Add the DRY dynamic N-gram anti-repetition sampler

The DRY (Do not Repeat Yourself) sampler is a dynamic N-gram
repetition penalty that negatively scores tokens that would extend
sequences that already appear in the context.

See this discussion for a motivation and explanation of the sampler:
oobabooga/text-generation-webui#5677

This implementation of DRY mostly aligns with the obabooga version
with a few modifications. It uses a more efficient linear scanning
algorithm to identify repetitions. It also supports multi-token
sequence breakers. As a limitation, this implementation reuses
the rep pen range parameter, rather than introducing a new range
just for the DRY sampler.

There is a separate change to lite.koboldai.net that exposes the DRY
sampler parameters to KoboldAI Lite, so none of the embed files have
been changed as part of this commit.

Update default DRY parameters to match lite

Improve DRY token debug logging

Replace `and` with `&&` to fix MSVC compile error

Little known fact: The C++98 standard defines `and` as an
alternative token for the `&&` operator (along with a bunch
of other digraphs). MSVC does not allow these without using
the /Za option or including the <iso646.h> header. Change to
the more standard operator to make this code more portable.

Fix MSVC compile error because log is not constexpr

Replace the compile-time computation with a floating-point
approximation of log(std::numeric_limits<float>::max()).

Remove unused llama sampler variables and clean up sequence breakers.

Remove KCPP_SAMPLER_DRY as a separate enum entry

The DRY sampler is effectively a repetition penalty and there
are very few reasons to apply it at a different place in sampler
order than the standard single-token penalty. There are also
multiple projects that have dependencies on the existing sampler
IDs, including KoboldAI, KoboldAI Lite, and Silly Tavern. In order
to minimize the impact of the dependencies of adding the DRY sampler
to koboldcpp, it makes the most sense to not add a new ID for now,
and instead to piggyback on KCPP_SAMPLER_REP_PEN. In the future
if we find a use case for splitting the application of rep pen and DRY
we can introduce a new enum entry then.

Add the dry_penalty_last_n to independently control DRY penalty range

This parameter follows the oobabooga semantics: it's optional, with a
default value of zero. Zero means that DRY should sample the entire
context. Otherwise, it's the number of tokens from the end of the
context that are scanned for repetitions.

common/common.h

@@ -113,6 +113,11 @@ struct gpt_params {
     int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
     float   mirostat_tau      = 5.00f; // target entropy
     float   mirostat_eta      = 0.10f; // learning rate
+    float   dry_multiplier    = 0.0f;  // penalty multiplier, 0.0 = disabled
+    float   dry_base          = 1.75f; // exponential base
+    int32_t dry_allowed_length = 2;    // repeated sequences longer than this are penalized
+    int32_t dry_penalty_last_n = 0;    // how many tokens to scan for repetitions (0 = entire context)
+    std::vector<std::string> dry_sequence_breakers; // DRY sequence breakers
 
     // DynaTemp!
     float   dynatemp_range     = 0.0f;  // enables DynaTemp if greater than 0. dynatemp_min = temperature - dt_range, dynatemp_max = temperature + dt_range

expose.h

@@ -5,6 +5,7 @@ const int stop_token_max = 16;
 const int ban_token_max = 16;
 const int tensor_split_max = 16;
 const int logit_bias_max = 16;
+const int dry_seq_break_max = 16;
 const int images_max = 4;
 
 // match kobold's sampler list and order
@@ -89,6 +90,11 @@ struct generation_inputs
     const int mirostat = 0;
     const float mirostat_eta = 0.0f;
     const float mirostat_tau = 0.0f;
+    const float dry_multiplier = 0.0f;
+    const float dry_base = 0.0f;
+    const int dry_allowed_length = 0;
+    const int dry_penalty_last_n = 0;
+    const char * dry_sequence_breakers[dry_seq_break_max] = {};
     const samplers sampler_order[KCPP_SAMPLER_MAX] = {};
     const int sampler_len = 0;
     const bool allow_eos_token = false;

gpttype_adapter.cpp

@@ -10,6 +10,7 @@
 #include <cmath>
 #include <time.h>
 #include <mutex>
+#include <unordered_map>
 #include "model_adapter.h"
 #include "otherarch.h"
 #include "grammar-parser.h"
@@ -110,6 +111,9 @@ static std::vector<std::string> stop_sequence;
 static std::vector<int> special_stop_sequence; //for stop sequences that don't have a string representation
 static std::vector<std::string> banned_tokens;
 static std::vector<int> banned_token_ids;
+static std::unordered_multimap<gpt_vocab::id, std::vector<gpt_vocab::id>> dry_sequence_breakers; // Multi-mapping from first token of sequence to tail of sequence (tail is empty for a single token)
+static std::vector<int> dry_repeat_count; // Indexed as last_n_tokens
+static std::unordered_map<gpt_vocab::id, int> dry_max_token_repeat;
 static std::vector<llama_token_data> top_picks;
 static int remaining_tokens = 0;
 static int stopper_unused_tokens = 0;
@@ -309,6 +313,70 @@ static void print_tok_vec_str(std::vector<int> &vec)
     printf("\n%s", get_tok_vec_str(vec).c_str());
 }
 
+// Find tokens that completely contain `str`, either as a single token, or as a sequence of tokens.
+// It's important to use a hash map for head tokens because some models have many of them.
+// For example, the Llama 3 tokenizer has 6570 tokens containing the period ('.') character.
+// Single tokens are allowed to extend past `str` at the front and back. This is to allow, for
+// instance, the token '.\n' to be a head for both '.' and '\n'. However if a head token
+// begins a multi-token sequence, the head can only extend past `str` at the beginning. The
+// tail tokens are generated by tokenizing the remainder.
+static void GetOverlappingTokenSequences(const std::string& str, std::unordered_multimap<gpt_vocab::id, std::vector<gpt_vocab::id>>& token_sequences) {
+    for(int v=0;v<n_vocab;++v)
+    {
+        std::string word = FileFormatTokenizeID(v, file_format, true);
+        if (word.find(str) != std::string::npos)
+        {
+            // The string is entirely contained within this single token.
+            // Ensure that token_sequences only contains one key-value-pair with an empty value.
+            auto its = token_sequences.equal_range(v);
+            bool empty = false;
+            for (auto it = its.first; it != its.second; ++it) {
+                if (it->second.empty()) {
+                    empty = true;
+                    break;
+                }
+            }
+            if (!empty) {
+                token_sequences.emplace(v, std::vector<gpt_vocab::id>());
+            }
+        } else {
+            // Check whether a prefix of the string overlaps with a suffix of the token.
+            // Just do a naive O(N^2) search.
+            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) {
+                    // We matched to the end of the string. Since `str` is not contained in `word`,
+                    // there must be trailing letters in `str`.
+                    std::vector<gpt_vocab::id> tokenization;
+                    TokenizeString(str.substr(i), tokenization, file_format, false);
+
+                    // Ensure we don't already have a duplicate matching tokenization.
+                    auto its = token_sequences.equal_range(v);
+                    bool found = false;
+                    for (auto it = its.first; it != its.second; ++it) {
+                        if (tokenization == it->second) {
+                            found = true;
+                            break;
+                        }
+                    }
+                    if (!found)
+                    {
+                        token_sequences.emplace(v, tokenization);
+                    }
+                }
+            }
+        }
+    }
+}
 
 llama_token sample_token(llama_token_data_array * candidates, std::mt19937 & rng)
 {
@@ -428,6 +496,201 @@ void sample_top_a(llama_token_data_array * candidates, float a, size_t min_keep)
     candidates->size = last_idx;
 }
 
+void sample_dry(int n_ctx, int penalty_range, float penalty_multiplier, float penalty_base, int allowed_length, const std::unordered_multimap<gpt_vocab::id, std::vector<gpt_vocab::id>>& restart_sequences, llama_token_data_array * candidates) {
+    if (penalty_multiplier == 0.0f || penalty_base == 0.0f) {
+        return;
+    }
+    if (penalty_range <= 0) {
+        penalty_range = n_ctx;
+    }
+    auto last_n_repeat = std::min(std::min((int)current_context_tokens.size(), penalty_range), n_ctx);
+    if (last_n_repeat <= allowed_length) {
+        return;
+    }
+    const llama_token * last_tokens = current_context_tokens.data() + current_context_tokens.size() - last_n_repeat;
+
+    dry_repeat_count.assign(last_n_repeat, 0);
+    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 worst-case O(N^2) for perverse restart sequences, but typically will be O(N) since
+    // most restart sequences are a single token and we use a hash table to check for head token.
+
+    int rep_limit = last_n_repeat;
+    for (size_t i = 0; i < last_n_repeat; ++i) {
+        size_t ix = last_n_repeat - 1 - i;
+        auto its = restart_sequences.equal_range(last_tokens[ix]);
+        if (its.first == restart_sequences.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 <= i) {
+                bool match = true;
+                for (size_t offset = 0; offset < seq_len; ++offset) {
+                    // The +1 when indexing `last_tokens` is because we already matched the head.
+                    if (it->second[offset] != last_tokens[ix + 1 + offset]) {
+                        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 = (int)i - longest_match;
+            break;
+        }
+    }
+    if (rep_limit <= 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
+    //
+    // This step is worst case O(N), since the Z-algorithm is linear.
+    //
+    // 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
+
+    {
+        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 && last_tokens[last - n] == last_tokens[last - (n+k)]) {
+                    ++n;
+                }
+                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 (dry_repeat_count[last - p] < right_part_len) {
+                    int n = std::min(dry_repeat_count[last - p], rep_limit);
+                    dry_repeat_count[last - k] = n;
+                } else {
+                    int i = rt + 1;
+                    while (i < last_n_repeat && last_tokens[last - i] == last_tokens[last - (i - k)]) {
+                        i += 1;
+                    }
+
+                    int n = std::min(i - k, rep_limit);
+                    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 (size_t i = 0; i < last_n_repeat - 1; ++i) {
+        int repeat_len = dry_repeat_count[i];
+        if (repeat_len >= 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.
+            gpt_vocab::id token = last_tokens[i + 1];
+            // Track the maximum sequence ending in this token.
+            const auto& it = dry_max_token_repeat.find(token);
+            if (it == dry_max_token_repeat.end() || it->second < repeat_len) {
+                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 (penalty_base > 1.000001f) {
+        max_exponent = FLOAT_MAX_LOG / std::log(penalty_base);
+    }
+
+    if (debugmode==1 && !dry_max_token_repeat.empty()) {
+        printf("DRY penalties [");
+    }
+    size_t count = 0;
+    for (const auto& kvp: dry_max_token_repeat) {
+        gpt_vocab::id token = kvp.first;
+        int repeat_exp = kvp.second - allowed_length;
+        if (max_exponent > 0 && repeat_exp > max_exponent) {
+            repeat_exp = max_exponent;
+        }
+        float penalty = penalty_multiplier * pow(penalty_base, repeat_exp);
+        if (debugmode==1)
+        {
+            std::string tokenizedstr = FileFormatTokenizeID(token, file_format);
+            ::utreplace(tokenizedstr, "\n", "\\n");
+            printf("%s(%s %.02f)", count == 0 ? "" : " ", RemoveBell(tokenizedstr).c_str(), penalty);
+        }
+        candidates->data[token].logit -= penalty;
+        ++count;
+    }
+    if (debugmode==1 && !dry_max_token_repeat.empty()) {
+        printf("]\n");
+    }
+}
+
 void sample_rep_pen(int n_ctx, int rep_pen_range, float rep_pen, float rep_pen_slope, float presence_penalty, llama_token_data_array * candidates_p)
 {
     auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), rep_pen_range), n_ctx);
@@ -543,7 +806,7 @@ void sample_grammar(FileFormat file_format, int32_t n_vocab, llama_token_data_ar
 }
 
 int SampleLogits(const float * logits, int n_ctx, int n_vocab, int rep_pen_range, float rep_pen, float rep_pen_slope, float presence_penalty, float top_k, float top_a, float top_p, float min_p, float typical_p, float tfs, float temp, std::mt19937 & rng,
-int mirostat, float mirostat_tau, float mirostat_eta, const std::vector<samplers> & sampler_order, llama_grammar * grammar, float dynatemp_range, float dynatemp_exponent, float smoothing_factor)
+int mirostat, float mirostat_tau, float mirostat_eta, float dry_multiplier, float dry_base, int dry_allowed_length, int dry_penalty_last_n, const std::vector<samplers> & sampler_order, llama_grammar * grammar, float dynatemp_range, float dynatemp_exponent, float smoothing_factor)
 {
     int id = 0;
     std::vector<llama_token_data> candidates;
@@ -619,6 +882,7 @@ int mirostat, float mirostat_tau, float mirostat_eta, const std::vector<samplers
                     break;
                 case KCPP_SAMPLER_REP_PEN:
                     sample_rep_pen(n_ctx, rep_pen_range, rep_pen, rep_pen_slope, presence_penalty, &candidates_p);
+                    sample_dry(n_ctx, dry_penalty_last_n, dry_multiplier, dry_base, dry_allowed_length, dry_sequence_breakers, &candidates_p);
                     break;
                 default:
                     printf("\nSampleLogits: Unknown Sampler : %d",sampler_order[i]);
@@ -1808,11 +2072,44 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
     kcpp_params->mirostat = inputs.mirostat;
     kcpp_params->mirostat_eta = inputs.mirostat_eta;
     kcpp_params->mirostat_tau = inputs.mirostat_tau;
+    kcpp_params->dry_multiplier = inputs.dry_multiplier;
+    kcpp_params->dry_base = inputs.dry_base;
+    kcpp_params->dry_allowed_length = inputs.dry_allowed_length;
+    kcpp_params->dry_penalty_last_n = inputs.dry_penalty_last_n;
     kcpp_params->dynatemp_range = inputs.dynatemp_range;
     kcpp_params->dynatemp_exponent = inputs.dynatemp_exponent;
     kcpp_params->n_ctx = inputs.max_context_length;
     kcpp_params->smoothing_factor = inputs.smoothing_factor;
 
+    // Parse dry sequence breakers / restart sequences
+    kcpp_params->dry_sequence_breakers.clear();
+    for(int x=0;x<dry_seq_break_max;++x) {
+        std::string word = inputs.dry_sequence_breakers[x];
+        if(word!="") {
+            kcpp_params->dry_sequence_breakers.push_back(word);
+        }
+    }
+    dry_sequence_breakers.clear();
+    if(kcpp_params->dry_sequence_breakers.size()>0) {
+        if(debugmode==1) {
+            printf("\nProcessing %zu dry break strings...",kcpp_params->dry_sequence_breakers.size());
+        }
+        for (const auto& sequence_break: kcpp_params->dry_sequence_breakers) {
+            GetOverlappingTokenSequences(sequence_break, dry_sequence_breakers);
+        }
+        if(debugmode==1) {
+            int trivial = 0, non_trivial = 0;
+            for (const auto& seq: dry_sequence_breakers) {
+                if (seq.second.empty()) {
+                    ++trivial;
+                } else {
+                    ++non_trivial;
+                }
+            }
+            printf("\nFound a total of %zu restart heads, %d trivial, %d non-trivial.\n", dry_sequence_breakers.size(), trivial, non_trivial);
+        }
+    }
+
     bool stream_sse = inputs.stream_sse;
 
     bool allow_regular_prints = (debugmode!=-1 && !inputs.quiet) || debugmode >= 1;
@@ -2329,7 +2626,10 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
 
             id = SampleLogits(logitsPtr, nctx, n_vocab, last_n_size, repeat_penalty, kcpp_params->rep_pen_slope, presence_penalty,
             top_k, top_a, top_p, min_p, typical_p, tfs_z, temp, rng,
-            kcpp_params->mirostat, kcpp_params->mirostat_tau, kcpp_params->mirostat_eta, sampler_order, grammar, dynatemp_range, dynatemp_exponent, smoothing_factor);
+            kcpp_params->mirostat, kcpp_params->mirostat_tau, kcpp_params->mirostat_eta,
+            kcpp_params->dry_multiplier, kcpp_params->dry_base,
+            kcpp_params->dry_allowed_length, kcpp_params->dry_penalty_last_n,
+            sampler_order, grammar, dynatemp_range, dynatemp_exponent, smoothing_factor);
 
             if (llama_ctx_v4) {
                 empcats_step_post(llama_ctx_v4, id );