server: add repeat penalty sigmoid

common/common.h

@@ -117,6 +117,7 @@ struct gpt_sampler_params {
     float   penalty_repeat    = 1.00f; // 1.0 = disabled
     float   penalty_freq      = 0.00f; // 0.0 = disabled
     float   penalty_present   = 0.00f; // 0.0 = disabled
+    float   penalty_repeat_sigmoid_growth = 0.00f; // 0.0 = disabled
     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

common/sampling.cpp

@@ -168,6 +168,7 @@ struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const st
                 params.penalty_repeat,
                 params.penalty_freq,
                 params.penalty_present,
+                params.penalty_repeat_sigmoid_growth,
                 params.penalize_nl,
                 params.ignore_eos));
 

examples/server/README.md

@@ -350,6 +350,8 @@ node index.js
 
     `frequency_penalty`: Repeat alpha frequency penalty. Default: `0.0`, which is disabled.
 
+    `repeat_penalty_sigmoid_growth`: Apply the sigmoid function to `repeat_penalty` within `repeat_last_n` range. The value of `1` means linear change in penalty from 1 to `repeat_penalty`. Higher values > 1 increase the difference in the resulting penalty between the first and the second half of the penalty range. Lower values < 1 change the resulting penalty slower in the middle of the range. Negative values will be changing the penalty in the same way, but from `repeat_penalty` to 1. Default: `0.0`, which is disabled.
+
     `mirostat`: Enable Mirostat sampling, controlling perplexity during text generation. Default: `0`, where `0` is disabled, `1` is Mirostat, and `2` is Mirostat 2.0.
 
     `mirostat_tau`: Set the Mirostat target entropy, parameter tau. Default: `5.0`

examples/server/server.cpp

@@ -898,6 +898,7 @@ struct server_context {
         slot.sparams.penalty_repeat    = json_value(data, "repeat_penalty",    default_sparams.penalty_repeat);
         slot.sparams.penalty_freq      = json_value(data, "frequency_penalty", default_sparams.penalty_freq);
         slot.sparams.penalty_present   = json_value(data, "presence_penalty",  default_sparams.penalty_present);
+        slot.sparams.penalty_repeat_sigmoid_growth = json_value(data, "repeat_penalty_sigmoid_growth",  default_sparams.penalty_repeat_sigmoid_growth);
         slot.sparams.mirostat          = json_value(data, "mirostat",          default_sparams.mirostat);
         slot.sparams.mirostat_tau      = json_value(data, "mirostat_tau",      default_sparams.mirostat_tau);
         slot.sparams.mirostat_eta      = json_value(data, "mirostat_eta",      default_sparams.mirostat_eta);
@@ -1239,6 +1240,7 @@ struct server_context {
             {"repeat_penalty",            slot.sparams.penalty_repeat},
             {"presence_penalty",          slot.sparams.penalty_present},
             {"frequency_penalty",         slot.sparams.penalty_freq},
+            {"repeat_penalty_sigmoid_growth", slot.sparams.penalty_repeat_sigmoid_growth},
             {"mirostat",                  slot.sparams.mirostat},
             {"mirostat_tau",              slot.sparams.mirostat_tau},
             {"mirostat_eta",              slot.sparams.mirostat_eta},

include/llama.h

@@ -1124,6 +1124,7 @@ extern "C" {
                                float   penalty_repeat,  // 1.0 = disabled
                                float   penalty_freq,    // 0.0 = disabled
                                float   penalty_present, // 0.0 = disabled
+                               float   penalty_repeat_sigmoid_growth, // 0.0 = disabled
                                 bool   penalize_nl,     // consider newlines as a repeatable token
                                 bool   ignore_eos);     // ignore the end-of-sequence token
 

src/llama-sampling.cpp

@@ -1381,6 +1381,7 @@ struct llama_sampler_penalties {
     const float   penalty_repeat;
     const float   penalty_freq;
     const float   penalty_present;
+    const float   penalty_repeat_sigmoid_growth;
 
     const bool    penalize_nl;
     const bool    ignore_eos;
@@ -1450,6 +1451,115 @@ static void llama_sampler_penalties_apply(struct llama_sampler * smpl, llama_tok
         }
     }
 
+    struct sigmoid {
+        protected:
+            bool enabled;
+            float growth;
+            bool use_mirrored;
+            const ring_buffer<llama_token> & last_tokens;
+            size_t last_tokens_size;
+            size_t penalty_last_n;
+            float token_x;
+            float y_min = 0;
+            float y_diff = 0;
+
+            inline float calc_sigmoid(float x) {
+                float y = 1 / (1 + exp((-x + 0.5) * growth));
+                return y;
+            }
+
+            inline float calc_sigmoid_inv_growth(float x) {
+                float y = 1 / (1 + exp((-x + 0.5) / growth));
+                return y;
+            }
+
+            // sigmoid mirrored by y=x
+            inline float calc_mirrored_sigmoid(float x) {
+                if ((x == 0 && growth > 0) || (x >= 1 && growth < 0)) {
+                    return 0;
+                }
+                if ((x == 0 && growth < 0) || (x >= 1 && growth > 0)) {
+                    return 1;
+                }
+                // the actual formula: y = 0.5 - log((1 - x) / x) / growth
+                // but we invert the growth to transform the initial (0;1) range to the (1;+inf) range
+                float inv_growth = 1 / growth;
+                float y = 0.5 - log((1 - x) / x) / inv_growth;
+                return y;
+            }
+
+            inline float calc_norm_coeff(float x) {
+                if (use_mirrored) {
+                    float norm_x = (x + y_min) * y_diff; // normalize x within a range of the non-mirrored sigmoid's y
+                    float y = calc_mirrored_sigmoid(norm_x);
+                    return y;
+                }
+
+                float y = calc_sigmoid(x);
+                float norm_y = (y - y_min) / y_diff;
+                return norm_y;
+            }
+
+            static inline float apply_norm_coeff(float coeff, float penalty) {
+                float initial_diff = penalty - 1;
+                float result_diff = initial_diff * coeff;
+                return 1 + result_diff;
+            }
+
+        public:
+            explicit sigmoid(
+                float growth,
+                const ring_buffer<llama_token> & last_tokens,
+                size_t penalty_last_n
+            ) :
+            enabled(growth != 0),
+            growth(growth),
+            use_mirrored(abs(growth) < 1),
+            last_tokens(last_tokens),
+            last_tokens_size(std::min(penalty_last_n, last_tokens.size())),
+            penalty_last_n(penalty_last_n),
+            token_x(1 / (float)penalty_last_n) {
+                if (!enabled) {
+                    return;
+                }
+                float y1;
+                float y2;
+                if (use_mirrored) {
+                    y1 = calc_sigmoid_inv_growth(0);
+                    y2 = calc_sigmoid_inv_growth(1);
+                } else {
+                    y1 = calc_sigmoid(0);
+                    y2 = calc_sigmoid(1);
+                }
+                y_min = std::min(y1, y2);
+                float y_max = std::max(y1, y2);
+                y_diff = y_max - y_min;
+            }
+
+            inline float apply(float penalty, llama_token token) {
+                if (!enabled) {
+                    return penalty;
+                }
+                // the position (from the end) within the penalty tokens array
+                size_t token_rindex = 0;
+                while (token_rindex < last_tokens_size) {
+                    if (last_tokens.rat(token_rindex) == token) {
+                        break; // must always break at some point, otherwise it's UB
+                    }
+                    token_rindex++;
+                }
+                // the position within the penalty range,
+                // it's 1-indexed, so the last token in the range will correspond to x=1
+                size_t token_pos = penalty_last_n - token_rindex;
+                float x = token_x * token_pos;
+                float coeff = calc_norm_coeff(x);
+                float resulting_penalty = apply_norm_coeff(coeff, penalty);
+                return resulting_penalty;
+            }
+    };
+
+    sigmoid penalty_repeat_sigmoid(ctx->penalty_repeat_sigmoid_growth, ctx->prev, ctx->penalty_last_n);
+
     // 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>;
@@ -1461,7 +1571,8 @@ static void llama_sampler_penalties_apply(struct llama_sampler * smpl, llama_tok
 
     // 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);
+        const auto token = cur_p->data[i].id;
+        const auto token_iter = token_count.find(token);
         if (token_iter == token_count.end()) {
             continue;
         }
@@ -1470,11 +1581,14 @@ static void llama_sampler_penalties_apply(struct llama_sampler * smpl, llama_tok
 
         // 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.
+        float applied_penalty_repeat;
         if (cur_p->data[i].logit <= 0) {
-            cur_p->data[i].logit *= ctx->penalty_repeat;
+            applied_penalty_repeat = ctx->penalty_repeat;
         } else {
-            cur_p->data[i].logit /= ctx->penalty_repeat;
+            applied_penalty_repeat = 1 / ctx->penalty_repeat;
         }
+        applied_penalty_repeat = penalty_repeat_sigmoid.apply(applied_penalty_repeat, token);
+        cur_p->data[i].logit *= applied_penalty_repeat;
 
         cur_p->data[i].logit -= float(count) * ctx->penalty_freq + float(count > 0) * ctx->penalty_present;
     }
@@ -1502,6 +1616,7 @@ static struct llama_sampler * llama_sampler_penalties_clone(const struct llama_s
             ctx->penalty_repeat,
             ctx->penalty_freq,
             ctx->penalty_present,
+            ctx->penalty_repeat_sigmoid_growth,
             ctx->penalize_nl,
             ctx->ignore_eos);
 
@@ -1536,6 +1651,7 @@ struct llama_sampler * llama_sampler_init_penalties(
         float penalty_repeat,
         float penalty_freq,
         float penalty_present,
+        float penalty_repeat_sigmoid_growth,
         bool penalize_nl,
         bool ignore_eos) {
     if (linefeed_id == LLAMA_TOKEN_NULL) {
@@ -1558,6 +1674,7 @@ struct llama_sampler * llama_sampler_init_penalties(
             /* .penalty_repeat  = */ penalty_repeat,
             /* .penalty_freq    = */ penalty_freq,
             /* .penalty_present = */ penalty_present,
+            /* .penalty_repeat_sigmoid_growth */ penalty_repeat_sigmoid_growth,
             /* .penalize_nl     = */ penalize_nl,
             /* .ignore_eos      = */ ignore_eos,
             /* .prev            = */ ring_buffer<llama_token>(penalty_last_n),

tests/test-sampling.cpp

@@ -149,7 +149,7 @@ static void test_penalties(
 
     llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
 
-    auto * sampler = llama_sampler_init_penalties(n_vocab, LLAMA_TOKEN_NULL, LLAMA_TOKEN_NULL, last_tokens.size(), repeat_penalty, alpha_frequency, alpha_presence, false, false);
+    auto * sampler = llama_sampler_init_penalties(n_vocab, LLAMA_TOKEN_NULL, LLAMA_TOKEN_NULL, last_tokens.size(), repeat_penalty, alpha_frequency, alpha_presence, 0.0, false, false);
 
     for (size_t i = 0; i < last_tokens.size(); i++) {
         llama_sampler_accept(sampler, last_tokens[i]);