llama-cpp-bin 9536.0.0__tar.gz → 9544.0.0__tar.gz

{llama_cpp_bin-9536.0.0 → llama_cpp_bin-9544.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-cpp-bin
-Version: 9536.0.0
+Version: 9544.0.0
 Summary: llama.cpp server binary built from source
 Author-email: Vladyslav Tkachenko <vladlearns@gmail.com>
 License: AGPL-3.0-or-later

{llama_cpp_bin-9536.0.0 → llama_cpp_bin-9544.0.0}/llama.cpp/common/chat.cpp

@@ -1625,8 +1625,17 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
     const std::string THINK_END       = "</think>";
     const std::string GEN_PROMPT      = "<|im_start|>assistant\n";
 
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    // Copy reasoning to the "thinking" field the template expects
+    auto adjusted_messages = json::array();
+    for (auto msg : inputs.messages) {
+        if (msg.contains("reasoning_content") && msg.at("reasoning_content").is_string()) {
+            msg["thinking"] = msg.at("reasoning_content");
+        }
+        adjusted_messages.push_back(msg);
+    }
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs, adjusted_messages);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, adjusted_messages);
     data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
     data.supports_thinking = true;
     data.preserved_tokens  = { TOOL_CALL_START, TOOL_CALL_END, THINK_START, THINK_END };
@@ -1639,7 +1648,9 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
     data.thinking_end_tag   = THINK_END;
 
     auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
-    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    // Gate by reasoning format and whether the template supports <think>
+    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE &&
+                             tmpl.source().find(THINK_START) != std::string::npos;
     auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
 
     if (inputs.has_continuation()) {
@@ -1658,7 +1669,7 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
         auto end = p.end();
 
         auto reasoning = p.eps();
-        if (extract_reasoning && inputs.enable_thinking) {
+        if (extract_reasoning) {
             reasoning = p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
         }
 

{llama_cpp_bin-9536.0.0 → llama_cpp_bin-9544.0.0}/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -6349,6 +6349,15 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     }
 #endif
 
+#if defined(VK_NV_cooperative_matrix2)
+    VkPhysicalDeviceCooperativeMatrix2FeaturesNV coopmat2_features {};
+    coopmat2_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_2_FEATURES_NV;
+    if (coopmat2_support) {
+        last_struct->pNext = (VkBaseOutStructure *)&coopmat2_features;
+        last_struct = (VkBaseOutStructure *)&coopmat2_features;
+    }
+#endif
+
     VkPhysicalDeviceCooperativeMatrixDecodeVectorFeaturesNV coopmat2_decode_vector_features {};
     coopmat2_decode_vector_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_DECODE_VECTOR_FEATURES_NV;
     if (coopmat2_decode_vector_support) {
@@ -6380,6 +6389,19 @@ static void ggml_vk_print_gpu_info(size_t idx) {
 #endif
                    && ggml_vk_khr_cooperative_matrix_support(props2.properties, driver_props, device_architecture);
 
+#if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+    coopmat2_support = coopmat2_support &&
+                       coopmat2_features.cooperativeMatrixWorkgroupScope &&
+                       coopmat2_features.cooperativeMatrixFlexibleDimensions &&
+                       coopmat2_features.cooperativeMatrixReductions &&
+                       coopmat2_features.cooperativeMatrixConversions &&
+                       coopmat2_features.cooperativeMatrixPerElementOperations &&
+                       coopmat2_features.cooperativeMatrixTensorAddressing &&
+                       coopmat2_features.cooperativeMatrixBlockLoads;
+#else
+    coopmat2_support = false;
+#endif
+
     coopmat2_decode_vector_support = coopmat2_decode_vector_support && coopmat2_decode_vector_features.cooperativeMatrixDecodeVector;
 #if !defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR_GLSLC_SUPPORT)
     coopmat2_decode_vector_support = false;

{llama_cpp_bin-9536.0.0 → llama_cpp_bin-9544.0.0}/llama.cpp/src/llama-context.cpp

@@ -341,7 +341,7 @@ llama_context::llama_context(
         // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
         bool pipeline_parallel =
             model.n_devices() > 1 &&
-            model.n_gpu_layers() > model.hparams.n_layer() &&
+            model.n_gpu_layers() > model.hparams.n_layer_all &&
             model.split_mode() == LLAMA_SPLIT_MODE_LAYER &&
             cparams.offload_kqv &&
             !model.has_tensor_overrides();
@@ -2351,7 +2351,7 @@ llm_graph_cb llama_context::graph_get_cb() const {
 
         // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
         // FIXME: fix in ggml_backend_sched
-        const bool full_offload = model.n_gpu_layers() > model.hparams.n_layer();
+        const bool full_offload = model.n_gpu_layers() > model.hparams.n_layer_all;
         if (ubatch.n_tokens < 32 || full_offload) {
             if (il != -1 && strcmp(name, "norm") == 0) {
                 const auto & dev_layer = model.dev_layer(il);

{llama_cpp_bin-9536.0.0 → llama_cpp_bin-9544.0.0}/llama.cpp/src/llama-model.cpp

@@ -1205,7 +1205,7 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
     const auto & use_mlock    = params.use_mlock;
     const auto & tensor_split = params.tensor_split;
 
-    const int n_layer = hparams.n_layer_all;
+    const int n_layer_all = hparams.n_layer_all;
     const int n_gpu_layers = this->n_gpu_layers();
 
     const bool use_mmap_buffer = true;
@@ -1262,10 +1262,10 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
         splits[i] /= split_sum;
     }
 
-    const int i_gpu_start = std::max(n_layer + 1 - n_gpu_layers, 0);
-    const int act_gpu_layers = devices.empty() ? 0 : std::min(n_gpu_layers, n_layer + 1);
+    const int i_gpu_start = std::max(n_layer_all + 1 - n_gpu_layers, 0);
+    const int act_gpu_layers = devices.empty() ? 0 : std::min(n_gpu_layers, n_layer_all + 1);
     auto get_layer_buft_list = [&](int il) -> llama_model::impl::layer_dev {
-        const bool is_swa = il < n_layer && hparams.is_swa(il);
+        const bool is_swa = il < n_layer_all && hparams.is_swa(il);
         if (il < i_gpu_start || (il - i_gpu_start) >= act_gpu_layers) {
             LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s, is_swa = %d\n", il, ggml_backend_dev_name(cpu_dev), is_swa);
             return {cpu_dev, &pimpl->cpu_buft_list};
@@ -1281,13 +1281,13 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
     pimpl->dev_input = { cpu_dev, &pimpl->cpu_buft_list };
 
     // assign the repeating layers to the devices according to the splits
-    pimpl->dev_layer.resize(n_layer);
-    for (int il = 0; il < n_layer; ++il) {
+    pimpl->dev_layer.resize(n_layer_all);
+    for (int il = 0; il < n_layer_all; ++il) {
         pimpl->dev_layer[il] = get_layer_buft_list(il);
     }
 
     // assign the output layer
-    pimpl->dev_output = get_layer_buft_list(n_layer);
+    pimpl->dev_output = get_layer_buft_list(n_layer_all);
 
     const auto TENSOR_NOT_REQUIRED = llama_model_loader::TENSOR_NOT_REQUIRED;
 
@@ -1303,14 +1303,14 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
             throw std::runtime_error("model has expert layers but no expert layers are used");
         }
 
-        layers.resize(n_layer);
+        layers.resize(n_layer_all);
 
         // call the per-model loading function
         load_arch_tensors(ml);
 
         // generic pass: load optional per-tensor/per-expert ".scale" tensors (e.g. NVFP4 scale2)
         // this avoids having to add scale loading to every architecture
-        for (int i = 0; i < n_layer; ++i) {
+        for (int i = 0; i < n_layer_all; ++i) {
             auto & layer = layers[i];
 
             // attention weight scales (per-tensor, shape {1})
@@ -1568,7 +1568,7 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
     }
 
     if (llama_supports_gpu_offload()) {
-        const int n_gpu = std::min(n_gpu_layers, n_layer);
+        const int n_gpu = std::min(n_gpu_layers, n_layer_all);
 
         int n_repeating = n_gpu;
         if (n_repeating > 0) {
@@ -1577,8 +1577,8 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
         }
         LLAMA_LOG_INFO("%s: offloading %d repeating layers to GPU\n", __func__, n_repeating);
 
-        const int max_backend_supported_layers = n_layer + 1;
-        const int max_offloadable_layers       = n_layer + 1;
+        const int max_backend_supported_layers = n_layer_all + 1;
+        const int max_offloadable_layers       = n_layer_all + 1;
 
         LLAMA_LOG_INFO("%s: offloaded %d/%d layers to GPU\n", __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers);
     }

{llama_cpp_bin-9536.0.0 → llama_cpp_bin-9544.0.0}/llama.cpp/tests/test-chat.cpp

@@ -1825,6 +1825,104 @@ static void test_convert_responses_to_chatcmpl() {
     }
 }
 
+// Shared LFM2 parser cases - all variants use one output format and parser
+static void test_lfm2_parser(const std::string & template_path, bool detailed_debug) {
+    auto tst = peg_tester(template_path, detailed_debug);
+
+    // Basic content only
+    tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+
+    // Single tool call without reasoning
+    tst.test("<|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|>")
+        .tools({ special_function_tool })
+        .expect(message_assist_call)
+        .run();
+
+    // Tool call with string argument
+    tst.test("<|tool_call_start|>[get_time(city=\"XYZCITY\")]<|tool_call_end|>")
+        .tools({ get_time_tool })
+        .expect(message_with_tool_calls("get_time", "{\"city\":\"XYZCITY\"}"))
+        .run();
+
+    // Python literals become JSON
+    tst.test("<|tool_call_start|>[toggle(enabled=True)]<|tool_call_end|>")
+        .tools({ toggle_tool })
+        .expect(message_with_tool_calls("toggle", R"({"enabled": true})"))
+        .run();
+
+    tst.test("<|tool_call_start|>[set_nullable(value=None)]<|tool_call_end|>")
+        .tools({ nullable_tool })
+        .expect(message_with_tool_calls("set_nullable", R"({"value": null})"))
+        .run();
+
+    // Nested Python literal
+    tst.test("<|tool_call_start|>[set_config(config={\"enabled\": True, \"count\": 3})]<|tool_call_end|>")
+        .tools({ config_tool })
+        .expect(message_with_tool_calls("set_config", R"({"config": {"enabled": true, "count": 3}})"))
+        .run();
+
+    // JSON literals are accepted too
+    tst.test("<|tool_call_start|>[set_config(config={\"enabled\": true, \"note\": null})]<|tool_call_end|>")
+        .tools({ config_tool })
+        .expect(message_with_tool_calls("set_config", R"({"config": {"enabled": true, "note": null}})"))
+        .run();
+
+    // Dotted function name with structured args
+    tst.test("<|tool_call_start|>[Calendar.create_event(title=\"demo\", participants=[\"Alice\", \"Bob\"], "
+             "metadata={\"priority\": \"high\", \"reminder\": true})]<|tool_call_end|>")
+        .tools({ calendar_create_event_tool })
+        .expect(message_with_tool_calls(
+            "Calendar.create_event",
+            R"({"title": "demo", "participants": ["Alice", "Bob"], "metadata": {"priority": "high", "reminder": true}})"))
+        .run();
+
+    // Markdown links stay content
+    tst.test("Use this format: [link text](url). Example: [Wikipedia](https://www.wikipedia.org).")
+        .tools({ get_time_tool })
+        .expect(simple_assist_msg("Use this format: [link text](url). Example: [Wikipedia](https://www.wikipedia.org)."))
+        .run();
+
+    // Python tool with multiline code in string
+    tst.test("<|tool_call_start|>[python(code=\"def hello():\\n    print('hey')\")]<|tool_call_end|>")
+        .tools({ python_tool })
+        .expect_tool_calls({
+            { "python", R"#({"code": "def hello():\\n    print('hey')"})#", "" }
+        })
+        .run();
+
+    // Content before tool call (no reasoning)
+    tst.test("Let me check the time.<|tool_call_start|>[get_time(city=\"Paris\")]<|tool_call_end|>")
+        .tools({ get_time_tool })
+        .expect(message_with_reasoning_content_and_multiple_tool_calls(
+            "", "Let me check the time.", { { "get_time", "{\"city\":\"Paris\"}" } }
+        ))
+        .run();
+
+    // Multiple tool calls (parallel)
+    tst.test("<|tool_call_start|>[special_function(arg1=1), special_function_with_opt(arg1=1, arg2=2)]<|tool_call_end|>")
+        .parallel_tool_calls(true)
+        .tools({ special_function_tool, special_function_tool_with_optional_param })
+        .expect_tool_calls({
+            { "special_function", R"({"arg1": 1})", {} },
+            { "special_function_with_opt", R"({"arg1": 1, "arg2": 2})", {} },
+        })
+        .run();
+
+    // Partial tool call (streaming)
+    tst.test("<|tool_call_start|>[special_function(arg1=")
+        .tools({ special_function_tool })
+        .is_partial(true)
+        .expect(simple_assist_msg("", "", "special_function", "{\"arg1\": "))
+        .run();
+
+    // Tool call with empty arguments
+    tst.test("<|tool_call_start|>[empty_args()]<|tool_call_end|>")
+        .tools({ empty_args_tool })
+        .expect(simple_assist_msg("", "", "empty_args", "{}"))
+        .run();
+
+}
+
 static void test_template_output_peg_parsers(bool detailed_debug) {
     LOG_DBG("%s\n", __func__);
 
@@ -4038,49 +4136,30 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .run();
     }
 
-    // LFM2-8B-A1B tests - uses <|tool_list_start|>/<|tool_list_end|> and <|tool_call_start|>[name(args)]<|tool_call_end|>
-    {
-        auto tst = peg_tester("models/templates/LFM2-8B-A1B.jinja", detailed_debug);
-
-        // Basic content only
-        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+    for (const char * tmpl : {
+             "models/templates/LFM2-8B-A1B.jinja",
+             "models/templates/LFM2.5-Instruct.jinja",
+             "models/templates/LFM2.5-8B-A1B.jinja",
+         }) {
+        test_lfm2_parser(tmpl, detailed_debug);
+    }
 
-        // Single tool call without reasoning
-        tst.test("<|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|>")
-            .tools({ special_function_tool })
-            .expect(message_assist_call)
-            .run();
+    // Thinking cases only apply to LFM2.5-8B-A1B, the one LFM2 template that emits <think>
+    {
+        auto tst = peg_tester("models/templates/LFM2.5-8B-A1B.jinja", detailed_debug);
 
-        // Tool call with string argument
-        tst.test("<|tool_call_start|>[get_time(city=\"XYZCITY\")]<|tool_call_end|>")
-            .tools({ get_time_tool })
-            .expect(message_with_tool_calls("get_time", "{\"city\":\"XYZCITY\"}"))
-            .run();
+        // Reasoning is parsed independent of enable_thinking
 
-        // Tool call with reasoning (enable_thinking=true)
+        // Tool call with reasoning
         tst.test("<think>I'm\nthinking</think><|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|>")
-            .enable_thinking(true)
             .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
             .tools({ special_function_tool })
             .expect(message_assist_call_thoughts)
             .run();
 
-        // Multiple tool calls (parallel)
-        tst.test("<|tool_call_start|>[special_function(arg1=1), special_function_with_opt(arg1=1, arg2=2)]<|tool_call_end|>")
-            .parallel_tool_calls(true)
-            .tools({
-                special_function_tool, special_function_tool_with_optional_param
-            })
-            .expect_tool_calls({
-                { "special_function", R"({"arg1": 1})", {} },
-                { "special_function_with_opt", R"({"arg1": 1, "arg2": 2})", {} },
-            })
-            .run();
-
         // Tool call with reasoning and content
         tst.test("<think>I need to call a function</think>"
                  "Let me check the time.<|tool_call_start|>[get_time(city=\"Paris\")]<|tool_call_end|>")
-            .enable_thinking(true)
             .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
             .tools({ get_time_tool })
             .expect(message_with_reasoning_content_and_multiple_tool_calls(
@@ -4088,32 +4167,9 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             ))
             .run();
 
-        // Python tool with multiline code in string
-        tst.test("<|tool_call_start|>[python(code=\"def hello():\\n    print('hey')\")]<|tool_call_end|>")
-            .tools({ python_tool })
-            .expect_tool_calls({
-                { "python", R"#({"code": "def hello():\\n    print('hey')"})#", "" }
-            })
-            .run();
-
-        // Partial tool call (streaming)
-        tst.test("<|tool_call_start|>[special_function(arg1=")
-            .tools({ special_function_tool })
-            .is_partial(true)
-            .expect(simple_assist_msg("", "", "special_function", "{\"arg1\": "))
-            .run();
-
-        // Tool call with empty arguments
-        tst.test("<|tool_call_start|>[empty_args()]<|tool_call_end|>")
-            .tools({ empty_args_tool })
-            .expect(simple_assist_msg("", "", "empty_args", "{}"))
-            .run();
-
-        // fake tool call marker in reasoning
-        tst.test(
-               "<think>Let me think about <|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|> hmm</think>"
-               "<|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|>")
-            .enable_thinking(true)
+        // Fake tool call marker inside reasoning is not parsed as a call
+        tst.test("<think>Let me think about <|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|> hmm</think>"
+                 "<|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|>")
             .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
             .tools({ special_function_tool })
             .expect_reasoning("Let me think about <|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|> hmm")
@@ -4122,127 +4178,21 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             })
             .run();
 
-        // Continuation tests
-        tst.test("world!\nWhat's up?")
-            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
-            .enable_thinking(true)
-            .messages({ message_user, message_assist_prefill_content })
-            .add_generation_prompt(false)
-            .continue_final_message(COMMON_CHAT_CONTINUATION_CONTENT)
-            .expect_reasoning("I'm thinking")
-            .expect_content("Hello, world!\nWhat's up?")
-            .run();
-
-        tst.test(" thinking</think>Hello, world!\nWhat's up?")
+        // enable_thinking=false still captures emitted reasoning
+        tst.test("<think>I'm\nthinking</think>Hello, world!\nWhat's up?")
+            .enable_thinking(false)
             .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
-            .enable_thinking(true)
-            .messages({ message_user, message_assist_prefill_reasoning })
-            .add_generation_prompt(false)
-            .continue_final_message(COMMON_CHAT_CONTINUATION_REASONING)
-            .expect_reasoning("I'm thinking")
-            .expect_content("Hello, world!\nWhat's up?")
-            .run();
-    }
-
-    // LFM2.5 tests - format <|tool_call_start|>[name(args)]<|tool_call_end|>
-    {
-        auto tst = peg_tester("models/templates/LFM2.5-Instruct.jinja", detailed_debug);
-
-        // Basic content only
-        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
-
-        // Single tool call without reasoning
-        tst.test("<|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|>")
-            .tools({ special_function_tool })
-            .expect(message_assist_call)
-            .run();
-
-        // Tool call with string argument
-        tst.test("<|tool_call_start|>[get_time(city=\"XYZCITY\")]<|tool_call_end|>")
-            .tools({ get_time_tool })
-            .expect(message_with_tool_calls("get_time", "{\"city\":\"XYZCITY\"}"))
-            .run();
-
-        // Python literals become JSON.
-        tst.test("<|tool_call_start|>[toggle(enabled=True)]<|tool_call_end|>")
-            .tools({ toggle_tool })
-            .expect(message_with_tool_calls("toggle", R"({"enabled": true})"))
-            .run();
-
-        tst.test("<|tool_call_start|>[set_nullable(value=None)]<|tool_call_end|>")
-            .tools({ nullable_tool })
-            .expect(message_with_tool_calls("set_nullable", R"({"value": null})"))
-            .run();
-
-        // Nested Python literal.
-        tst.test("<|tool_call_start|>[set_config(config={\"enabled\": True, \"count\": 3})]<|tool_call_end|>")
-            .tools({ config_tool })
-            .expect(message_with_tool_calls("set_config", R"({"config": {"enabled": true, "count": 3}})"))
-            .run();
-
-        // JSON literals are accepted too.
-        tst.test("<|tool_call_start|>[set_config(config={\"enabled\": true, \"note\": null})]<|tool_call_end|>")
-            .tools({ config_tool })
-            .expect(message_with_tool_calls("set_config", R"({"config": {"enabled": true, "note": null}})"))
-            .run();
-
-        // Dotted function name with structured args.
-        tst.test("<|tool_call_start|>[Calendar.create_event(title=\"demo\", participants=[\"Alice\", \"Bob\"], "
-                 "metadata={\"priority\": \"high\", \"reminder\": true})]<|tool_call_end|>")
-            .tools({ calendar_create_event_tool })
-            .expect(message_with_tool_calls(
-                "Calendar.create_event",
-                R"({"title": "demo", "participants": ["Alice", "Bob"], "metadata": {"priority": "high", "reminder": true}})"))
-            .run();
-
-        // Markdown links stay content.
-        tst.test("Use this format: [link text](url). Example: [Wikipedia](https://www.wikipedia.org).")
-            .tools({ get_time_tool })
-            .expect(simple_assist_msg("Use this format: [link text](url). Example: [Wikipedia](https://www.wikipedia.org)."))
+            .expect(message_assist_thoughts)
             .run();
 
-        // Tool call with reasoning (enable_thinking=true)
         tst.test("<think>I'm\nthinking</think><|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|>")
-            .enable_thinking(true)
+            .enable_thinking(false)
             .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
             .tools({ special_function_tool })
             .expect(message_assist_call_thoughts)
             .run();
 
-        // Multiple tool calls (parallel)
-        tst.test("<|tool_call_start|>[special_function(arg1=1), special_function_with_opt(arg1=1, arg2=2)]<|tool_call_end|>")
-            .parallel_tool_calls(true)
-            .tools({
-                special_function_tool, special_function_tool_with_optional_param
-            })
-            .expect_tool_calls({
-                { "special_function", R"({"arg1": 1})", {} },
-                { "special_function_with_opt", R"({"arg1": 1, "arg2": 2})", {} },
-            })
-            .run();
-
-        // Tool call with content before tool call
-        tst.test("Let me check the time.<|tool_call_start|>[get_time(city=\"Paris\")]<|tool_call_end|>")
-            .tools({ get_time_tool })
-            .expect(message_with_reasoning_content_and_multiple_tool_calls(
-                "", "Let me check the time.", { { "get_time", "{\"city\":\"Paris\"}" } }
-            ))
-            .run();
-
-        // Partial tool call (streaming)
-        tst.test("<|tool_call_start|>[special_function(arg1=")
-            .tools({ special_function_tool })
-            .is_partial(true)
-            .expect(simple_assist_msg("", "", "special_function", "{\"arg1\": "))
-            .run();
-
-        // Tool call with empty arguments
-        tst.test("<|tool_call_start|>[empty_args()]<|tool_call_end|>")
-            .tools({ empty_args_tool })
-            .expect(simple_assist_msg("", "", "empty_args", "{}"))
-            .run();
-
-        // Continuation tests
+        // Continuation: prefill content
         tst.test("world!\nWhat's up?")
             .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
             .enable_thinking(true)
@@ -4253,6 +4203,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .expect_content("Hello, world!\nWhat's up?")
             .run();
 
+        // Continuation: prefill reasoning
         tst.test(" thinking</think>Hello, world!\nWhat's up?")
             .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
             .enable_thinking(true)
@@ -5478,18 +5429,25 @@ static void test_template_generation_prompt() {
         check(tmpls, continuation_reasoning(), "<|im_assistant|>assistant<|im_middle|><think>I'm");
     }
 
-    {
-        auto tmpls = read_templates("models/templates/LFM2-8B-A1B.jinja");
+    for (const char * tmpl : {
+             "models/templates/LFM2-8B-A1B.jinja",
+             "models/templates/LFM2.5-Instruct.jinja",
+             "models/templates/LFM2.5-8B-A1B.jinja",
+         }) {
+        auto tmpls = read_templates(tmpl);
         check(tmpls, basic(),                  "<|im_start|>assistant\n");
         check(tmpls, continuation_content(),   "<|im_start|>assistant\n<think>I'm thinking</think>Hello, ");
         check(tmpls, continuation_reasoning(), "<|im_start|>assistant\n<think>I'm");
     }
 
     {
-        auto tmpls = read_templates("models/templates/LFM2.5-Instruct.jinja");
-        check(tmpls, basic(),                  "<|im_start|>assistant\n");
-        check(tmpls, continuation_content(),   "<|im_start|>assistant\n<think>I'm thinking</think>Hello, ");
-        check(tmpls, continuation_reasoning(), "<|im_start|>assistant\n<think>I'm");
+        // 8B-A1B renders prior-turn reasoning via the "thinking" field
+        auto tmpls = read_templates("models/templates/LFM2.5-8B-A1B.jinja");
+        common_chat_templates_inputs inputs;
+        inputs.messages              = { message_user, message_assist_call_thoughts, tool_msg };
+        inputs.add_generation_prompt = true;
+        auto params = common_chat_templates_apply(tmpls.get(), inputs);
+        assert_contains(params.prompt, "<think>I'm\nthinking</think>");
     }
 
     {

{llama_cpp_bin-9536.0.0 → llama_cpp_bin-9544.0.0}/llama.cpp/tools/completion/completion.cpp

@@ -33,12 +33,8 @@
 #endif
 
 static llama_context           ** g_ctx;
-static llama_model             ** g_model;
 static common_sampler          ** g_smpl;
 static common_params            * g_params;
-static std::vector<llama_token> * g_input_tokens;
-static std::ostringstream       * g_output_ss;
-static std::vector<llama_token> * g_output_tokens;
 static bool is_interacting  = false;
 static bool need_insert_eot = false;
 
@@ -136,7 +132,6 @@ int llama_completion(int argc, char ** argv) {
     llama_context * ctx = nullptr;
     common_sampler * smpl = nullptr;
 
-    g_model = &model;
     g_ctx = &ctx;
     g_smpl = &smpl;
 
@@ -549,9 +544,9 @@ int llama_completion(int argc, char ** argv) {
     int n_consumed         = 0;
     int n_session_consumed = 0;
 
-    std::vector<int>   input_tokens;  g_input_tokens  = &input_tokens;
-    std::vector<int>   output_tokens; g_output_tokens = &output_tokens;
-    std::ostringstream output_ss;     g_output_ss     = &output_ss;
+    std::vector<int>   input_tokens;
+    std::vector<int>   output_tokens;
+    std::ostringstream output_ss;
     std::ostringstream assistant_ss; // for storing current assistant message, used in conversation mode
 
     // the first thing we will do is to output the prompt, so set color accordingly
@@ -989,7 +984,7 @@ int llama_completion(int argc, char ** argv) {
         LOG("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
         session_tokens.insert(session_tokens.end(), embd.begin(), embd.end());
         llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
-        LOG_INF("saved final session to %s, n_tokens = %ld\n", path_session.data(), session_tokens.size());
+        LOG_INF("saved final session to %s, n_tokens = %zu\n", path_session.data(), session_tokens.size());
 
     }
 

{llama_cpp_bin-9536.0.0 → llama_cpp_bin-9544.0.0}/llama.cpp/tools/mtmd/clip-graph.h

@@ -37,6 +37,9 @@ struct clip_graph {
     float kq_scale; // TODO: maybe move this to hparams
     const clip_flash_attn_type flash_attn_type;
 
+    // TODO [QWEN_VIDEO]: improve this in the future
+    int n_batch = 1;
+
     ggml_context_ptr ctx0_ptr;
     ggml_context * ctx0;
     ggml_cgraph * gf;