llama: update to commit 71e90e88 (#10192)

llama/llama.cpp/src/llama-chat.cpp

@@ -4,6 +4,7 @@
 
 #include <map>
 #include <sstream>
+#include <algorithm>
 
 #if __cplusplus >= 202000L
     #define LU8(x) (const char*)(u8##x)
@@ -58,6 +59,9 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "granite",           LLM_CHAT_TEMPLATE_GRANITE           },
     { "gigachat",          LLM_CHAT_TEMPLATE_GIGACHAT          },
     { "megrez",            LLM_CHAT_TEMPLATE_MEGREZ            },
+    { "yandex",            LLM_CHAT_TEMPLATE_YANDEX            },
+    { "bailing",           LLM_CHAT_TEMPLATE_BAILING           },
+    { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -167,6 +171,12 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_GIGACHAT;
     } else if (tmpl_contains("<|role_start|>")) {
         return LLM_CHAT_TEMPLATE_MEGREZ;
+    } else if (tmpl_contains(" Ассистент:")) {
+        return LLM_CHAT_TEMPLATE_YANDEX;
+    } else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("'HUMAN'")) {
+        return LLM_CHAT_TEMPLATE_BAILING;
+    } else if (tmpl_contains("<|header_start|>") && tmpl_contains("<|header_end|>")) {
+        return LLM_CHAT_TEMPLATE_LLAMA4;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -566,7 +576,51 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|role_start|>assistant<|role_end|>";
         }
-    } else {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_YANDEX) {
+        // Yandex template ("\n\n" is defined as EOT token)
+
+        ss << "<s>";
+
+        for (size_t i = 0; i < chat.size(); i++) {
+            std::string role(chat[i]->role);
+            if (role == "user") {
+                ss << " Пользователь: " << chat[i]->content << "\n\n";
+            } else if (role == "assistant") {
+                ss << " Ассистент: " << chat[i]->content << "\n\n";
+            }
+        }
+
+        // Add generation prompt if needed
+        if (add_ass) {
+            ss << " Ассистент:[SEP]";
+        }
+    }  else if (tmpl == LLM_CHAT_TEMPLATE_BAILING) {
+        // Bailing (Ling) template
+        for (auto message : chat) {
+            std::string role(message->role);
+
+            if (role == "user") {
+                role = "HUMAN";
+            } else {
+                std::transform(role.begin(), role.end(), role.begin(), ::toupper);
+            }
+
+            ss << "<role>" << role << "</role>" << message->content;
+        }
+
+        if (add_ass) {
+            ss << "<role>ASSISTANT</role>";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_LLAMA4) {
+        // Llama 4
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|header_start|>" << role << "<|header_end|>\n\n" << trim(message->content) << "<|eot|>";
+        }
+        if (add_ass) {
+            ss << "<|header_start|>assistant<|header_end|>\n\n";
+        }
+    }  else {
         // template not supported
         return -1;
     }
@@ -584,4 +638,3 @@ int32_t llama_chat_builtin_templates(const char ** output, size_t len) {
     }
     return (int32_t) LLM_CHAT_TEMPLATES.size();
 }
-

llama/llama.cpp/src/llama-chat.h

@@ -38,6 +38,9 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_GRANITE,
     LLM_CHAT_TEMPLATE_GIGACHAT,
     LLM_CHAT_TEMPLATE_MEGREZ,
+    LLM_CHAT_TEMPLATE_YANDEX,
+    LLM_CHAT_TEMPLATE_BAILING,
+    LLM_CHAT_TEMPLATE_LLAMA4,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

llama/llama.cpp/src/llama-cparams.h

@@ -30,6 +30,7 @@ struct llama_cparams {
     bool flash_attn;
     bool no_perf;
     bool cross_attn;
+    bool warmup;
 
     enum llama_pooling_type pooling_type;
 

llama/llama.cpp/src/llama-grammar.h

@@ -3,6 +3,7 @@
 #include "llama.h"
 
 #include <map>
+#include <regex>
 #include <string>
 #include <vector>
 
@@ -105,6 +106,11 @@ struct llama_grammar_parser {
     void print(FILE * file);
 };
 
+struct llama_grammar_trigger_pattern {
+    std::string pattern;
+    std::regex  regex;
+};
+
 struct llama_grammar {
     // note: allow null vocab for testing (not great)
     const llama_vocab * vocab;
@@ -122,7 +128,10 @@ struct llama_grammar {
     bool                     awaiting_trigger = false; // Initialized to true for lazy grammars only
     std::string              trigger_buffer;           // Output buffered by lazy grammar. Will be cleared once trigger is found.
     std::vector<llama_token> trigger_tokens;           // Tokens that trigger a lazy grammar, or tokens to force printing of (even if special).
-    std::vector<std::string> trigger_words;
+    std::vector<llama_grammar_trigger_pattern>
+                             trigger_patterns;         // Regular expressions that trigger a lazy grammar. Must be a full match of the entire generated
+                                                       // string, and the grammar will be given the string from the first match group onwards.
+
 };
 
 //
@@ -141,8 +150,8 @@ struct llama_grammar * llama_grammar_init_impl(
                       const char * grammar_str,
                       const char * grammar_root,
                               bool lazy,
-                     const char ** trigger_words,
-                            size_t num_trigger_words,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns,
                const llama_token * trigger_tokens,
                             size_t num_trigger_tokens);
 

llama/llama.cpp/src/llama-hparams.cpp

@@ -2,8 +2,6 @@
 
 #include "ggml.h"
 
-#include <algorithm>
-
 uint32_t llama_hparams::n_head(uint32_t il) const {
     if (il < n_layer) {
         return n_head_arr[il];
@@ -80,6 +78,14 @@ bool llama_hparams::n_bskcn(uint32_t n, uint32_t il) const {
     GGML_ABORT("fatal error");
 }
 
+bool llama_hparams::is_swa(uint32_t il) const {
+    if (il < n_layer) {
+        return n_swa > 0 && n_swa_pattern > 0 && il % n_swa_pattern < (n_swa_pattern - 1);
+    }
+
+    GGML_ABORT("fatal error");
+}
+
 bool llama_hparams::cross_attention_layers(uint32_t il) const {
     return std::find(cross_attn_layers.begin(), cross_attn_layers.end(), il) != cross_attn_layers.end();
-}
\ No newline at end of file
+}

llama/llama.cpp/src/llama-io.cpp

+#include "llama-io.h"
+
+void llama_io_write_i::write_string(const std::string & str) {
+    uint32_t str_size = str.size();
+
+    write(&str_size,  sizeof(str_size));
+    write(str.data(), str_size);
+}
+
+void llama_io_read_i::read_string(std::string & str) {
+    uint32_t str_size;
+    read_to(&str_size, sizeof(str_size));
+
+    str.assign((const char *) read(str_size), str_size);
+}

llama/llama.cpp/src/llama-memory.cpp

+#include "llama-memory.h"