sampling: add Top-nσ sampler (#11223)

* initial sampling changes:

* completed top nsigma sampler implementation

* apply parameter to only llama-cli

* updated readme

* added tests and fixed nsigma impl

* cleaned up pr

* format

* format

* format

* removed commented tests

* cleanup pr and remove explicit floats

* added top-k sampler to improve performance

* changed sigma to float

* fixed string format to float

* Update src/llama-sampling.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update common/sampling.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update src/llama-sampling.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update src/llama-sampling.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update src/llama-sampling.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update src/llama-sampling.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* added llama_sampler_init

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

src/llama-sampling.cpp

@@ -1698,6 +1698,73 @@ struct llama_sampler * llama_sampler_init_penalties(
     );
 }
 
+// top-n-sigma
+
+struct llama_sampler_top_n_sigma {
+    const float n;
+};
+
+static const char * llama_sampler_top_n_sigma_name(const struct llama_sampler * /*smpl*/) {
+    return "top-n-sigma";
+}
+
+static void llama_sampler_top_n_sigma_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    const auto * ctx = (llama_sampler_top_n_sigma *) smpl->ctx;
+
+    // find max logit and calculate mean
+    float max = cur_p->data[0].logit;
+    float logits_sum = 0;
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        if (cur_p->data[i].logit > max) {
+            max = cur_p->data[i].logit;
+        }
+        logits_sum += cur_p->data[i].logit;
+    }
+    float mean = logits_sum/cur_p->size;
+
+    // calculate standard deviation
+    float acc = 0;
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        acc += pow(cur_p->data[i].logit - mean, 2);
+    }
+    float std = sqrt(acc/cur_p->size);
+
+    //apply mask
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        if (cur_p->data[i].logit < max - (ctx->n * std)) {
+            cur_p->data[i].logit = -INFINITY;
+        }
+    }
+    llama_sampler_softmax_impl(cur_p);
+}
+
+static struct llama_sampler * llama_sampler_top_n_sigma_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_top_n_sigma *) smpl->ctx;
+    return llama_sampler_init_top_n_sigma(ctx->n);
+}
+
+static void llama_sampler_top_n_sigma_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_top_n_sigma *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_top_n_sigma_i = {
+    /* .name   = */ llama_sampler_top_n_sigma_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_top_n_sigma_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_top_n_sigma_clone,
+    /* .free   = */ llama_sampler_top_n_sigma_free,
+};
+
+struct llama_sampler * llama_sampler_init_top_n_sigma(float n) {
+    return llama_sampler_init(
+        /* .iface = */ &llama_sampler_top_n_sigma_i,
+        /* .ctx   = */ new llama_sampler_top_n_sigma {
+            /* .n = */ n,
+        }
+    );
+}
+
 // DRY
 
 struct llama_sampler_dry {