chat : enable tool call in thinking for DS4 (#26269 )

vulkan: add POOL_1D op (#25431 )
* vulkan : add pool1d push constants and pipeline field Declared data structures needed for POOL1D OP, which are the vk_op_pool1d_push_constants struct and pipeline_pool1d_f32 field. * vulkan : add pool1d compute shader Added pool1d.comp for Vulkan backend mirroring the existing pool2d shader. * vulkan : add full GGML_OP_POOL_1D support Added pipeline creation and op dispatch for 1D pooling in the Vulkan backend. * vulkan : fix pool1d shader logic Registered pool1d_f32 in vulkan-shaders-gen.cpp and fixed tensor dimension indices and avg pool scale. * vulkan : fix pool1d end boundary crash and expand test coverage Fixed an issue where the shader crashed when the end boundary was negative when k0 < p0. Also, added more test cases related to this fix.
2026-08-01 07:28:39 +00:00 · 2026-08-01 00:13:07 -05:00 · 2026-07-31 16:48:58 +02:00 · 2026-07-31 16:26:37 +02:00 · 2026-07-31 15:30:19 +02:00
12 changed files with 288 additions and 111 deletions
@@ -1943,18 +1943,6 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
        adjusted_messages = deepseek_v4_sort_tool_results(inputs.messages);
    }

-    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.thinking_start_tag = "<think>";
-    data.thinking_end_tags  = {"</think>"};
-    data.preserved_tokens   = {
-        "｜DSML｜",
-        "<think>",
-        "</think>",
-    };
-
    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
@@ -1972,6 +1960,18 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
    const std::string PARAM_END    = "</" + DSML + "parameter>";
    const std::string GEN_PROMPT   = "<｜Assistant｜>";

+    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.thinking_start_tag = THINK_START;
+    data.thinking_end_tags  = {THINK_END, FC_START};
+    data.preserved_tokens   = {
+        DSML,
+        THINK_START,
+        THINK_END,
+    };
+
    if (inputs.has_continuation()) {
        const auto & msg = inputs.continue_msg;

@@ -1983,13 +1983,101 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
        data.prompt += data.generation_prompt;
    }

+    bool require_tools   = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+    bool has_tool_calls = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+
    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
        auto generation_prompt = p.literal(GEN_PROMPT);
-        auto end = p.end();
+        auto end               = p.end();
+
+        // build tool call section first since we might need it in reasoning
+        auto tool_choice = p.choice();
+        if (has_tool_calls) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string  name     = function.at("name");
+                auto         params   = function.contains("parameters") ? function.at("parameters") : json::object();
+                const auto & props    = params.contains("properties") ? params.at("properties") : json::object();
+
+                std::set<std::string> required;
+                if (params.contains("required")) {
+                    params.at("required").get_to(required);
+                }
+
+                auto schema_info = common_schema_info();
+                schema_info.resolve_refs(params);
+
+                std::vector<common_peg_parser> required_parsers;
+                std::vector<common_peg_parser> optional_parsers;
+                for (const auto & [param_name, param_schema] : props.items()) {
+                    bool is_required = required.find(param_name) != required.end();
+                    bool is_string   = schema_info.resolves_to_string(param_schema);
+
+                    auto arg = p.tool_arg(
+                        p.tool_arg_open(p.literal(PARAM_START + " name=\"") + p.tool_arg_name(p.literal(param_name)) +
+                                        p.literal("\" string=\"" + std::string(is_string ? "true" : "false") + "\">")) +
+                        (is_string ?
+                             p.tool_arg_string_value(p.until(PARAM_END)) :
+                             p.tool_arg_json_value(p.schema(p.json(), "tool-" + name + "-arg-" + param_name + "-schema",
+                                                            param_schema, false))) +
+                        p.tool_arg_close(p.literal(PARAM_END)));
+
+                    auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
+                    if (is_required) {
+                        required_parsers.push_back(named_arg);
+                    } else {
+                        optional_parsers.push_back(named_arg);
+                    }
+                }
+
+                common_peg_parser args_seq = p.eps();
+                for (size_t i = 0; i < required_parsers.size(); i++) {
+                    if (i > 0) {
+                        args_seq = args_seq + p.space();
+                    }
+                    args_seq = args_seq + required_parsers[i];
+                }
+
+                if (!optional_parsers.empty()) {
+                    common_peg_parser any_opt = p.choice();
+                    for (const auto & opt : optional_parsers) {
+                        any_opt |= opt;
+                    }
+                    args_seq = args_seq + p.repeat(p.space() + any_opt, 0, -1);
+                }
+
+                common_peg_parser invoke_body = args_seq;
+                auto              func_parser = p.tool(p.tool_open(p.literal(INVOKE_START + " name=\"") +
+                                                                   p.tool_name(p.literal(name)) + p.literal("\">\n")) +
+                                                       invoke_body + p.space() + p.tool_close(p.literal(INVOKE_END)));
+
+                tool_choice |= p.rule("tool-" + name, func_parser);
+            });
+        }
+
+        common_peg_parser tool_calls = p.eps();
+        if (inputs.parallel_tool_calls) {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice +
+                p.zero_or_more(p.space() + tool_choice) + p.space() + p.literal(FC_END));
+        } else {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice + p.space() + p.literal(FC_END));
+        }

        auto reasoning = p.eps();
+        auto reasoning_with_tc = p.eps();
+        auto obligatory_tool_calls = tool_calls;
+        bool allow_reasoning_with_tc = false;
+
+        if (!require_tools) {
+            tool_calls = p.optional(tool_calls);
+        }
+
        if (extract_reasoning && inputs.enable_thinking) {
            reasoning = p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
+            reasoning_with_tc = THINK_START + p.reasoning(p.until_one_of({ FC_START, THINK_END })) + obligatory_tool_calls;
+            allow_reasoning_with_tc = true;
        } else if (extract_reasoning) {
            // Thinking disabled but reasoning extraction requested: the generation prompt
            // contains an empty <think></think> pair (V3.2) or a bare </think> (V4) that
@@ -2007,101 +2095,19 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
            return generation_prompt + reasoning + response_format + end;
        }

-        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+        if (!has_tool_calls) {
            return generation_prompt + reasoning + p.content(p.rest()) + end;
        }

-        auto tool_choice = p.choice();
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool.at("function");
-            std::string  name     = function.at("name");
-            auto params   = function.contains("parameters") ? function.at("parameters") : json::object();
-            const auto & props    = params.contains("properties") ? params.at("properties") : json::object();
-
-            std::set<std::string> required;
-            if (params.contains("required")) {
-                params.at("required").get_to(required);
-            }
-
-            auto schema_info = common_schema_info();
-            schema_info.resolve_refs(params);
-
-            std::vector<common_peg_parser> required_parsers;
-            std::vector<common_peg_parser> optional_parsers;
-            for (const auto & [param_name, param_schema] : props.items()) {
-                bool is_required = required.find(param_name) != required.end();
-                bool is_string   = schema_info.resolves_to_string(param_schema);
-
-                auto arg = p.tool_arg(
-                    p.tool_arg_open(
-                        p.literal(PARAM_START + " name=\"") +
-                        p.tool_arg_name(p.literal(param_name)) +
-                        p.literal("\" string=\"" + std::string(is_string ? "true" : "false") + "\">")) +
-                    (is_string
-                         ? p.tool_arg_string_value(p.until(PARAM_END))
-                         : p.tool_arg_json_value(p.schema(p.json(),
-                                                          "tool-" + name + "-arg-" + param_name + "-schema",
-                                                          param_schema, false))) +
-                    p.tool_arg_close(p.literal(PARAM_END)));
-
-                auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
-                if (is_required) {
-                    required_parsers.push_back(named_arg);
-                } else {
-                    optional_parsers.push_back(named_arg);
-                }
-            }
-
-            common_peg_parser args_seq = p.eps();
-            for (size_t i = 0; i < required_parsers.size(); i++) {
-                if (i > 0) {
-                    args_seq = args_seq + p.space();
-                }
-                args_seq = args_seq + required_parsers[i];
-            }
-
-            if (!optional_parsers.empty()) {
-                common_peg_parser any_opt = p.choice();
-                for (const auto & opt : optional_parsers) {
-                    any_opt |= opt;
-                }
-                args_seq = args_seq + p.repeat(p.space() + any_opt, 0, -1);
-            }
-
-            common_peg_parser invoke_body = args_seq;
-            auto func_parser = p.tool(
-                p.tool_open(p.literal(INVOKE_START + " name=\"") +
-                            p.tool_name(p.literal(name)) + p.literal("\">\n")) +
-                invoke_body + p.space() +
-                p.tool_close(p.literal(INVOKE_END)));
-
-            tool_choice |= p.rule("tool-" + name, func_parser);
-        });
-
-        auto require_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-
-        common_peg_parser tool_calls = p.eps();
-        if (inputs.parallel_tool_calls) {
-            tool_calls = p.trigger_rule("tool-call",
-                p.literal(FC_START) + p.space() + tool_choice +
-                p.zero_or_more(p.space() + tool_choice) + p.space() + p.literal(FC_END));
-        } else {
-            tool_calls = p.trigger_rule("tool-call",
-                p.literal(FC_START) + p.space() + tool_choice + p.space() + p.literal(FC_END));
-        }
-
-        if (!require_tools) {
-            tool_calls = p.optional(tool_calls);
-        }
-
-        auto content_before_tools = p.content(p.until(FC_START));
-        return generation_prompt + reasoning + content_before_tools + tool_calls + end;
+        auto content_before_tools = p.negate(p.literal(THINK_START)) + p.content(p.until(FC_START));
+        return allow_reasoning_with_tc ? generation_prompt + (reasoning_with_tc | (reasoning + content_before_tools + tool_calls)) + end :
+            generation_prompt + reasoning + content_before_tools + tool_calls + end;
    });

    data.parser = parser.save();

    if (include_grammar) {
-        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
+        data.grammar_lazy = has_tools && !require_tools;
        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
            foreach_function(inputs.tools, [&](const json & tool) {
                const auto & function = tool.at("function");
@@ -998,6 +998,7 @@ struct vk_device_struct {
    vk_pipeline pipeline_snake_f32;
    vk_pipeline pipeline_snake_f16;
    vk_pipeline pipeline_snake_bf16;
+    vk_pipeline pipeline_pool1d_f32;
    vk_pipeline pipeline_pool2d_f32;
    vk_pipeline pipeline_rwkv_wkv6_f32;
    vk_pipeline pipeline_rwkv_wkv7_f32;
@@ -1687,6 +1688,17 @@ struct vk_op_snake_push_constants {
    uint32_t ne1;
 };

+struct vk_op_pool1d_push_constants {
+    uint32_t IL;
+    uint32_t OL;
+    uint32_t OC;
+    uint32_t pelements;
+    uint32_t op;
+    int32_t k0;
+    int32_t s0;
+    int32_t p0;
+};
+
 struct vk_op_pool2d_push_constants {
    uint32_t IW; uint32_t IH;
    uint32_t OW; uint32_t OH;
@@ -1934,6 +1946,7 @@ struct ggml_vk_garbage_collector {
 static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx, vk_context subctx);
 static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested = nullptr);
 static void ggml_pipeline_allocate_descriptor_sets(ggml_backend_vk_context * ctx);
+static bool ggml_vk_intel_windows_driver_equals_or_newer_than(uint32_t driver_version, uint32_t threshold_major, uint32_t threshold_minor);

 static bool vk_memory_logger_enabled = false;

@@ -5552,8 +5565,11 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
    ggml_vk_create_pipeline(device, device->pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);

    ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    // Intel Arc B390 was observed segfaulting with this shader.
-    if (device->subgroup_basic && device->subgroup_shuffle && device->vendor_id != VK_VENDOR_ID_INTEL) {
+    // Intel Windows driver older than 32.0.101.8860 will crash when using fwht kernels on Xe2+ GPUS so we gate that here
+    const bool can_use_fwht = device->driver_id != vk::DriverId::eIntelProprietaryWindows ||
+        device->architecture != vk_device_architecture::INTEL_XE2 ||
+        (device->architecture == vk_device_architecture::INTEL_XE2 && ggml_vk_intel_windows_driver_equals_or_newer_than(device->properties.driverVersion, 101, 8860));
+    if (can_use_fwht && device->subgroup_basic && device->subgroup_shuffle) {
        int idx = 0;
        for (uint32_t n : {64, 128, 256, 512}) {
            if (device->subgroup_size <= n) {
@@ -5561,8 +5577,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
            }
            ++idx;
        }
-    } else if (device->driver_id != vk::DriverId::eIntelProprietaryWindows) {
-        // Disabled on Intel Windows due to a driver bug: https://github.com/ggml-org/llama.cpp/pull/23964#issuecomment-4598226147
+    } else if (can_use_fwht) {
        int idx = 0;
        for (uint32_t n : {64, 128, 256, 512}) {
            const uint32_t block_size = std::min(device->subgroup_size, n);
@@ -5619,6 +5634,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
    ggml_vk_create_pipeline(device, device->pipeline_snake_f16,  "snake_f16",  snake_f16_len,  snake_f16_data,  "main", 4, sizeof(vk_op_snake_push_constants), {256, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_snake_bf16, "snake_bf16", snake_bf16_len, snake_bf16_data, "main", 4, sizeof(vk_op_snake_push_constants), {256, 1, 1}, {}, 1);

+    ggml_vk_create_pipeline(device, device->pipeline_pool1d_f32, "pool1d_f32", pool1d_f32_len, pool1d_f32_data, "main", 2, sizeof(vk_op_pool1d_push_constants), {512, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_pool2d_f32, "pool2d_f32", pool2d_f32_len, pool2d_f32_data, "main", 2, sizeof(vk_op_pool2d_push_constants), {512, 1, 1}, {}, 1);

    ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv6_f32, "rwkv_wkv6_f32", rwkv_wkv6_f32_len, rwkv_wkv6_f32_data, "main", 7, sizeof(vk_op_rwkv_wkv6_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
@@ -11332,6 +11348,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
            case GGML_TYPE_BF16: return ctx->device->pipeline_col2im_1d_bf16;
            default:             return nullptr;
        }
+    case GGML_OP_POOL_1D:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_pool1d_f32;
+        }
+        return nullptr;
    case GGML_OP_POOL_2D:
        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
            return ctx->device->pipeline_pool2d_f32;
@@ -11854,6 +11875,13 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
        {
            elements = { uint32_t(dst->ne[0]), uint32_t(dst->ne[1]), 1 };
        } break;
+    case GGML_OP_POOL_1D:
+        {
+            const uint32_t N = dst->ne[3] * dst->ne[2];
+            const uint32_t OC = dst->ne[1];
+            const uint32_t OL = dst->ne[0];
+            elements = { N * OC * OL, 1, 1};
+        } break;
    case GGML_OP_POOL_2D:
        {
            const uint32_t N = dst->ne[3];
@@ -13773,6 +13801,29 @@ static void ggml_vk_snake_dispatch_fused(ggml_backend_vk_context * ctx, vk_conte
    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { x_buf, a_buf, inv_b_buf, dst_buf }, pc, elements);
 }

+static void ggml_vk_pool_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    uint32_t op = static_cast<uint32_t>(dst->op_params[0]);
+    const int32_t k0 = dst->op_params[1];
+    const int32_t s0 = dst->op_params[2];
+    const int32_t p0 = dst->op_params[3];
+
+    const uint32_t IL = src0->ne[0];
+
+    const uint32_t N = dst->ne[3] * dst->ne[2];
+
+    const uint32_t OC = dst->ne[1];
+    const uint32_t OL = dst->ne[0];
+
+    const uint32_t parallel_elements = N * OC * OL;
+
+    ggml_vk_op_f32<vk_op_pool1d_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_POOL_1D, {
+        IL, OL, OC,
+        parallel_elements,
+        op,
+        k0, s0, p0,
+    });
+}
+
 static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
    uint32_t op = static_cast<uint32_t>(dst->op_params[0]);
    const int32_t k1 = dst->op_params[1];
@@ -15284,6 +15335,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
    case GGML_OP_CONV_TRANSPOSE_1D:
        ggml_vk_conv_transpose_1d(ctx, compute_ctx, src0, src1, node);

+        break;
+    case GGML_OP_POOL_1D:
+        ggml_vk_pool_1d(ctx, compute_ctx, src0, node);
+
        break;
    case GGML_OP_POOL_2D:
        ggml_vk_pool_2d(ctx, compute_ctx, src0, node);
@@ -18001,6 +18056,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
        case GGML_OP_CONV_2D_DW:
            return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16)
                && op->src[1]->type == GGML_TYPE_F32;
+        case GGML_OP_POOL_1D:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
        case GGML_OP_POOL_2D:
            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
        case GGML_OP_RWKV_WKV6:
@@ -18466,6 +18523,22 @@ static uint32_t ggml_vk_intel_shader_core_count(const vk::PhysicalDevice& vkdev)
    }
 }

+static bool ggml_vk_intel_windows_driver_equals_or_newer_than(uint32_t driver_version, uint32_t threshold_major, uint32_t threshold_minor) {
+#if defined(_WIN32)
+    // Intel Windows encodes xxx.yyyy as [31:14].[13:0].
+    const uint32_t major = driver_version >> 14;
+    const uint32_t minor = driver_version & 0x3fff;
+
+    return major > threshold_major || (major == threshold_major && minor >= threshold_minor);
+#else
+    GGML_UNUSED(driver_version);
+    GGML_UNUSED(threshold_major);
+    GGML_UNUSED(threshold_minor);
+    return true;
+#endif
+}
+
+
 // checks

 #ifdef GGML_VULKAN_CHECK_RESULTS
@@ -18920,6 +18993,13 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
            const int32_t oc     = tensor->op_params[1];
            const int32_t p0     = tensor->op_params[2];
            tensor_clone = ggml_col2im_1d(ggml_ctx, src_clone[0], stride, oc, p0);
+        } else if (tensor->op == GGML_OP_POOL_1D) {
+            enum ggml_op_pool op = static_cast<ggml_op_pool>(tensor->op_params[0]);
+            const int32_t k0 = tensor->op_params[1];
+            const int32_t s0 = tensor->op_params[2];
+            const int32_t p0 = tensor->op_params[3];
+
+            tensor_clone = ggml_pool_1d(ggml_ctx, src_clone[0], op, k0, s0, p0);
        } else if (tensor->op == GGML_OP_POOL_2D) {
            enum ggml_op_pool op = static_cast<ggml_op_pool>(tensor->op_params[0]);
            const int32_t k0 = tensor->op_params[1];
@@ -0,0 +1,65 @@
+#version 450
+
+#include "types.glsl"
+
+#extension GL_EXT_shader_16bit_storage : require
+
+layout(push_constant) uniform parameter {
+    uint IL;
+    uint OL;
+    uint OC;
+    uint pelements;
+    uint op;
+    int k0;
+    int s0;
+    int p0;
+} p;
+
+#define BLOCK_SIZE 512
+#define FLT_MAX 3.402823466e+38F
+#define OP_POOL_MAX 0u
+#define OP_POOL_AVG 1u
+
+layout (local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout(binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout(binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.x;
+    if (idx >= p.pelements) {
+        return;
+    }
+
+    const uint nc = idx / p.OL;
+    const uint cur_ol = idx % p.OL;
+
+    const int start = int(cur_ol) * p.s0 - p.p0;
+    const int bl = max(start, 0);
+    const int el = min(max(start + p.k0, 0), int(p.IL));
+
+    const int window_size = el - bl;
+    const float scale = window_size > 0 ? 1.0 / float(window_size) : 0.0;
+    float res;
+
+    if (p.op == OP_POOL_AVG) {
+        res = 0.0;
+    } else if (p.op == OP_POOL_MAX) {
+        res = -FLT_MAX;
+    } else {
+        return;
+    }
+
+    #pragma unroll
+    for (uint i = bl; i < el; i++) {
+        const float cur = D_TYPE(data_a[nc * p.IL + i]);
+
+        if (p.op == OP_POOL_AVG) {
+            res += cur * scale;
+        } else if (p.op == OP_POOL_MAX) {
+            res = max(res, cur);
+        }
+    }
+
+    data_d[nc * p.OL + cur_ol] = res;
+}
@@ -1052,6 +1052,7 @@ void process_shaders() {
    string_to_spv("snake_f16",  "snake.comp", {{"DATA_A_F16", "1"},  {"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
    string_to_spv("snake_bf16", "snake.comp", {{"DATA_A_BF16", "1"}, {"DATA_D_BF16", "1"}, {"A_TYPE", "uint16_t"},  {"D_TYPE", "uint16_t"}});

+    string_to_spv("pool1d_f32", "pool1d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
    string_to_spv("pool2d_f32", "pool2d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));

    string_to_spv("rwkv_wkv6_f32", "wkv6.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
@@ -353,6 +353,7 @@ class Keys:
        class Attention:
            HEAD_COUNT      = "clip.vision.attention.head_count"
            HEAD_COUNT_KV   = "clip.vision.attention.head_count_kv" # used by mimovl (GQA)
+            HEAD_DIM        = "clip.vision.attention.head_dim" # set when qkv width != n_embd
            LAYERNORM_EPS   = "clip.vision.attention.layer_norm_epsilon"

        class Projector:
@@ -1226,6 +1226,9 @@ class GGUFWriter:
    def add_vision_head_count_kv(self, value: int) -> None:
        self.add_uint32(Keys.ClipVision.Attention.HEAD_COUNT_KV, value)

+    def add_vision_head_dim(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.Attention.HEAD_DIM, value)
+
    def add_vision_attention_layernorm_eps(self, value: float) -> None:
        self.add_float32(Keys.ClipVision.Attention.LAYERNORM_EPS, value)

@@ -8192,9 +8192,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {

    for (ggml_type type_input : {GGML_TYPE_F32}) {
        for (ggml_op_pool pool_type : {GGML_OP_POOL_AVG, GGML_OP_POOL_MAX}) {
-            for (int k0 : {1, 3}) {
-                for (int s0 : {1, 2}) {
-                    for (int p0 : {0, 1}) {
+            for (int k0 : {1, 2, 3}) {
+                for (int s0 : {1, 2, 3}) {
+                    for (int p0 : {0, 1, 2, 3}) {
                        test_cases.emplace_back(new test_pool1d(pool_type, type_input, { 10,  3, 2, 1 }, k0, s0, p0));
                        test_cases.emplace_back(new test_pool1d(pool_type, type_input, { 11,  1, 3, 2 }, k0, s0, p0));
                        test_cases.emplace_back(new test_pool1d(pool_type, type_input, { 128, 2, 1, 3 }, k0, s0, p0));
@@ -4227,6 +4227,20 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
            .expect_reasoning("I'm thinking")
            .expect_content("Hello, world!\nWhat's up?")
            .run();
+
+        tst.test(
+            "Let me check the time\n\n"
+            "<｜DSML｜tool_calls>\n"
+            "<｜DSML｜invoke name=\"get_time\">\n"
+            "<｜DSML｜parameter name=\"city\" string=\"true\">Tokyo</｜DSML｜parameter>\n"
+            "</｜DSML｜invoke>\n"
+            "</｜DSML｜tool_calls>") // no </think> after the TC close because the grammar will immediately constrain it to end
+            .enable_thinking(true)
+            .reasoning_format(COMMON_REASONING_FORMAT_DEEPSEEK)
+            .tools({ get_time_tool })
+            .expect_reasoning("Let me check the time")
+            .expect_tool_calls({ { "get_time", R"({"city": "Tokyo"})", {} } })
+            .run();
    }

    // GLM-4.6 tests - format: <tool_call>function_name\n<arg_key>...</arg_key>\n<arg_value>...</arg_value>\n</tool_call>
@@ -41,6 +41,7 @@
 #define KEY_PROJ_DIM            "clip.%s.projection_dim"
 #define KEY_N_HEAD              "clip.%s.attention.head_count"
 #define KEY_N_HEAD_KV           "clip.%s.attention.head_count_kv"
+#define KEY_N_EMBD_HEAD         "clip.%s.attention.head_dim"
 #define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
 #define KEY_FEATURE_LAYERS      "clip.%s.feature_layer"

@@ -54,6 +54,8 @@ struct clip_hparams {
    int32_t projection_dim = 0;
    int32_t n_head = 0;
    int32_t n_head_kv = 0;
+    // 0 = derive from n_embd; set when qkv width != n_embd
+    int32_t n_embd_head = 0;
    int32_t n_layer = 0;
    int32_t n_merge = 1; // number of patch merges **per-side**

@@ -253,7 +253,7 @@ clip_graph::clip_graph(clip_ctx * ctx, const clip_image_f32 & img) :
        n_embd(hparams.n_embd),
        n_head(hparams.n_head),
        n_head_kv(hparams.n_head_kv),
-        d_head(n_head > 0 ? n_embd / n_head : 0),
+        d_head(hparams.n_embd_head > 0 ? hparams.n_embd_head : (n_head > 0 ? n_embd / n_head : 0)),
        n_layer(hparams.n_layer),
        n_mmproj_embd(clip_n_mmproj_embd(ctx)),
        eps(hparams.eps),
@@ -372,13 +372,13 @@ ggml_tensor * clip_graph::build_vit(
                /* nb1    */ ggml_row_size(cur->type, d_head),
                /* nb2    */ cur->nb[1],
                /* nb3    */ cur->nb[1] * n_pos,
-                /* offset */ ggml_row_size(cur->type, n_embd));
+                /* offset */ ggml_row_size(cur->type, n_head * d_head));

                Vcur = ggml_view_4d(ctx0, cur, d_head, n_head, n_pos, B,
                /* nb1    */ ggml_row_size(cur->type, d_head),
                /* nb2    */ cur->nb[1],
                /* nb3    */ cur->nb[1] * n_pos,
-                /* offset */ ggml_row_size(cur->type, 2 * n_embd));
+                /* offset */ ggml_row_size(cur->type, 2 * n_head * d_head));

                if (layer.q_norm) {
                    GGML_ASSERT(layer.q_norm->ne[0] == Qcur->ne[0]);
@@ -1190,6 +1190,7 @@ struct clip_model_loader {
            const char * prefix = is_vision ? "vision" : "audio";
            get_u32(string_format(KEY_N_EMBD,         prefix), hparams.n_embd);
            get_u32(string_format(KEY_N_HEAD,         prefix), hparams.n_head);
+            get_u32(string_format(KEY_N_EMBD_HEAD,    prefix), hparams.n_embd_head, false);
            get_u32(string_format(KEY_N_FF,           prefix), hparams.n_ff);
            get_u32(string_format(KEY_N_BLOCK,        prefix), hparams.n_layer);
            get_u32(string_format(KEY_PROJ_DIM,       prefix), hparams.projection_dim);
@@ -9,6 +9,7 @@
 #include "peg-parser.h"

 #include <fstream>
+#include <iterator>
 #include <numeric>
 #include <optional>
 #include <sstream>
@@ -398,7 +399,7 @@ int main(int argc, char ** argv) {
        if (std::optional<common_chat_params> spec_tmpl =
                common_chat_try_specialized_template(chat_template, template_source, params)) {
            LOG_ERR("\n");
-            LOG_ERR("This template uses a specialized parser, analysis results will not be available.");
+            LOG_ERR("This template uses a specialized parser, analysis results will not be available.\n");
            parser_data = *spec_tmpl;
        } else {
            // Render template scenarios if requested
@@ -426,7 +427,9 @@ int main(int argc, char ** argv) {
                // Generate Parser
                parser_data = autoparser::peg_generator::generate_parser(chat_template, params, analysis);
            }
+        }

+        if (!std::empty(parser_data.parser)) {
            LOG_ERR("\n=== Generated Parser ===\n");
            common_peg_arena arena;
            arena.load(parser_data.parser);
Author	SHA1	Message	Date
Piotr Wilkin (ilintar)andGitHub	ddd4ec1428	chat : enable tool call in thinking for DS4 (#26269 )	2026-08-01 00:13:07 -05:00
Anand PatilandGitHub	876a432116	vulkan: add POOL_1D op (#25431 ) * vulkan : add pool1d push constants and pipeline field Declared data structures needed for POOL1D OP, which are the vk_op_pool1d_push_constants struct and pipeline_pool1d_f32 field. * vulkan : add pool1d compute shader Added pool1d.comp for Vulkan backend mirroring the existing pool2d shader. * vulkan : add full GGML_OP_POOL_1D support Added pipeline creation and op dispatch for 1D pooling in the Vulkan backend. * vulkan : fix pool1d shader logic Registered pool1d_f32 in vulkan-shaders-gen.cpp and fixed tensor dimension indices and avg pool scale. * vulkan : fix pool1d end boundary crash and expand test coverage Fixed an issue where the shader crashed when the end boundary was negative when k0 < p0. Also, added more test cases related to this fix.	2026-07-31 16:48:58 +02:00
Masato NakasakaandGitHub	eb41d503ba	vulkan: Introduce driver version check for Windows Intel GPU to mitigate crashing (#25192 ) * Removed crash guard for Intel Crash fixed from driver 32.0.101.8860 * Added driver version check for windows * Change to convert from driverVersion rather than string * No need to use signed * Refactor * allow GPU other than Xe2+ * adjusted function body position	2026-07-31 16:26:37 +02:00
Xuan-Son Nguyen GitHub Daniel Han	db7d8b24b5	mtmd: add n_embd_head (#26342 ) Co-authored-by: Daniel Han <unslothai@gmail.com>	2026-07-31 15:30:19 +02:00