Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	.devops/llama-server.Dockerfile
#	README.md
#	flake.lock
#	ggml/src/ggml-vulkan.cpp
#	ggml/src/vulkan-shaders/concat.comp
#	ggml/src/vulkan-shaders/pad.comp
#	ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
#	scripts/sync-ggml-am.sh
#	scripts/sync-ggml.last
#	src/llama.cpp
#	tests/test-backend-ops.cpp

src/llama-vocab.cpp

@@ -18,17 +18,14 @@
 //
 
 static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
-    std::string result;
-    for (size_t pos = 0; ; pos += search.length()) {
-        auto new_pos = s.find(search, pos);
-        if (new_pos == std::string::npos) {
-            result += s.substr(pos, s.size() - pos);
-            break;
-        }
-        result += s.substr(pos, new_pos - pos) + replace;
-        pos = new_pos;
+    if (search.empty()) {
+        return; // Avoid infinite loop if 'search' is an empty string
+    }
+    size_t pos = 0;
+    while ((pos = s.find(search, pos)) != std::string::npos) {
+        s.replace(pos, search.length(), replace);
+        pos += replace.length();
     }
-    s = std::move(result);
 }
 
 LLAMA_ATTRIBUTE_FORMAT(1, 2)
@@ -1044,6 +1041,9 @@ struct llm_tokenizer_ugm {
      * the best tokenization.
     */
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
+        // get current size of output (for reversal later)
+        size_t output_size = output.size();
+
         // normalize the input first
         std::string normalized;
         normalize(text, &normalized);
@@ -1123,7 +1123,7 @@ struct llm_tokenizer_ugm {
         }
 
         // reverse the output since we added tokens starting from the end of the input
-        std::reverse(output.begin(), output.end());
+        std::reverse(output.begin() + output_size, output.end());
     }
 
 private:
@@ -1701,7 +1701,8 @@ llama_token_attr llama_token_get_attr_impl(const struct llama_vocab & vocab, lla
 bool llama_token_is_eog_impl(const struct llama_vocab & vocab, llama_token token) {
     return token != -1 && (
         token == llama_token_eos_impl(vocab) ||
-        token == llama_token_eot_impl(vocab)
+        token == llama_token_eot_impl(vocab) ||
+        token == llama_token_eom_impl(vocab)
     );
 }
 
@@ -1757,6 +1758,10 @@ llama_token llama_token_eot_impl(const struct llama_vocab & vocab) {
     return vocab.special_eot_id;
 }
 
+llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
+    return vocab.special_eom_id;
+}
+
 int32_t llama_tokenize_impl(
     const struct llama_vocab & vocab,
                   const char * text,

src/llama-vocab.h

@@ -45,6 +45,7 @@ struct llama_vocab {
     id special_suffix_id = -1;
     id special_middle_id = -1;
     id special_eot_id    = -1; // TODO: move above after "eos_id", and here add "file separator" token
+    id special_eom_id    = -1;
 
     // tokenizer flags
     bool tokenizer_add_space_prefix = false;
@@ -101,6 +102,7 @@ llama_token llama_token_prefix_impl(const struct llama_vocab & vocab);
 llama_token llama_token_middle_impl(const struct llama_vocab & vocab);
 llama_token llama_token_suffix_impl(const struct llama_vocab & vocab);
 llama_token llama_token_eot_impl   (const struct llama_vocab & vocab);
+llama_token llama_token_eom_impl   (const struct llama_vocab & vocab);
 
 int32_t llama_tokenize_impl(
         const struct llama_vocab & vocab,

src/llama.cpp

@@ -357,6 +357,7 @@ enum llm_kv {
     LLM_KV_TOKENIZER_SUFFIX_ID,
     LLM_KV_TOKENIZER_MIDDLE_ID,
     LLM_KV_TOKENIZER_EOT_ID,
+    LLM_KV_TOKENIZER_EOM_ID,
 
     LLM_KV_ADAPTER_TYPE,
     LLM_KV_ADAPTER_LORA_ALPHA,
@@ -454,6 +455,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_TOKENIZER_SUFFIX_ID,            "tokenizer.ggml.suffix_token_id"          },
     { LLM_KV_TOKENIZER_MIDDLE_ID,            "tokenizer.ggml.middle_token_id"          },
     { LLM_KV_TOKENIZER_EOT_ID,               "tokenizer.ggml.eot_token_id"             },
+    { LLM_KV_TOKENIZER_EOM_ID,               "tokenizer.ggml.eom_token_id"             },
 
     { LLM_KV_ADAPTER_TYPE,                  "adapter.type"       },
     { LLM_KV_ADAPTER_LORA_ALPHA,            "adapter.lora.alpha" },
@@ -5623,6 +5625,7 @@ static void llm_load_vocab(
             { LLM_KV_TOKENIZER_SUFFIX_ID, vocab.special_suffix_id },
             { LLM_KV_TOKENIZER_MIDDLE_ID, vocab.special_middle_id },
             { LLM_KV_TOKENIZER_EOT_ID,    vocab.special_eot_id    },
+            { LLM_KV_TOKENIZER_EOM_ID,    vocab.special_eom_id    },
         };
 
         for (const auto & it : special_token_types) {
@@ -5675,6 +5678,17 @@ static void llm_load_vocab(
                 }
             }
         }
+
+        // find EOM token: "<|eom_id|>"
+        //
+        // TODO: convert scripts should provide this token through the KV metadata LLAMA_KV_TOKENIZER_EOM_ID
+        //       for now, we apply this workaround to find the EOM token based on its text
+        if (vocab.special_eom_id == -1) {
+            const auto & t = vocab.token_to_id.find("<|eom_id|>");
+            if (t != vocab.token_to_id.end()) {
+                vocab.special_eom_id = t->second;
+            }
+        }
     }
 
     // build special tokens cache