cui-llama.rn 1.2.3 → 1.2.6

package/cpp/llama-sampling.cpp

@@ -63,6 +63,30 @@ static void llama_log_softmax(float * array, size_t size) {
 }
 */
 
+static void llama_sampler_temp_impl(llama_token_data_array * cur_p, float temp) {
+    if (temp <= 0.0f) {
+        // find the token with the highest logit and set the rest to -inf
+        size_t max_i = 0;
+        float  max_l = cur_p->data[0].logit;
+
+        for (size_t i = 1; i < cur_p->size; ++i) {
+            if (cur_p->data[i    ].logit > max_l) {
+                cur_p->data[max_i].logit = -INFINITY;
+                max_i = i;
+                max_l = cur_p->data[i].logit;
+            } else {
+                cur_p->data[i].logit = -INFINITY;
+            }
+        }
+
+        return;
+    }
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        cur_p->data[i].logit /= temp;
+    }
+}
+
 static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) {
     LM_GGML_ASSERT(cur_p->size > 0);
 
@@ -89,7 +113,7 @@ static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) {
 }
 
 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
+    // TODO: move bucket sort to separate function so that top_p/typical/softmax first is equally fast
     // if (k >= (int32_t)cur_p->size) {
     //     return;
     // }
@@ -428,6 +452,9 @@ static const char * llama_sampler_dist_name(const struct llama_sampler * /*smpl*
 
 static void llama_sampler_dist_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_dist *) smpl->ctx;
+
+    llama_sampler_softmax_impl(cur_p);
+
     cur_p->selected = llama_sample_dist(cur_p, ctx->rng);
 }
 
@@ -707,245 +734,6 @@ struct llama_sampler * llama_sampler_init_min_p(float p, size_t 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_sampler_softmax_impl(cur_p);
-    
-    // unsorted iteration
-    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 < 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(cur_p-> data[i]);
-            }
-        }
-        // if there is only one or no top_tokens, do not truncate
-
-        if (top_tokens.size() <= 1) {
-            return;
-        }
-
-        // sort top_tokens
-        std::sort(top_tokens.begin(), top_tokens.end(), [](const llama_token_data & a, const llama_token_data & b) {
-            return a.logit < b.logit;
-        });
-            
-        // insert top_tokens with probability. Always insert lowest top_token
-        low_tokens.push_back(top_tokens[0]);
-        std::uniform_real_distribution<float> random_float(0.0 , 1.0);
-        for (size_t i = 1; i < top_tokens.size(); ++i) {
-            if(random_float(rng) <= xtc_probability) {
-                low_tokens.push_back(top_tokens[i]);
-            }
-        }
-        if(low_tokens.size() >= min_keep) {
-           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 (!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;
-        }
-
-        // find last token over threshold
-
-        size_t last_index = 0;
-
-        for (; last_index < cur_p -> size; ++last_index) {
-            if(cur_p -> data[last_index].p < xtc_threshold) {
-                break;
-            }
-        }
-        
-        // check if only 1 token above threshold
-        if(last_index <= 1) {
-            return;
-        }
-        last_index--;
-        // items beyond safe index will be ignored
-        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(cur_p-> data[i], cur_p->data[safe_index - 1]);
-                safe_index--;
-                if (cur_p-> sorted) {
-                    cur_p -> sorted = false;
-                }
-            }
-        }
-        cur_p -> size = safe_index;
-    }
-}
-
-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);
-}
-
-static void llama_sampler_xtc_free(struct llama_sampler * smpl) {
-    delete (const llama_sampler_xtc *) smpl->ctx;
-}
-
-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
-        },
-    };
-}
-
-// tail-free
-
-struct llama_sampler_tail_free {
-    const float  z;
-    const size_t min_keep;
-};
-
-static const char * llama_sampler_tail_free_name(const struct llama_sampler * /*smpl*/) {
-    return "tail-free";
-}
-
-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_sampler_softmax_impl(cur_p);
-
-    // Compute the first and second derivatives
-    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] = 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];
-    }
-
-    // Calculate absolute value of second derivatives
-    for (size_t i = 0; i < second_derivatives.size(); ++i) {
-        second_derivatives[i] = std::abs(second_derivatives[i]);
-    }
-
-    // Normalize the second derivatives
-    {
-        const float second_derivatives_sum = std::accumulate(second_derivatives.begin(), second_derivatives.end(), 0.0f);
-
-        if (second_derivatives_sum > 1e-6f) {
-            for (float & value : second_derivatives) {
-                value /= second_derivatives_sum;
-            }
-        } else {
-            for (float & value : second_derivatives) {
-                value = 1.0f / second_derivatives.size();
-            }
-        }
-    }
-
-    float cum_sum = 0.0f;
-    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 > ctx->z && i >= ctx->min_keep) {
-            last_idx = i;
-            break;
-        }
-    }
-
-    // Resize the output vector to keep only the tokens above the tail location
-    cur_p->size = last_idx;
-}
-
-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;
-}
-
-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 {
@@ -1057,9 +845,8 @@ static const char * llama_sampler_temp_name(const struct llama_sampler * /*smpl*
 
 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;
-    }
+
+    llama_sampler_temp_impl(cur_p, ctx->temp);
 }
 
 static struct llama_sampler * llama_sampler_temp_clone(const struct llama_sampler * smpl) {
@@ -1106,6 +893,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
     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;
 
         // no need to do anything if there is only one (or zero) candidates
@@ -1143,9 +931,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
     #endif
 
         // Apply the dynamically calculated temperature scaling
-        for (size_t i = 0; i < cur_p->size; ++i) {
-            cur_p->data[i].logit /= dyn_temp;
-        }
+        llama_sampler_temp_impl(cur_p, dyn_temp);
 
         // Re-compute softmax probabilities after scaling logits with dynamic temperature
         const double max_l_double = cur_p->data[0].logit;
@@ -1169,9 +955,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
         }
     #endif
     } else {
-        for (size_t i = 0; i < cur_p->size; ++i) {
-            cur_p->data[i].logit /= ctx->temp;
-        }
+        llama_sampler_temp_impl(cur_p, ctx->temp);
     }
 }
 
@@ -1204,6 +988,101 @@ struct llama_sampler * llama_sampler_init_temp_ext(float temp, float delta, floa
     };
 }
 
+// xtc
+
+struct llama_sampler_xtc {
+    const float    probability;
+    const float    threshold;
+    const size_t   min_keep;
+
+    const uint32_t seed;
+    uint32_t       seed_cur;
+
+    std::mt19937   rng;
+};
+
+static const char * llama_sampler_xtc_name(const struct llama_sampler * /*smpl*/) {
+    return "xtc";
+}
+
+static void llama_sample_xtc_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_xtc *) smpl->ctx;
+
+    if (ctx->probability <= 0.0f
+        || ctx->threshold > 0.5f
+        || cur_p->size < 2) {
+        return;
+    }
+
+    std::uniform_real_distribution<float> distribution(0.0f, 1.0f);
+    float chance = distribution(ctx->rng);
+    if (chance > ctx->probability) return;
+
+    // in case it's not sorted/recalculated yet
+    llama_sampler_softmax_impl(cur_p);
+
+    int pos_last = 0;
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        if (cur_p->data[i].p >= ctx->threshold) {
+            pos_last = i;
+        } else break;
+    }
+
+    if (cur_p->size - pos_last >= ctx->min_keep && pos_last > 0) {
+        cur_p->data += pos_last;
+        cur_p->size -= pos_last;
+    }
+}
+
+static struct llama_sampler * llama_sampler_xtc_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_xtc *) smpl->ctx;
+    auto * result = llama_sampler_init_xtc(ctx->probability, ctx->threshold, ctx->min_keep, ctx->seed);
+
+    // copy the state
+    {
+        auto * result_ctx = (llama_sampler_xtc *) result->ctx;
+
+        result_ctx->rng = ctx->rng;
+    }
+
+    return result;
+}
+
+static void llama_sampler_xtc_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_xtc *) smpl->ctx;
+}
+
+static void llama_sampler_xtc_reset(struct llama_sampler * smpl) {
+    auto * ctx = (llama_sampler_xtc *) smpl->ctx;
+    ctx->seed_cur = get_rng_seed(ctx->seed);
+    ctx->rng.seed(ctx->seed_cur);
+}
+
+static struct llama_sampler_i llama_sampler_xtc_i = {
+    /* .name   = */ llama_sampler_xtc_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sample_xtc_apply,
+    /* .reset  = */ llama_sampler_xtc_reset,
+    /* .clone  = */ llama_sampler_xtc_clone,
+    /* .free   = */ llama_sampler_xtc_free,
+};
+
+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 {
+        /* .iface = */ &llama_sampler_xtc_i,
+        /* .ctx   = */ new llama_sampler_xtc {
+            /* .probability   = */ p,
+            /* .threshold     = */ t,
+            /* .min_keep      = */ min_keep,
+            /* .seed          = */ seed,
+            /* .seed_cur      = */ seed_cur,
+            /* .rng           = */ std::mt19937(seed_cur),
+        },
+    };
+}
+
 // mirostat
 
 struct llama_sampler_mirostat {
@@ -1710,6 +1589,397 @@ struct llama_sampler * llama_sampler_init_penalties(
     };
 }
 
+// DRY
+
+struct llama_sampler_dry {
+    int32_t total_context_size;
+
+    const float   dry_multiplier;
+    const float   dry_base;
+    const int32_t dry_allowed_length;
+    const int32_t dry_penalty_last_n;
+
+    std::unordered_multimap<llama_token, std::vector<llama_token>> dry_processed_breakers;
+    std::vector<int> dry_repeat_count;
+    std::unordered_map<llama_token, int> dry_max_token_repeat;
+    ring_buffer<llama_token> last_tokens;
+};
+
+// Ported from Koboldcpp, original PR: https://github.com/LostRuins/koboldcpp/pull/982 (Original author: pi6am)
+static void get_overlapping_token_sequences(const llama_vocab & vocab, const std::string& str, std::unordered_multimap<llama_token, std::vector<llama_token>>& token_sequences, int max_tail_len = -1) {
+    for (llama_token token_id = 0; token_id < (llama_token)vocab.n_vocab; token_id++) {
+        std::string word = llama_detokenize(vocab, {token_id}, true);
+        if (word.find(str) != std::string::npos) {
+            token_sequences.emplace(token_id, std::vector<llama_token>());
+        } else {
+            size_t word_len = word.size(), str_len = str.size();
+            size_t pos = -1;
+            while ((pos = word.find(str[0], pos + 1)) != std::string::npos) {
+                bool match = true;
+                size_t i;
+                for (i = 1; i < str_len && i + pos < word_len; ++i) {
+                    if (word[pos + i] != str[i]) {
+                        match = false;
+                        break;
+                    }
+                }
+                if (match) {
+                    std::vector<llama_token> tokenization = llama_tokenize_internal(vocab, str.substr(i), false, false);
+                    if (max_tail_len >= 0 && tokenization.size() > (size_t)max_tail_len) {
+                        tokenization.resize(max_tail_len);
+                    }
+
+                    // Ensure we don't already have a duplicate matching tokenization
+                    auto its = token_sequences.equal_range(token_id);
+                    bool found = false;
+                    for (auto it = its.first; it != its.second; ++it) {
+                        if (tokenization == it->second) {
+                            found = true;
+                            break;
+                        }
+                    }
+                    if (!found) {
+                        token_sequences.emplace(token_id, tokenization);
+                    }
+                }
+            }
+        }
+    }
+}
+
+static const char * llama_sampler_dry_name(const struct llama_sampler * /*smpl*/) {
+    return "dry";
+}
+
+static void llama_sampler_dry_accept(struct llama_sampler * smpl, llama_token token) {
+    auto * ctx = (llama_sampler_dry *) smpl->ctx;
+    if (ctx->dry_multiplier == 0.0f || ctx->dry_base < 1.0f || ctx->dry_penalty_last_n == 0) {
+        return;
+    }
+
+    ctx->last_tokens.push_back(token);
+}
+
+// Ported from Koboldcpp, original PR: https://github.com/LostRuins/koboldcpp/pull/982 (Original author: pi6am)
+static void llama_sampler_dry_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_dry *) smpl->ctx;
+
+    if (ctx->dry_multiplier == 0.0f || ctx->dry_base < 1.0f || ctx->dry_penalty_last_n == 0) {
+        return;
+    }
+
+    int32_t effective_dry_penalty_last_n = (ctx->dry_penalty_last_n == -1) ? ctx->total_context_size : std::max(ctx->dry_penalty_last_n, 0);
+    int last_n_repeat = std::min(std::min((int)ctx->last_tokens.size(), effective_dry_penalty_last_n), ctx->total_context_size);
+
+    if (last_n_repeat <= ctx->dry_allowed_length) {
+        return;
+    }
+
+    ctx->dry_repeat_count.assign(last_n_repeat, 0);
+    ctx->dry_max_token_repeat.clear();
+
+    // Step 1: Look for restart sequences to limit the maximum repetition length.
+    // Work backwards through the context looking for any token that begins a restart sequence.
+    //
+    // The collection `restart_sequences` is a mapping from a "head" token to all "tail"
+    // sequences that together comprise a restart sequence. This allows us to quickly check
+    // whether each token is the head of a complete sequence. Most restart sequences are actually
+    // a single token, and for these the "tail" is an empty vector.
+    //
+    // If the token is a "head", test all restart sequences that begin with this token
+    // (there will often only be one sequence for each token, but if sequences like 'aaaq1' and
+    // 'aaa1' are used as restart strings, both could start with 'aaa' when tokenized). The
+    // longest matching sequence (if any) is used to limit the maximum repetition length.
+    //
+    // Note that in the case case of a short sequence contained in a longer one, this might fail to
+    // find the smallest value for `rep_limit`. For example, if 'amniotic' and 'ni' are both used as
+    // restart sequences, 'ni' will be found first, and since it's shorter it will fail to suppress
+    // 'otic'. This is a minor issue since fully contained restart sequences are likely to be rare.
+    //
+    // This is theoretically worst-case O(N^2) for arbitrary restart sequences, which is why we
+    // have already clamped the maximum tail sequence length when generating `restart_sequences`.
+    // With clamping, this scan is O(N) in the context length.
+
+    int rep_limit = last_n_repeat;
+    for (int i = 0; i < last_n_repeat; ++i) {
+        llama_token token = ctx->last_tokens.rat(i);
+        auto its = ctx->dry_processed_breakers.equal_range(token);
+        if (its.first == ctx->dry_processed_breakers.end()) {
+            continue;
+        }
+        int longest_match = -1;
+        for (auto it = its.first; it != its.second; ++it) {
+            // Note that (*it) does not contain the head character, so seq_len will be
+            // the restart sequence length minus 1.
+            // In the common case of a single-token restart sequence, (*it) will be empty
+            // and we will trivially match.
+            int seq_len = (int)it->second.size();
+            if (seq_len > longest_match && seq_len <= (int)i) {
+                bool match = true;
+                for (int offset = 0; offset < seq_len; ++offset) {
+                    // The -1 when indexing `last_tokens` is because we already matched the head.
+                    if (it->second[offset] != ctx->last_tokens.rat(i - offset - 1)) {
+                        match = false;
+                        break;
+                    }
+                }
+                if (match) {
+                    longest_match = seq_len;
+                }
+            }
+        }
+        if (longest_match >= 0) {
+            // We found a restart sequence starting `i` tokens from the end and continuing for
+            // `longest_match` tokens.
+            rep_limit = i - longest_match;
+            break;
+        }
+    }
+    if (rep_limit < ctx->dry_allowed_length) {
+        return;
+    }
+
+    // Step 2: Iterate in reverse over the last N tokens of the context, using the "Z-algorithm" (in
+    // the reverse direction) to efficiently compute the positions and lengths of suffixes appearing
+    // elsewhere in the context. We limit the suffix length to `rep_limit` to respect restart sequences.
+    //
+    // This algorithm is not currently documented on Wikipedia, but there is a clear description here:
+    // https://ivanyu.me/blog/2014/10/15/z-algorithm/
+    //
+    // The code below is adapted from the public domain implementation by the same author here:
+    // https://github.com/ivanyu/string-algorithms/blob/master/z_algorithm.py
+    //
+    // Example:
+    // Last N tokens: a b c c b c y a b c
+    // Repeat counts: 0 0 3 1 0 2 0 0 0 0
+    //                    ^
+    //   This `3` means that the last three tokens of the context (a b c) also appear here.
+    //
+    // This step is worst case O(N) since the Z-algorithm is linear, despite the appearance of nested
+    // for/while loops. This can be seen by observing that the `lt` and `rt` bounds are set after each
+    // repeated suffix is detected (i.e. after each while loop when n > 0). These bound variables
+    // ensure that the inner while loops only examine each token in the context once as the outer
+    // for loop iterates over the context.
+
+    {
+        const int last = last_n_repeat - 1;
+        int rt = 0, lt = 0;
+
+        for (int k = 1; k < last_n_repeat; ++k) {
+            if (k > rt) {
+                // If k is outside the current Z-box, do naive computation.
+                int n = 0;
+                while (n + k < last_n_repeat && ctx->last_tokens.rat(n) == ctx->last_tokens.rat(n+k)) {
+                    ++n;
+                }
+                ctx->dry_repeat_count[last - k] = std::min(n, rep_limit);
+                if (n > 0) {
+                    lt = k;
+                    rt = k+n-1;
+                }
+            } else {
+                // If k is inside the current Z-box, consider two cases.
+
+                int p = k - lt; // Pair index.
+                int right_part_len = rt - k + 1;
+
+                if (ctx->dry_repeat_count[last - p] < right_part_len) {
+                    int n = std::min(ctx->dry_repeat_count[last - p], rep_limit);
+                    ctx->dry_repeat_count[last - k] = n;
+                } else {
+                    int i = rt + 1;
+                    while (i < last_n_repeat && ctx->last_tokens.rat(i) == ctx->last_tokens.rat(i - k)) {
+                        i += 1;
+                    }
+
+                    int n = std::min(i - k, rep_limit);
+                    ctx->dry_repeat_count[last - k] = n;
+                    lt = k;
+                    rt = i - 1;
+                }
+            }
+        }
+    }
+
+    // Step 3: Iterate over dry_repeat_count and last_tokens, examining the maximum repeat length
+    // that would be generated by emitting each new token that would extend a sequence.
+    //
+    // Following the same example as above:
+    // Last N tokens: a b c c b c y a b c
+    // Repeat counts: 0 0 3 1 0 2 0 0 0 0
+    //
+    // For each non-zero, look ahead one token. This token, if emitted, would extend the repetition.
+    // c: 3 -> 4 (from `a b c` to `a b c c`)
+    // b: 1 -> 2 (from `c` to `c b`)
+    // y: 2 -> 3 (from `b c` to `b c y`)
+
+    for (int i = 0; i < last_n_repeat - 1; ++i) {
+        int repeat_len = ctx->dry_repeat_count[i];
+        if (repeat_len >= ctx->dry_allowed_length) {
+            // This token ends a repeat, so the next token would continue one.
+            // By convention, the value of `repeat_len` only includes the tokens currently
+            // in the context, not the new token that would be added.
+            llama_token token = ctx->last_tokens.rat(last_n_repeat - 2 - i);
+            // Track the maximum sequence ending in this token.
+            const auto& it = ctx->dry_max_token_repeat.find(token);
+            if (it == ctx->dry_max_token_repeat.end() || it->second < repeat_len) {
+                ctx->dry_max_token_repeat[token] = repeat_len;
+            }
+        }
+    }
+
+    // Step 4: Apply logit penalties based on the maximum repeat length for relevant tokens.
+
+    // Prevent floating point overflow in `pow(penalty_base, exponent)` by clamping to `max_exponent`.
+    // Compute it from `penalty_base` and the approximate log of `std::numeric_limits<float>::max()`
+    const float FLOAT_MAX_LOG = 88.7228391f;
+    int max_exponent = 0;
+    if (ctx->dry_base > 1.000001f) {
+        max_exponent = FLOAT_MAX_LOG / std::log(ctx->dry_base);
+    }
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        const auto& af_kvp = ctx->dry_max_token_repeat.find(cur_p->data[i].id);
+        if (af_kvp != ctx->dry_max_token_repeat.end()) {
+            // Check all sequence breakers starting with this token
+            auto range = ctx->dry_processed_breakers.equal_range(cur_p->data[i].id);
+            bool is_single_token_breaker = false;
+
+            for (auto it = range.first; it != range.second; ++it) {
+                if (it->second.empty()) {
+                    is_single_token_breaker = true;
+                    break;
+                }
+            }
+
+            // Apply penalty only if it's not a single-token sequence breaker
+            if (!is_single_token_breaker) {
+                int repeat_exp = af_kvp->second - ctx->dry_allowed_length;
+                if (max_exponent > 0 && repeat_exp > max_exponent) {
+                    repeat_exp = max_exponent;
+                }
+                float penalty = ctx->dry_multiplier * std::pow(ctx->dry_base, repeat_exp);
+                cur_p->data[i].logit -= penalty;
+            }
+        }
+    }
+
+    cur_p->sorted = false;
+}
+
+static void llama_sampler_dry_reset(struct llama_sampler * smpl) {
+    auto * ctx = (llama_sampler_dry *) smpl->ctx;
+    ctx->last_tokens.clear();
+    ctx->dry_repeat_count.clear();
+    ctx->dry_max_token_repeat.clear();
+}
+
+static struct llama_sampler * llama_sampler_dry_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (llama_sampler_dry *) smpl->ctx;
+
+    // nullptr is passed as vocab because it is only needed for raw sequence breaker processing, which we have already done and will be copying
+    auto * result = llama_sampler_init_dry(nullptr, ctx->dry_multiplier, ctx->dry_base, ctx->dry_allowed_length, ctx->dry_penalty_last_n, NULL, 0);
+    // Copy the state, including the processed breakers
+    {
+        auto * result_ctx = (llama_sampler_dry *) result->ctx;
+        result_ctx->dry_processed_breakers = ctx->dry_processed_breakers;
+        result_ctx->dry_repeat_count = ctx->dry_repeat_count;
+        result_ctx->dry_max_token_repeat = ctx->dry_max_token_repeat;
+        result_ctx->last_tokens = ctx->last_tokens;
+    }
+
+    return result;
+}
+
+static void llama_sampler_dry_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_dry *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_dry_i = {
+    /* .name   = */ llama_sampler_dry_name,
+    /* .accept = */ llama_sampler_dry_accept,
+    /* .apply  = */ llama_sampler_dry_apply,
+    /* .reset  = */ llama_sampler_dry_reset,
+    /* .clone  = */ llama_sampler_dry_clone,
+    /* .free   = */ llama_sampler_dry_free,
+};
+
+struct llama_sampler * llama_sampler_init_dry_impl(const struct llama_vocab & vocab, int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const char** seq_breakers, size_t num_breakers) {
+    int32_t effective_dry_penalty_last_n = (dry_penalty_last_n == -1) ? context_size : std::max(dry_penalty_last_n, 0);
+    std::unordered_multimap<llama_token, std::vector<llama_token>> processed_breakers;
+    const int MAX_CHAR_LEN = 40;
+    const int MAX_SEQ_LEN = 20;
+
+    const bool dry_enabled = (dry_multiplier != 0.0f && dry_base >= 1.0f && dry_penalty_last_n != 0);
+
+    if (dry_enabled && seq_breakers != nullptr && num_breakers > 0) {
+        // Process sequence breakers
+        for (size_t i = 0; i < num_breakers; ++i) {
+            if (seq_breakers[i] == nullptr || std::strlen(seq_breakers[i]) == 0) {
+                LLAMA_LOG_WARN("skipping null or empty DRY sequence breaker at index %zu\n", i);
+                continue;
+            }
+
+            std::string sequence_break(seq_breakers[i]);
+            if (sequence_break.empty()) {
+                LLAMA_LOG_WARN("skipping empty DRY sequence breaker\n");
+                continue;
+            }
+
+            if (sequence_break.size() > MAX_CHAR_LEN) {
+                LLAMA_LOG_WARN("truncating DRY sequence breaker to %d characters\n", MAX_CHAR_LEN);
+                sequence_break.resize(MAX_CHAR_LEN);
+            }
+
+            get_overlapping_token_sequences(vocab, sequence_break, processed_breakers, MAX_SEQ_LEN);
+        }
+    }
+
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_dry_i,
+        /* .ctx   = */ new llama_sampler_dry {
+            /* .total_context_size     = */ context_size,
+            /* .dry_multiplier         = */ dry_multiplier,
+            /* .dry_base               = */ dry_base,
+            /* .dry_allowed_length     = */ dry_allowed_length,
+            /* .dry_penalty_last_n     = */ dry_penalty_last_n,
+            /* .dry_processed_breakers = */ std::move(processed_breakers),
+            /* .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
+struct llama_sampler * llama_sampler_init_dry_testing(int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const std::vector<std::vector<llama_token>>& seq_breakers) {
+    llama_vocab dummy_vocab;
+    auto * result = llama_sampler_init_dry_impl(dummy_vocab, context_size, dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, NULL, 0);
+    auto * ctx = (llama_sampler_dry *) result->ctx;
+
+    // Process the token-based sequence breakers
+    ctx->dry_processed_breakers.clear();
+    if (seq_breakers.empty()) {
+        LLAMA_LOG_WARN("empty DRY sequence breakers list in llama_sampler_init_dry_testing\n");
+    } else {
+        for (const auto& breaker : seq_breakers) {
+            if (breaker.empty()) {
+                LLAMA_LOG_WARN("skipping DRY empty sequence breaker\n");
+                continue;
+            }
+            llama_token head_token = breaker[0];
+            std::vector<llama_token> tail_tokens(breaker.begin() + 1, breaker.end());
+            ctx->dry_processed_breakers.emplace(head_token, std::move(tail_tokens));
+        }
+
+        if (ctx->dry_processed_breakers.empty()) {
+            LLAMA_LOG_WARN("no valid DRY sequence breakers processed in llama_sampler_init_dry_testing\n");
+        }
+    }
+
+    return result;
+}
+
 // logit-bias
 
 struct llama_sampler_logit_bias {
@@ -1789,6 +2059,229 @@ struct llama_sampler * llama_sampler_init_logit_bias(
     };
 }
 
+// infill
+
+//#define LM_GGML_DEBUG_SAMPLER_INFILL
+
+struct llama_sampler_infill {
+    const struct llama_vocab * vocab;
+
+    std::vector<char> buf0;
+    std::vector<char> buf1;
+};
+
+static const char * llama_sampler_infill_name(const struct llama_sampler * /*smpl*/) {
+    return "infill";
+}
+
+static void llama_sampler_infill_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_infill *) smpl->ctx;
+
+    llama_sampler_softmax_impl(cur_p);
+
+#if defined(LM_GGML_DEBUG_SAMPLER_INFILL)
+#define LOG_DBG_CUR LLAMA_LOG_DEBUG
+#else
+#define LOG_DBG_CUR(...)
+#endif
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        LOG_DBG_CUR("%s: cur_p[%3zu] = { id: %6d, p: %.6f, logit: %6.3f }\n", __func__, i, cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
+    }
+
+    float p_txt_sum = 0.0f;
+    float p_eog_sum = 0.0f;
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        if (llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id)) {
+            p_eog_sum += cur_p->data[i].p;
+        } else {
+            p_txt_sum += cur_p->data[i].p;
+        }
+    }
+
+    const float rat = p_eog_sum == 0.0 ? INFINITY : p_txt_sum / p_eog_sum; LM_GGML_UNUSED(rat);
+
+    LOG_DBG_CUR("%s: p_txt_sum = %.2f, p_eog_sum = %.2f, rat = %.2f, n = %zu\n", __func__, p_txt_sum, p_eog_sum, rat, cur_p->size);
+
+    if (3*p_eog_sum*cur_p->size > p_txt_sum) {
+        LOG_DBG_CUR("%s: the ratio p_txt/p_eog = %.2f is too low -> sampling EOG\n", __func__, p_txt_sum/p_eog_sum);
+
+        // keep just the EOG tokens
+        const auto size_org = cur_p->size;
+
+        cur_p->size = 0;
+
+        float p_sum = 0.0f;
+
+        for (size_t i = 0; i < size_org; ++i) {
+            if (llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id)) {
+                p_sum += cur_p->data[i].p;
+
+                cur_p->data[cur_p->size++] = cur_p->data[i];
+            }
+        }
+
+        // normalize probs
+        for (size_t i = 0; i < cur_p->size; ++i) {
+            cur_p->data[i].p /= p_sum;
+        }
+
+        return;
+    }
+
+    size_t n_combined = 0; LM_GGML_UNUSED(n_combined);
+
+    // combine tokens with common prefix
+    for (size_t i0 = 0; i0 < cur_p->size; ++i0) {
+        for (size_t i1 = 0; i1 < cur_p->size; ++i1) {
+            if (cur_p->data[i0].logit == -INFINITY) {
+                break;
+            }
+
+            if (i0 == i1 || cur_p->data[i1].logit == -INFINITY) {
+                continue;
+            }
+
+            int len0 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i0].id, ctx->buf0.data(), ctx->buf0.size(), 0, false);
+            if (len0 < 0) {
+                ctx->buf0.resize(len0);
+                len0 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i0].id, ctx->buf0.data(), ctx->buf0.size(), 0, false);
+                assert(len0 > 0);
+            }
+
+            int len1 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i1].id, ctx->buf1.data(), ctx->buf1.size(), 0, false);
+            if (len1 < 0) {
+                ctx->buf1.resize(len1);
+                len1 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i1].id, ctx->buf1.data(), ctx->buf1.size(), 0, false);
+                assert(len1 > 0);
+            }
+
+            // token i0 is a prefix of token i1
+            if (len0 > 0 && len0 <= len1 && memcmp(ctx->buf0.data(), ctx->buf1.data(), len0) == 0) {
+                int dst = i0;
+                int src = i1;
+
+                // merge into the token with higher probability
+                if (cur_p->data[i1].p > cur_p->data[i0].p) {
+                    std::swap(dst, src);
+                }
+
+                cur_p->data[dst].p += cur_p->data[src].p;
+                cur_p->data[src].logit = -INFINITY;
+                cur_p->data[src].p     = 0.0f;
+
+                n_combined++;
+            }
+        }
+    }
+
+    size_t n_non_eog = 0;
+
+    size_t size_org = cur_p->size;
+
+    float p_sum = 0.0f;
+    float thold = 0.2f;
+
+    cur_p->size = 0;
+
+    LOG_DBG_CUR("%s: n_combined = %zu, applying thold = %.3f\n", __func__, n_combined, thold);
+
+    for (size_t i = 0; i < size_org; ++i) {
+        const bool is_eog = llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id);
+
+        if (cur_p->data[i].p < thold && !is_eog) {
+            continue;
+        }
+
+        if (!is_eog) {
+            ++n_non_eog;
+        }
+
+        p_sum += cur_p->data[i].p;
+
+        // keep this token
+        cur_p->data[cur_p->size++] = cur_p->data[i];
+    }
+
+    LOG_DBG_CUR("%s: n_non_eog = %zu\n", __func__, n_non_eog);
+
+    // if no non-EOG tokens are left -> reduce cur_p to single EOT token
+    if (n_non_eog == 0) {
+        cur_p->size = 1;
+        cur_p->data[0].id = llama_token_eot_impl(*ctx->vocab);
+        cur_p->data[0].logit = 1.0f;
+
+        return;
+    }
+
+    // normalize probs
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        cur_p->data[i].p /= p_sum;
+
+        LOG_DBG_CUR("%s: cur_p[%3zu] = { id: %6d, p: %.6f, logit: %6.3f }\n", __func__, i, cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
+    }
+
+    size_org = cur_p->size;
+    p_sum = 0.0f;
+    thold = 1.0/(n_non_eog + 1);
+
+    cur_p->size = 0;
+
+    LOG_DBG_CUR("%s: applying thold = %.3f\n", __func__, thold);
+
+    for (size_t i = 0; i < size_org; ++i) {
+        const bool is_eog = llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id);
+
+        if (cur_p->data[i].p < thold && !is_eog) {
+            continue;
+        }
+
+        p_sum += cur_p->data[i].p;
+
+        cur_p->data[cur_p->size++] = cur_p->data[i];
+    }
+
+    // normalize probs
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        cur_p->data[i].p /= p_sum;
+
+        LOG_DBG_CUR("%s: cur_p[%3zu] = { id: %6d, p: %.6f, logit: %6.3f }\n", __func__, i, cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
+    }
+
+#undef LOG_DBG_CUR
+}
+
+static struct llama_sampler * llama_sampler_infill_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_infill *) smpl->ctx;
+    return llama_sampler_init_infill_impl(*ctx->vocab);
+}
+
+static void llama_sampler_infill_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_infill *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_infill_i = {
+    /* .name   = */ llama_sampler_infill_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_infill_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_infill_clone,
+    /* .free   = */ llama_sampler_infill_free,
+};
+
+struct llama_sampler * llama_sampler_init_infill_impl(
+        const struct llama_vocab & vocab) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_infill_i,
+        /* .ctx   = */ new llama_sampler_infill {
+            /* .vocab = */ &vocab,
+            /* .buf0 = */ std::vector<char>(512),
+            /* .buf1 = */ std::vector<char>(512),
+        },
+    };
+}
+
 // utils
 
 uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl) {