llama-cpp-bin 9628.0.0__tar.gz → 9637.0.0__tar.gz

{llama_cpp_bin-9628.0.0 → llama_cpp_bin-9637.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-cpp-bin
-Version: 9628.0.0
+Version: 9637.0.0
 Summary: llama.cpp server binary built from source
 Author-email: Vladyslav Tkachenko <vladlearns@gmail.com>
 License: AGPL-3.0-or-later

{llama_cpp_bin-9628.0.0 → llama_cpp_bin-9637.0.0}/llama.cpp/common/chat.cpp

@@ -1979,6 +1979,146 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
     return data;
 }
 
+// Cohere2 MoE (a.k.a. "North Code") parser.
+//
+// The assistant turn is fully marker-wrapped:
+//   <|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>
+//     <|START_THINKING|>{reasoning}<|END_THINKING|>
+//     then EITHER content:    <|START_TEXT|>{content}<|END_TEXT|>
+//          OR     tool calls: <|START_ACTION|>[
+//                                 {"tool_call_id": "0", "tool_name": "f", "parameters": {...}}, ...
+//                             ]<|END_ACTION|>
+//   <|END_OF_TURN_TOKEN|>
+//
+// The generation prompt forces a leading <|START_THINKING|> (when reasoning is enabled, which is
+// the template default), so the model's output continues from *inside* the thinking block. The
+// parser literal therefore only covers the stable <|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|> prefix
+// and the reasoning rule consumes the <|START_THINKING|> ... <|END_THINKING|> markers itself,
+// regardless of whether they came from the generation prompt or the generated text.
+static common_chat_params common_chat_params_init_cohere2moe(const common_chat_template &          tmpl,
+                                                              const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    const std::string TURN_START    = "<|START_OF_TURN_TOKEN|>";
+    const std::string TURN_END      = "<|END_OF_TURN_TOKEN|>";
+    const std::string CHATBOT       = "<|CHATBOT_TOKEN|>";
+    const std::string USER          = "<|USER_TOKEN|>";
+    const std::string SYSTEM        = "<|SYSTEM_TOKEN|>";
+    const std::string THINK_START   = "<|START_THINKING|>";
+    const std::string THINK_END     = "<|END_THINKING|>";
+    const std::string TEXT_START    = "<|START_TEXT|>";
+    const std::string TEXT_END      = "<|END_TEXT|>";
+    const std::string ACTION_START  = "<|START_ACTION|>";
+    const std::string ACTION_END    = "<|END_ACTION|>";
+    const std::string RESULT_START  = "<|START_TOOL_RESULT|>";
+    const std::string RESULT_END    = "<|END_TOOL_RESULT|>";
+
+    // Stable prefix of the generation prompt that precedes the (forced) <|START_THINKING|> marker.
+    const std::string GEN_PREFIX = TURN_START + CHATBOT;
+
+    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt  = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking  = true;
+    data.thinking_start_tag = THINK_START;
+    data.thinking_end_tag   = THINK_END;
+    data.preserved_tokens   = {
+        TURN_START, TURN_END, CHATBOT, USER, SYSTEM,
+        THINK_START, THINK_END,
+        TEXT_START, TEXT_END,
+        ACTION_START, ACTION_END,
+        RESULT_START, RESULT_END,
+    };
+
+    // Split the rendered prompt into per-role message spans. Tool results are rendered with the
+    // system token followed by <|START_TOOL_RESULT|>, so the "tool" delimiter must be listed before
+    // the plain "system" one (it is a strict superset, and the role split tries delimiters in order).
+    data.message_spans = common_chat_split_by_role(data.prompt, {
+        { "assistant", GEN_PREFIX },
+        { "user",      TURN_START + USER },
+        { "tool",      TURN_START + SYSTEM + RESULT_START },
+        { "system",    TURN_START + SYSTEM },
+    });
+
+    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
+    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = GEN_PREFIX + THINK_START + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += THINK_END + TEXT_START + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto generation_prompt = p.literal(GEN_PREFIX);
+        auto end               = p.end();
+
+        // The thinking block is always present (the generation prompt forces <|START_THINKING|>).
+        // When extracting reasoning, capture its body; otherwise keep the whole block (markers
+        // included) inline as content, matching reasoning_format=NONE conventions.
+        common_peg_parser reasoning = p.eps();
+        if (extract_reasoning) {
+            reasoning = p.optional(p.literal(THINK_START) +
+                                   p.reasoning(p.until_one_of({ THINK_END, TEXT_START, ACTION_START })) +
+                                   p.optional(p.literal(THINK_END)));
+        } else {
+            reasoning = p.optional(p.content(p.literal(THINK_START) +
+                                             p.until_one_of({ THINK_END, TEXT_START, ACTION_START }) +
+                                             p.optional(p.literal(THINK_END))));
+        }
+
+        auto text_content = p.literal(TEXT_START) + p.content(p.until(TEXT_END)) + p.optional(p.literal(TEXT_END));
+
+        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+            return generation_prompt + reasoning + text_content + p.optional(p.literal(TURN_END)) + end;
+        }
+
+        auto require_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+        // <|START_ACTION|>[ {"tool_call_id": "0", "tool_name": "f", "parameters": {...}}, ... ]<|END_ACTION|>
+        auto tool_calls = p.standard_json_tools(ACTION_START, ACTION_END, inputs.tools, inputs.parallel_tool_calls,
+                                                /* force_tool_calls = */ true,
+                                                /* name_key         = */ "tool_name",
+                                                /* args_key         = */ "parameters",
+                                                /* array_wrapped    = */ true,
+                                                /* function_is_key  = */ false,
+                                                /* call_id_key      = */ "",
+                                                /* gen_call_id_key  = */ "tool_call_id",
+                                                /* parameters_order = */ { "tool_call_id", "tool_name", "parameters" });
+
+        // Content and tool calls are mutually exclusive in this format.
+        common_peg_parser body = require_tools ? tool_calls : p.choice({ tool_calls, text_content });
+
+        return generation_prompt + reasoning + body + p.optional(p.literal(TURN_END)) + end;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, ACTION_START }
+        };
+    }
+
+    return data;
+}
+
 namespace workaround {
 
 static void map_developer_role_to_system(json & messages) {
@@ -2227,6 +2367,15 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
         return common_chat_params_init_kimi_k2(tmpl, params);
     }
 
+    // Cohere2 MoE / North Code - marker-wrapped format with <|START_TEXT|> content and
+    // <|START_ACTION|> JSON tool calls. <|START_TEXT|> is unique to this template (the older
+    // Command-R templates use <|START_RESPONSE|>).
+    if (src.find("<|START_TEXT|>") != std::string::npos &&
+        src.find("<|START_ACTION|>") != std::string::npos) {
+        LOG_DBG("Using specialized template: Cohere2 MoE\n");
+        return common_chat_params_init_cohere2moe(tmpl, params);
+    }
+
     if (is_lfm2_template(src)) {
         LOG_DBG("Using specialized template: LFM2\n");
         return common_chat_params_init_lfm2(tmpl, params, /* tool_list_tokens = */ true);

{llama_cpp_bin-9628.0.0 → llama_cpp_bin-9637.0.0}/llama.cpp/common/jinja/runtime.cpp

@@ -316,12 +316,22 @@ value filter_expression::execute_impl(context & ctx) {
 
     JJ_DEBUG("Applying filter to %s", input->type().c_str());
 
+    auto set_filter_alias = [](auto & filter_id) {
+        if (filter_id == "count") {
+            filter_id = "length";
+        } else if (filter_id == "d") {
+            filter_id = "default";
+        } else if (filter_id == "e") {
+            filter_id = "escape";
+        } else if (filter_id == "trim") {
+            filter_id = "strip";
+        }
+    };
+
     if (is_stmt<identifier>(filter)) {
         auto filter_id = cast_stmt<identifier>(filter)->val;
 
-        if (filter_id == "trim") {
-            filter_id = "strip"; // alias
-        }
+        set_filter_alias(filter_id);
         JJ_DEBUG("Applying filter '%s' to %s", filter_id.c_str(), input->type().c_str());
         // TODO: Refactor filters so this coercion can be done automatically
         if (!input->is_undefined() && !is_val<value_string>(input) && (
@@ -345,9 +355,7 @@ value filter_expression::execute_impl(context & ctx) {
         }
         auto filter_id = cast_stmt<identifier>(call->callee)->val;
 
-        if (filter_id == "trim") {
-            filter_id = "strip"; // alias
-        }
+        set_filter_alias(filter_id);
         JJ_DEBUG("Applying filter '%s' with arguments to %s", filter_id.c_str(), input->type().c_str());
         func_args args(ctx);
         for (const auto & arg_expr : call->args) {

{llama_cpp_bin-9628.0.0 → llama_cpp_bin-9637.0.0}/llama.cpp/src/llama-vocab.cpp

@@ -2280,7 +2280,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 clean_spaces = false;
                 ignore_merges = true;
             } else if (
-                tokenizer_pre == "tiny_aya") {
+                tokenizer_pre == "tiny_aya" ||
+                tokenizer_pre == "cohere2moe") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_TINY_AYA;
                 clean_spaces = false;
             } else if (

{llama_cpp_bin-9628.0.0 → llama_cpp_bin-9637.0.0}/llama.cpp/tests/test-chat.cpp

@@ -2644,6 +2644,100 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .run();
     }
 
+    {
+        // Cohere2 MoE (North Code) - dedicated parser.
+        // Marker-wrapped format: <|START_THINKING|>...<|END_THINKING|> then either
+        // <|START_TEXT|>...<|END_TEXT|> (content) or <|START_ACTION|>[json]<|END_ACTION|> (tools).
+        // The generation prompt forces a leading <|START_THINKING|>, so model output begins inside
+        // the thinking block: test inputs start with the reasoning body, not the <|START_THINKING|> tag.
+        auto tst = peg_tester("models/templates/Cohere2MoE.jinja", detailed_debug);
+
+        // Content with reasoning, extracted.
+        tst.test("I'm\nthinking<|END_THINKING|><|START_TEXT|>Hello, world!\nWhat's up?<|END_TEXT|>")
+            .reasoning_format(COMMON_REASONING_FORMAT_DEEPSEEK)
+            .expect(message_assist_thoughts)
+            .run();
+
+        // Content with reasoning, reasoning_format=NONE -> thinking kept inline in content (markers preserved).
+        tst.test("I'm\nthinking<|END_THINKING|><|START_TEXT|>Hello, world!\nWhat's up?<|END_TEXT|>")
+            .expect(message_assist_thoughts_unparsed_r7b)
+            .run();
+
+        // Content with empty thinking block.
+        tst.test("<|END_THINKING|><|START_TEXT|>Hello, world!\nWhat's up?<|END_TEXT|>")
+            .reasoning_format(COMMON_REASONING_FORMAT_DEEPSEEK)
+            .expect(message_assist)
+            .run();
+
+        // Single tool call with reasoning.
+        tst.test(
+               "I'm\nthinking<|END_THINKING|>"
+               "<|START_ACTION|>[\n"
+               "    {\"tool_call_id\": \"0\", \"tool_name\": \"special_function\", \"parameters\": {\"arg1\": 1}}\n"
+               "]<|END_ACTION|>")
+            .reasoning_format(COMMON_REASONING_FORMAT_DEEPSEEK)
+            .tools({ special_function_tool })
+            .expect(message_assist_thoughts_call_idx)
+            .run();
+
+        // Single tool call, empty thinking block (no reasoning content).
+        tst.test(
+               "<|END_THINKING|>"
+               "<|START_ACTION|>[\n"
+               "    {\"tool_call_id\": \"0\", \"tool_name\": \"special_function\", \"parameters\": {\"arg1\": 1}}\n"
+               "]<|END_ACTION|>")
+            .reasoning_format(COMMON_REASONING_FORMAT_DEEPSEEK)
+            .tools({ special_function_tool })
+            .expect(message_assist_call_idx)
+            .run();
+
+        // Tool call with an array argument (todo_list).
+        tst.test(
+               "<|END_THINKING|>"
+               "<|START_ACTION|>[\n"
+               "    {\"tool_call_id\": \"0\", \"tool_name\": \"todo_list\", \"parameters\": {\"todos\": [\"buy milk\", \"walk dog\"]}}\n"
+               "]<|END_ACTION|>")
+            .reasoning_format(COMMON_REASONING_FORMAT_DEEPSEEK)
+            .tools({ todo_list })
+            .expect(simple_assist_msg("", "", "todo_list", "{\"todos\": [\"buy milk\", \"walk dog\"]}", "0"))
+            .run();
+
+        // Parallel tool calls with reasoning.
+        tst.test(
+               "I'm\nthinking<|END_THINKING|>"
+               "<|START_ACTION|>[\n"
+               "    {\"tool_call_id\": \"0\", \"tool_name\": \"special_function\", \"parameters\": {\"arg1\": 1}},\n"
+               "    {\"tool_call_id\": \"1\", \"tool_name\": \"python\", \"parameters\": {\"code\": \"print('hey')\"}}\n"
+               "]<|END_ACTION|>")
+            .reasoning_format(COMMON_REASONING_FORMAT_DEEPSEEK)
+            .parallel_tool_calls(true)
+            .tools({ special_function_tool, python_tool })
+            .expect_reasoning("I'm\nthinking")
+            .expect_tool_calls({
+                { "special_function", R"({"arg1": 1})", "0" },
+                { "python", "{\"code\": \"print('hey')\"}", "1" },
+            })
+            .run();
+
+        // Tools available but the model answers with content instead of calling a tool.
+        tst.test("I'm\nthinking<|END_THINKING|><|START_TEXT|>Hello, world!\nWhat's up?<|END_TEXT|>")
+            .reasoning_format(COMMON_REASONING_FORMAT_DEEPSEEK)
+            .tools({ special_function_tool })
+            .expect(message_assist_thoughts)
+            .run();
+
+        // Partial tool call (streaming): name/id resolved before arguments arrive.
+        tst.test(
+               "I'm\nthinking<|END_THINKING|>"
+               "<|START_ACTION|>[\n"
+               "    {\"tool_call_id\": \"0\", \"tool_name\": \"special_function\", ")
+            .reasoning_format(COMMON_REASONING_FORMAT_DEEPSEEK)
+            .tools({ special_function_tool })
+            .is_partial(true)
+            .expect(message_assist_thoughts_partial_call)
+            .run();
+    }
+
     {
         // Google Gemma 2 2B - does not support tool calling
         auto tst = peg_tester("models/templates/google-gemma-2-2b-it.jinja");

{llama_cpp_bin-9628.0.0 → llama_cpp_bin-9637.0.0}/llama.cpp/tests/test-jinja.cpp

@@ -601,8 +601,8 @@ static void test_filters(testing & t) {
         "hello jinja"
     );
 
-    test_template(t, "length list",
-        "{{ items|length }}",
+    test_template(t, "length (count alias) list",
+        "{{ items|count }}",
         {{"items", json::array({1, 2, 3})}},
         "3"
     );
@@ -711,8 +711,8 @@ static void test_filters(testing & t) {
         "fallback"
     );
 
-    test_template(t, "default with falsy value",
-        "{{ ''|default('fallback', true) }}",
+    test_template(t, "default (d alias) with falsy value",
+        "{{ ''|d('fallback', true) }}",
         json::object(),
         "fallback"
     );

{llama_cpp_bin-9628.0.0 → llama_cpp_bin-9637.0.0}/llama.cpp/tools/cli/cli.cpp

@@ -97,11 +97,18 @@ struct cli_context {
                 task.params.chat_parser_params.parser.load(chat_params.parser);
             }
 
+            // Copy the preserved tokens into the sampling params
+            const llama_vocab * vocab = llama_model_get_vocab(
+                llama_get_model(ctx_server.get_llama_context()));
+            for (const auto & token : chat_params.preserved_tokens) {
+                auto ids = common_tokenize(vocab, token, false, true);
+                if (ids.size() == 1) {
+                    task.params.sampling.preserved_tokens.insert(ids[0]);
+                }
+            }
+
             // reasoning budget sampler
             if (!chat_params.thinking_end_tag.empty()) {
-                const llama_vocab * vocab = llama_model_get_vocab(
-                    llama_get_model(ctx_server.get_llama_context()));
-
                 task.params.sampling.reasoning_budget_tokens = defaults.sampling.reasoning_budget_tokens;
                 task.params.sampling.generation_prompt = chat_params.generation_prompt;
 

{llama_cpp_bin-9628.0.0 → llama_cpp_bin-9637.0.0}/src/llama_cpp_bin/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '9628.0.0'
-__version_tuple__ = version_tuple = (9628, 0, 0)
+__version__ = version = '9637.0.0'
+__version_tuple__ = version_tuple = (9637, 0, 0)
 
-__commit_id__ = commit_id = 'g230bc8dab'
+__commit_id__ = commit_id = 'g3c4c0b8d3'

{llama_cpp_bin-9628.0.0 → llama_cpp_bin-9637.0.0}/src/llama_cpp_bin.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-cpp-bin
-Version: 9628.0.0
+Version: 9637.0.0
 Summary: llama.cpp server binary built from source
 Author-email: Vladyslav Tkachenko <vladlearns@gmail.com>
 License: AGPL-3.0-or-later