Initial XTC commit

Adds XTC sampler, not activated by default, but recommended settings by default.

common/arg.cpp

@@ -966,6 +966,27 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
             params.sparams.min_p = std::stof(value);
         }
     ).set_sparam());
+    add_opt(llama_arg(
+        {"--xtc-p"}, "N",
+        format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sparams.xtc_p),
+        [](gpt_params & params, const std::string & value) {
+            params.sparams.xtc_p = std::stof(value);
+        }
+    ).set_sparam());
+    add_opt(llama_arg(
+        {"--xtc-t"}, "N",
+        format("xtc threshold (default: %.1f, 0.0 = disabled)", (double)params.sparams.xtc_t),
+        [](gpt_params & params, const std::string & value) {
+            params.sparams.xtc_t = std::stof(value);
+        }
+    ).set_sparam());
+    add_opt(llama_arg(
+        {"--xtc-t-max"}, "N",
+        format("xtc upper threshold (default: %.1f, 0.0 = disabled)", (double)params.sparams.xtc_t_max),
+        [](gpt_params & params, const std::string & value) {
+            params.sparams.xtc_t_max = std::stof(value);
+        }
+    ).set_sparam());
     add_opt(llama_arg(
         {"--tfs"}, "N",
         format("tail free sampling, parameter z (default: %.1f, 1.0 = disabled)", (double)params.sparams.tfs_z),

common/build-info.cpp

@@ -0,0 +1,4 @@
+int LLAMA_BUILD_NUMBER = 0;
+char const *LLAMA_COMMIT = "unknown";
+char const *LLAMA_COMPILER = "cc (GCC) 14.1.0";
+char const *LLAMA_BUILD_TARGET = "x86_64-w64-mingw32";

common/common.cpp

@@ -2060,6 +2060,9 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
     fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
     fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
+    fprintf(stream, "xtc_p: %f # default: 0.0\n", sparams.xtc_p);
+    fprintf(stream, "xtc_t: %f # default: 0.0\n", sparams.xtc_t);
+    fprintf(stream, "xtc_t_max: %f # default: 0.0\n", sparams.xtc_t_max);
     fprintf(stream, "typ_p: %f # default: 1.0\n", sparams.typ_p);
     fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
     fprintf(stream, "display_prompt: %s # default: true\n", params.display_prompt ? "true" : "false");

common/common.h

@@ -90,6 +90,7 @@ enum gpt_sampler_type {
     GPT_SAMPLER_TYPE_TFS_Z       = 4,
     GPT_SAMPLER_TYPE_TYPICAL_P   = 5,
     GPT_SAMPLER_TYPE_TEMPERATURE = 6,
+    GPT_SAMPLER_TYPE_XTC         = 7,
 };
 
 // dimensionality reduction methods, used by cvector-generator
@@ -108,6 +109,9 @@ struct gpt_sampler_params {
     int32_t top_k             = 40;    // <= 0 to use vocab size
     float   top_p             = 0.95f; // 1.0 = disabled
     float   min_p             = 0.05f; // 0.0 = disabled
+    float   xtc_p             = 0.50f; // 0.0 = disabled
+    float   xtc_t             = 0.10f; // 1.0 = disabled
+    float   xtc_t_max         = 1.00f; // 0.0 = disabled
     float   tfs_z             = 1.00f; // 1.0 = disabled
     float   typ_p             = 1.00f; // typical_p, 1.0 = disabled
     float   temp              = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities

common/sampling.cpp

@@ -130,10 +130,10 @@ std::string gpt_sampler_params::print() const {
 
     snprintf(result, sizeof(result),
             "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
-            "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, typical_p = %.3f, temp = %.3f\n"
+            "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, xtc_p = %.3f, xtc_t = %.3f, xtc_t_max = %.3f, typical_p = %.3f, temp = %.3f\n"
             "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
             penalty_last_n, penalty_repeat, penalty_freq, penalty_present,
-            top_k, tfs_z, top_p, min_p, typ_p, temp,
+            top_k, tfs_z, top_p, min_p, xtc_p, xtc_t, xtc_t_max, typ_p, temp,
             mirostat, mirostat_eta, mirostat_tau);
 
     return std::string(result);
@@ -184,6 +184,9 @@ struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const st
                     case GPT_SAMPLER_TYPE_MIN_P:
                         llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
                         break;
+                    case GPT_SAMPLER_TYPE_XTC:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_p, params.xtc_t, params.xtc_t_max, params.min_keep));
+                        break;
                     case GPT_SAMPLER_TYPE_TFS_Z:
                         llama_sampler_chain_add(result->chain, llama_sampler_init_tail_free(params.tfs_z, params.min_keep));
                         break;
@@ -372,6 +375,7 @@ char gpt_sampler_type_to_chr(enum gpt_sampler_type cnstr) {
         case GPT_SAMPLER_TYPE_TOP_P:       return 'p';
         case GPT_SAMPLER_TYPE_MIN_P:       return 'm';
         case GPT_SAMPLER_TYPE_TEMPERATURE: return 't';
+        case GPT_SAMPLER_TYPE_XTC:         return 'x';
         default : return '?';
     }
 }
@@ -384,6 +388,7 @@ std::string gpt_sampler_type_to_str(enum gpt_sampler_type cnstr) {
         case GPT_SAMPLER_TYPE_TOP_P:       return "top_p";
         case GPT_SAMPLER_TYPE_MIN_P:       return "min_p";
         case GPT_SAMPLER_TYPE_TEMPERATURE: return "temperature";
+        case GPT_SAMPLER_TYPE_XTC:         return "xtc";
         default : return "";
     }
 }
@@ -396,6 +401,7 @@ std::vector<gpt_sampler_type> gpt_sampler_types_from_names(const std::vector<std
         { "min_p",       GPT_SAMPLER_TYPE_MIN_P },
         { "tfs_z",       GPT_SAMPLER_TYPE_TFS_Z },
         { "temperature", GPT_SAMPLER_TYPE_TEMPERATURE },
+        { "xtc",         GPT_SAMPLER_TYPE_XTC },
     };
 
     // since samplers names are written multiple ways
@@ -441,7 +447,8 @@ std::vector<gpt_sampler_type> gpt_sampler_types_from_chars(const std::string & c
         { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TYPICAL_P),   GPT_SAMPLER_TYPE_TYPICAL_P },
         { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TOP_P),       GPT_SAMPLER_TYPE_TOP_P },
         { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_MIN_P),       GPT_SAMPLER_TYPE_MIN_P },
-        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TEMPERATURE), GPT_SAMPLER_TYPE_TEMPERATURE }
+        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TEMPERATURE), GPT_SAMPLER_TYPE_TEMPERATURE },
+        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_XTC),         GPT_SAMPLER_TYPE_XTC }
     };
 
     std::vector<gpt_sampler_type> samplers;

include/llama.h

@@ -1093,6 +1093,9 @@ extern "C" {
     /// @details Dynamic temperature implementation (a.k.a. entropy) described in the paper https://arxiv.org/abs/2309.02772.
     LLAMA_API struct llama_sampler * llama_sampler_init_temp_ext   (float   t, float   delta, float exponent);
 
+    /// @details XTC sampler as described in https://github.com/oobabooga/text-generation-webui/pull/6335
+    LLAMA_API struct llama_sampler * llama_sampler_init_xtc        (float   p, float   t, float     t_max, size_t min_keep);
+
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.

src/llama-sampling.cpp

@@ -1059,6 +1059,89 @@ 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 float  threshold_max;
+    const size_t min_keep;
+};
+
+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) {
+    const auto * ctx = (llama_sampler_xtc *) smpl->ctx;
+
+    if (ctx->probability <= 0.0f || ctx->threshold <= 0.0f || cur_p->size <= 1 || ctx->min_keep <= 2) {
+        return;
+    }
+
+    std::random_device rd;
+    float chance = (float)(rd()%100)/100;
+    if (chance > ctx->probability) return;
+    // in case it's not sorted/recalculated yet
+    llama_sampler_softmax_impl(cur_p);
+
+    int removed = 0;
+    // going through all candidates from back to front, easier to keep the last of probables
+    for (int i = (cur_p->size - 1); i >= 0; --i) {
+        if (cur_p->data[i].p >= ctx->threshold && cur_p->data[i].p <= ctx->threshold_max) {
+            if (removed == 0 || chance <= ctx->probability) {
+                ++removed;
+                if (removed >= 2) {
+                    // .logits are used for sorting and calculating .p in llama_sample_softmax_impl
+                    cur_p->data[i].logit = -999.0f;
+                    chance = (float)(rd()%100)/100;
+                }
+            }
+        }
+    }
+
+    if (removed >= 2) {
+        // sorting with new logits, ex-last probable will be the first anyway
+        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;
+
+        // resizing now that penalized tokens are at the back
+        cur_p->size = cur_p->size - removed + 1;
+    }
+}
+
+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->probability, ctx->threshold, ctx->threshold_max, ctx->min_keep);
+}
+
+static void llama_sampler_xtc_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_xtc *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_xtc_i = {
+    /* .name   = */ llama_sampler_xtc_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sample_xtc_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_xtc_clone,
+    /* .free   = */ llama_sampler_xtc_free,
+};
+
+struct llama_sampler * llama_sampler_init_xtc(float p, float t, float t_max, size_t min_keep) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_xtc_i,
+        /* .ctx   = */ new llama_sampler_xtc {
+            /* .probability   = */ p,
+            /* .threshold     = */ t,
+            /* .threshold_max = */ t_max,
+            /* .min_keep      = */ min_keep,
+        },
+    };
+}
+
 // mirostat
 
 struct llama_sampler_mirostat {