std::vector<llama_token_data> top_tokens, low_tokens;
 
         // split candidates into two arrays for low and high tokens
-        for (size_t i = 0; i < candidates->size; ++i) {
-            if (candidates->data[i].logit >= xtc_threshold) {
-                top_tokens.push_back(candidates->data[i]);
+        for (size_t i = 0; i < cur_p->size; ++i) {
+            if (cur_p->data[i].logit >= xtc_threshold) {
+                top_tokens.push_back(cur_p->data[i]);
             } else {
-                low_tokens.push_back(candidates-> data[i]);
+                low_tokens.push_back(cur_p-> data[i]);
             }
         }
         // if there is only one or no top_tokens, do not truncate
@@ -212,28 +728,28 @@ void llama_sample_xtc_impl(struct llama_sampling * smpl, llama_token_data_array
             }
         }
         if(low_tokens.size() >= min_keep) {
-           memcpy(candidates->data, low_tokens.data(), low_tokens.size()*sizeof(llama_token_data));
-            candidates->size = low_tokens.size();
+           memcpy(cur_p->data, low_tokens.data(), low_tokens.size()*sizeof(llama_token_data));
+            cur_p->size = low_tokens.size();
             xtc_applied = true;
         }
     }
     // sorted iteration
-    
+
     if (!xtc_applied) {
         // Sort the logits in descending order
-        if (!candidates->sorted) {
-            std::sort(candidates->data, candidates->data + candidates->size, [](const llama_token_data & a, const llama_token_data & b) {
+        if (!cur_p->sorted) {
+            std::sort(cur_p->data, cur_p->data + cur_p->size, [](const llama_token_data & a, const llama_token_data & b) {
                 return a.logit > b.logit;
             });
-            candidates->sorted = true;
+            cur_p->sorted = true;
         }
 
         // find last token over threshold
 
         size_t last_index = 0;
 
-        for (; last_index < candidates -> size; ++last_index) {
-            if(candidates -> data[last_index].p < xtc_threshold) {
+        for (; last_index < cur_p -> size; ++last_index) {
+            if(cur_p -> data[last_index].p < xtc_threshold) {
                 break;
             }
         }
@@ -244,102 +760,80 @@ void llama_sample_xtc_impl(struct llama_sampling * smpl, llama_token_data_array
         }
         last_index--;
         // items beyond safe index will be ignored
-        size_t safe_index = candidates -> size;
+        size_t safe_index = cur_p -> size;
 
         // remove tokens until last threshold item
         std::uniform_real_distribution<float> random_float(0.0 , 1.0);
         for (size_t i = 0; i < last_index; i++) {
             if(random_float(rng) < xtc_probability) {
-                std::swap(candidates-> data[i], candidates->data[safe_index - 1]);
+                std::swap(cur_p-> data[i], cur_p->data[safe_index - 1]);
                 safe_index--;
-                if (candidates-> sorted) {
-                    candidates -> sorted = false;
+                if (cur_p-> sorted) {
+                    cur_p -> sorted = false;
                 }
             }
         }
-        candidates -> size = safe_index;
-    }
-
-    if (smpl) {
-        smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
+        cur_p -> size = safe_index;
     }
 }
 
-void llama_sample_min_p_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float p, size_t min_keep) {
-    if (p <= 0.0f || !candidates->size) {
-        return;
-    }
-
-    const int64_t t_start_sample_us = lm_ggml_time_us();
-
-    bool min_p_applied = false;
-
-    // if the candidates aren't sorted, try the unsorted implementation first
-    if (!candidates->sorted) {
-        std::vector<llama_token_data> filtered_tokens;
-
-        float max_logit = -FLT_MAX;
-        for (size_t i = 0; i < candidates->size; ++i) {
-            max_logit = std::max(max_logit, candidates->data[i].logit);
-        }
-        const float min_logit = max_logit + logf(p); // min logit for p_i >= p * p_max
-
-        for (size_t i = 0; i < candidates->size; ++i) {
-            if (candidates->data[i].logit >= min_logit) {
-                filtered_tokens.push_back(candidates->data[i]);
-            }
-        }
-
-        // if we have enough values the operation was a success
-        if (filtered_tokens.size() >= min_keep) {
-            memcpy(candidates->data, filtered_tokens.data(), filtered_tokens.size()*sizeof(llama_token_data));
-            candidates->size = filtered_tokens.size();
-            min_p_applied = true;
-        }
-    }
+static struct llama_sampler * llama_sampler_xtc_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_xtc *) smpl->ctx;
+    return llama_sampler_init_xtc(ctx->xtc_p, ctx->xtc_t, ctx->min_keep, ctx->seed);
+}
 
-    // if the candidates are sorted or the unsorted implementation failed, use this implementation
-    if (!min_p_applied) {
-        // Sort the logits in descending order
-        if (!candidates->sorted) {
-            std::sort(candidates->data, candidates->data + candidates->size, [](const llama_token_data & a, const llama_token_data & b) {
-                return a.logit > b.logit;
-            });
-            candidates->sorted = true;
-        }
+static void llama_sampler_xtc_free(struct llama_sampler * smpl) {
+    delete (const llama_sampler_xtc *) smpl->ctx;
+}
 
-        const float min_logit = candidates->data[0].logit + logf(p); // min logit for p_i >= p * p_max
-        size_t i = 1; // first token always matches
+static struct llama_sampler_i llama_sampler_xtc_i = {
+    /* .name   = */ llama_sampler_xtc_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_xtc_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_xtc_clone,
+    /* .free   = */ llama_sampler_xtc_free,
+};
+
+struct llama_sampler * llama_sampler_init_xtc(float xtc_p, float xtc_t, size_t min_keep, uint32_t seed) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_xtc_i,
+        /* .ctx   = */ new llama_sampler_xtc {
+            /* .seed        = */ seed,
+            /* .rng         = */ std::mt19937(seed),
+            /* .xtc_p       = */ xtc_p,
+            /* .xtc_t       = */ xtc_t,
+            /* .min_keep    = */ min_keep
+        },
+    };
+}
 
-        for (; i < candidates->size; ++i) {
-            if (candidates->data[i].logit < min_logit && i >= min_keep) {
-                break; // prob too small
-            }
-        }
+// tail-free
 
-        // Resize the output vector to keep only the matching tokens
-        candidates->size = i;
-    }
+struct llama_sampler_tail_free {
+    const float  z;
+    const size_t min_keep;
+};
 
-    if (smpl) {
-        smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
-    }
+static const char * llama_sampler_tail_free_name(const struct llama_sampler * /*smpl*/) {
+    return "tail-free";
 }
 
-void llama_sample_tail_free_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float z, size_t min_keep) {
-    if (z >= 1.0f || candidates->size <= 2) {
+static void llama_sampler_tail_free_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    const auto * ctx = (llama_sampler_tail_free *) smpl->ctx;
+
+    if (ctx->z >= 1.0f || cur_p->size <= 2) {
         return;
     }
 
-    llama_sample_softmax_impl((struct llama_sampling *) nullptr, candidates);
-    const int64_t t_start_sample_us = lm_ggml_time_us();
+    llama_sampler_softmax_impl(cur_p);
 
     // Compute the first and second derivatives
-    std::vector<float> first_derivatives(candidates->size - 1);
-    std::vector<float> second_derivatives(candidates->size - 2);
+    std::vector<float> first_derivatives(cur_p->size - 1);
+    std::vector<float> second_derivatives(cur_p->size - 2);
 
     for (size_t i = 0; i < first_derivatives.size(); ++i) {
-        first_derivatives[i] = candidates->data[i].p - candidates->data[i + 1].p;
+        first_derivatives[i] = cur_p->data[i].p - cur_p->data[i + 1].p;
     }
     for (size_t i = 0; i < second_derivatives.size(); ++i) {
         second_derivatives[i] = first_derivatives[i] - first_derivatives[i + 1];
@@ -366,51 +860,86 @@ void llama_sample_tail_free_impl(struct llama_sampling * smpl, llama_token_data_
     }
 
     float cum_sum = 0.0f;
-    size_t last_idx = candidates->size;
+    size_t last_idx = cur_p->size;
     for (size_t i = 0; i < second_derivatives.size(); ++i) {
         cum_sum += second_derivatives[i];
 
         // Check if the running sum is greater than z or if we have kept at least min_keep tokens
-        if (cum_sum > z && i >= min_keep) {
+        if (cum_sum > ctx->z && i >= ctx->min_keep) {
             last_idx = i;
             break;
         }
     }
 
     // Resize the output vector to keep only the tokens above the tail location
-    candidates->size = last_idx;
+    cur_p->size = last_idx;
+}
 
-    if (smpl) {
-        smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
-    }
+static struct llama_sampler * llama_sampler_tail_free_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_tail_free *) smpl->ctx;
+    return llama_sampler_init_tail_free(ctx->z, ctx->min_keep);
+}
+
+static void llama_sampler_tail_free_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_tail_free *) smpl->ctx;
 }
 
-void llama_sample_typical_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float p, size_t min_keep) {
+static struct llama_sampler_i llama_sampler_tail_free_i = {
+    /* .name   = */ llama_sampler_tail_free_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_tail_free_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_tail_free_clone,
+    /* .free   = */ llama_sampler_tail_free_free,
+};
+
+struct llama_sampler * llama_sampler_init_tail_free(float z, size_t min_keep) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_tail_free_i,
+        /* .ctx   = */ new llama_sampler_tail_free {
+            /* .z        = */ z,
+            /*. min_keep = */ min_keep,
+        },
+    };
+}
+
+// typical
+
+struct llama_sampler_typical {
+    const float  p;
+    const size_t min_keep;
+};
+
+static const char * llama_sampler_typical_name(const struct llama_sampler * /*smpl*/) {
+    return "typical";
+}
+
+static void llama_sampler_typical_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    const auto * ctx = (llama_sampler_typical *) smpl->ctx;
+
     // Reference implementation:
     // https://github.com/huggingface/transformers/compare/main...cimeister:typical-sampling:typical-pr
-    if (p >= 1.0f) {
+    if (ctx->p >= 1.0f) {
         return;
     }
 
     // Compute the softmax of logits and calculate entropy
-    llama_sample_softmax_impl((struct llama_sampling *) nullptr, candidates);
-
-    const int64_t t_start_sample_us = lm_ggml_time_us();
+    llama_sampler_softmax_impl(cur_p);
 
     float entropy = 0.0f;
-    for (size_t i = 0; i < candidates->size; ++i) {
-        entropy += -candidates->data[i].p * logf(candidates->data[i].p);
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        entropy += -cur_p->data[i].p * logf(cur_p->data[i].p);
     }
 
     // Compute the absolute difference between negative log probability and entropy for each candidate
     std::vector<float> shifted_scores;
-    for (size_t i = 0; i < candidates->size; ++i) {
-        float shifted_score = fabsf(-logf(candidates->data[i].p) - entropy);
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        float shifted_score = fabsf(-logf(cur_p->data[i].p) - entropy);
         shifted_scores.push_back(shifted_score);
     }
 
     // Sort tokens based on the shifted_scores and their corresponding indices
-    std::vector<size_t> indices(candidates->size);
+    std::vector<size_t> indices(cur_p->size);
     std::iota(indices.begin(), indices.end(), 0);
 
     std::sort(indices.begin(), indices.end(), [&](size_t a, size_t b) {
@@ -423,197 +952,250 @@ void llama_sample_typical_impl(struct llama_sampling * smpl, llama_token_data_ar
 
     for (size_t i = 0; i < indices.size(); ++i) {
         size_t idx = indices[i];
-        cum_sum += candidates->data[idx].p;
+        cum_sum += cur_p->data[idx].p;
 
         // Check if the running sum is greater than typical or if we have kept at least min_keep tokens
-        if (cum_sum > p && i >= min_keep - 1) {
+        if (cum_sum > ctx->p && i >= ctx->min_keep - 1) {
             last_idx = i + 1;
             break;
         }
     }
 
     // Resize the output vector to keep only the locally typical tokens
-    std::vector<llama_token_data> new_candidates;
+    std::vector<llama_token_data> cur_p_new;
     for (size_t i = 0; i < last_idx; ++i) {
         size_t idx = indices[i];
-        new_candidates.push_back(candidates->data[idx]);
+        cur_p_new.push_back(cur_p->data[idx]);
     }
 
-    // Replace the data in candidates with the new_candidates data
-    std::copy(new_candidates.begin(), new_candidates.end(), candidates->data);
-    candidates->size = new_candidates.size();
-    candidates->sorted = false;
+    // Replace the data in cur_p with the cur_p_new data
+    std::copy(cur_p_new.begin(), cur_p_new.end(), cur_p->data);
+    cur_p->size = cur_p_new.size();
+    cur_p->sorted = false;
+}
 
-    if (smpl) {
-        smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
-    }
+static struct llama_sampler * llama_sampler_typical_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_typical *) smpl->ctx;
+    return llama_sampler_init_typical(ctx->p, ctx->min_keep);
 }
 
-void llama_sample_entropy_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float min_temp, float max_temp, float exponent_val) {
-    const int64_t t_start_sample_us = lm_ggml_time_us();
+static void llama_sampler_typical_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_typical *) smpl->ctx;
+}
 
-    // no need to do anything if there is only one (or zero) candidates
-    if(candidates->size <= 1) {
-        return;
-    }
+static struct llama_sampler_i llama_sampler_typical_i = {
+    /* .name   = */ llama_sampler_typical_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_typical_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_typical_clone,
+    /* .free   = */ llama_sampler_typical_free,
+};
+
+struct llama_sampler * llama_sampler_init_typical(float p, size_t min_keep) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_typical_i,
+        /* .ctx   = */ new llama_sampler_typical {
+            /* .p        = */ p,
+            /* .min_keep = */ min_keep,
+        },
+    };
+}
 
-    // Calculate maximum possible entropy
-    float max_entropy = -logf(1.0f / candidates->size);
+// temp
 
-    llama_sample_softmax_impl((struct llama_sampling *) nullptr, candidates);
+struct llama_sampler_temp {
+    const float temp;
+};
 
-    // Calculate entropy of the softmax probabilities
-    float entropy = 0.0f;
-    for (size_t i = 0; i < candidates->size; ++i) {
-        float prob = candidates->data[i].p;
-        if (prob > 0.0f) { // Ensure no log(0)
-            entropy -= prob * logf(prob);
-        }
-    }
+static const char * llama_sampler_temp_name(const struct llama_sampler * /*smpl*/) {
+    return "temp";
+}
 
-    // Normalize the entropy (max_entropy cannot be 0 here because we checked candidates->size != 1 above)
-    float normalized_entropy = entropy / max_entropy;
+static void llama_sampler_temp_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    const auto * ctx = (llama_sampler_temp *) smpl->ctx;
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        cur_p->data[i].logit /= ctx->temp;
+    }
+}
 
-    // Map the normalized entropy to the desired temperature range using the power function
-    float dyn_temp = min_temp + (max_temp - min_temp) * powf(normalized_entropy, exponent_val);
+static struct llama_sampler * llama_sampler_temp_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_temp *) smpl->ctx;
+    return llama_sampler_init_temp(ctx->temp);
+}
 
-#ifdef DEBUG
-    LLAMA_LOG_INFO("Your text maxtemp value is: %f\n", max_temp);
-    LLAMA_LOG_INFO("Entropy: %f\n", entropy);
-    LLAMA_LOG_INFO("Max Possible Entropy: %f\n", max_entropy);
-    LLAMA_LOG_INFO("Normalized Entropy: %f\n", normalized_entropy);
-    LLAMA_LOG_INFO("Exponent: %f\n", exponent_val);
-    LLAMA_LOG_INFO("Dynamic Temperature (dyn_temp): %f\n", dyn_temp);
-#endif
+static void llama_sampler_temp_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_temp *) smpl->ctx;
+}
 
-    // Apply the dynamically calculated temperature scaling
-    for (size_t i = 0; i < candidates->size; ++i) {
-        candidates->data[i].logit /= dyn_temp;
-    }
+static struct llama_sampler_i llama_sampler_temp_i = {
+    /* .name   = */ llama_sampler_temp_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_temp_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_temp_clone,
+    /* .free   = */ llama_sampler_temp_free,
+};
+
+struct llama_sampler * llama_sampler_init_temp(float temp) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_temp_i,
+        /* .ctx   = */ new llama_sampler_temp {
+            /*.temp = */ temp,
+        },
+    };
+}
 
-    // Re-compute softmax probabilities after scaling logits with dynamic temperature
-    double max_l_double = candidates->data[0].logit;
-    double cum_sum_double = 0.0;
-    for (size_t i = 0; i < candidates->size; ++i) {
-        double p = exp(candidates->data[i].logit - max_l_double);
-        candidates->data[i].p = p; // Store the scaled probability
-        cum_sum_double += p;
-    }
-    for (size_t i = 0; i < candidates->size; ++i) {
-        candidates->data[i].p /= cum_sum_double; // Re-normalize the probabilities
-    }
+// temp-ext
 
-#ifdef DEBUG
-    // Print the updated top 25 probabilities after temperature scaling
-    LLAMA_LOG_INFO("\nUpdated Top 25 Probabilities After Dynamic Temperature Scaling (in percentages):\n");
-    for (size_t i = 0; i < 25 && i < candidates->size; ++i) {
-        LLAMA_LOG_INFO("Token %zu: %f%%\n", i + 1, candidates->data[i].p * 100.0f);
-    }
-#endif
+struct llama_sampler_temp_ext {
+    const float temp;
+    const float delta;
+    const float exponent;
+};
 
-    if (smpl) {
-        smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
-    }
+static const char * llama_sampler_temp_ext_name(const struct llama_sampler * /*smpl*/) {
+    return "temp-ext";
 }
 
-void llama_sample_temp_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float temp) {
-    const int64_t t_start_sample_us = lm_ggml_time_us();
+static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    const auto * ctx = (llama_sampler_temp_ext *) smpl->ctx;
+    if (ctx->delta > 0) {
+        const float min_temp = std::max(0.0f, ctx->temp - ctx->delta);
+        const float max_temp = ctx->temp + ctx->delta;
+        float exponent_val = ctx->exponent;
 
-    for (size_t i = 0; i < candidates->size; ++i) {
-        candidates->data[i].logit /= temp;
-    }
+        // no need to do anything if there is only one (or zero) candidates
+        if (cur_p->size <= 1) {
+            return;
+        }
 
-    if (smpl) {
-        smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
-    }
-}
+        // Calculate maximum possible entropy
+        float max_entropy = -logf(1.0f / cur_p->size);
 
-void llama_sample_repetition_penalties_impl(
-        struct llama_sampling * smpl,
-       llama_token_data_array * candidates,
-            const llama_token * last_tokens,
-                       size_t   penalty_last_n,
-                       float   penalty_repeat,
-                       float   penalty_freq,
-                       float   penalty_present) {
-    if (penalty_last_n == 0 || (penalty_repeat == 1.0f && penalty_freq == 0.0f && penalty_present == 0.0f)) {
-        return;
-    }
+        llama_sampler_softmax_impl(cur_p);
 
-    const int64_t t_start_sample_us = lm_ggml_time_us();
+        // Calculate entropy of the softmax probabilities
+        float entropy = 0.0f;
+        for (size_t i = 0; i < cur_p->size; ++i) {
+            float prob = cur_p->data[i].p;
+            if (prob > 0.0f) { // Ensure no log(0)
+                entropy -= prob * logf(prob);
+            }
+        }
 
-    // Create a frequency map to count occurrences of each token in last_tokens
-    std::unordered_map<llama_token, int> token_count;
-    for (size_t i = 0; i < penalty_last_n; ++i) {
-        token_count[last_tokens[i]]++;
-    }
+        // Normalize the entropy (max_entropy cannot be 0 here because we checked cur_p->size != 1 above)
+        float normalized_entropy = entropy / max_entropy;
 
-    // Apply frequency and presence penalties to the candidates
-    for (size_t i = 0; i < candidates->size; ++i) {
-        const auto token_iter = token_count.find(candidates->data[i].id);
-        if (token_iter == token_count.end()) {
-            continue;
+        // Map the normalized entropy to the desired temperature range using the power function
+        float dyn_temp = min_temp + (max_temp - min_temp) * powf(normalized_entropy, exponent_val);
+
+    #ifdef DEBUG
+        LLAMA_LOG_INFO("Your text maxtemp value is: %f\n", max_temp);
+        LLAMA_LOG_INFO("Entropy: %f\n", entropy);
+        LLAMA_LOG_INFO("Max Possible Entropy: %f\n", max_entropy);
+        LLAMA_LOG_INFO("Normalized Entropy: %f\n", normalized_entropy);
+        LLAMA_LOG_INFO("Exponent: %f\n", exponent_val);
+        LLAMA_LOG_INFO("Dynamic Temperature (dyn_temp): %f\n", dyn_temp);
+    #endif
+
+        // Apply the dynamically calculated temperature scaling
+        for (size_t i = 0; i < cur_p->size; ++i) {
+            cur_p->data[i].logit /= dyn_temp;
         }
 
-        const int count = token_iter->second;
+        // Re-compute softmax probabilities after scaling logits with dynamic temperature
+        const double max_l_double = cur_p->data[0].logit;
 
-        // The academic publication that described this technique actually just only divided, but that would cause tokens with negative logits to become more likely, which is obviously wrong.
-        // This is common fix for this problem, which is to multiply by the penalty instead of dividing.
-        if (candidates->data[i].logit <= 0) {
-            candidates->data[i].logit *= penalty_repeat;
-        } else {
-            candidates->data[i].logit /= penalty_repeat;
+        double cum_sum_double = 0.0;
+        for (size_t i = 0; i < cur_p->size; ++i) {
+            double p = exp(cur_p->data[i].logit - max_l_double);
+            cur_p->data[i].p = p; // Store the scaled probability
+            cum_sum_double += p;
+        }
+
+        for (size_t i = 0; i < cur_p->size; ++i) {
+            cur_p->data[i].p /= cum_sum_double; // Re-normalize the probabilities
         }
 
-        candidates->data[i].logit -= float(count) * penalty_freq + float(count > 0) * penalty_present;
+    #ifdef DEBUG
+        // Print the updated top 25 probabilities after temperature scaling
+        LLAMA_LOG_INFO("\nUpdated Top 25 Probabilities After Dynamic Temperature Scaling (in percentages):\n");
+        for (size_t i = 0; i < 25 && i < cur_p->size; ++i) {
+            LLAMA_LOG_INFO("Token %zu: %f%%\n", i + 1, cur_p->data[i].p * 100.0f);
+        }
+    #endif
+    } else {
+        for (size_t i = 0; i < cur_p->size; ++i) {
+            cur_p->data[i].logit /= ctx->temp;
+        }
     }
+}
 
-    candidates->sorted = false;
+static struct llama_sampler * llama_sampler_temp_ext_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_temp_ext *) smpl->ctx;
+    return llama_sampler_init_temp_ext(ctx->temp, ctx->delta, ctx->exponent);
+}
 
-    if (smpl) {
-        smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
-    }
+static void llama_sampler_temp_ext_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_temp_ext *) smpl->ctx;
 }
 
-void llama_sample_apply_guidance_impl(
-        struct llama_sampling * smpl,
-                        float * logits,
-                        float * logits_guidance,
-                        float   scale) {
-    LM_GGML_ASSERT(smpl);
+static struct llama_sampler_i llama_sampler_temp_ext_i = {
+    /* .name   = */ llama_sampler_temp_ext_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_temp_ext_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_temp_ext_clone,
+    /* .free   = */ llama_sampler_temp_ext_free,
+};
+
+struct llama_sampler * llama_sampler_init_temp_ext(float temp, float delta, float exponent) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_temp_ext_i,
+        /* .ctx   = */ new llama_sampler_temp_ext {
+            /* .temp     = */ temp,
+            /* .delta    = */ delta,
+            /* .exponent = */ exponent,
+        },
+    };
+}
 
-    const auto t_start_sample_us = lm_ggml_time_us();
-    const auto n_vocab = smpl->n_vocab;
+// mirostat
 
-    llama_log_softmax(logits, n_vocab);
-    llama_log_softmax(logits_guidance, n_vocab);
+struct llama_sampler_mirostat {
+    const int32_t n_vocab;
 
-    for (int i = 0; i < n_vocab; ++i) {
-              auto & l = logits[i];
-        const auto & g = logits_guidance[i];
+    const uint32_t seed;
 
-        l = scale * (l - g) + g;
-    }
+    const float tau;
+    const float eta;
 
-    smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
-}
+    const int32_t m;
 
-llama_token llama_sample_token_mirostat_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float tau, float eta, int32_t m, float * mu) {
-    LM_GGML_ASSERT(smpl);
+    float mu;
 
-    const int32_t n_vocab = float(smpl->n_vocab);
+    std::mt19937 rng;
 
-    int64_t t_start_sample_us = lm_ggml_time_us();
+    std::vector<float> probs;
+};
+
+static const char * llama_sampler_mirostat_name(const struct llama_sampler * /*smpl*/) {
+    return "mirostat";
+}
 
-    llama_sample_softmax_impl((struct llama_sampling *) nullptr, candidates);
+static void llama_sampler_mirostat_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_mirostat *) smpl->ctx;
+
+    llama_sampler_softmax_impl(cur_p);
 
     // Estimate s_hat using the most probable m tokens
     float s_hat = 0.0;
     float sum_ti_bi = 0.0;
     float sum_ti_sq = 0.0;
-    for (size_t i = 0; i < size_t(m - 1) && i < candidates->size - 1; ++i) {
+    for (size_t i = 0; i < size_t(ctx->m - 1) && i < cur_p->size - 1; ++i) {
         float t_i = logf(float(i + 2) / float(i + 1));
-        float b_i = logf(candidates->data[i].p / candidates->data[i + 1].p);
+        float b_i = logf(cur_p->data[i].p / cur_p->data[i + 1].p);
         sum_ti_bi += t_i * b_i;
         sum_ti_sq += t_i * t_i;
     }
@@ -621,109 +1203,532 @@ llama_token llama_sample_token_mirostat_impl(struct llama_sampling * smpl, llama
 
     // Compute k from the estimated s_hat and target surprise value
     float epsilon_hat = s_hat - 1;
-    float k = powf((epsilon_hat * powf(2, *mu)) / (1 - powf(n_vocab, -epsilon_hat)), 1 / s_hat);
+    float k = powf((epsilon_hat * powf(2, ctx->mu)) / (1 - powf(ctx->n_vocab, -epsilon_hat)), 1 / s_hat);
 
-    // Sample the next word X using top-k sampling
-    llama_sample_top_k_impl((struct llama_sampling *) nullptr, candidates, int(k), 1);
-    smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
-    llama_token X = llama_sample_token_impl(smpl, candidates);
-    t_start_sample_us = lm_ggml_time_us();
+    llama_sampler_top_k_impl(cur_p, std::max(int(k), 1));
+    llama_sampler_softmax_impl(cur_p);
 
-    // Compute error as the difference between observed surprise and target surprise value
-    size_t X_idx = std::distance(candidates->data, std::find_if(candidates->data, candidates->data + candidates->size, [&](const llama_token_data & candidate) {
-        return candidate.id == X;
-    }));
-    float observed_surprise = -log2f(candidates->data[X_idx].p);
-    float e = observed_surprise - tau;
+    const int idx = llama_sample_dist(cur_p, ctx->rng, ctx->probs);
+
+    cur_p->selected = idx;
+
+    float observed_surprise = -log2f(cur_p->data[idx].p);
+    float e = observed_surprise - ctx->tau;
 
     // Update mu using the learning rate and error
-    *mu = *mu - eta * e;
+    ctx->mu = ctx->mu - ctx->eta * e;
+}
+
+static struct llama_sampler * llama_sampler_mirostat_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_mirostat *) smpl->ctx;
+    auto * result = llama_sampler_init_mirostat(ctx->n_vocab, ctx->seed, ctx->tau, ctx->eta, ctx->m);
+
+    // copy the state
+    {
+        auto * result_ctx = (llama_sampler_mirostat *) smpl->ctx;
+
+        result_ctx->mu  = ctx->mu;
+        result_ctx->rng = ctx->rng;
+    }
+
+    return result;
+}
+
+static void llama_sampler_mirostat_reset(struct llama_sampler * smpl) {
+    auto * ctx = (llama_sampler_mirostat *) smpl->ctx;
+    ctx->mu = 2.0f*ctx->tau;
+    ctx->rng = std::mt19937(ctx->seed);
+}
+
+static void llama_sampler_mirostat_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_mirostat *) smpl->ctx;
+}
 
-    smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
-    return X;
+static struct llama_sampler_i llama_sampler_mirostat_i = {
+    /* .name   = */ llama_sampler_mirostat_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_mirostat_apply,
+    /* .reset  = */ llama_sampler_mirostat_reset,
+    /* .clone  = */ llama_sampler_mirostat_clone,
+    /* .free   = */ llama_sampler_mirostat_free,
+};
+
+struct llama_sampler * llama_sampler_init_mirostat(int32_t n_vocab, uint32_t seed, float tau, float eta, int32_t m) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_mirostat_i,
+        /* .ctx   = */ new llama_sampler_mirostat {
+            /* .n_vocab = */ n_vocab,
+            /* .seed    = */ seed,
+            /* .tau     = */ tau,
+            /* .eta     = */ eta,
+            /* .m       = */ m,
+            /* .mu      = */ 2.0f*tau,
+            /* .rng     = */ std::mt19937(seed),
+            /* .probs   = */ {},
+        },
+    };
 }
 
-llama_token llama_sample_token_mirostat_v2_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float tau, float eta, float * mu) {
-    int64_t t_start_sample_us;
-    t_start_sample_us = lm_ggml_time_us();
+// mirostat v2
+
+struct llama_sampler_mirostat_v2 {
+    const uint32_t seed;
 
-    llama_sample_softmax_impl(smpl, candidates);
+    const float tau;
+    const float eta;
+
+    float mu;
+
+    std::mt19937 rng;
+
+    std::vector<float> probs;
+};
+
+static const char * llama_sampler_mirostat_v2_name(const struct llama_sampler * /*smpl*/) {
+    return "mirostat-v2";
+}
+
+static void llama_sampler_mirostat_v2_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_mirostat_v2 *) smpl->ctx;
+
+    llama_sampler_softmax_impl(cur_p);
 
     // Truncate the words with surprise values greater than mu
-    candidates->size = std::distance(candidates->data, std::find_if(candidates->data, candidates->data + candidates->size, [&](const llama_token_data & candidate) {
-        return -log2f(candidate.p) > *mu;
+    cur_p->size = std::distance(cur_p->data, std::find_if(cur_p->data, cur_p->data + cur_p->size, [&](const llama_token_data & candidate) {
+        return -log2f(candidate.p) > ctx->mu;
     }));
 
-    if (candidates->size == 0) {
-        candidates->size = 1;
-    }
-
-    if (smpl) {
-        smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
+    if (cur_p->size == 0) {
+        cur_p->size = 1;
     }
 
     // Normalize the probabilities of the remaining words
-    llama_sample_softmax_impl(smpl, candidates);
+    llama_sampler_softmax_impl(cur_p);
 
-    // Sample the next word X from the remaining words
-    llama_token X = llama_sample_token_impl(smpl, candidates);
-    t_start_sample_us = lm_ggml_time_us();
+    const int idx = llama_sample_dist(cur_p, ctx->rng, ctx->probs);
 
-    // Compute error as the difference between observed surprise and target surprise value
-    size_t X_idx = std::distance(candidates->data, std::find_if(candidates->data, candidates->data + candidates->size, [&](const llama_token_data & candidate) {
-        return candidate.id == X;
-    }));
-    float observed_surprise = -log2f(candidates->data[X_idx].p);
-    float e = observed_surprise - tau;
+    cur_p->selected = idx;
+
+    float observed_surprise = -log2f(cur_p->data[idx].p);
+    float e = observed_surprise - ctx->tau;
 
     // Update mu using the learning rate and error
-    *mu = *mu - eta * e;
+    ctx->mu = ctx->mu - ctx->eta * e;
+}
+
+static void llama_sampler_mirostat_v2_reset(struct llama_sampler * smpl) {
+    auto * ctx = (llama_sampler_mirostat_v2 *) smpl->ctx;
+    ctx->mu = 2.0f*ctx->tau;
+    ctx->rng = std::mt19937(ctx->seed);
+}
+
+static struct llama_sampler * llama_sampler_mirostat_v2_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_mirostat_v2 *) smpl->ctx;
+
+    auto * result = llama_sampler_init_mirostat_v2(ctx->seed, ctx->tau, ctx->eta);
+
+    // copy the state
+    {
+        auto * result_ctx = (llama_sampler_mirostat_v2 *) result->ctx;
 
-    if (smpl) {
-        smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
+        result_ctx->mu  = ctx->mu;
+        result_ctx->rng = ctx->rng;
     }
-    return X;
+
+    return result;
 }
 
-llama_token llama_sample_token_greedy_impl(struct llama_sampling * smpl, llama_token_data_array * candidates) {
-    const int64_t t_start_sample_us = lm_ggml_time_us();
+static void llama_sampler_mirostat_v2_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_mirostat_v2 *) smpl->ctx;
+}
 
-    // Find max element
-    auto * max_iter = std::max_element(candidates->data, candidates->data + candidates->size, [](const llama_token_data & a, const llama_token_data & b) {
-        return a.logit < b.logit;
-    });
+static struct llama_sampler_i llama_sampler_mirostat_v2_i = {
+    /* .name   = */ llama_sampler_mirostat_v2_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_mirostat_v2_apply,
+    /* .reset  = */ llama_sampler_mirostat_v2_reset,
+    /* .clone  = */ llama_sampler_mirostat_v2_clone,
+    /* .free   = */ llama_sampler_mirostat_v2_free,
+};
+
+struct llama_sampler * llama_sampler_init_mirostat_v2(uint32_t seed, float tau, float eta) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_mirostat_v2_i,
+        /* .ctx   = */ new llama_sampler_mirostat_v2 {
+            /* .seed  = */ seed,
+            /* .tau   = */ tau,
+            /* .eta   = */ eta,
+            /* .mu    = */ 2.0f*tau,
+            /* .rng   = */ std::mt19937(seed),
+            /* .probs = */ {},
+        },
+    };
+}
 
-    llama_token result = max_iter->id;
-    if (smpl) {
-        smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
-        smpl->n_sample++;
+// grammar
+
+struct llama_sampler_grammar {
+    const struct llama_vocab * vocab;
+
+    std::string grammar_str;
+    std::string grammar_root;
+
+    struct llama_grammar * grammar;
+};
+
+static const char * llama_sampler_grammar_name(const struct llama_sampler * /*smpl*/) {
+    return "grammar";
+}
+
+static void llama_sampler_grammar_accept_impl(struct llama_sampler * smpl, llama_token token) {
+    auto * ctx = (llama_sampler_grammar *) smpl->ctx;
+    if (ctx->grammar) {
+        llama_grammar_accept_impl(*ctx->grammar, token);
     }
+}
+
+static void llama_sampler_grammar_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_grammar *) smpl->ctx;
+    if (ctx->grammar) {
+        llama_grammar_apply_impl(*ctx->grammar, cur_p);
+    }
+}
+
+static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
+    auto * ctx = (llama_sampler_grammar *) smpl->ctx;
+    if (!ctx->grammar) {
+        return;
+    }
+
+    auto * grammar_new = llama_grammar_init_impl(ctx->grammar->vocab, ctx->grammar_str.c_str(), ctx->grammar_root.c_str());
+
+    llama_grammar_free_impl(ctx->grammar);
+    ctx->grammar = grammar_new;
+}
+
+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);
+
+    // copy the state
+    {
+        auto * result_ctx = (llama_sampler_grammar *) result->ctx;
+
+        if (ctx->grammar) {
+            result_ctx->grammar_str  = ctx->grammar_str;
+            result_ctx->grammar_root = ctx->grammar_root;
+
+            result_ctx->grammar = llama_grammar_clone_impl(*ctx->grammar);
+        }
+    }
+
     return result;
 }
 
-llama_token llama_sample_token_with_rng_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, std::mt19937 & rng) {
-    LM_GGML_ASSERT(smpl);
+static void llama_sampler_grammar_free(struct llama_sampler * smpl) {
+    const auto * ctx = (llama_sampler_grammar *) smpl->ctx;
 
-    const int64_t t_start_sample_us = lm_ggml_time_us();
-    llama_sample_softmax_impl((struct llama_sampling *) nullptr, candidates);
+    if (ctx->grammar) {
+        llama_grammar_free_impl(ctx->grammar);
+    }
 
-    std::vector<float> probs;
-    probs.reserve(candidates->size);
-    for (size_t i = 0; i < candidates->size; ++i) {
-        probs.push_back(candidates->data[i].p);
+    delete ctx;
+}
+
+static struct llama_sampler_i llama_sampler_grammar_i = {
+    /* .name   = */ llama_sampler_grammar_name,
+    /* .accept = */ llama_sampler_grammar_accept_impl,
+    /* .apply  = */ llama_sampler_grammar_apply,
+    /* .reset  = */ llama_sampler_grammar_reset,
+    /* .clone  = */ llama_sampler_grammar_clone,
+    /* .free   = */ llama_sampler_grammar_free,
+};
+
+struct llama_sampler * llama_sampler_init_grammar_impl(const struct llama_vocab & vocab, const char * grammar_str, const char * grammar_root) {
+    auto * ctx = new llama_sampler_grammar;
+
+    if (grammar_str != nullptr && grammar_str[0] != '\0') {
+        *ctx = {
+            /* .vocab        = */ &vocab,
+            /* .grammar_str  = */ grammar_str,
+            /* .grammar_root = */ grammar_root,
+            /* .grammar      = */ llama_grammar_init_impl(&vocab, grammar_str, grammar_root),
+        };
+    } else {
+        *ctx = {
+            /* .vocab        = */ &vocab,
+            /* .grammar_str  = */ {},
+            /* .grammar_root = */ {},
+            /* .grammar      = */ nullptr,
+        };
+    }
+
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_grammar_i,
+        /* .ctx   = */ ctx,
+    };
+}
+
+// penalties
+
+struct llama_sampler_penalties {
+    const int32_t     n_vocab;
+    const llama_token special_eos_id;
+    const llama_token linefeed_id;
+
+    const int32_t penalty_last_n;
+    const float   penalty_repeat;
+    const float   penalty_freq;
+    const float   penalty_present;
+
+    const bool    penalize_nl;
+    const bool    ignore_eos;
+
+    ring_buffer<llama_token> prev;
+};
+
+static const char * llama_sampler_penalties_name(const struct llama_sampler * /*smpl*/) {
+    return "penalties";
+}
+
+static void llama_sampler_penalties_accept(struct llama_sampler * smpl, llama_token token) {
+    auto * ctx = (llama_sampler_penalties *) smpl->ctx;
+    if (ctx->penalty_last_n == 0) {
+        return;
+    }
+
+    ctx->prev.push_back(token);
+}
+
+static void llama_sampler_penalties_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_penalties *) smpl->ctx;
+
+    if (ctx->ignore_eos) {
+        assert(ctx->special_eos_id >= 0);
+
+        // optimistically check if the candidates are not yet sorted/shuffled/truncated
+        if (cur_p->size > (size_t) ctx->special_eos_id && cur_p->data[ctx->special_eos_id].id == ctx->special_eos_id) {
+            cur_p->data[ctx->special_eos_id].logit = -INFINITY;
+        } else {
+            // else, search for the special EOS token
+            for (size_t i = 0; i < cur_p->size; ++i) {
+                if (cur_p->data[i].id == ctx->special_eos_id) {
+                    cur_p->data[i].logit = -INFINITY;
+                    break;
+                }
+            }
+        }
+    }
+
+    if ((ctx->penalty_last_n == 0) ||
+        (ctx->penalty_repeat == 1.0f && ctx->penalty_freq == 0.0f && ctx->penalty_present == 0.0f)) {
+        return;
+    }
+
+    bool nl_found = false;
+    size_t nl_idx = 0;
+    float nl_logit = -INFINITY;
+    if (!ctx->penalize_nl) {
+        assert(ctx->linefeed_id >= 0);
+
+        // optimistically check if the candidates are not yet sorted/shuffled/truncated
+        if (cur_p->size > (size_t) ctx->linefeed_id && cur_p->data[ctx->linefeed_id].id == ctx->linefeed_id) {
+            nl_found = true;
+            nl_idx = ctx->linefeed_id;
+            nl_logit = cur_p->data[ctx->linefeed_id].logit;
+        } else {
+            // else, search for the linefeed token
+            for (size_t i = 0; i < cur_p->size; ++i) {
+                if (cur_p->data[i].id == ctx->linefeed_id) {
+                    nl_found = true;
+                    nl_idx = i;
+                    nl_logit = cur_p->data[i].logit;
+                    break;
+                }
+            }
+        }
     }
 
-    std::discrete_distribution<> dist(probs.begin(), probs.end());
-    int idx = dist(rng);
+    // Create a frequency map to count occurrences of each token in last_tokens
+    // TODO: optimize this by maintaining the token count in the sampler context
+    using llama_token_cnt = std::unordered_map<llama_token, int>;
+    llama_token_cnt token_count;
 
-    llama_token result = candidates->data[idx].id;
+    for (int i = 0; i < std::min<int>(ctx->penalty_last_n, ctx->prev.size()); ++i) {
+        token_count[ctx->prev.rat(i)]++;
+    }
 
-    smpl->t_sample_us += lm_ggml_time_us() - t_start_sample_us;
-    smpl->n_sample++;
+    // Apply frequency and presence penalties to the cur_p
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        const auto token_iter = token_count.find(cur_p->data[i].id);
+        if (token_iter == token_count.end()) {
+            continue;
+        }
+
+        const int count = token_iter->second;
+
+        // The academic publication that described this technique actually just only divided, but that would cause tokens with negative logits to become more likely, which is obviously wrong.
+        // This is common fix for this problem, which is to multiply by the penalty instead of dividing.
+        if (cur_p->data[i].logit <= 0) {
+            cur_p->data[i].logit *= ctx->penalty_repeat;
+        } else {
+            cur_p->data[i].logit /= ctx->penalty_repeat;
+        }
+
+        cur_p->data[i].logit -= float(count) * ctx->penalty_freq + float(count > 0) * ctx->penalty_present;
+    }
+
+    cur_p->sorted = false;
+
+    if (!ctx->penalize_nl && nl_found) {
+        // restore the logit of the newline token if it was penalized
+        cur_p->data[nl_idx].logit = nl_logit;
+    }
+}
+
+static void llama_sampler_penalties_reset(struct llama_sampler * smpl) {
+    auto * ctx = (llama_sampler_penalties *) smpl->ctx;
+    ctx->prev.clear();
+}
+
+static struct llama_sampler * llama_sampler_penalties_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_penalties *) smpl->ctx;
+    auto * result = llama_sampler_init_penalties(
+            ctx->n_vocab,
+            ctx->special_eos_id,
+            ctx->linefeed_id,
+            ctx->penalty_last_n,
+            ctx->penalty_repeat,
+            ctx->penalty_freq,
+            ctx->penalty_present,
+            ctx->penalize_nl,
+            ctx->ignore_eos);
+
+    // copy the state
+    {
+        auto * result_ctx = (llama_sampler_penalties *) result->ctx;
+
+        result_ctx->prev = ctx->prev;
+    }
 
     return result;
 }
 
-llama_token llama_sample_token_impl(struct llama_sampling * smpl, llama_token_data_array * candidates) {
-    return llama_sample_token_with_rng_impl(smpl, candidates, smpl->rng);
+static void llama_sampler_penalties_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_penalties *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_penalties_i = {
+    /* .name   = */ llama_sampler_penalties_name,
+    /* .accept = */ llama_sampler_penalties_accept,
+    /* .apply  = */ llama_sampler_penalties_apply,
+    /* .reset  = */ llama_sampler_penalties_reset,
+    /* .clone  = */ llama_sampler_penalties_clone,
+    /* .free   = */ llama_sampler_penalties_free,
+};
+
+struct llama_sampler * llama_sampler_init_penalties(
+        int32_t n_vocab,
+        llama_token special_eos_id,
+        llama_token linefeed_id,
+        int32_t penalty_last_n,
+        float penalty_repeat,
+        float penalty_freq,
+        float penalty_present,
+        bool penalize_nl,
+        bool ignore_eos) {
+    if (linefeed_id == LLAMA_TOKEN_NULL) {
+        penalize_nl = true;
+    }
+
+    if (special_eos_id == LLAMA_TOKEN_NULL) {
+        ignore_eos = false;
+    }
+
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_penalties_i,
+        /* .ctx   = */ new llama_sampler_penalties {
+            /* .n_vocab         = */ n_vocab,
+            /* .special_eos_id  = */ special_eos_id,
+            /* .linefeed_id     = */ linefeed_id,
+            /* .penalty_last_n  = */ penalty_last_n,
+            /* .penalty_repeat  = */ penalty_repeat,
+            /* .penalty_freq    = */ penalty_freq,
+            /* .penalty_present = */ penalty_present,
+            /* .penalize_nl     = */ penalize_nl,
+            /* .ignore_eos      = */ ignore_eos,
+            /* .prev            = */ ring_buffer<llama_token>(penalty_last_n),
+        },
+    };
+}
+
+// logit-bias
+
+struct llama_sampler_logit_bias {
+    const int32_t n_vocab;
+
+    const std::vector<llama_logit_bias> logit_bias;
+
+    std::vector<llama_logit_bias> to_search;
+};
+
+static const char * llama_sampler_logit_bias_name(const struct llama_sampler * /*smpl*/) {
+    return "logit-bias";
+}
+
+static void llama_sampler_logit_bias_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_logit_bias *) smpl->ctx;
+
+    ctx->to_search.clear();
+
+    // update the candidates that have not been shuffled in the vocabulary (i.e. idx == id)
+    for (const auto & lb : ctx->logit_bias) {
+        if (lb.token >= 0 && cur_p->size > (size_t) lb.token && cur_p->data[lb.token].id == lb.token) {
+            cur_p->data[lb.token].logit += lb.bias;
+        } else {
+            ctx->to_search.push_back(lb);
+        }
+    }
+
+    // search for the remaining candidates that were not found in the previous step
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        for (const auto & lb : ctx->to_search) {
+            if (cur_p->data[i].id == lb.token) {
+                cur_p->data[i].logit += lb.bias;
+                break;
+            }
+        }
+    }
+}
+static struct llama_sampler * llama_sampler_logit_bias_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_logit_bias *) smpl->ctx;
+    return llama_sampler_init_logit_bias(ctx->n_vocab, ctx->logit_bias.size(), ctx->logit_bias.data());
+}
+
+static void llama_sampler_logit_bias_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_logit_bias *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_logit_bias_i = {
+    /* .name   = */ llama_sampler_logit_bias_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_logit_bias_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_logit_bias_clone,
+    /* .free   = */ llama_sampler_logit_bias_free,
+};
+
+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 {
+        /* .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  = */ {},
+        },
+    };
 }