XTC: added xtc_threshold_max parameter as an upper limit

* 1.0 by default, so doesn't affect anything
* can be used to eliminate tokens within a range if you are sure that some top tokens are not clichéd (in finetuned models, for example)

base/llama-addon.cpp

@@ -36,8 +36,8 @@
 #include <type_traits>
 #include <unordered_map>
 
-void llama_sample_xtc_addon(struct llama_context * ctx, llama_token_data_array * candidates, float xtc_probability, float xtc_threshold, bool xtc_probability_once, int xtc_min, size_t min_keep) {
-    if (xtc_probability <= 0.0f || xtc_threshold <= 0.0f || candidates->size <= 1) {
+void llama_sample_xtc_addon(struct llama_context * ctx, llama_token_data_array * candidates, float xtc_probability, float xtc_threshold, float xtc_threshold_max, bool xtc_probability_once, int xtc_min, size_t min_keep) {
+    if (xtc_probability <= 0.0f || xtc_threshold <= 0.0f || xtc_threshold_max == xtc_threshold || xtc_min < 1 || candidates->size <= 1) {
         return;
     }
 
@@ -52,7 +52,7 @@ void llama_sample_xtc_addon(struct llama_context * ctx, llama_token_data_array *
     size_t removed = 0;
     // going through all candidates to correctly trigget the effect
     for (size_t i = 0; i < candidates->size; ++i) {
-        if (candidates->data[i].p >= xtc_threshold) {
+        if (candidates->data[i].p >= xtc_threshold && candidates->data[i].p <= xtc_threshold_max) {
                 if (id_first == -1) {
                     id_first = i;
                     ++removed;

base/llama-addon.h

@@ -7,6 +7,7 @@
           llama_token_data_array * candidates,
                            float   xtc_probability,
                            float   xtc_threshold,
+                           float   xtc_threshold_max,
                            bool    xtc_probability_once,
                            int     xtc_min,
                           size_t   min_keep);

base/sampling.cpp

@@ -146,6 +146,7 @@ void sampler_queue(
     const float       p_step               = params.p_step;
     const float       xtc_probability      = params.xtc_probability;
     const float       xtc_threshold        = params.xtc_threshold;
+    const float       xtc_threshold_max    = params.xtc_threshold_max;
     const float       xtc_probability_once = params.xtc_probability_once;
     const float       xtc_min              = params.xtc_min;
     const std::string samplers_sequence    = params.samplers_sequence;
@@ -158,7 +159,7 @@ void sampler_queue(
             case 'p': llama_sample_top_p        (ctx_main, &cur_p, top_p,     min_keep); break;
             case 'm': llama_sample_min_p_addon  (ctx_main, &cur_p, min_p,     min_keep); break;
             case 's': llama_sample_p_step_addon (ctx_main, &cur_p, p_step,    min_keep); break;
-            case 'x': llama_sample_xtc_addon    (ctx_main, &cur_p, xtc_probability, xtc_threshold, xtc_probability_once, xtc_min, min_keep); break;
+            case 'x': llama_sample_xtc_addon    (ctx_main, &cur_p, xtc_probability, xtc_threshold, xtc_threshold_max, xtc_probability_once, xtc_min, min_keep); break;
             case 't': {
                 if (dynatemp_range>0)
                 {

base/sampling.h

@@ -59,6 +59,7 @@ typedef struct llama_sampling_params {
     int32_t     dry_penalty_last_n    = -1;                 // DRY last n tokens to penalize (0 = disable penalty, -1 = context size)
     float       xtc_probability       = 0.5; // probability of removing a top token
     float       xtc_threshold         = 0.1; // minimum tokens probablitity for this to run
+    float       xtc_threshold_max     = 1.0; // maximum tokens probablitity for this to run
     bool        xtc_probability_once  = false; // should we calculate chances one or for each token
     int         xtc_min               = 2; // minimum number of penalizeable tokens
     std::string samplers_sequence     = "kfypmts"; // top_k, tail_free, typical_p, top_p, min_p, temp, p_step

chat_plain.h

@@ -557,7 +557,7 @@ class chat
                     case 'f': result += name_tfs_z; if (params.sparams.tfs_z != paramsDefault.sparams.tfs_z) result += std::format("={:.3f}",params.sparams.tfs_z); break;
                     case 'y': result += name_typical_p; if (params.sparams.typical_p != paramsDefault.sparams.typical_p) result += std::format("={:.3f}",params.sparams.typical_p); break;
                     case 's': result += name_p_step; if (params.sparams.p_step != paramsDefault.sparams.p_step) result += std::format("={:.3f}",params.sparams.p_step); break;
-                    case 'x': result += std::format("xtc={:.3f}-{:.03f}%",params.sparams.xtc_threshold,params.sparams.xtc_probability); break;
+                    case 'x': result += std::format("xtc={:.3f}-{:.3f}({}%/{})",params.sparams.xtc_threshold,params.sparams.xtc_threshold_max,params.sparams.xtc_probability*100,params.sparams.xtc_min); if (params.sparams.xtc_probability_once) result += "once"; else result += "each"; break;
                     case 'p': result += name_top_p; if (params.sparams.top_p != paramsDefault.sparams.top_p) result += std::format("={:.3f}",params.sparams.top_p); break;
                     case 'm': result += name_min_p; if (params.sparams.min_p != paramsDefault.sparams.min_p) result += std::format("={:.3f}",params.sparams.min_p); break;
                     case 't': {

include/jsonParams.h

@@ -519,6 +519,7 @@ static void getParamsFromJson(nlohmann::json& config, gpt_params& params, bool h
     if (checkJNum(config, "tfs_z")) params.sparams.tfs_z = config["tfs_z"];
     if (checkJNum(config, "xtc_probability")) params.sparams.xtc_probability = config["xtc_probability"];
     if (checkJNum(config, "xtc_threshold")) params.sparams.xtc_threshold = config["xtc_threshold"];
+    if (checkJNum(config, "xtc_threshold_max")) params.sparams.xtc_threshold_max = config["xtc_threshold_max"];
     if (checkJNum(config, "xtc_min")) params.sparams.xtc_min = config["xtc_min"];
     if (checkJBool(config, "xtc_probability_once")) params.sparams.xtc_probability_once = config["xtc_probability_once"];
 

thread_chat.h

@@ -1313,6 +1313,7 @@ struct configurableChat{
         } else if (params.sparams.p_step != paramsDefault.sparams.p_step) modelConfig[model]["p_step"] = params.sparams.p_step;
         if (params.sparams.xtc_probability != paramsDefault.sparams.xtc_probability) modelConfig[model]["xtc_probability"] = params.sparams.xtc_probability;
         if (params.sparams.xtc_threshold != paramsDefault.sparams.xtc_threshold) modelConfig[model]["xtc_threshold"] = params.sparams.xtc_threshold;
+        if (params.sparams.xtc_threshold_max != paramsDefault.sparams.xtc_threshold_max) modelConfig[model]["xtc_threshold_max"] = params.sparams.xtc_threshold_max;
         if (params.sparams.xtc_min != paramsDefault.sparams.xtc_min) modelConfig[model]["xtc_min"] = params.sparams.xtc_min;
         if (params.sparams.xtc_probability_once != paramsDefault.sparams.xtc_probability_once) modelConfig[model]["xtc_probability_once"] = params.sparams.xtc_probability_once;
         // penalties
@@ -1424,6 +1425,7 @@ struct configurableChat{
         if (params.sparams.p_step != paramsDefault.sparams.p_step) newCard["p_step"] = params.sparams.p_step;
         if (params.sparams.xtc_probability != paramsDefault.sparams.xtc_probability) newCard["xtc_probability"] = params.sparams.xtc_probability;
         if (params.sparams.xtc_threshold != paramsDefault.sparams.xtc_threshold) newCard["xtc_threshold"] = params.sparams.xtc_threshold;
+        if (params.sparams.xtc_threshold_max != paramsDefault.sparams.xtc_threshold_max) newCard["xtc_threshold_max"] = params.sparams.xtc_threshold_max;
         if (params.sparams.xtc_min != paramsDefault.sparams.xtc_min) newCard["xtc_min"] = params.sparams.xtc_min;
         if (params.sparams.xtc_probability_once != paramsDefault.sparams.xtc_probability_once) newCard["xtc_probability_once"] = params.sparams.xtc_probability_once;
         //penalties