vulkan: fast path for walsh-hadamard transform (#23687)

* vulkan: fast path for walsh-hadamard transform

* disable for intel due to segfault

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -860,6 +860,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_argsort_large_f32[num_argsort_pipelines];
     vk_pipeline pipeline_topk_f32[num_topk_pipelines];
     vk_pipeline pipeline_sum_rows_f32;
+    vk_pipeline pipeline_fwht_f32[4];
     vk_pipeline pipeline_cumsum_f32;
     vk_pipeline pipeline_cumsum_small_f32;
     vk_pipeline pipeline_cumsum_multipass1_f32;
@@ -1150,6 +1151,13 @@ struct vk_op_push_constants {
     float param4;
 };
 
+struct vk_op_fwht_push_constants {
+    uint32_t n_rows;
+    uint32_t src_offset;
+    uint32_t dst_offset;
+    float scale;
+};
+
 struct vk_op_count_experts_push_constants {
     uint32_t ne00;
     uint32_t ne01;
@@ -2055,6 +2063,15 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
     GGML_UNUSED(src3);
 }
 
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_fwht_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
+    p.src_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    p.dst_offset = get_misalign_bytes(ctx, dst)  / ggml_type_size(dst->type);
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src2);
+    GGML_UNUSED(src3);
+}
+
 struct ggml_backend_vk_buffer_context {
     vk_device_ref device;
     vk_buffer dev_buffer;
@@ -2095,9 +2112,9 @@ void vk_memory_logger::log_deallocation(vk_buffer_ref buf_ref) {
     const bool device = bool(buf->memory_property_flags & vk::MemoryPropertyFlagBits::eDeviceLocal);
     std::string type = device ? "device" : "host";
     auto it = allocations.find(buf->buffer);
-    total_device -= device ? it->second : 0;
-    total_host -= device ? 0 : it->second;
     if (it != allocations.end()) {
+        total_device -= device ? it->second : 0;
+        total_host -= device ? 0 : it->second;
         VK_LOG_MEMORY(buf->device->name << ": -" << format_size(it->second) << " " << type << " at " << buf->buffer << ". Total device: " << format_size(total_device) << ", total host: " << format_size(total_host));
         allocations.erase(it);
     } else {
@@ -4982,6 +4999,16 @@ static void ggml_vk_load_shaders(vk_device& device) {
     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) {
+        int idx = 0;
+        for (uint32_t n : {64, 128, 256, 512}) {
+            if (device->subgroup_size <= n) {
+                ggml_vk_create_pipeline(device, device->pipeline_fwht_f32[idx], "fwht_f32", fwht_f32_len, fwht_f32_data, "main", 2, sizeof(vk_op_fwht_push_constants), {1, 1, 1}, { device->subgroup_size, n }, 1, true, true, device->subgroup_size);
+            }
+            ++idx;
+        }
+    }
 
     const uint32_t cumsum_elem_per_thread = (device->vendor_id == VK_VENDOR_ID_AMD || device->vendor_id == VK_VENDOR_ID_INTEL) ? 2 : 4;
     ggml_vk_create_pipeline(device, device->pipeline_cumsum_f32,       "cumsum_f32", cumsum_f32_len, cumsum_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { 256, device->subgroup_size, cumsum_elem_per_thread }, 1, true, true, device->subgroup_size);
@@ -7233,7 +7260,7 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont
                             const uint64_t s_off = buf_offset + i3*nb3 + i2*nb2 + i1*nb1;
                             const uint64_t d_off = offset + i3*dstnb3 + i2*dstnb2 + i1*dstnb1;
                             for (uint64_t i0 = 0; i0 < ne0; i0++) {
-                                slices.push_back({ s_off + i1*nb0, d_off + i0*dstnb0, dstnb0 });
+                                slices.push_back({ s_off + i0*nb0, d_off + i0*dstnb0, dstnb0 });
                             }
                         }
                     }
@@ -8741,6 +8768,68 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
         }, pc, { (uint32_t)ne03, (uint32_t)ne01, (uint32_t)ne12 });
 }
 
+static int ggml_vk_fwht_pipeline_idx(int64_t n) {
+    switch (n) {
+        case 64:  return 0;
+        case 128: return 1;
+        case 256: return 2;
+        case 512: return 3;
+        default:  return -1;
+    }
+}
+
+static bool ggml_vk_can_use_fwht(const ggml_backend_vk_context * ctx, const ggml_tensor * src1, const ggml_tensor * dst) {
+    if (ctx->num_additional_fused_ops != 0) {
+        return false;
+    }
+
+    if (ggml_get_op_params_i32(dst, 1) != GGML_HINT_SRC0_IS_HADAMARD) {
+        return false;
+    }
+
+    const int idx = ggml_vk_fwht_pipeline_idx(src1->ne[0]);
+    if (idx < 0 || ctx->device->pipeline_fwht_f32[idx] == nullptr) {
+        return false;
+    }
+
+    if (src1->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
+        return false;
+    }
+
+    if (!ggml_is_contiguous(src1)) {
+        return false;
+    }
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    return true;
+}
+
+static void ggml_vk_fwht(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src, ggml_tensor * dst) {
+    const int idx = ggml_vk_fwht_pipeline_idx(src->ne[0]);
+    vk_pipeline pipeline = ctx->device->pipeline_fwht_f32[idx];
+
+    const uint32_t rows_per_workgroup = 4;
+    const uint32_t n_rows = (uint32_t)ggml_nrows(src);
+    const uint32_t max_workgroups_x = ctx->device->properties.limits.maxComputeWorkGroupCount[0];
+
+    const uint32_t total_workgroups = CEIL_DIV(n_rows, rows_per_workgroup);
+    const uint32_t workgroups_x = std::min(total_workgroups, max_workgroups_x);
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+
+    const vk_subbuffer src_buf = ggml_vk_tensor_subbuffer(ctx, src, true);
+    const vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst, true);
+
+    vk_op_fwht_push_constants pc = {
+        n_rows,
+        0,
+        0,
+        1.0f / std::sqrt((float)src->ne[0]),
+    };
+    init_pushconst_tensor_offsets(ctx, pc, src, nullptr, nullptr, nullptr, dst);
+
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src_buf, dst_buf }, pc, { workgroups_x, 1, 1 });
+}
+
 static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx) {
     ggml_tensor * dst = cgraph->nodes[node_idx];
     ggml_tensor * src0 = dst->src[0];
@@ -8774,6 +8863,8 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, c
 
             m_offset += cur_M_size;
         }
+    } else if (ggml_vk_can_use_fwht(ctx, src1, dst)) {
+        ggml_vk_fwht(ctx, subctx, src1, dst);
     } else if (src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && dst->ne[1] == 1 &&
         // detect 0213 permutation, and batch size of 1
         src0->nb[0] <= src0->nb[2] &&

ggml/src/ggml-vulkan/vulkan-shaders/fwht.comp

@@ -0,0 +1,69 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : require
+#extension GL_KHR_shader_subgroup_basic : enable
+#extension GL_KHR_shader_subgroup_shuffle : enable
+
+layout(local_size_x_id = 0, local_size_y = 4, local_size_z = 1) in;
+
+layout(constant_id = 0) const uint WARP_SIZE = 32;
+layout(constant_id = 1) const uint N = 128;
+
+layout(push_constant) uniform parameter
+{
+    uint n_rows;
+    uint src_offset;
+    uint dst_offset;
+    float scale;
+};
+
+layout(binding = 0, std430) readonly buffer A { float data_a[]; };
+layout(binding = 1, std430) writeonly buffer D { float data_d[]; };
+
+const uint EL_W = N / WARP_SIZE;
+
+void main() {
+    const uint lane = gl_SubgroupInvocationID;
+    for (uint row = gl_WorkGroupID.x * gl_WorkGroupSize.y + gl_SubgroupID;
+            row < n_rows;
+            row += gl_NumWorkGroups.x * gl_WorkGroupSize.y) {
+        const uint row_offset = row * N;
+
+        float reg[EL_W];
+
+        [[unroll]]
+        for (uint i = 0; i < EL_W; ++i) {
+            reg[i] = data_a[src_offset + row_offset + i * WARP_SIZE + lane] * scale;
+        }
+
+        [[unroll]]
+        for (uint h = 1; h < WARP_SIZE; h <<= 1) {
+            [[unroll]]
+            for (uint j = 0; j < EL_W; ++j) {
+                const float val = reg[j];
+                const float val2 = subgroupShuffleXor(val, h);
+                reg[j] = (lane & h) == 0 ? val + val2 : val2 - val;
+            }
+        }
+
+        [[unroll]]
+        for (uint h = WARP_SIZE; h < N; h <<= 1) {
+            const uint step = h / WARP_SIZE;
+            [[unroll]]
+            for (uint j = 0; j < EL_W; j += 2 * step) {
+                [[unroll]]
+                for (uint k = 0; k < step; ++k) {
+                    const float x = reg[j + k];
+                    const float y = reg[j + k + step];
+                    reg[j + k] = x + y;
+                    reg[j + k + step] = x - y;
+                }
+            }
+        }
+
+        [[unroll]]
+        for (uint i = 0; i < EL_W; ++i) {
+            data_d[dst_offset + row_offset + i * WARP_SIZE + lane] = reg[i];
+        }
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -934,6 +934,7 @@ void process_shaders() {
 
     string_to_spv("argmax_f32", "argmax.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "int"}}));
     string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    string_to_spv("fwht_f32", "fwht.comp", {});
     string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}}));
     string_to_spv("cumsum_f32", "cumsum.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("cumsum_multipass1_f32", "cumsum_multipass1.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));

models/templates/ibm-granite-granite-4.1.jinja

@@ -0,0 +1,114 @@
+{%- set tools_system_message_prefix = 'You are a helpful assistant with access to the following tools. You may call one or more tools to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>'  %}
+{%- set tools_system_message_suffix = '\n</tools>\n\nFor each tool call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call>. If a tool does not exist in the provided list of tools, notify the user that you do not have the ability to fulfill the request.' %}
+{%- set documents_system_message_prefix = 'You are a helpful assistant with access to the following documents. You may use one or more documents to assist with the user query.\n\nYou are given a list of documents within <documents></documents> XML tags:\n<documents>' %}
+{%- set documents_system_message_suffix = '\n</documents>\n\nWrite the response to the user\'s input by strictly aligning with the facts in the provided documents. If the information needed to answer the question is not available in the documents, inform the user that the question cannot be answered based on the available data.' %}
+{%- if available_tools is defined and available_tools %}
+    {%- set tools = available_tools %}
+{%- endif %}
+{%- set ns = namespace(tools_system_message=tools_system_message_prefix,
+                       documents_system_message=documents_system_message_prefix,
+                       system_message=''
+                       ) %}
+{%- if tools %}
+    {%- for tool in tools %}
+        {%- set ns.tools_system_message = ns.tools_system_message + '\n' + (tool | tojson) %}
+    {%- endfor %}
+    {%- set ns.tools_system_message = ns.tools_system_message + tools_system_message_suffix %}
+{%- else %}
+    {%- set ns.tools_system_message = '' %}
+{%- endif %}
+{%- if documents %}
+    {%- for document in documents %}
+        {%- set ns.documents_system_message = ns.documents_system_message + '\n' + (document | tojson) %}
+    {%- endfor %}
+    {%- set ns.documents_system_message = ns.documents_system_message + documents_system_message_suffix %}
+{%- else %}
+    {%- set ns.documents_system_message = '' %}
+{%- endif %}
+{%- if messages[0].role == 'system' %}
+    {%- if messages[0].content is string %}
+        {%- set ns.system_message = messages[0].content %}
+    {%- elif messages[0].content is iterable %}
+        {%- for entry in messages[0].content %}
+            {%- if entry.type== 'text' %}
+                {%- if ns.system_message != '' %}
+                    {%- set ns.system_message = ns.system_message + '\n' %}
+                {%- endif %}
+                {%- set ns.system_message = ns.system_message + entry.text %}
+            {%- endif %}
+        {%- endfor %}
+    {%- endif %}
+    {%- if tools and documents %}
+        {%- set ns.system_message = ns.system_message + '\n\n' +  ns.tools_system_message + '\n\n' + ns.documents_system_message %}
+    {%- elif tools %}
+        {%- set ns.system_message = ns.system_message + '\n\n' + ns.tools_system_message %}
+    {%- elif documents %}
+        {%- set ns.system_message = ns.system_message + '\n\n' + ns.documents_system_message %}
+    {%- endif %}
+{%- else %}
+    {%- if tools and documents %}
+        {%- set ns.system_message = ns.tools_system_message + '\n\n' + ns.documents_system_message  %}
+    {%- elif tools %}
+        {%- set ns.system_message = ns.tools_system_message %}
+    {%- elif documents %}
+        {%- set ns.system_message = ns.documents_system_message %}
+    {%- endif %}
+{%- endif %}
+{%- if ns.system_message %}
+    {{- '<|start_of_role|>system<|end_of_role|>' + ns.system_message + '<|end_of_text|>\n' }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = namespace(val='') %}
+    {%- if message.content is string %}
+        {%- set content.val = message.content %}
+    {%- else %}
+        {%- if message.content is iterable %}
+            {%- for entry in message.content %}
+                {%- if entry.type== 'text' %}
+                    {%- if content.val != '' %}
+                        {%- set content.val = content.val + '\n' %}
+                    {%- endif %}
+                    {%- set content.val = content.val + entry.text %}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+    {%- endif %}
+    {%- if (message.role == 'user') or (message.role == 'system' and not loop.first) %}
+        {{- '<|start_of_role|>' + message.role + '<|end_of_role|>' + content.val + '<|end_of_text|>\n' }}
+    {%- elif message.role == 'assistant' %}
+        {{- '<|start_of_role|>' + message.role + '<|end_of_role|>' + content.val }}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content.val) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|end_of_text|>\n' }}
+    {%- elif message.role == 'tool' %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != 'tool') %}
+            {{- '<|start_of_role|>user<|end_of_role|>' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content.val }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != 'tool') %}
+            {{- '<|end_of_text|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|start_of_role|>assistant<|end_of_role|>' }}
+{%- endif %}

src/llama-chat.cpp

@@ -62,6 +62,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "rwkv-world",        LLM_CHAT_TEMPLATE_RWKV_WORLD        },
     { "granite",           LLM_CHAT_TEMPLATE_GRANITE_3_X       },
     { "granite-4.0",       LLM_CHAT_TEMPLATE_GRANITE_4_0       },
+    { "granite-4.1",       LLM_CHAT_TEMPLATE_GRANITE_4_1       },
     { "gigachat",          LLM_CHAT_TEMPLATE_GIGACHAT          },
     { "megrez",            LLM_CHAT_TEMPLATE_MEGREZ            },
     { "yandex",            LLM_CHAT_TEMPLATE_YANDEX            },
@@ -194,7 +195,10 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_RWKV_WORLD;
     } else if (tmpl_contains("<|start_of_role|>")) {
         if (tmpl_contains("<tool_call>") || tmpl_contains("<tools>")) {
-            return LLM_CHAT_TEMPLATE_GRANITE_4_0;
+            if (tmpl_contains("g4_default_system_message")) {
+                return LLM_CHAT_TEMPLATE_GRANITE_4_0;
+            }
+            return LLM_CHAT_TEMPLATE_GRANITE_4_1;
         }
         return LLM_CHAT_TEMPLATE_GRANITE_3_X;
     } else if (tmpl_contains("message['role'] + additional_special_tokens[0] + message['content'] + additional_special_tokens[1]")) {
@@ -651,6 +655,20 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|start_of_role|>assistant<|end_of_role|>";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GRANITE_4_1) {
+        // IBM Granite 4.1 template
+        for (const auto & message : chat) {
+            std::string role(message->role);
+            if (role == "assistant_tool_call") {
+                ss << "<|start_of_role|>assistant<|end_of_role|><|tool_call|>";
+            } else {
+                ss << "<|start_of_role|>" << role << "<|end_of_role|>";
+            }
+            ss << message->content << "<|end_of_text|>\n";
+        }
+        if (add_ass) {
+            ss << "<|start_of_role|>assistant<|end_of_role|>";
+        }
     } else if (tmpl == LLM_CHAT_TEMPLATE_GIGACHAT) {
         // GigaChat template
         bool has_system = !chat.empty() && std::string(chat[0]->role) == "system";

src/llama-chat.h

@@ -41,6 +41,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_RWKV_WORLD,
     LLM_CHAT_TEMPLATE_GRANITE_3_X,
     LLM_CHAT_TEMPLATE_GRANITE_4_0,
+    LLM_CHAT_TEMPLATE_GRANITE_4_1,
     LLM_CHAT_TEMPLATE_GIGACHAT,
     LLM_CHAT_TEMPLATE_MEGREZ,
     LLM_CHAT_TEMPLATE_YANDEX,

tests/test-backend-ops.cpp

@@ -8318,6 +8318,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 128, 1, 128));
     test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 64, 1, 64));
     test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 256, 1, 256));
+    test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 512, 1, 512));
     test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 128, 32, 128));
     test_cases.emplace_back(new test_mul_mat_hadamard(GGML_TYPE_F32, GGML_TYPE_F32, 128, 4, 128, {2, 3}));
 

tests/test-chat-template.cpp

@@ -618,6 +618,16 @@ int main_automated_tests(void) {
         },
         {
             /* .name= */ "ibm-granite/granite-4.0 (tool call)",
+            /* .template_str= */ "{%- for message in messages %}\n    {%- if message['role'] == 'assistant_tool_call' %}\n    {{- '<|start_of_role|>assistant<|end_of_role|><|tool_call|>' + message['content'] + '<|end_of_text|>\\n' }}\n    {%- else %}\n    {{- '<|start_of_role|>' + message['role'] + '<|end_of_role|>' + message['content'] + '<|end_of_text|>\\n' }}\n    {%- endif %}\n    {%- if loop.last and add_generation_prompt %}\n    {{- '<|start_of_role|>assistant<|end_of_role|>' }}\n    {%- endif %}\n{%- endfor %}\n{# <tool_call> <tools> g4_default_system_message #}",
+            /* .expected_output= */       "<|start_of_role|>system<|end_of_role|>You are a helpful assistant<|end_of_text|>\n<|start_of_role|>user<|end_of_role|>Hello<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|>Hi there<|end_of_text|>\n<|start_of_role|>user<|end_of_role|>Who are you<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|>   I am an assistant   <|end_of_text|>\n<|start_of_role|>user<|end_of_role|>Another question<|end_of_text|>\n<|start_of_role|>user<|end_of_role|>What is the weather?<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|><|tool_call|><tool_call>\n{\"name\": \"get_weather\", \"arguments\": {\"location\": \"NYC\"}}\n</tool_call><|end_of_text|>\n<|start_of_role|>tool_response<|end_of_role|>{\"temperature\": 72}<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|>",
+            /* .expected_output_jinja= */ "",
+            /* .bos_token= */ "",
+            /* .eos_token= */ "",
+            /* .supported_with_jinja= */ true,
+            /* .extra_conversation= */ {{"user", "What is the weather?"}, {"assistant_tool_call", "<tool_call>\n{\"name\": \"get_weather\", \"arguments\": {\"location\": \"NYC\"}}\n</tool_call>"}, {"tool_response", "{\"temperature\": 72}"}},
+        },
+        {
+            /* .name= */ "ibm-granite/granite-4.1 (tool call)",
             /* .template_str= */ "{%- for message in messages %}\n    {%- if message['role'] == 'assistant_tool_call' %}\n    {{- '<|start_of_role|>assistant<|end_of_role|><|tool_call|>' + message['content'] + '<|end_of_text|>\\n' }}\n    {%- else %}\n    {{- '<|start_of_role|>' + message['role'] + '<|end_of_role|>' + message['content'] + '<|end_of_text|>\\n' }}\n    {%- endif %}\n    {%- if loop.last and add_generation_prompt %}\n    {{- '<|start_of_role|>assistant<|end_of_role|>' }}\n    {%- endif %}\n{%- endfor %}\n{# <tool_call> <tools> #}",
             /* .expected_output= */       "<|start_of_role|>system<|end_of_role|>You are a helpful assistant<|end_of_text|>\n<|start_of_role|>user<|end_of_role|>Hello<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|>Hi there<|end_of_text|>\n<|start_of_role|>user<|end_of_role|>Who are you<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|>   I am an assistant   <|end_of_text|>\n<|start_of_role|>user<|end_of_role|>Another question<|end_of_text|>\n<|start_of_role|>user<|end_of_role|>What is the weather?<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|><|tool_call|><tool_call>\n{\"name\": \"get_weather\", \"arguments\": {\"location\": \"NYC\"}}\n</tool_call><|end_of_text|>\n<|start_of_role|>tool_response<|end_of_role|>{\"temperature\": 72}<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|>",
             /* .expected_output_jinja= */ "",

tests/test-chat.cpp

@@ -2914,6 +2914,21 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .run();
     }
 
+    {
+        // IBM Granite 4.1 (same format as 4.0)
+        auto tst = peg_tester("models/templates/ibm-granite-granite-4.1.jinja", detailed_debug);
+
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+
+        tst.test(
+               "<tool_call>\n"
+               "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+               "</tool_call>")
+            .tools({ special_function_tool })
+            .expect(message_assist_call)
+            .run();
+    }
+
     {
         // ByteDance-Seed-OSS (reasoning and tool calling model)
         auto tst = peg_tester("models/templates/ByteDance-Seed-OSS.jinja", detailed_debug);