llama-cpp-bin 9542.0.0__tar.gz → 9548.0.0__tar.gz

{llama_cpp_bin-9542.0.0 → llama_cpp_bin-9548.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-cpp-bin
-Version: 9542.0.0
+Version: 9548.0.0
 Summary: llama.cpp server binary built from source
 Author-email: Vladyslav Tkachenko <vladlearns@gmail.com>
 License: AGPL-3.0-or-later

{llama_cpp_bin-9542.0.0 → llama_cpp_bin-9548.0.0}/llama.cpp/common/arg.cpp

@@ -444,7 +444,7 @@ bool common_params_handle_models(common_params & params, llama_example curr_ex)
     opts.offline         = params.offline;
     opts.skip_download   = params.skip_download;
     opts.download_mtp    = spec_type_draft_mtp;
-    opts.download_mmproj = !params.no_mmproj;
+    opts.download_mmproj = !params.no_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty();
 
     // sub-models (draft, mmproj, vocoder) are explicitly specified by the user,
     // so we should not auto-discover mtp/mmproj siblings for them

{llama_cpp_bin-9542.0.0 → llama_cpp_bin-9548.0.0}/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-9542.0.0 → llama_cpp_bin-9548.0.0}/llama.cpp/common/speculative.cpp

@@ -58,10 +58,10 @@ static bool common_speculative_are_compatible(
     const llama_vocab * vocab_tgt = llama_model_get_vocab(model_tgt);
     const llama_vocab * vocab_dft = llama_model_get_vocab(model_dft);
 
-    const bool vocab_type_tgt = llama_vocab_type(vocab_tgt);
+    const auto vocab_type_tgt = llama_vocab_type(vocab_tgt);
     LOG_DBG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt);
 
-    const bool vocab_type_dft = llama_vocab_type(vocab_dft);
+    const auto vocab_type_dft = llama_vocab_type(vocab_dft);
     LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
 
     if (vocab_type_tgt != vocab_type_dft) {

{llama_cpp_bin-9542.0.0 → llama_cpp_bin-9548.0.0}/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-9542.0.0 → llama_cpp_bin-9548.0.0}/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;

{llama_cpp_bin-9542.0.0 → llama_cpp_bin-9548.0.0}/llama.cpp/tools/mtmd/clip-impl.h

@@ -480,10 +480,6 @@ struct clip_image_u8 {
         buf[idx + 2] = rgb[2];
     }
 
-    size_t n_pixels() const {
-        return (size_t) nx * (size_t) ny;
-    }
-
     size_t n_elements() const {
         return n_pixels() * 3;
     }
@@ -492,10 +488,16 @@ struct clip_image_u8 {
     std::vector<uint8_t> buf;
     int nx = 0;
     int ny = 0;
+
+    size_t n_pixels() const {
+        return (size_t) nx * (size_t) ny;
+    }
 };
 
 // For images, buf.size() == nx*ny*3
 //     Memory layout: RGBRGBRGB...
+// For seq, buf.size() == nx*ny*3*nt
+//     Memory layout: RGBRGB...RGBRGB... (nt times)
 // For audio, only one channel is used, buf.size() == nx*ny
 //     nx will be n_frames and ny will be n_mel
 struct clip_image_f32 {
@@ -544,10 +546,6 @@ struct clip_image_f32 {
         }
     }
 
-    size_t n_pixels() const {
-        return (size_t) nx_ * (size_t) ny_;
-    }
-
     size_t n_elements() const {
         return n_pixels() * 3;
     }
@@ -580,6 +578,10 @@ struct clip_image_f32 {
     std::vector<float> buf;
     int nx_ = 0;
     int ny_ = 0;
+
+    size_t n_pixels() const {
+        return (size_t) nx_ * (size_t) ny_;
+    }
 };
 
 //
@@ -627,6 +629,7 @@ static void clip_log_internal(enum ggml_log_level level, const char * format, ..
     va_end(args);
 }
 
+#define LOG_TRC(...) clip_log_internal(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
 #define LOG_DBG(...) clip_log_internal(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
 #define LOG_INF(...) clip_log_internal(GGML_LOG_LEVEL_INFO,  __VA_ARGS__)
 #define LOG_WRN(...) clip_log_internal(GGML_LOG_LEVEL_WARN,  __VA_ARGS__)

{llama_cpp_bin-9542.0.0 → llama_cpp_bin-9548.0.0}/llama.cpp/tools/mtmd/clip.cpp

@@ -527,7 +527,7 @@ ggml_tensor * clip_graph::build_inp() {
 }
 
 ggml_tensor * clip_graph::build_inp_raw(int channels) {
-    ggml_tensor * inp_raw = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, img.nx(), img.ny(), channels);
+    ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, img.nx(), img.ny(), channels, n_batch);
     ggml_set_name(inp_raw, "inp_raw");
     ggml_set_input(inp_raw);
     return inp_raw;
@@ -848,8 +848,6 @@ ggml_tensor * clip_graph::build_patch_merge_permute(ggml_tensor * cur, int scale
 }
 
 static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch & imgs) {
-    GGML_ASSERT(imgs.entries.size() == 1 && "n_batch > 1 is not supported");
-
     const clip_image_f32 & img = *imgs.entries[0];
     std::unique_ptr<clip_graph> builder;
 
@@ -1009,6 +1007,9 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             GGML_ABORT("missing cgraph builder");
     }
 
+    // TODO [QWEN_VIDEO]: improve this in the future
+    builder->n_batch = imgs.entries.size();
+
     return builder->build();
 }
 
@@ -3479,12 +3480,15 @@ bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f3
 
 bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs_c_ptr, float * vec) {
     const clip_image_f32_batch & imgs = *imgs_c_ptr;
-    int batch_size = imgs.entries.size();
+    int n_batch_cur = imgs.entries.size();
+
+    // maximum supported batch size, usually == 2 for qwen-vl-based models
+    int n_batch_max = clip_model_n_batch_max(ctx);
 
     // TODO @ngxson : implement batch size > 1 as a loop
     //                we don't need true batching support because the cgraph will gonna be big anyway
-    if (batch_size != 1) {
-        return false; // only support batch size of 1
+    if (n_batch_cur > n_batch_max) {
+        return false;
     }
 
     // if buffers are not allocated, we need to do a warmup run to allocate them
@@ -3555,18 +3559,20 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         // └─────┘ │
         //   ──────┘ x B
 
-        for (size_t i = 0; i < imgs.entries.size(); i++) {
-            const int nx = imgs.entries[i]->nx();
-            const int ny = imgs.entries[i]->ny();
-            const int n = nx * ny;
+        // IMPORTANT: [QWEN_VIDEO] the batch dim is currently used for temporal dim in Qwen-VL models
+        // All entries must have the same spatial size (enforced by can_batch_with() during merging)
+        {
+            const int nx = imgs.entries[0]->nx();
+            const int ny = imgs.entries[0]->ny();
+            const int n  = nx * ny;
 
-            for (int b = 0; b < batch_size; b++) {
+            for (int b = 0; b < n_batch_cur; b++) {
                 const auto & buf = imgs.entries[b]->get_ro_buf();
                 float * batch_entry = inp_raw.data() + b * (3*n);
                 for (int y = 0; y < ny; y++) {
                     for (int x = 0; x < nx; x++) {
-                        size_t base_src = 3*(y * nx + x); // idx of the first channel
-                        size_t base_dst =    y * nx + x;  // idx of the first channel
+                        size_t base_src = 3*(y * nx + x);
+                        size_t base_dst =    y * nx + x;
                         batch_entry[      base_dst] = buf[base_src    ];
                         batch_entry[1*n + base_dst] = buf[base_src + 1];
                         batch_entry[2*n + base_dst] = buf[base_src + 2];
@@ -4549,6 +4555,17 @@ bool clip_has_audio_encoder(const struct clip_ctx * ctx) {
     return ctx->model.modality == CLIP_MODALITY_AUDIO;
 }
 
+int clip_model_n_batch_max(const struct clip_ctx * ctx) {
+    switch (ctx->proj_type()) {
+        case PROJECTOR_TYPE_QWEN2VL:
+        case PROJECTOR_TYPE_QWEN25VL:
+        case PROJECTOR_TYPE_QWEN3VL:
+            return 2;
+        default:
+            return 1;
+    }
+}
+
 //
 // API used internally with mtmd
 //

{llama_cpp_bin-9542.0.0 → llama_cpp_bin-9548.0.0}/llama.cpp/tools/mtmd/clip.h

@@ -20,6 +20,12 @@ struct clip_image_size {
     bool operator==(const clip_image_size & other) const {
         return width == other.width && height == other.height;
     }
+    bool operator!=(const clip_image_size & other) const {
+        return !(*this == other);
+    }
+    int area() const {
+        return width * height;
+    }
 };
 
 struct clip_image_f32;
@@ -101,6 +107,8 @@ bool clip_is_llava(const struct clip_ctx * ctx);
 bool clip_has_vision_encoder(const struct clip_ctx * ctx);
 bool clip_has_audio_encoder(const struct clip_ctx * ctx);
 
+int clip_model_n_batch_max(const struct clip_ctx * ctx);
+
 std::map<ggml_backend_dev_t, size_t> clip_get_mem_usage(const struct clip_ctx * ctx);
 
 struct clip_cap {

{llama_cpp_bin-9542.0.0 → llama_cpp_bin-9548.0.0}/llama.cpp/tools/mtmd/models/models.h

@@ -31,10 +31,11 @@ struct clip_graph_pixtral : clip_graph {
 struct clip_graph_qwen2vl : clip_graph {
     clip_graph_qwen2vl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
     ggml_cgraph * build() override;
+    ggml_tensor * build_inp_with_temporal_merge();
 };
 
-struct clip_graph_qwen3vl : clip_graph {
-    clip_graph_qwen3vl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+struct clip_graph_qwen3vl : clip_graph_qwen2vl {
+    clip_graph_qwen3vl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph_qwen2vl(ctx, img) {}
     ggml_cgraph * build() override;
 };