hip: compile debug builds with -O2 on hip to avoid a compiler bug (#20392)

Co-authored-by: Mishusha <pmv26021975@gmail.com>

common/chat.cpp

@@ -1354,6 +1354,77 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
     return data;
 }
 
+static common_chat_params common_chat_params_init_gigachat_v3(
+        const common_chat_template & tmpl,
+        const autoparser::templates_params & inputs) {
+
+    common_chat_params data;
+
+    data.prompt            = common_chat_template_direct_apply(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = false;
+    data.preserved_tokens  = {
+        "<|message_sep|>\n\n",
+        "<|role_sep|>\n",
+    };
+
+    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
+    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+    auto tool_call_start_prefix = "<|message_sep|>\n\nfunction call<|role_sep|>\n";
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            // Build a choice of all available tools
+            auto tool_choice = p.choice();
+            for (const auto & tool : inputs.tools) {
+                const auto & function = tool.at("function");
+                std::string name = function.at("name");
+                const auto & schema = function.at("parameters");
+
+                auto tool_name = p.json_member("name", "\"" + p.tool_name(p.literal(name)) + "\"");
+                auto tool_args = p.json_member("arguments", p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema)));
+
+                auto tool_open = p.tool_open(p.literal("{") << tool_name);
+
+                tool_choice |= p.rule("tool-" + name, tool_open << "," << tool_args << "}");
+            }
+
+            // Define the tool call structure
+            auto min_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
+            auto max_calls = 1; // parallel toolcalls are not supported
+            auto tool_call = p.rule("tool-call", p.literal(tool_call_start_prefix) + tool_choice);
+            auto tool_calls = p.trigger_rule("tool-call-root", p.repeat(tool_call, /* min = */ min_calls, /* max = */ max_calls));
+
+            return p.content(p.until("<|message_sep|>\n\n")) << tool_calls;
+        }
+
+        // Content only parser
+        include_grammar = false;
+        return p.content(p.rest());
+
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto schema = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            {COMMON_GRAMMAR_TRIGGER_TYPE_WORD, tool_call_start_prefix}
+        };
+    }
+    return data;
+}
+
 namespace workaround {
 
 static void map_developer_role_to_system(json & messages) {
@@ -1525,6 +1596,15 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
         return common_chat_params_init_lfm2(tmpl, params);
     }
 
+    // GigaChatV3 format detection
+    if (src.find("<|role_sep|>") != std::string::npos &&
+        src.find("<|message_sep|>") != std::string::npos &&
+        src.find("<|function_call|>") == std::string::npos
+    ) {
+        LOG_DBG("Using specialized template: GigaChatV3\n");
+        return common_chat_params_init_gigachat_v3(tmpl, params);
+    }
+
     try {
         LOG_DBG("Using differential autoparser\n");
         struct autoparser::autoparser autoparser;

common/common.h

@@ -926,7 +926,7 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 // MoE utils
 //
 
-const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate)_(ch|)exps";
+const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate|gate_up)_(ch|)exps";
 
 inline std::string llm_ffn_exps_block_regex(int idx) {
     return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);

convert_hf_to_gguf.py

@@ -5062,7 +5062,7 @@ class Phi2Model(TextModel):
         self.gguf_writer.add_add_bos_token(False)
 
 
-@ModelBase.register("Phi3ForCausalLM")
+@ModelBase.register("Phi3ForCausalLM", "Phi4ForCausalLMV")
 class Phi3MiniModel(TextModel):
     model_arch = gguf.MODEL_ARCH.PHI3
 
@@ -5237,6 +5237,129 @@ class Phi3MiniModel(TextModel):
         yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32))
         yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32))
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith(("model.vision_tower.", "vision_tower.", "model.mm_projector.", "mm_projector.")):
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Phi4ForCausalLMV")
+class Phi4VisionMmprojModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+
+        self.vision_total_layers = int(self.find_vparam(self.n_block_keys))
+        if self.vision_total_layers < 2:
+            raise ValueError(
+                f"Phi-4 vision mmproj conversion requires at least 2 vision layers, got {self.vision_total_layers}"
+            )
+
+        # Phi-4 uses SigLIP2 hidden_states[-2], so export one fewer encoder block and
+        # drop post-layernorm/head weights. This makes the GGUF runtime output match
+        # the feature map consumed by the patched siglip.cpp Phi-4 projector path.
+        self.vision_export_layers = self.vision_total_layers - 1
+        self.vision_last_layer_idx = self.vision_total_layers - 1
+
+        for key in self.n_block_keys:
+            if key in self.hparams_vision:
+                self.hparams_vision[key] = self.vision_export_layers
+                break
+
+        self.block_count = self.vision_export_layers
+        self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.MMPROJ, self.block_count)
+
+        patch_size = self.preprocessor_config.get("patch_size")
+        if patch_size is None:
+            raise KeyError("Phi-4 vision mmproj conversion requires patch_size in preprocessor_config.json")
+
+        self.hparams_vision["patch_size"] = patch_size
+
+        pos_emb_name = next(
+            (
+                name for name in self.model_tensors
+                if name.endswith("vision_model.embeddings.position_embedding.weight")
+            ),
+            None,
+        )
+        if pos_emb_name is None:
+            raise KeyError("Phi-4 vision mmproj conversion could not find position_embedding.weight")
+
+        pos_emb_shape = self.model_tensors[pos_emb_name]().shape
+        base_grid_tokens = int(pos_emb_shape[0])
+        grid_side = math.isqrt(base_grid_tokens)
+        if grid_side * grid_side != base_grid_tokens:
+            raise ValueError(f"Unexpected Phi-4 position embedding shape: {tuple(pos_emb_shape)}")
+
+        self.hparams_vision["image_size"] = grid_side * patch_size
+
+        min_num_patches = self.preprocessor_config.get("min_num_patches", self.global_config.get("min_num_patches"))
+        max_num_patches = self.preprocessor_config.get("max_num_patches", self.global_config.get("max_num_patches"))
+        if min_num_patches is None or max_num_patches is None:
+            raise KeyError("Phi-4 vision mmproj conversion requires min_num_patches and max_num_patches")
+
+        self.min_pixels = int(min_num_patches) * patch_size * patch_size
+        self.max_pixels = int(max_num_patches) * patch_size * patch_size
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.PHI4)
+        self.gguf_writer.add_vision_min_pixels(self.min_pixels)
+        self.gguf_writer.add_vision_max_pixels(self.max_pixels)
+        self.gguf_writer.add_vision_use_gelu(True)
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("layer_norm_eps", 1e-6))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith(("model.vision_tower.vision_tower.", "vision_tower.")):
+            if ".vision_model.head." in name:
+                return
+
+            new_name = name.replace("model.vision_tower.vision_tower.", "vision_tower.")
+
+            if ".vision_model.post_layernorm." in new_name:
+                return
+
+            if bid is not None and bid == self.vision_last_layer_idx:
+                return
+
+            if new_name.endswith("vision_model.embeddings.patch_embedding.weight"):
+                assert self.hparams_vision is not None
+                if data_torch.ndim != 2:
+                    raise ValueError(f"Unexpected Phi-4 patch embedding shape: {tuple(data_torch.shape)}")
+
+                patch_area = self.hparams_vision["patch_size"] ** 2
+                in_features = data_torch.shape[1]
+                if in_features % patch_area != 0:
+                    raise ValueError(
+                        f"Phi-4 patch embedding input dim {in_features} is not divisible by patch area {patch_area}"
+                    )
+
+                num_channels = in_features // patch_area
+                patch_size = self.hparams_vision["patch_size"]
+                data_torch = data_torch.view(data_torch.shape[0], patch_size, patch_size, num_channels)
+                data_torch = data_torch.permute(0, 3, 1, 2)
+
+            yield from super().modify_tensors(data_torch, new_name, bid)
+            return
+
+        if name.startswith(("model.mm_projector.", "mm_projector.")):
+            local_name = name
+            local_name = local_name.replace("model.mm_projector.", "")
+            local_name = local_name.replace("mm_projector.", "")
+
+            if not (local_name.startswith("0.") or local_name.startswith("2.")):
+                return
+
+            suffix = ".bias" if local_name.endswith(".bias") else ".weight"
+            mm_idx = int(local_name.split(".", maxsplit=1)[0])
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.V_MMPROJ, mm_idx, suffix=suffix), data_torch)
+            return
+
+        return
+
 
 @ModelBase.register("PhiMoEForCausalLM")
 class PhiMoeModel(Phi3MiniModel):

ggml/src/ggml-hip/CMakeLists.txt

@@ -11,6 +11,10 @@ endif()
 list(APPEND CMAKE_PREFIX_PATH  ${ROCM_PATH})
 list(APPEND CMAKE_PREFIX_PATH "${ROCM_PATH}/lib64/cmake")
 
+if (NOT DEFINED CMAKE_HIP_FLAGS_DEBUG)
+    set(CMAKE_HIP_FLAGS_DEBUG "-g -O2")
+endif()
+
 # CMake on Windows doesn't support the HIP language yet
 if (WIN32)
     set(CXX_IS_HIPCC TRUE)

gguf-py/gguf/constants.py

@@ -3881,6 +3881,7 @@ class VisionProjectorType:
     GEMMA3 = "gemma3"
     GEMMA3NV = "gemma3nv"
     GEMMA3NA = "gemma3na"
+    PHI4 = "phi4"
     IDEFICS3 = "idefics3"
     PIXTRAL = "pixtral"
     LLAMA4 = "llama4"

models/templates/GigaChat3-10B-A1.8B.jinja

@@ -0,0 +1,355 @@
+{#--------TOOL RENDERING FUNCTIONS---------#}
+
+{#---------------------------------------------------------------
+  Converts JSON Schema (dict) to a TypeScript type definition
+----------------------------------------------------------------#}
+{%- macro json_schema_to_typescript(schema, indent="") -%}
+	{%- set ADDITIONAL_JSON_KEYS = ['format', 'maxItems', 'maximum', 'minItems', 'minimum', 'pattern'] -%}
+	{%- set ty = schema.get("type") -%}
+
+	{# ---------------- OBJECT ---------------- #}
+	{%- if ty == "object" -%}
+		{{- "{\n" -}}
+
+		{# Start building property list #}
+		{%- set props = schema.get("properties", {}) -%}
+		{%- set required = schema.get("required", []) -%}
+		{%- set has_additional_props = schema.get("additionalProperties") is defined -%}
+		{%- set additional_props_type = none -%}
+		{%- if has_additional_props -%}
+			{%- if schema.additionalProperties == true -%}
+				{%- set additional_props_type = {'type': 'any'} -%}
+			{%- elif schema.additionalProperties is mapping -%}
+				{%- set additional_props_type = schema.additionalProperties -%}
+			{%- endif -%}
+		{%- endif -%}
+
+		{%- for key, val in props.items() -%}
+			{# ---------- Description Comments ---------- #}
+			{%- if "description" in val -%}
+				{%- for line in val['description'].split('\n') -%}
+					{%- if line.strip() -%}
+						{{- indent + '// ' + line + '\n' -}}
+					{%- endif -%}
+				{%- endfor -%}
+			{%- endif -%}
+
+			{# ---------- Additional JSON Keys ---------- #}
+			{%- for add_key, add_val in val.items() -%}
+				{%- if add_key in ADDITIONAL_JSON_KEYS -%}
+					{%- if add_val is string -%}
+						{{- indent + '// ' + add_key + ': "' + add_val + '"' + '\n'  -}}
+					{%- else -%}
+						{{- indent + '// ' + add_key + ': ' ~ add_val ~ '\n' -}}
+					{%- endif -%}
+				{%- endif -%}
+			{%- endfor -%}
+
+			{# ---------- Property Definition ---------- #}
+            {%- set type_str = json_schema_to_typescript(
+                val, 
+                indent + "  "
+            ) -%}
+
+			{{- indent + key + ('' if key in required else '?') + ': ' + type_str + ',' -}}
+
+			{%- if "default" in val or "defalut_value" in val -%}
+				{%- set default = val.get("default", val.get("defalut_value")) -%}
+				{%- if default is string -%}
+					{{- ' // default: "' + default + '"' -}}
+				{%- else -%}
+					{{- ' // default: ' ~ default -}}
+				{%- endif -%}
+			{%- endif -%}
+
+			{{- "\n" -}}
+		{%- endfor -%}
+
+		{# Handle additionalProperties as index signature #}
+		{%- if has_additional_props and additional_props_type is not none -%}
+			{%- set additional_type_str = json_schema_to_typescript(
+				additional_props_type,
+				indent + "  "
+			) -%}
+			{{- indent + '[key: string]: ' + additional_type_str + '\n' -}}
+		{%- endif -%}
+
+		{{- indent[: (indent|length - "  "|length) ] + '}' -}}
+
+	{# ---------------- STRING ---------------- #}
+	{%- elif ty == "string" -%}
+		{%- if schema.get("enum") -%}
+            {%- set ns = namespace(enum = []) -%}
+            {%- for en in schema['enum'] -%}
+                {%- set ns.enum = ns.enum + ['"' ~ en ~ '"'] -%}
+            {%- endfor -%}
+			{{- ns.enum | join(' | ') -}}
+		{%- elif schema.get("format", "none") in ['date-time', 'date'] -%}
+			{{- 'Date' -}}
+		{%- else -%}
+			{{- 'string' -}}
+		{%- endif -%}
+
+	{# ---------------- NUMBER / INTEGER ---------------- #}
+	{%- elif ty in ["number", "integer"] -%}
+		{%- if schema.get("enum") -%}
+			{{- schema.enum | join(' | ') -}}
+		{%- else -%}
+			{{- 'number' -}}
+		{%- endif -%}
+
+	{# ---------------- BOOLEAN ---------------- #}
+	{%- elif ty == "boolean" -%}
+		{{- 'boolean' -}}
+
+	{# ---------------- ARRAY ---------------- #}
+	{%- elif ty == "array" -%}
+		{%- if "items" in schema -%}
+			{{- json_schema_to_typescript(schema['items'], indent) + '[]' -}}
+		{%- else -%}
+			{{- 'Array<any>' -}}
+		{%- endif -%}
+
+	{# ---------------- FALLBACK ---------------- #}
+	{%- else -%}
+		{{- 'any' -}}
+	{%- endif -%}
+{%- endmacro -%}
+
+{#---------------------------------------------------------------
+  Renders a namespace and its tool definitions in TypeScript style
+----------------------------------------------------------------#}
+
+{%- macro render_tool_namespace(namespace_name, tools) -%}
+	{%- set ns = namespace(sections = ['namespace ' ~ namespace_name ~ ' {']) -%}
+
+	{%- for tool in tools -%}
+		{%- if tool.function -%}
+			{%- set tool = tool.function -%}
+		{%- endif -%}
+
+		{%- set ns_tool = namespace(content_lines=[]) -%}
+
+		{# ---------- TOOL DESCRIPTION ---------- #}
+		{%- if tool.get('description') -%}
+			{%- for line in tool['description'].split('\n') -%}
+				{%- if line.strip() -%}
+					{%- set ns_tool.content_lines = ns_tool.content_lines + ['// ' ~ line] -%}
+				{%- endif -%}
+			{%- endfor -%}
+		{%- endif -%}
+
+		{# ---------- TOOL SIGNATURE ---------- #}
+		{%- set main_body = "" -%}
+		{%- set params = tool.get("parameters") -%}
+		{%- if params and params.get("properties") -%}
+			{%- set param_type = json_schema_to_typescript(params, "  ") -%}
+			{%- set main_body = 'type ' ~ tool.name ~ ' = (_: ' ~ param_type ~ ') => ' -%}
+		{%- else -%}
+			{%- set main_body = 'type ' ~ tool.name ~ ' = () => ' -%}
+		{%- endif -%}
+
+		{# ---------- RETURN TYPE ---------- #}
+		{%- set return_params = tool.get("return_parameters") -%}
+		{%- if return_params and return_params.get("properties") -%}
+			{%- set return_type = json_schema_to_typescript(return_params, "  ") -%}
+			{%- set main_body = main_body ~ return_type -%}
+		{%- else -%}
+			{%- set main_body = main_body ~ 'any' -%}
+		{%- endif -%}
+
+		{%- set main_body = main_body ~ ';\n' -%}
+
+		{%- set ns_tool.content_lines = ns_tool.content_lines + [main_body] -%}
+
+		{# ---------- ADD TOOL TO SECTIONS ---------- #}
+		{%- set ns.sections = ns.sections + [ns_tool.content_lines | join('\n')] -%}
+	{%- endfor -%}
+
+	{%- set ns.sections = ns.sections + ['} // namespace ' ~ namespace_name] -%}
+
+	{{- ns.sections | join('\n') -}}
+{%- endmacro -%}
+
+
+{# ----------- MESSAGE RENDERING HELPER FUNCTIONS ------------ #}
+
+{%- macro render_role_message(message, role=None) -%}
+    {%- if not role -%}
+        {%- set role = message["role"] -%}
+    {%- endif -%}
+
+    {%- set message_content = message['content'] or '' -%}
+    {%- if message_content is not string -%}
+        {%- set message_content = message_content | tojson(ensure_ascii=False) -%}
+    {%- endif -%}
+
+    {{- role + add_tokens.role_sep + message_content + add_tokens.message_sep -}}
+
+{%- endmacro -%}
+
+
+{%- macro render_function_call(message) -%}
+    {%- set call = message['content'] -%}
+    {%- if call.function -%}
+        {%- set call = call.function -%}
+    {%- endif -%}
+
+    {%- set arguments = call['arguments'] -%}
+	{%- if arguments is not string -%}
+		{%- set arguments = arguments| tojson(ensure_ascii=False) -%}
+	{%- endif -%}
+
+    {{- render_role_message(
+        {
+            'role': 'function call',
+            'content': '{"name": "' ~ call['name'] ~ '", "arguments": ' ~ arguments ~ '}'
+        }
+    ) -}}
+{%- endmacro -%}
+
+{# ----- SPECIAL TOKENS ----- #}
+
+{%- set add_tokens = namespace(
+    role_sep="<|role_sep|>\n", 
+    message_sep="<|message_sep|>\n\n"
+) -%}
+
+{# ----- DEFAULT DEVSYSTEM ----- #}
+
+{%- set DEVSYSTEM -%}
+<role_description>
+Description of the roles available in the dialog.
+
+`developer system`
+A message added by Sber before the main dialog. It has the highest priority and sets global, non-overridable conditions (for example, conversation rules, the safety policy, the assistant's overall response style, etc.).
+
+`system`
+A system instruction added by developers or by the user, but with a lower priority than `developer system`. It usually describes the assistant's instructions, a specific response style, and other conditions for this particular dialog.
+
+`user`
+A message or request from the user. The assistant follows it if it does not conflict with higher-priority instructions (see <instruction_priority>).
+
+`user memory`
+A sequence of the most up-to-date long-term facts about the user at the time of their request, presented as a JSON list of strings. Facts are listed in chronological order, meaning newer facts are appended to the end of the sequence. When facts are changed or deleted, records of previous facts remain in the sequence. The assistant saves facts using a function and uses them in accordance with the <memory_guidelines> block below.
+
+`added files`
+Metadata about files available for use in the dialog, presented in JSON format. It contains the following keys: id (a unique file identifier), name (file name), type (file type).
+
+`assistant`
+The assistant's reply to the user's request. If the system instruction or the user does not set additional rules for `assistant`, this reply must comply with the instructions in the <assistant_guidelines> block below. The list of functions available to call is contained in `function descriptions`. The name of the required function and its arguments will be generated next by the `function call` role. In its replies, the assistant follows the instructions in accordance with <instruction_priority>.
+
+`function descriptions`
+Function descriptions in TypeScript format. A function is a special tool (or a set of instructions) that the assistant can call to perform specific actions, computations, or obtain data needed to solve the user's task. Each function description contains blocks with the name, description, and arguments. Sometimes the description contains separate blocks with return parameters and usage examples that illustrate the correct call and arguments.
+
+`function call`
+The function that `assistant` calls based on the dialog context, and its arguments. The function is invoked in strict accordance with the instructions in the <function_usage> block.
+
+`function result`
+The result of the last function call.
+</role_description>
+
+<available_modalities>
+The assistant can work with the following modalities: text, available functions.
+</available_modalities>
+
+<instruction_priority>
+If instructions from different roles conflict within the dialog context, observe the following priorities:  
+`developer system` > `system` > `user` > `function descriptions` > `function result` > `user memory`
+</instruction_priority>
+
+<function_usage>
+Basic instructions for working with functions.
+
+Only call those functions that are described in `function descriptions`.
+
+Call available functions when, according to their description, such a call will help provide a more complete and/or accurate answer to the user's request. Fill in function arguments using information from the dialog context. If a function could help answer the request but a required argument is missing from the context, ask the user for the missing data before calling the function. If a necessary function is unavailable or an error occurs, briefly inform the user and, if possible, suggest an alternative.
+</function_usage>
+
+<memory_guidelines>
+Rules for using facts in long-term memory:
+
+If there is no message under the `user memory` role in the dialog, this is equivalent to the absence of long-term facts about the user in memory. In that case, information about the user is limited to the current dialog, and no new facts should be saved.
+</memory_guidelines>
+
+<assistant_guidelines>
+You are a helpful assistant.
+
+# Instructions
+- Strictly follow the instruction priority.
+- Maintain a logical chain of reasoning when answering the user's question.
+- For complex questions (for example, STEM), try to answer in detail unless the system message or dialog context limits the response length.
+- Be helpful, truthful, and avoid unsafe or prohibited content in your responses.
+- Try to reply in the language in which the user asked their question.
+</assistant_guidelines>
+
+A dialog will follow below.
+The dialog may include various roles described in the <role_description> block.
+Each turn begins with the role name and a special token that marks the end of the role's full name, and ends with a special end-of-turn token.
+Your task is to continue the dialog from the last specified role in accordance with the dialog context.
+{%- endset -%}
+
+
+{#- ---------------------- RENDERING STARTS HERE ---------------------- -#}
+
+
+{# ----- RENDER BOS TOKEN ----- #}
+{{- bos_token -}}
+
+
+{# ----- RENDER DEVSYSTEM ----- #}
+{{- render_role_message({"role": "developer system", "content": DEVSYSTEM}) -}}
+
+{# ----- RENDER SYSTEM IF PRESENT ----- #}
+{%- if messages and messages[0]['role'] == 'system' -%}
+    {{- render_role_message(messages[0]) -}}
+    {%- set messages = messages[1:] -%}
+{%- endif -%}
+
+{# ----- RENDER TOOLS ----- #}
+{%- if tools -%}
+    {%- set tools_content = (
+        render_tool_namespace('functions', tools) 
+        + "\n\n"
+    ) -%}
+    {{- render_role_message({'role': 'function descriptions', 'content': tools_content}) -}}
+{%- endif -%}
+
+{# ----- MAIN MESSAGE LOOP ----- #}
+{%- for message in messages -%}
+
+    {# ----- TOOL MESSAGE -------#}
+    {%- if message['role'] == 'tool' -%}
+        {{- render_role_message(message, 'function result') -}}
+
+
+    {# ----- ASSISTANT MESSAGE ----- #}
+    {%- elif message['role'] == 'assistant' -%}
+
+        {# ----- FUNCTION CALL PART CHECKING: SINGLE CALL SETUP ----- #}
+        {%- if message.tool_calls is defined and message.tool_calls -%}
+            {%- set function_call = message.tool_calls[0] -%}
+        {%- else -%}
+            {%- set function_call = None -%}
+        {%- endif -%}
+
+        {# ----- MAIN ASSISTANT RENDERING ----- #}
+
+        {{- render_role_message({'role': 'assistant', 'content': message.content}) -}}
+        {%- if function_call -%}
+            {{- render_function_call({'role': 'function call', 'content': function_call}) -}}
+        {%- endif -%}
+
+        
+    {# ----- OTHER MESSAGES ----- #}
+    {%- else -%}
+        {{- render_role_message(message) -}}
+    {%- endif -%}
+
+    {# ----- ADDING GENERATION PROMPT ----- #}
+
+    {%- if loop.last and add_generation_prompt and message['role'] != 'assistant' -%}
+        {{- 'assistant' + add_tokens.role_sep -}}
+    {%- endif -%}
+
+{%- endfor -%}

models/templates/GigaChat3.1-10B-A1.8B.jinja

@@ -0,0 +1,339 @@
+{#--------TOOL RENDERING FUNCTIONS---------#}
+
+{#---------------------------------------------------------------
+  Converts JSON Schema (dict) to a TypeScript type definition
+----------------------------------------------------------------#}
+{%- macro json_schema_to_typescript(schema, indent="") -%}
+	{%- set ADDITIONAL_JSON_KEYS = ['format', 'maxItems', 'maximum', 'minItems', 'minimum', 'pattern'] -%}
+	{%- set ty = schema.get("type") -%}
+
+	{# ---------------- OBJECT ---------------- #}
+	{%- if ty == "object" -%}
+		{{- "{\n" -}}
+
+		{# Start building property list #}
+		{%- set props = schema.get("properties", {}) -%}
+		{%- set required = schema.get("required", []) -%}
+		{%- set has_additional_props = schema.get("additionalProperties") is defined -%}
+		{%- set additional_props_type = none -%}
+		{%- if has_additional_props -%}
+			{%- if schema.additionalProperties == true -%}
+				{%- set additional_props_type = {'type': 'any'} -%}
+			{%- elif schema.additionalProperties is mapping -%}
+				{%- set additional_props_type = schema.additionalProperties -%}
+			{%- endif -%}
+		{%- endif -%}
+
+		{%- for key, val in props.items() -%}
+			{# ---------- Description Comments ---------- #}
+			{%- if "description" in val -%}
+				{%- for line in val['description'].split('\n') -%}
+					{%- if line.strip() -%}
+						{{- indent + '// ' + line + '\n' -}}
+					{%- endif -%}
+				{%- endfor -%}
+			{%- endif -%}
+
+			{# ---------- Additional JSON Keys ---------- #}
+			{%- for add_key, add_val in val.items() -%}
+				{%- if add_key in ADDITIONAL_JSON_KEYS -%}
+					{%- if add_val is string -%}
+						{{- indent + '// ' + add_key + ': "' + add_val + '"' + '\n'  -}}
+					{%- else -%}
+						{{- indent + '// ' + add_key + ': ' ~ add_val ~ '\n' -}}
+					{%- endif -%}
+				{%- endif -%}
+			{%- endfor -%}
+
+			{# ---------- Property Definition ---------- #}
+            {%- set type_str = json_schema_to_typescript(
+                val, 
+                indent + "  "
+            ) -%}
+
+			{{- indent + key + ('' if key in required else '?') + ': ' + type_str + ',' -}}
+
+			{%- if "default" in val or "defalut_value" in val -%}
+				{%- set default = val.get("default", val.get("defalut_value")) -%}
+				{%- if default is string -%}
+					{{- ' // default: "' + default + '"' -}}
+				{%- else -%}
+					{{- ' // default: ' ~ default -}}
+				{%- endif -%}
+			{%- endif -%}
+
+			{{- "\n" -}}
+		{%- endfor -%}
+
+		{# Handle additionalProperties as index signature #}
+		{%- if has_additional_props and additional_props_type is not none -%}
+			{%- set additional_type_str = json_schema_to_typescript(
+				additional_props_type,
+				indent + "  "
+			) -%}
+			{{- indent + '[key: string]: ' + additional_type_str + '\n' -}}
+		{%- endif -%}
+
+		{{- indent[: (indent|length - "  "|length) ] + '}' -}}
+
+	{# ---------------- STRING ---------------- #}
+	{%- elif ty == "string" -%}
+		{%- if schema.get("enum") -%}
+            {%- set ns = namespace(enum = []) -%}
+            {%- for en in schema['enum'] -%}
+                {%- set ns.enum = ns.enum + ['"' ~ en ~ '"'] -%}
+            {%- endfor -%}
+			{{- ns.enum | join(' | ') -}}
+		{%- elif schema.get("format", "none") in ['date-time', 'date'] -%}
+			{{- 'Date' -}}
+		{%- else -%}
+			{{- 'string' -}}
+		{%- endif -%}
+
+	{# ---------------- NUMBER / INTEGER ---------------- #}
+	{%- elif ty in ["number", "integer"] -%}
+		{%- if schema.get("enum") -%}
+			{{- schema.enum | join(' | ') -}}
+		{%- else -%}
+			{{- 'number' -}}
+		{%- endif -%}
+
+	{# ---------------- BOOLEAN ---------------- #}
+	{%- elif ty == "boolean" -%}
+		{{- 'boolean' -}}
+
+	{# ---------------- ARRAY ---------------- #}
+	{%- elif ty == "array" -%}
+		{%- if "items" in schema -%}
+			{{- json_schema_to_typescript(schema['items'], indent) + '[]' -}}
+		{%- else -%}
+			{{- 'Array<any>' -}}
+		{%- endif -%}
+
+	{# ---------------- FALLBACK ---------------- #}
+	{%- else -%}
+		{{- 'any' -}}
+	{%- endif -%}
+{%- endmacro -%}
+
+{#---------------------------------------------------------------
+  Renders a namespace and its tool definitions in TypeScript style
+----------------------------------------------------------------#}
+
+{%- macro render_tool_namespace(namespace_name, tools) -%}
+	{%- set ns = namespace(sections = ['namespace ' ~ namespace_name ~ ' {']) -%}
+
+	{%- for tool in tools -%}
+		{%- if tool.function -%}
+			{%- set tool = tool.function -%}
+		{%- endif -%}
+
+		{%- set ns_tool = namespace(content_lines=[]) -%}
+
+		{# ---------- TOOL DESCRIPTION ---------- #}
+		{%- if tool.get('description') -%}
+			{%- for line in tool['description'].split('\n') -%}
+				{%- if line.strip() -%}
+					{%- set ns_tool.content_lines = ns_tool.content_lines + ['// ' ~ line] -%}
+				{%- endif -%}
+			{%- endfor -%}
+		{%- endif -%}
+
+		{# ---------- TOOL SIGNATURE ---------- #}
+		{%- set main_body = "" -%}
+		{%- set params = tool.get("parameters") -%}
+		{%- if params and params.get("properties") -%}
+			{%- set param_type = json_schema_to_typescript(params, "  ") -%}
+			{%- set main_body = 'type ' ~ tool.name ~ ' = (_: ' ~ param_type ~ ') => ' -%}
+		{%- else -%}
+			{%- set main_body = 'type ' ~ tool.name ~ ' = () => ' -%}
+		{%- endif -%}
+
+		{# ---------- RETURN TYPE ---------- #}
+		{%- set return_params = tool.get("return_parameters") -%}
+		{%- if return_params and return_params.get("properties") -%}
+			{%- set return_type = json_schema_to_typescript(return_params, "  ") -%}
+			{%- set main_body = main_body ~ return_type -%}
+		{%- else -%}
+			{%- set main_body = main_body ~ 'any' -%}
+		{%- endif -%}
+
+		{%- set main_body = main_body ~ ';\n' -%}
+
+		{%- set ns_tool.content_lines = ns_tool.content_lines + [main_body] -%}
+
+		{# ---------- ADD TOOL TO SECTIONS ---------- #}
+		{%- set ns.sections = ns.sections + [ns_tool.content_lines | join('\n')] -%}
+	{%- endfor -%}
+
+	{%- set ns.sections = ns.sections + ['} // namespace ' ~ namespace_name] -%}
+
+	{{- ns.sections | join('\n') -}}
+{%- endmacro -%}
+
+
+{# ----------- MESSAGE RENDERING HELPER FUNCTIONS ------------ #}
+
+{%- macro render_function_call(call) -%}
+    {%- if call.function -%}
+        {%- set call = call.function -%}
+    {%- endif -%}
+
+    {%- set arguments = call['arguments'] -%}
+	{%- if arguments is not string -%}
+		{%- set arguments = arguments| tojson(ensure_ascii=False) -%}
+	{%- endif -%}
+
+    {{- '{"name": "' ~ call['name'] ~ '", "arguments": ' ~ arguments ~ '}' -}}
+{%- endmacro -%}
+
+
+{%- macro render_role_message(message, role=None) -%}
+    {%- if not role -%}
+        {%- set role = message["role"] -%}
+    {%- endif -%}
+
+    {%- set message_content = message['content'] or '' -%}
+    {%- if message_content is not string -%}
+        {%- set message_content = message_content | tojson(ensure_ascii=False) -%}
+    {%- endif -%}
+
+    {{- role + add_tokens.role_sep + message_content -}}
+
+    {%- if message.tool_calls is defined and message.tool_calls -%}
+        {{- add_tokens.function_call + render_function_call(message.tool_calls[0]) -}}
+    {%- endif -%}
+
+    {{- add_tokens.message_sep -}}
+
+{%- endmacro -%}
+
+
+
+{# ----- SPECIAL TOKENS ----- #}
+
+{%- set add_tokens = namespace(
+    role_sep="<|role_sep|>\n", 
+    message_sep="<|message_sep|>\n\n",
+    function_call="<|function_call|>"
+) -%}
+
+{# ----- DEFAULT DEVSYSTEM ----- #}
+
+{%- set DEVSYSTEM -%}
+<role_description>
+Description of the roles available in the dialog.
+
+`developer system`
+A message added by Sber before the main dialog. It has the highest priority and sets global, non-overridable conditions (for example, conversation rules, the safety policy, the assistant's overall response style, etc.).
+
+`system`
+A system instruction added by developers or by the user, but with a lower priority than `developer system`. It usually describes the assistant's instructions, a specific response style, and other conditions for this particular dialog.
+
+`user`
+A message or request from the user. The assistant follows it if it does not conflict with higher-priority instructions (see <instruction_priority>).
+
+`user memory`
+A sequence of the most up-to-date long-term facts about the user at the time of their request, presented as a JSON list of strings. Facts are listed in chronological order, meaning newer facts are appended to the end of the sequence. When facts are changed or deleted, records of previous facts remain in the sequence. The assistant saves facts using a function and uses them in accordance with the <memory_guidelines> block below.
+
+`added files`
+Metadata about files available for use in the dialog, presented in JSON format. It contains the following keys: id (a unique file identifier), name (file name), type (file type).
+
+`assistant`
+The assistant's reply to the user's request. If the system instruction or the user does not set additional rules for `assistant`, this reply must comply with the instructions in the <assistant_guidelines> block below. The list of functions available to call is contained in `function descriptions`. The name of the required function and its arguments will be generated next by the `function call` role. In its replies, the assistant follows the instructions in accordance with <instruction_priority>.
+
+`function descriptions`
+Function descriptions in TypeScript format. A function is a special tool (or a set of instructions) that the assistant can call to perform specific actions, computations, or obtain data needed to solve the user's task. Each function description contains blocks with the name, description, and arguments. Sometimes the description contains separate blocks with return parameters and usage examples that illustrate the correct call and arguments.
+
+`function call`
+The function that `assistant` calls based on the dialog context, and its arguments. The function is invoked in strict accordance with the instructions in the <function_usage> block.
+
+`function result`
+The result of the last function call.
+</role_description>
+
+<available_modalities>
+The assistant can work with the following modalities: text, available functions.
+</available_modalities>
+
+<instruction_priority>
+If instructions from different roles conflict within the dialog context, observe the following priorities:  
+`developer system` > `system` > `user` > `function descriptions` > `function result` > `user memory`
+</instruction_priority>
+
+<function_usage>
+Basic instructions for working with functions.
+
+Only call those functions that are described in `function descriptions`.
+
+Call available functions when, according to their description, such a call will help provide a more complete and/or accurate answer to the user's request. Fill in function arguments using information from the dialog context. If a function could help answer the request but a required argument is missing from the context, ask the user for the missing data before calling the function. If a necessary function is unavailable or an error occurs, briefly inform the user and, if possible, suggest an alternative.
+</function_usage>
+
+<memory_guidelines>
+Rules for using facts in long-term memory:
+
+If there is no message under the `user memory` role in the dialog, this is equivalent to the absence of long-term facts about the user in memory. In that case, information about the user is limited to the current dialog, and no new facts should be saved.
+</memory_guidelines>
+
+<assistant_guidelines>
+You are a helpful assistant.
+
+# Instructions
+- Strictly follow the instruction priority.
+- Maintain a logical chain of reasoning when answering the user's question.
+- For complex questions (for example, STEM), try to answer in detail unless the system message or dialog context limits the response length.
+- Be helpful, truthful, and avoid unsafe or prohibited content in your responses.
+- Try to reply in the language in which the user asked their question.
+</assistant_guidelines>
+
+A dialog will follow below.
+The dialog may include various roles described in the <role_description> block.
+Each turn begins with the role name and a special token that marks the end of the role's full name, and ends with a special end-of-turn token.
+Your task is to continue the dialog from the last specified role in accordance with the dialog context.
+{%- endset -%}
+
+
+{#- ---------------------- RENDERING STARTS HERE ---------------------- -#}
+
+
+{# ----- RENDER BOS TOKEN ----- #}
+{{- bos_token -}}
+
+
+{# ----- RENDER DEVSYSTEM ----- #}
+{{- render_role_message({"role": "developer system", "content": DEVSYSTEM}) -}}
+
+{# ----- RENDER SYSTEM IF PRESENT ----- #}
+{%- if messages and messages[0]['role'] == 'system' -%}
+    {{- render_role_message(messages[0]) -}}
+    {%- set messages = messages[1:] -%}
+{%- endif -%}
+
+{# ----- RENDER TOOLS ----- #}
+{%- if tools -%}
+    {%- set tools_content = (
+        render_tool_namespace('functions', tools) 
+        + "\n\n"
+    ) -%}
+    {{- render_role_message({'role': 'function descriptions', 'content': tools_content}) -}}
+{%- endif -%}
+
+{# ----- MAIN MESSAGE LOOP ----- #}
+{%- for message in messages -%}
+
+    {# ----- TOOL MESSAGE -------#}
+    {%- if message['role'] == 'tool' -%}
+        {{- render_role_message(message, 'function result') -}}
+        
+    {# ----- OTHER MESSAGES ----- #}
+    {%- else -%}
+        {{- render_role_message(message) -}}
+    {%- endif -%}
+
+    {# ----- ADDING GENERATION PROMPT ----- #}
+
+    {%- if loop.last and add_generation_prompt and message['role'] != 'assistant' -%}
+        {{- 'assistant' + add_tokens.role_sep -}}
+    {%- endif -%}
+
+{%- endfor -%}

src/llama-graph.cpp

@@ -900,7 +900,8 @@ ggml_tensor * llm_graph_context::build_cvec(
 
 ggml_tensor * llm_graph_context::build_lora_mm(
           ggml_tensor * w,
-          ggml_tensor * cur) const {
+          ggml_tensor * cur,
+          ggml_tensor * w_s) const {
     ggml_tensor * res = ggml_mul_mat(ctx0, w, cur);
 
     for (const auto & lora : *loras) {
@@ -921,6 +922,10 @@ ggml_tensor * llm_graph_context::build_lora_mm(
         res = ggml_add(ctx0, res, ab_cur);
     }
 
+    if (w_s) {
+        res = ggml_mul(ctx0, res, w_s);
+    }
+
     return res;
 }
 

src/llama-graph.h

@@ -764,10 +764,11 @@ struct llm_graph_context {
              ggml_tensor * cur,
                      int   il) const;
 
-    // do mat_mul, while optionally apply lora
+    // do mat_mul, while optionally apply lora and per-tensor scale
     ggml_tensor * build_lora_mm(
               ggml_tensor * w,
-              ggml_tensor * cur) const;
+              ggml_tensor * cur,
+              ggml_tensor * w_s = nullptr) const;
 
     // do mat_mul_id, while optionally apply lora
     ggml_tensor * build_lora_mm_id(

src/models/bitnet.cpp

@@ -29,10 +29,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
         // self-attention
         {
             // compute Q and K and RoPE them
-            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-            if (model.layers[il].wq_s) {
-                Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_s);
-            }
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
             cb(Qcur, "Qcur", il);
             if (model.layers[il].bq) {
                 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
@@ -40,10 +37,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
             }
 
             // B1.K
-            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-            if (model.layers[il].wk_s) {
-                Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_s);
-            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
             cb(Kcur, "Kcur", il);
             if (model.layers[il].bk) {
                 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
@@ -51,10 +45,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
             }
 
             // B1.V
-            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-            if (model.layers[il].wv_s) {
-                Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_s);
-            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
             cb(Vcur, "Vcur", il);
             if (model.layers[il].bv) {
                 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
@@ -90,10 +81,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
                     LLM_NORM_RMS, il);
             cb(cur, "attn_sub_norm", il);
 
-            cur = build_lora_mm(model.layers[il].wo, cur);
-            if (model.layers[il].wo_s) {
-                cur = ggml_mul(ctx0, cur, model.layers[il].wo_s);
-            }
+            cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
             if (model.layers[il].bo) {
                 cur = ggml_add(ctx0, cur, model.layers[il].bo);
             }
@@ -127,10 +115,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
                 LLM_NORM_RMS, il);
         cb(cur, "ffn_sub_norm", il);
 
-        cur = build_lora_mm(model.layers[il].ffn_down, cur);
-        if (model.layers[il].ffn_down_s) {
-            cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_s);
-        }
+        cur = build_lora_mm(model.layers[il].ffn_down, cur, model.layers[il].ffn_down_s);
         cb(cur, "ffn_down", il);
 
         cur = ggml_add(ctx0, cur, ffn_inp);

src/models/llama.cpp

@@ -43,28 +43,19 @@ llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_gra
             ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
 
             // compute Q and K and RoPE them
-            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-            if (model.layers[il].wq_s) {
-                Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_s);
-            }
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
             cb(Qcur, "Qcur", il);
             if (model.layers[il].bq) {
                 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
                 cb(Qcur, "Qcur", il);
             }
-            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-            if (model.layers[il].wk_s) {
-                Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_s);
-            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
             cb(Kcur, "Kcur", il);
             if (model.layers[il].bk) {
                 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
                 cb(Kcur, "Kcur", il);
             }
-            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-            if (model.layers[il].wv_s) {
-                Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_s);
-            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
             cb(Vcur, "Vcur", il);
             if (model.layers[il].bv) {
                 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);

src/models/qwen3.cpp

@@ -30,22 +30,13 @@ llm_build_qwen3::llm_build_qwen3(const llama_model & model, const llm_graph_para
         // self-attention
         {
             // compute Q and K and RoPE them
-            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-            if (model.layers[il].wq_s) {
-                Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_s);
-            }
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
             cb(Qcur, "Qcur", il);
 
-            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-            if (model.layers[il].wk_s) {
-                Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_s);
-            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
             cb(Kcur, "Kcur", il);
 
-            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-            if (model.layers[il].wv_s) {
-                Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_s);
-            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
             cb(Vcur, "Vcur", il);
 
             Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);

src/models/qwen3moe.cpp

@@ -30,22 +30,13 @@ llm_build_qwen3moe::llm_build_qwen3moe(const llama_model & model, const llm_grap
         // self_attention
         {
             // compute Q and K and RoPE them
-            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-            if (model.layers[il].wq_s) {
-                Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_s);
-            }
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
             cb(Qcur, "Qcur", il);
 
-            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-            if (model.layers[il].wk_s) {
-                Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_s);
-            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
             cb(Kcur, "Kcur", il);
 
-            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-            if (model.layers[il].wv_s) {
-                Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_s);
-            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
             cb(Vcur, "Vcur", il);
 
             Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);

tests/test-chat.cpp

@@ -2765,6 +2765,42 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .run();
 
     }
+
+    // GigaChat V3
+    {
+        auto tst = peg_tester("models/templates/GigaChat3-10B-A1.8B.jinja", detailed_debug);
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("<|message_sep|>\n\nfunction call<|role_sep|>\n{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}")
+            .tools({ special_function_tool })
+            .expect(message_assist_call)
+            .run();
+
+        tst.test(
+            "Hello, world!\nWhat's up?"
+            "<|message_sep|>\n\nfunction call<|role_sep|>\n{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}"
+        )
+            .tools({ special_function_tool })
+            .expect(message_assist_call_content)
+            .run();
+    }
+
+    // GigaChat V3.1
+    {
+        auto tst = peg_tester("models/templates/GigaChat3.1-10B-A1.8B.jinja", detailed_debug);
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("<|function_call|>{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}")
+            .tools({ special_function_tool })
+            .expect(message_assist_call)
+            .run();
+
+        tst.test(
+            "Hello, world!\nWhat's up?"
+            "<|function_call|>{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}"
+        )
+            .tools({ special_function_tool })
+            .expect(message_assist_call_content)
+            .run();
+    }
 }
 
 // Test the developer role to system workaround with a simple mock template

tools/mtmd/clip-impl.h

@@ -216,6 +216,7 @@ enum projector_type {
     PROJECTOR_TYPE_GEMMA3,
     PROJECTOR_TYPE_GEMMA3NV,
     PROJECTOR_TYPE_GEMMA3NA,
+    PROJECTOR_TYPE_PHI4,
     PROJECTOR_TYPE_IDEFICS3,
     PROJECTOR_TYPE_PIXTRAL,
     PROJECTOR_TYPE_QWEN25VL,
@@ -253,6 +254,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
     { PROJECTOR_TYPE_GEMMA3NV,  "gemma3nv"},
     { PROJECTOR_TYPE_GEMMA3NA,  "gemma3na"},
+    { PROJECTOR_TYPE_PHI4,      "phi4"},
     { PROJECTOR_TYPE_IDEFICS3,  "idefics3"},
     { PROJECTOR_TYPE_PIXTRAL,   "pixtral"},
     { PROJECTOR_TYPE_ULTRAVOX,  "ultravox"},

tools/mtmd/clip.cpp

@@ -792,6 +792,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         case PROJECTOR_TYPE_IDEFICS3:
         case PROJECTOR_TYPE_LFM2:
         case PROJECTOR_TYPE_JANUS_PRO:
+        case PROJECTOR_TYPE_PHI4:
             {
                 builder = std::make_unique<clip_graph_siglip>(ctx, img);
             } break;
@@ -1144,6 +1145,13 @@ struct clip_model_loader {
                         // ref: https://huggingface.co/LiquidAI/LFM2.5-VL-1.6B/blob/main/processor_config.json
                         hparams.set_limit_image_tokens(64, 256);
                     } break;
+                case PROJECTOR_TYPE_PHI4:
+                    {
+                        hparams.n_merge = 1;
+                        get_u32(KEY_IMAGE_MIN_PIXELS, hparams.image_min_pixels);
+                        get_u32(KEY_IMAGE_MAX_PIXELS, hparams.image_max_pixels);
+                        hparams.set_warmup_n_tokens(16*16);
+                    } break;
                 case PROJECTOR_TYPE_PIXTRAL:
                 case PROJECTOR_TYPE_LIGHTONOCR:
                     {
@@ -1841,6 +1849,13 @@ struct clip_model_loader {
                     model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
                     model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"));
                 } break;
+            case PROJECTOR_TYPE_PHI4:
+                {
+                    model.mm_0_w = get_tensor(string_format(TN_LLAVA_PROJ, 0, "weight"));
+                    model.mm_0_b = get_tensor(string_format(TN_LLAVA_PROJ, 0, "bias"));
+                    model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
+                    model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
+                } break;
             case PROJECTOR_TYPE_LFM2A:
                 {
                     for (int i : {0, 2, 3, 5, 6}) {
@@ -3157,6 +3172,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
                 res_imgs->entries.push_back(std::move(img_f32));
             } break;
 
+        case PROJECTOR_TYPE_PHI4:
         case PROJECTOR_TYPE_PIXTRAL:
         case PROJECTOR_TYPE_LIGHTONOCR:
             {
@@ -3383,6 +3399,7 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
         case PROJECTOR_TYPE_MLP:
         case PROJECTOR_TYPE_MLP_NORM:
         case PROJECTOR_TYPE_JANUS_PRO:
+        case PROJECTOR_TYPE_PHI4:
             {
                 // do nothing
             } break;
@@ -3884,6 +3901,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         case PROJECTOR_TYPE_VOXTRAL:
         case PROJECTOR_TYPE_MUSIC_FLAMINGO:
         case PROJECTOR_TYPE_JANUS_PRO:
+        case PROJECTOR_TYPE_PHI4:
         case PROJECTOR_TYPE_COGVLM:
             {
                 // do nothing
@@ -4013,6 +4031,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         case PROJECTOR_TYPE_LDPV2:
             return ctx->model.mm_model_peg_0_b->ne[0];
         case PROJECTOR_TYPE_MLP:
+        case PROJECTOR_TYPE_PHI4:
         case PROJECTOR_TYPE_PIXTRAL:
         case PROJECTOR_TYPE_LIGHTONOCR:
             return ctx->model.mm_2_w->ne[1];

tools/mtmd/models/siglip.cpp

@@ -4,7 +4,7 @@ ggml_cgraph * clip_graph_siglip::build() {
     ggml_tensor * inp = build_inp();
 
     ggml_tensor * learned_pos_embd = model.position_embeddings;
-    if (proj_type == PROJECTOR_TYPE_LFM2) {
+    if (proj_type == PROJECTOR_TYPE_LFM2 || proj_type == PROJECTOR_TYPE_PHI4) {
         learned_pos_embd = resize_position_embeddings();
     }
 
@@ -75,6 +75,14 @@ ggml_cgraph * clip_graph_siglip::build() {
             hparams.ffn_op,
             -1);
 
+    } else if (proj_type == PROJECTOR_TYPE_PHI4) {
+        cur = build_ffn(cur,
+            model.mm_0_w, model.mm_0_b,
+            nullptr, nullptr,
+            model.mm_2_w, model.mm_2_b,
+            FFN_GELU,
+            -1);
+
     } else {
         GGML_ABORT("SigLIP: Unsupported projector type");
     }

tools/mtmd/mtmd.cpp

@@ -290,6 +290,9 @@ struct mtmd_context {
             img_beg = "<|vision_start|>";
             img_end = "<|vision_end|>";
 
+        } else if (proj == PROJECTOR_TYPE_PHI4) {
+            // Phi-4 uses media marker insertion only. Keep image boundary text empty.
+
         } else if (proj == PROJECTOR_TYPE_LLAMA4) {
             // (more details in mtmd_context constructor)
             img_beg = "<|image_start|>";