Merge branch 'master' into gg/llama-kv-cache

ggml-ci

src/llama-grammar.cpp

@@ -1186,7 +1186,7 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
                     return;
                 }
             }
-            LLAMA_LOG_DEBUG("Grammar still awaiting trigger after token %d (`%s`) (buffer: `%s`)\n", token, piece.c_str(), grammar.trigger_buffer.c_str());
+            LLAMA_LOG_DEBUG("Grammar still awaiting trigger after token %d (`%s`)\n", token, piece.c_str());
             return;
         }
     }

src/llama-grammar.h

@@ -116,7 +116,7 @@ struct llama_grammar {
     llama_partial_utf8 partial_utf8;
 
     // lazy grammars wait for trigger words or tokens before constraining the sampling.
-    // we still ahve trigger_tokens for non-lazy grammars to force printing of special trigger tokens.
+    // we still have trigger_tokens for non-lazy grammars to force printing of special trigger tokens.
     // (useful e.g. for tool_choice=required)
     bool                     lazy             = false;
     bool                     awaiting_trigger = false; // Initialized to true for lazy grammars only

src/llama-impl.h

@@ -6,13 +6,13 @@
 #include <vector>
 
 #ifdef __GNUC__
-#ifdef __MINGW32__
-#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#    if defined(__MINGW32__) && !defined(__clang__)
+#        define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#    else
+#        define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#    endif
 #else
-#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
-#endif
-#else
-#define LLAMA_ATTRIBUTE_FORMAT(...)
+#    define LLAMA_ATTRIBUTE_FORMAT(...)
 #endif
 
 //

src/llama-kv-cache.h

@@ -57,7 +57,7 @@ struct llama_kv_cache {
     bool can_shift = false;
 
     // Note: The value of head isn't only used to optimize searching
-    // for a free KV slot. llama_decode_internal also uses it, so it
+    // for a free KV slot. llama_decode_impl also uses it, so it
     // cannot be freely changed after a slot has been allocated.
     uint32_t head = 0;
     uint32_t size = 0;

src/llama-mmap.h

@@ -1,5 +1,6 @@
 #pragma once
 
+#include <cstdint>
 #include <memory>
 #include <vector>
 

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 {

src/unicode.cpp

@@ -708,7 +708,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
     const auto cpts = unicode_cpts_from_utf8(text);
 
     // generate a "collapsed" representation of the text, where all codepoints are replaced by a single byte
-    // ref: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2081479935
+    // ref: https://github.com/ggml-org/llama.cpp/pull/6920#issuecomment-2081479935
     std::string text_collapsed;
     if (need_collapse) {
         // collapse all unicode categories