server: skip unused log lines on router mode (#24463)

* vocab : refactor normalizer flags into options struct, add strip_accents

* Update src/llama-vocab.h

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update src/llama-vocab.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

.devops/cpu.Dockerfile

@@ -53,7 +53,7 @@ LABEL org.opencontainers.image.created=$BUILD_DATE \
       org.opencontainers.image.source=$IMAGE_SOURCE
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl \
+    && apt-get install -y libgomp1 curl ffmpeg \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/cuda.Dockerfile

@@ -59,7 +59,7 @@ LABEL org.opencontainers.image.created=$BUILD_DATE \
       org.opencontainers.image.source=$IMAGE_SOURCE
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl \
+    && apt-get install -y libgomp1 curl ffmpeg \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/intel.Dockerfile

@@ -57,11 +57,21 @@ LABEL org.opencontainers.image.created=$BUILD_DATE \
       org.opencontainers.image.url=$IMAGE_URL \
       org.opencontainers.image.source=$IMAGE_SOURCE
 
-ARG IGC_VERSION=v2.20.5
-ARG IGC_VERSION_FULL=2_2.20.5+19972
-ARG COMPUTE_RUNTIME_VERSION=25.40.35563.10
-ARG COMPUTE_RUNTIME_VERSION_FULL=25.40.35563.10-0
-ARG IGDGMM_VERSION=22.8.2
+#Following versions are for multiple GPUs, since 26.x has known issue:
+#   https://github.com/ggml-org/llama.cpp/issues/21747,
+#   https://github.com/intel/compute-runtime/issues/921.
+#ARG IGC_VERSION=v2.20.5
+#ARG IGC_VERSION_FULL=2_2.20.5+19972
+#ARG COMPUTE_RUNTIME_VERSION=25.40.35563.10
+#ARG COMPUTE_RUNTIME_VERSION_FULL=25.40.35563.10-0
+#ARG IGDGMM_VERSION=22.8.2
+
+
+ARG IGC_VERSION=v2.34.4
+ARG IGC_VERSION_FULL=2_2.34.4+21428
+ARG COMPUTE_RUNTIME_VERSION=26.18.38308.1
+ARG COMPUTE_RUNTIME_VERSION_FULL=26.18.38308.1-0
+ARG IGDGMM_VERSION=22.10.0
 RUN mkdir /tmp/neo/ && cd /tmp/neo/ \
   && wget https://github.com/intel/intel-graphics-compiler/releases/download/$IGC_VERSION/intel-igc-core-${IGC_VERSION_FULL}_amd64.deb \
   && wget https://github.com/intel/intel-graphics-compiler/releases/download/$IGC_VERSION/intel-igc-opencl-${IGC_VERSION_FULL}_amd64.deb \
@@ -75,7 +85,7 @@ RUN mkdir /tmp/neo/ && cd /tmp/neo/ \
   && dpkg --install *.deb
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl \
+    && apt-get install -y libgomp1 curl ffmpeg \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/musa.Dockerfile

@@ -64,7 +64,7 @@ LABEL org.opencontainers.image.created=$BUILD_DATE \
       org.opencontainers.image.source=$IMAGE_SOURCE
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl \
+    && apt-get install -y libgomp1 curl ffmpeg \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/openvino.Dockerfile

@@ -107,7 +107,7 @@ LABEL org.opencontainers.image.created=$BUILD_DATE \
       org.opencontainers.image.source=$IMAGE_SOURCE
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 libtbb12 curl wget ocl-icd-libopencl1 \
+    && apt-get install -y libgomp1 libtbb12 curl wget ffmpeg ocl-icd-libopencl1 \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/rocm.Dockerfile

@@ -76,7 +76,7 @@ LABEL org.opencontainers.image.created=$BUILD_DATE \
       org.opencontainers.image.source=$IMAGE_SOURCE
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl \
+    && apt-get install -y libgomp1 curl ffmpeg \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/vulkan.Dockerfile

@@ -49,7 +49,7 @@ LABEL org.opencontainers.image.created=$BUILD_DATE \
       org.opencontainers.image.source=$IMAGE_SOURCE
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl libvulkan1 mesa-vulkan-drivers \
+    && apt-get install -y libgomp1 curl ffmpeg libvulkan1 mesa-vulkan-drivers \
     libglvnd0 libgl1 libglx0 libegl1 libgles2 \
     && apt autoremove -y \
     && apt clean -y \

.devops/zendnn.Dockerfile

@@ -46,7 +46,7 @@ LABEL org.opencontainers.image.created=$BUILD_DATE \
       org.opencontainers.image.source=$IMAGE_SOURCE
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 libnuma1 curl \
+    && apt-get install -y libgomp1 libnuma1 curl ffmpeg \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.github/workflows/build-webgpu.yml

@@ -35,6 +35,29 @@ env:
   LLAMA_ARG_LOG_TIMESTAMPS: 1
 
 jobs:
+  format:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v6
+
+      - name: Install clang-format 22
+        run: |
+          wget -qO- https://apt.llvm.org/llvm-snapshot.gpg.key |
+            sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc > /dev/null
+          sudo add-apt-repository -y \
+            "deb http://apt.llvm.org/noble/ llvm-toolchain-noble-22 main"
+          sudo apt-get update
+          sudo apt-get install -y clang-format-22
+
+      - name: Check formatting
+        run: |
+          find ggml/src/ggml-webgpu \
+            -type f \( -name '*.cpp' -o -name '*.hpp' -o -name '*.h' \) \
+            -print0 |
+            xargs -0 clang-format-22 --dry-run --Werror
+
   macos:
     runs-on: macos-latest
 

.github/workflows/docker.yml

@@ -82,8 +82,8 @@ jobs:
             { "tag": "cpu", "dockerfile": ".devops/s390x.Dockerfile", "platforms": "linux/s390x", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04-s390x" },
             { "tag": "cuda cuda12", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "12.8.1", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
             { "tag": "cuda cuda12", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "12.8.1", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04-arm" },
-            { "tag": "cuda13", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "13.1.1", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
-            { "tag": "cuda13", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "13.1.1", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04-arm" },
+            { "tag": "cuda13", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "13.3.0", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
+            { "tag": "cuda13", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "13.3.0", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04-arm" },
             { "tag": "musa", "dockerfile": ".devops/musa.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
             { "tag": "intel", "dockerfile": ".devops/intel.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
             { "tag": "vulkan", "dockerfile": ".devops/vulkan.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04" },

.github/workflows/release.yml

@@ -504,7 +504,7 @@ jobs:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
-    runs-on: windows-2025
+    runs-on: windows-2025-vs2026
 
     permissions:
       actions: write
@@ -535,12 +535,12 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: release-windows-2025-${{ matrix.arch }}-cpu
+          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
 
       - name: Build
         shell: cmd
         run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
+          call "C:\Program Files\Microsoft Visual Studio\18\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -D CMAKE_TOOLCHAIN_FILE=cmake/${{ matrix.arch }}-windows-llvm.cmake ^
             -DLLAMA_BUILD_BORINGSSL=ON ^
@@ -554,12 +554,12 @@ jobs:
       - name: ccache-clear
         uses: ./.github/actions/ccache-clear
         with:
-          key: release-windows-2025-${{ matrix.arch }}-cpu
+          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
 
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.44.35112\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
+          Copy-Item "C:\Program Files\Microsoft Visual Studio\18\Enterprise\VC\Redist\MSVC\14.51.36231\debug_nonredist\${{ matrix.arch }}\Microsoft.VC145.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
           7z a -snl llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts

.github/workflows/winget.yml

@@ -17,7 +17,7 @@ jobs:
 
       - name: Install komac
         run: |
-          cargo binstall komac@2.15.0 -y
+          cargo binstall komac@2.16.0 -y
 
       - name: Find latest release
         id: find_latest_release

common/arg.cpp

@@ -444,7 +444,7 @@ bool common_params_handle_models(common_params & params, llama_example curr_ex)
     opts.offline         = params.offline;
     opts.skip_download   = params.skip_download;
     opts.download_mtp    = spec_type_draft_mtp;
-    opts.download_mmproj = !params.no_mmproj;
+    opts.download_mmproj = !params.no_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty();
 
     // sub-models (draft, mmproj, vocoder) are explicitly specified by the user,
     // so we should not auto-discover mtp/mmproj siblings for them
@@ -1360,7 +1360,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         {"--cache-idle-slots"},
         {"--no-cache-idle-slots"},
-        "save and clear idle slots on new task (default: enabled, requires unified KV and cache-ram)",
+        "save idle slots to the prompt cache on new task, and clear them when using unified KV (default: enabled, requires cache-ram)",
         [](common_params & params, bool value) {
             params.cache_idle_slots = value;
         }
@@ -1615,7 +1615,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("samplers that will be used for generation in the order, separated by \';\'\n(default: %s)", sampler_type_names.c_str()),
         [](common_params & params, const std::string & value) {
             const auto sampler_names = string_split<std::string>(value, ';');
-            params.sampling.samplers = common_sampler_types_from_names(sampler_names, true);
+            params.sampling.samplers = common_sampler_types_from_names(sampler_names);
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS;
         }
     ).set_sampling());
@@ -2221,8 +2221,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ_OFFLOAD"));
     add_opt(common_arg(
-        {"--image", "--audio"}, "FILE",
-        "path to an image or audio file. use with multimodal models, use comma-separated values for multiple files\n",
+        {"--image", "--audio", "--video"}, "FILE",
+        "path to an image, audio, or video file. use with multimodal models, use comma-separated values for multiple files\n",
         [](common_params & params, const std::string & value) {
             for (const auto & item : parse_csv_row(value)) {
                 params.image.emplace_back(item);
@@ -3333,6 +3333,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             common_log_set_file(common_log_main(), value.c_str());
         }
     ).set_env("LLAMA_ARG_LOG_FILE"));
+    add_opt(common_arg(
+        {"--log-prompts-dir"}, "PATH",
+        "Log prompts to directory (only used for debugging, default: disabled)",
+        [](common_params & params, const std::string & value) {
+            params.path_prompts_log_dir = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
     add_opt(common_arg(
         {"--log-colors"}, "[on|off|auto]",
         "Set colored logging ('on', 'off', or 'auto', default: 'auto')\n"

common/chat-peg-parser.cpp

@@ -87,6 +87,8 @@ static std::string normalize_quotes_to_json(const std::string & input) {
     bool in_single_quoted = false;
     bool in_double_quoted = false;
 
+    auto is_word_char = [](char ch) { return std::isalnum(static_cast<unsigned char>(ch)) || ch == '_'; };
+
     for (size_t i = 0; i < input.size(); ++i) {
         char c = input[i];
 
@@ -151,6 +153,29 @@ static std::string normalize_quotes_to_json(const std::string & input) {
                 in_single_quoted = true;
                 result += '"';
             }
+        } else if (!in_single_quoted && !in_double_quoted && (c == 'T' || c == 'F' || c == 'N') &&
+                   (i == 0 || !is_word_char(input[i - 1]))) {
+            // Python literals -> JSON; prefix match keeps streamed partials monotonic.
+            static constexpr std::pair<std::string_view, std::string_view> literals[] = {
+                { "True", "true" }, { "False", "false" }, { "None", "null" },
+            };
+            size_t n = 0;
+            while (i + n < input.size() && is_word_char(input[i + n])) {
+                ++n;
+            }
+            std::string_view token(input.data() + i, n);
+            bool matched = false;
+            for (const auto & [py, js] : literals) {
+                if (py.substr(0, n) == token) {
+                    result += js.substr(0, n);
+                    i += n - 1;
+                    matched = true;
+                    break;
+                }
+            }
+            if (!matched) {
+                result += c;
+            }
         } else {
             result += c;
         }
@@ -353,12 +378,8 @@ void common_chat_peg_mapper::map(const common_peg_ast_node & node) {
             }
             value_to_add += escape_json_string_inner(value_content);
         } else if (!value_content.empty()) {
-            // For potential containers, normalize Python-style single quotes to JSON double quotes
-            bool is_potential_container = value_content[0] == '[' || value_content[0] == '{';
-            if (is_potential_container) {
-                value_content = normalize_container_value(value_content);
-            }
-            value_to_add += value_content;
+            // Pythonic scalars/containers -> JSON.
+            value_to_add += normalize_container_value(value_content);
         }
 
         args_target() += value_to_add;
@@ -466,11 +487,34 @@ common_peg_parser common_chat_peg_builder::standard_constructed_tools(
     return force_tool_calls ? section : optional(section);
 }
 
+// Like python_value(), but the leaf also accepts JSON-cased true/false/null, used by LFM2/LFM2.5
+common_peg_parser common_chat_peg_builder::python_or_json_value() {
+    return rule("python-or-json-value", [this]() {
+        auto ws    = space();
+        auto value = python_or_json_value();
+
+        auto member  = sequence({ python_string(), ws, literal(":"), ws, value });
+        auto members = sequence({ member, zero_or_more(sequence({ ws, literal(","), ws, member })) });
+        auto dict    = rule("python-or-json-dict", [&]() {
+            return sequence({ literal("{"), ws, choice({ literal("}"), sequence({ members, ws, literal("}") }) }), ws });
+        });
+
+        auto elements = sequence({ value, zero_or_more(sequence({ literal(","), ws, value })) });
+        auto array    = rule("python-or-json-array", [&]() {
+            return sequence({ literal("["), ws, choice({ literal("]"), sequence({ elements, ws, literal("]") }) }), ws });
+        });
+
+        return choice({ dict, array, python_string(), python_number(),
+                        python_bool(), python_null(), json_bool(), json_null() });
+    });
+}
+
 // Python-style tool calls: name(arg1="value1", arg2=123)
 // Used only by LFM2 for now, so we don't merge it into autoparser
 common_peg_parser common_chat_peg_builder::python_style_tool_calls(
     const ordered_json & tools,
-    bool                 parallel_tool_calls) {
+    bool                 parallel_tool_calls,
+    bool                 allow_json_literals) {
     if (!tools.is_array() || tools.empty()) {
         return eps();
     }
@@ -504,7 +548,7 @@ common_peg_parser common_chat_peg_builder::python_style_tool_calls(
                 if (is_string_type) {
                     arg_value_parser = string_value_parser;
                 } else {
-                    arg_value_parser = tool_arg_value(python_value());
+                    arg_value_parser = tool_arg_value(allow_json_literals ? python_or_json_value() : python_value());
                 }
 
                 // Full argument: name="value" or name=value

common/chat-peg-parser.h

@@ -132,9 +132,13 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     // Helper for Python-style function call format: name(arg1="value1", arg2=123)
     // Used by LFM2 and similar templates
     common_peg_parser python_style_tool_calls(const nlohmann::ordered_json & tools,
-                                              bool                           parallel_tool_calls);
+                                              bool                           parallel_tool_calls,
+                                              bool                           allow_json_literals);
 
   private:
+    // Python values plus JSON true/false/null.
+    common_peg_parser python_or_json_value();
+
     // Implementation helpers for standard_json_tools — one per JSON tool call layout mode
     common_peg_parser build_json_tools_function_is_key(const nlohmann::ordered_json & tools,
                                                        const std::string &            args_key,
@@ -195,4 +199,3 @@ struct tagged_peg_parser {
 
 tagged_peg_parser build_tagged_peg_parser(
     const std::function<common_peg_parser(common_peg_parser_builder & builder)> & fn);
-

common/chat.cpp

@@ -1608,42 +1608,52 @@ static common_chat_params common_chat_params_init_kimi_k2(const common_chat_temp
     return data;
 }
 
-// LFM2 format: uses <|tool_list_start|>[...]<|tool_list_end|> in system prompt
-// and <|tool_call_start|>[name(arg="val")]<|tool_call_end|> for tool calls.
-// - Reasoning: <think>{reasoning}</think> (optional)
-// - Content: text before a tool call (optional)
-// - Tool calls: Python-style, e.g. [function_name(arg1="value1", arg2="value2")]
-//   Tool calls can appear multiple times (parallel tool calls supported)
-static common_chat_params common_chat_params_init_lfm2(const common_chat_template &    tmpl,
-                                                       const autoparser::generation_params & inputs) {
+// LFM2/LFM2.5 parser. Tool calls are almost Python-style and parallel-capable
+// (except dotted names and JSON literals true/false/null).
+// Always wrapped in <|tool_call_start|>[name(args)]<|tool_call_end|> with optional <think> reasoning.
+// tool_list_tokens preserves LFM2 system tool-list markers.
+static common_chat_params common_chat_params_init_lfm2(const common_chat_template &          tmpl,
+                                                       const autoparser::generation_params & inputs,
+                                                       bool tool_list_tokens) {
     common_chat_params data;
 
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking = true;
-    data.preserved_tokens  = {
-        "<|tool_list_start|>",
-        "<|tool_list_end|>",
-        "<|tool_call_start|>",
-        "<|tool_call_end|>",
-        "<think>",
-        "</think>",
-    };
-
-    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
-    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
-
     const std::string TOOL_CALL_START = "<|tool_call_start|>";
     const std::string TOOL_CALL_END   = "<|tool_call_end|>";
+    const std::string TOOL_LIST_START = "<|tool_list_start|>";
+    const std::string TOOL_LIST_END   = "<|tool_list_end|>";
     const std::string THINK_START     = "<think>";
     const std::string THINK_END       = "</think>";
     const std::string GEN_PROMPT      = "<|im_start|>assistant\n";
 
+    // Copy reasoning to the "thinking" field the template expects
+    auto adjusted_messages = json::array();
+    for (auto msg : inputs.messages) {
+        if (msg.contains("reasoning_content") && msg.at("reasoning_content").is_string()) {
+            msg["thinking"] = msg.at("reasoning_content");
+        }
+        adjusted_messages.push_back(msg);
+    }
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs, adjusted_messages);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, adjusted_messages);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = true;
+    data.preserved_tokens  = { TOOL_CALL_START, TOOL_CALL_END, THINK_START, THINK_END };
+    if (tool_list_tokens) {
+        data.preserved_tokens.push_back(TOOL_LIST_START);
+        data.preserved_tokens.push_back(TOOL_LIST_END);
+    }
+
     data.thinking_start_tag = THINK_START;
     data.thinking_end_tag   = THINK_END;
 
+    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
+    // Gate by reasoning format and whether the template supports <think>
+    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE &&
+                             tmpl.source().find(THINK_START) != std::string::npos;
+    auto include_grammar   = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+
     if (inputs.has_continuation()) {
         const auto & msg = inputs.continue_msg;
 
@@ -1660,17 +1670,21 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
         auto end = p.end();
 
         auto reasoning = p.eps();
-        if (extract_reasoning && inputs.enable_thinking) {
+        if (extract_reasoning) {
             reasoning = p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
         }
 
         if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+            if (has_response_format) {
+                auto response_format = p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema));
+                return generation_prompt + reasoning + response_format + end;
+            }
             return generation_prompt + reasoning + p.content(p.rest()) + end;
         }
         auto tool_calls = p.rule("tool-calls",
             p.trigger_rule("tool-call",
                 p.literal(TOOL_CALL_START) +
-                p.python_style_tool_calls(inputs.tools, inputs.parallel_tool_calls) +
+                p.python_style_tool_calls(inputs.tools, inputs.parallel_tool_calls, /* allow_json_literals = */ true) +
                 p.literal(TOOL_CALL_END)
             )
         );
@@ -1683,13 +1697,17 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
     data.parser = parser.save();
 
     if (include_grammar) {
-        data.grammar_lazy = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
         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);
             });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
             parser.build_grammar(builder, data.grammar_lazy);
         });
 
@@ -1697,93 +1715,6 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
             { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, TOOL_CALL_START }
         };
     }
-    return data;
-}
-
-// LFM2.5 format: uses plain "List of tools: [...]" in system prompt, no wrapper tokens.
-// Tool calls are bare [name(arg="val")], though model may optionally emit <|tool_call_start|>.
-// - Reasoning: <think>{reasoning}</think> (optional)
-// - Content: text before a tool call (optional)
-// - Tool calls: Python-style, e.g. [function_name(arg1="value1", arg2="value2")]
-//   Tool calls can appear multiple times (parallel tool calls supported)
-static common_chat_params common_chat_params_init_lfm2_5(const common_chat_template &    tmpl,
-                                                         const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking = true;
-    data.preserved_tokens  = {
-        "<|tool_call_start|>",
-        "<|tool_call_end|>",
-        "<think>",
-        "</think>",
-    };
-
-    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
-    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
-
-    const std::string THINK_START     = "<think>";
-    const std::string THINK_END       = "</think>";
-    const std::string GEN_PROMPT      = "<|im_start|>assistant\n";
-
-    data.thinking_start_tag = THINK_START;
-    data.thinking_end_tag   = THINK_END;
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.literal(GEN_PROMPT);
-        auto end = p.end();
-
-        auto reasoning = p.eps();
-        if (extract_reasoning && inputs.enable_thinking) {
-            reasoning = p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
-        }
-
-        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
-            return generation_prompt + reasoning + p.content(p.rest()) + end;
-        }
-
-        auto tool_calls = p.rule("tool-calls",
-            p.trigger_rule("tool-call",
-                p.python_style_tool_calls(inputs.tools, inputs.parallel_tool_calls)
-            )
-        );
-
-        auto content = p.content(p.until_one_of({"<|tool_call_start|>", "["}));
-        auto maybe_start = p.optional(p.literal("<|tool_call_start|>"));
-        return generation_prompt + reasoning + content + maybe_start + tool_calls + end;
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = 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);
-        });
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const std::string name = tool.at("function").at("name");
-            data.grammar_triggers.push_back({ COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "[" + name + "(" });
-        });
-    }
 
     return data;
 }
@@ -2298,14 +2229,14 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
 
     if (is_lfm2_template(src)) {
         LOG_DBG("Using specialized template: LFM2\n");
-        return common_chat_params_init_lfm2(tmpl, params);
+        return common_chat_params_init_lfm2(tmpl, params, /* tool_list_tokens = */ true);
     }
 
     // LFM2.5 format detection: template uses plain "List of tools: [...]" with no special tokens
     if (src.find("List of tools: [") != std::string::npos &&
         src.find("<|tool_list_start|>") == std::string::npos) {
         LOG_DBG("Using specialized template: LFM2.5\n");
-        return common_chat_params_init_lfm2_5(tmpl, params);
+        return common_chat_params_init_lfm2(tmpl, params, /* tool_list_tokens = */ false);
     }
 
     // GigaChatV3 format detection

common/common.cpp

@@ -1148,7 +1148,7 @@ static void common_init_sampler_from_model(
         if (llama_model_meta_val_str(model, llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE), buf, sizeof(buf)) > 0) {
             const std::vector<std::string> sampler_names = string_split<std::string>(std::string(buf), ';');
             if (!sampler_names.empty()) {
-                sparams.samplers = common_sampler_types_from_names(sampler_names, true);
+                sparams.samplers = common_sampler_types_from_names(sampler_names);
             }
         }
     }

common/common.h

@@ -489,6 +489,7 @@ struct common_params {
     std::string input_prefix         = ""; // string to prefix user inputs with                             // NOLINT
     std::string input_suffix         = ""; // string to suffix user inputs with                             // NOLINT
     std::string logits_file          = ""; // file for saving *all* logits                                  // NOLINT
+    std::string path_prompts_log_dir = ""; // directory with logged prompts                                 // NOLINT
 
     // llama-debug specific options
     std::string logits_output_dir = "data"; // directory for saving logits output files                     // NOLINT
@@ -571,7 +572,7 @@ struct common_params {
     struct common_params_model mmproj;
     bool mmproj_use_gpu = true;     // use GPU for multimodal model
     bool no_mmproj = false;         // explicitly disable multimodal model
-    std::vector<std::string> image; // path to image file(s)
+    std::vector<std::string> image; // path to image file(s) ; TODO: change the name to "media"
     int image_min_tokens = -1;
     int image_max_tokens = -1;
 

common/sampling.cpp

@@ -769,54 +769,63 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
     }
 }
 
-std::vector<common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
-    std::unordered_map<std::string, common_sampler_type> sampler_canonical_name_map {
-        { "dry",         COMMON_SAMPLER_TYPE_DRY },
-        { "top_k",       COMMON_SAMPLER_TYPE_TOP_K },
-        { "top_p",       COMMON_SAMPLER_TYPE_TOP_P },
-        { "top_n_sigma", COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
-        { "typ_p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
-        { "min_p",       COMMON_SAMPLER_TYPE_MIN_P },
-        { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
-        { "xtc",         COMMON_SAMPLER_TYPE_XTC },
-        { "infill",      COMMON_SAMPLER_TYPE_INFILL },
-        { "penalties",   COMMON_SAMPLER_TYPE_PENALTIES },
-        { "adaptive_p",  COMMON_SAMPLER_TYPE_ADAPTIVE_P },
-    };
-
-    // since samplers names are written multiple ways
-    // make it ready for both system names and input names
-    std::unordered_map<std::string, common_sampler_type> sampler_alt_name_map {
-        { "top-k",       COMMON_SAMPLER_TYPE_TOP_K },
-        { "top-p",       COMMON_SAMPLER_TYPE_TOP_P },
-        { "top-n-sigma", COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
-        { "nucleus",     COMMON_SAMPLER_TYPE_TOP_P },
-        { "typical-p",   COMMON_SAMPLER_TYPE_TYPICAL_P },
-        { "typical",     COMMON_SAMPLER_TYPE_TYPICAL_P },
-        { "typ-p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
-        { "typ",         COMMON_SAMPLER_TYPE_TYPICAL_P },
-        { "min-p",       COMMON_SAMPLER_TYPE_MIN_P },
-        { "temp",        COMMON_SAMPLER_TYPE_TEMPERATURE },
-        { "adaptive-p",  COMMON_SAMPLER_TYPE_ADAPTIVE_P },
-    };
+std::vector<common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names) {
+    // sampler names can be written multiple ways; generate aliases from canonical names
+    static const auto sampler_name_map = []{
+        // canonical sampler name mapping
+        std::unordered_map<std::string, common_sampler_type> canonical_name_map {
+            { "dry",         COMMON_SAMPLER_TYPE_DRY         },
+            { "top_k",       COMMON_SAMPLER_TYPE_TOP_K       },
+            { "top_p",       COMMON_SAMPLER_TYPE_TOP_P       },
+            { "top_n_sigma", COMMON_SAMPLER_TYPE_TOP_N_SIGMA },
+            { "typ_p",       COMMON_SAMPLER_TYPE_TYPICAL_P   },
+            { "min_p",       COMMON_SAMPLER_TYPE_MIN_P       },
+            { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
+            { "xtc",         COMMON_SAMPLER_TYPE_XTC         },
+            { "infill",      COMMON_SAMPLER_TYPE_INFILL      },
+            { "penalties",   COMMON_SAMPLER_TYPE_PENALTIES   },
+            { "adaptive_p",  COMMON_SAMPLER_TYPE_ADAPTIVE_P  }
+        };
+        std::unordered_map<std::string, common_sampler_type> alias_name_map;
+        for (const auto & entry : canonical_name_map) {
+            const std::string & canonical = entry.first;
+            if (canonical.find('_') == std::string::npos) {
+                continue;
+            }
+            // kebab-case: "top-k", "min-p", etc.
+            {
+                std::string kebab_case = canonical;
+                std::replace(kebab_case.begin(), kebab_case.end(), '_', '-');
+                alias_name_map.insert({kebab_case, entry.second});
+            }
+            // no dash: "topk", "minp", etc.
+            {
+                std::string no_dash = canonical;
+                no_dash.erase(std::remove(no_dash.begin(), no_dash.end(), '_'), no_dash.end());
+                alias_name_map.insert({no_dash, entry.second});
+            }
+        }
+        // misc. aliases
+        alias_name_map.insert({"nucleus", COMMON_SAMPLER_TYPE_TOP_P});
+        alias_name_map.insert({"temp",    COMMON_SAMPLER_TYPE_TEMPERATURE});
+        alias_name_map.insert({"typ",     COMMON_SAMPLER_TYPE_TYPICAL_P});
+        // include aliases + canonical names in the complete mapping
+        alias_name_map.merge(canonical_name_map);
+        return alias_name_map;
+    }();
 
     std::vector<common_sampler_type> samplers;
     samplers.reserve(names.size());
 
     for (const auto & name : names) {
-        auto sampler = sampler_canonical_name_map.find(name);
-        if (sampler != sampler_canonical_name_map.end()) {
+        std::string name_lower = name;
+        std::transform(name_lower.begin(), name_lower.end(), name_lower.begin(), ::tolower);
+        auto sampler = sampler_name_map.find(name_lower);
+        if (sampler != sampler_name_map.end()) {
             samplers.push_back(sampler->second);
             continue;
         }
-        if (allow_alt_names) {
-            sampler = sampler_alt_name_map.find(name);
-            if (sampler != sampler_alt_name_map.end()) {
-                samplers.push_back(sampler->second);
-                continue;
-            }
-        }
-        LOG_WRN("%s: unable to match sampler by name '%s'\n", __func__, name.c_str());
+        LOG_WRN("%s: unable to match sampler by name '%s'\n", __func__, name_lower.c_str());
     }
 
     return samplers;

common/sampling.h

@@ -109,7 +109,7 @@ std::string common_sampler_prev_str(common_sampler * gsmpl, llama_context * ctx,
 char        common_sampler_type_to_chr(enum common_sampler_type cnstr);
 std::string common_sampler_type_to_str(enum common_sampler_type cnstr);
 
-std::vector<enum common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
+std::vector<enum common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names);
 std::vector<enum common_sampler_type> common_sampler_types_from_chars(const std::string & chars);
 
 llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab,

common/speculative.cpp

@@ -3,13 +3,14 @@
 #include "common.h"
 #include "ggml.h"
 #include "llama.h"
-#include "../src/llama-ext.h" // staging API: llama_set_embeddings_nextn / llama_get_embeddings_nextn_ith (used by MTP)
 #include "log.h"
 #include "ngram-cache.h"
 #include "ngram-map.h"
 #include "ngram-mod.h"
 #include "sampling.h"
 
+#include "../src/llama-ext.h" // staging API: llama_set_embeddings_nextn / llama_get_embeddings_nextn_ith (used by MTP)
+
 #include <algorithm>
 #include <cassert>
 #include <cstring>
@@ -58,10 +59,10 @@ static bool common_speculative_are_compatible(
     const llama_vocab * vocab_tgt = llama_model_get_vocab(model_tgt);
     const llama_vocab * vocab_dft = llama_model_get_vocab(model_dft);
 
-    const bool vocab_type_tgt = llama_vocab_type(vocab_tgt);
+    const auto vocab_type_tgt = llama_vocab_type(vocab_tgt);
     LOG_DBG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt);
 
-    const bool vocab_type_dft = llama_vocab_type(vocab_dft);
+    const auto vocab_type_dft = llama_vocab_type(vocab_dft);
     LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
 
     if (vocab_type_tgt != vocab_type_dft) {
@@ -418,6 +419,8 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
 
     int32_t n_embd = 0;
 
+    bool is_mem_shared = false;
+
     // Per-sequence cross-batch carryover: pair (h_p, x_{p+1}) at MTP pos p+1.
     // The last h-row of one process() call needs the first token of the NEXT
     // call to pair with, so it's stashed here until that next call fires.
@@ -444,7 +447,9 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         auto * ctx_dft = this->params.ctx_dft;
         GGML_ASSERT(ctx_tgt && ctx_dft && "MTP requires ctx_tgt and ctx_dft to be set");
 
-        n_embd = llama_model_n_embd(llama_get_model(ctx_dft));
+        n_embd = llama_model_n_embd_out(llama_get_model(ctx_dft));
+        GGML_ASSERT(n_embd == llama_model_n_embd(llama_get_model(ctx_tgt)) &&
+                "MTP input row width must match the target h_nextn width");
 
         LOG_INF("%s: adding speculative implementation 'draft-mtp'\n", __func__);
         LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%.2f, n_embd=%d, backend_sampling=%d\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min, n_embd, (int) this->params.backend_sampling);
@@ -490,6 +495,8 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         llama_set_embeddings_nextn(ctx_tgt, true, /*masked*/ false);
         llama_set_embeddings_nextn(ctx_dft, true, /*masked*/ true);
 
+        is_mem_shared = llama_get_ctx_other(ctx_dft) == ctx_tgt;
+
         pending_h.assign(n_seq, std::vector<float>(n_embd, 0.0f));
 
         i_batch_beg.assign(n_seq, -1);
@@ -526,9 +533,11 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         if (N <= 0) {
             return;
         }
+
         auto * ctx_dft = this->params.ctx_dft;
         const llama_pos pos_max = llama_memory_seq_pos_max(llama_get_memory(ctx_dft), seq_id);
-        if (pos_max < N - 1) {
+
+        if (pos_max < N - 1 && !is_mem_shared) {
             LOG_WRN("%s: ctx_dft pos_max=%d < N-1=%d - "
                     "process() hook may not have run on every prefill ubatch "
                     "(need_embd / logits=1 on every prompt position?). "
@@ -571,48 +580,42 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
 
         const size_t row_bytes = (size_t) n_embd * sizeof(float);
 
-        common_batch_clear(batch);
+        // if kv is shared with target (e.g Gemma4), then we can skip this catch-up decode
+        if (!is_mem_shared) {
+            common_batch_clear(batch);
 
-        for (int k = 0; k < n_tokens; ++k) {
-            common_batch_add(batch, batch_in.token[k], batch_in.pos[k], { batch_in.seq_id[k][0] }, 0);
-        }
-
-        // shift the tgt embeddings to the right by one position
-        // assumes that the tokens in the batch are sequential for each sequence
-        // i.e. we cannot have seq_id like this: [0, 0, 0, 1, 1, 0, 1, 1]
-        //                                                       ^--- this is a problem
-        // TODO:this is generally true, but would be nice to assert it
-        {
-            const float * h_tgt = llama_get_embeddings_nextn(ctx_tgt);
-            std::memcpy(batch.embd + (size_t) 1 * n_embd, h_tgt, row_bytes * (n_tokens-1));
-
-            //{
-            //    // string with seq_ids in the batch
-            //    std::stringstream ss;
-            //    for (int i = 0; i < n_tokens; ++i) {
-            //        ss << batch_in.seq_id[i][0] << ",";
-            //    }
-            //    LOG_WRN("%s: batch_in.seq_id = %s\n", __func__, ss.str().c_str());
-            //}
-        }
-
-        // fill the pending embeddings from a previous run
-        auto set_h = [&](int idx, const float * h_row) {
-            std::memcpy(batch.embd + (size_t) idx * n_embd, h_row, row_bytes);
-        };
-
-        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
-            if (i_batch_beg[seq_id] < 0) {
-                continue;
+            for (int k = 0; k < n_tokens; ++k) {
+                common_batch_add(batch, batch_in.token[k], batch_in.pos[k], { batch_in.seq_id[k][0] }, 0);
             }
 
-            set_h(i_batch_beg[seq_id], pending_h[seq_id].data());
-        }
+            // shift the tgt embeddings to the right by one position
+            // assumes that the tokens in the batch are sequential for each sequence
+            // i.e. we cannot have seq_id like this: [0, 0, 0, 1, 1, 0, 1, 1]
+            //                                                       ^--- this is a problem
+            // TODO:this is generally true, but would be nice to assert it
+            {
+                const float * h_tgt = llama_get_embeddings_nextn(ctx_tgt);
+                std::memcpy(batch.embd + (size_t) 1 * n_embd, h_tgt, row_bytes * (n_tokens-1));
+            }
 
-        const int32_t rc = llama_decode(ctx_dft, batch);
-        if (rc != 0) {
-            LOG_ERR("%s: llama_decode(ctx_dft) failed rc=%d (pos=%d)\n", __func__, (int) rc, (int) batch_in.pos[0]);
-            return false;
+            // fill the pending embeddings from a previous run
+            auto set_h = [&](int idx, const float * h_row) {
+                std::memcpy(batch.embd + (size_t) idx * n_embd, h_row, row_bytes);
+            };
+
+            for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
+                if (i_batch_beg[seq_id] < 0) {
+                    continue;
+                }
+
+                set_h(i_batch_beg[seq_id], pending_h[seq_id].data());
+            }
+
+            const int32_t rc = llama_decode(ctx_dft, batch);
+            if (rc != 0) {
+                LOG_ERR("%s: llama_decode(ctx_dft) failed rc=%d (pos=%d)\n", __func__, (int) rc, (int) batch_in.pos[0]);
+                return false;
+            }
         }
 
         for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
@@ -721,7 +724,13 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
                     continue;
                 }
 
-                common_batch_add(batch, id, dp.n_past + i + 1, { seq_id }, true);
+                if (is_mem_shared) {
+                    // note: with shared memory (e.g. Gemma4 assistants) we use the same position for all draft tokens
+                    // ref: https://github.com/huggingface/transformers/blob/effde20942e3f82a1b97449f60b3a48c5ff96145/docs/source/en/model_doc/gemma4_assistant.md?plain=1#L36-L37
+                    common_batch_add(batch, id, dp.n_past, { seq_id }, true);
+                } else {
+                    common_batch_add(batch, id, dp.n_past + i + 1, { seq_id }, true);
+                }
                 std::memcpy(batch.embd + n_embd*(batch.n_tokens - 1), h_row, row_bytes);
             }
 
@@ -834,7 +843,8 @@ struct common_speculative_impl_ngram_map_k : public common_speculative_impl {
     common_speculative_impl_ngram_map_k(
             const common_ngram_map & config,
             uint32_t n_seq)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K, n_seq)
+        : common_speculative_impl(config.key_only ? COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K
+            : COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, n_seq)
     {
         for (uint32_t i = 0; i < n_seq; i++) {
             this->config.push_back(config);

conversion/__init__.py

@@ -75,9 +75,11 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "Gemma3TextModel": "gemma",
     "Gemma3nForCausalLM": "gemma",
     "Gemma3nForConditionalGeneration": "gemma",
+    "Gemma4AssistantForCausalLM": "gemma",
     "Gemma4ForConditionalGeneration": "gemma",
     "Gemma4ForCausalLM": "gemma",
     "Gemma4UnifiedForConditionalGeneration": "gemma",
+    "Gemma4UnifiedAssistantForCausalLM": "gemma",
     "GemmaForCausalLM": "gemma",
     "Glm4ForCausalLM": "glm",
     "Glm4MoeForCausalLM": "glm",
@@ -253,6 +255,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
     "Glm4vMoeForConditionalGeneration": "qwen3vl",
     "GlmOcrForConditionalGeneration": "qwen3vl",
     "GlmasrModel": "ultravox",
+    "Granite4VisionForConditionalGeneration": "granite",
     "GraniteSpeechForConditionalGeneration": "granite",
     "HunYuanVLForConditionalGeneration": "hunyuan",
     "Idefics3ForConditionalGeneration": "smolvlm",

conversion/gemma.py

@@ -785,6 +785,26 @@ class Gemma4UnifiedModel(Gemma4Model):
             self.gguf_writer.add_suppress_tokens(suppress_tokens)
 
 
+@ModelBase.register("Gemma4AssistantForCausalLM", "Gemma4UnifiedAssistantForCausalLM")
+class Gemma4AssistantModel(Gemma4Model):
+    model_arch = gguf.MODEL_ARCH.GEMMA4_ASSISTANT
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+
+        if "masked_embedding" in name:
+            logger.debug(f"Skipping get tensor {name!r} in safetensors so that convert can end normally.")
+            return None
+
+        return super().filter_tensors(item)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_embedding_length_out(self.hparams["backbone_hidden_size"])
+        self.gguf_writer.add_nextn_predict_layers(self.block_count)
+
+
 @ModelBase.register("Gemma4ForConditionalGeneration")
 class Gemma4VisionAudioModel(MmprojModel):
     has_audio_encoder = True
@@ -812,10 +832,11 @@ class Gemma4VisionAudioModel(MmprojModel):
         self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("layer_norm_eps", 1e-6))
 
         # audio params
-        assert self.hparams_audio is not None
-        self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.GEMMA4A)
-        self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["feat_in"])
-        self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams_audio.get("layer_norm_eps", 1e-6))
+        if self.has_audio_encoder:
+            assert self.hparams_audio is not None
+            self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.GEMMA4A)
+            self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["feat_in"])
+            self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams_audio.get("layer_norm_eps", 1e-6))
 
     def is_audio_tensor(self, name: str) -> bool:
         return "audio_tower" in name or "embed_audio" in name

conversion/granite.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import re
 from typing import Any, Callable, Iterable, TYPE_CHECKING
 
 import torch
@@ -13,7 +14,7 @@ from .llama import LlamaModel
 from .mamba import Mamba2Model
 
 
-@ModelBase.register("GraniteForCausalLM", "GraniteSpeechForConditionalGeneration")
+@ModelBase.register("GraniteForCausalLM")
 class GraniteModel(LlamaModel):
     """Conversion for IBM's GraniteForCausalLM"""
     model_arch = gguf.MODEL_ARCH.GRANITE
@@ -46,11 +47,29 @@ class GraniteModel(LlamaModel):
             self.gguf_writer.add_logit_scale(logits_scale)
             logger.info("gguf: (granite) logits_scale = %s", logits_scale)
 
+        # If being used as the base for Granite4 Vision, add deepstack_layer_arr
+        if self.hparams.get("spatial_target_layers") or self.hparams.get("deepstack_layer_map"):
+            normalized_projector_map = Granite4VisionMmprojModel.get_normalized_projector_map(self.hparams)
+            deepstack_mapping_arr = [-1 for _ in range(self.block_count)] # Populate with -1 sentinels
+            for proj_idx, (_, llm_layer, _, _) in enumerate(normalized_projector_map):
+                # Skip the first projector which is handled as the base embedding
+                # stream like normal
+                if proj_idx == 0:
+                    continue
+                deepstack_mapping_arr[llm_layer] = proj_idx
+            self.gguf_writer.add_deepstack_mapping(deepstack_mapping_arr)
+
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
         name, gen = item
-        if name.startswith("encoder."):
-            return None
+        # Skip multimodal tensors
+        if (
+            name.startswith(("encoder."))
+            or "image_" in name
+            or "layerwise_projectors" in name
+            or "spatial_projectors" in name
+        ):
+            return
         return super().filter_tensors(item)
 
 
@@ -241,7 +260,8 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
         assert self.d_inner % d_head == 0, f"SSM inner size {self.d_inner} not a multiple of head dim {d_head}"
 
     def set_vocab(self):
-        self.hparams["pad_vocab_size_multiple"] = 8
+        # For models with no ssm layers, don't pad for mamba2
+        self.hparams["pad_vocab_size_multiple"] = 8 if self._ssm_layers else 1
         Mamba2Model.set_vocab(self)
 
 
@@ -326,3 +346,133 @@ class GraniteSpeechMmprojModel(MmprojModel):
                 data_torch = data_torch.squeeze(1)
 
         yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Granite4VisionForConditionalGeneration")
+class Granite4VisionMmprojModel(MmprojModel):
+    has_vision_encoder = True
+    has_audio_encoder = False
+
+    @staticmethod
+    def get_normalized_projector_map(global_config: dict) -> list[tuple[int, int, str, int]]:
+        """Normalize both deepstack and spatial projector maps to the form:
+        (vision_layer, llm_layer, <type>, type_index)
+
+        This is then used to populate the following mappings:
+        - vision_feature_layers (mmproj hparam): ordered list of all
+          vision_layer values where order corresponds with the order of the
+          stacked projector tensors
+          NOTE: Values may appear multiple times for spatial projectors
+        - tensor_prefix_map (mmproj tensors): mapping from tensor prefixes to
+          the index of the corresponding projector in the stacked tensors
+        - deepstack_layer_arr (llm hparam): per-text-layer array indicating
+          which input vision feature should be injected at that layer
+          (-1 if none)
+
+        Output: (vision_layer, llm_layer, <type>, type_index)
+        """
+        deepstack_map = global_config.get("deepstack_layer_map", [])  # [[vis_layer, llm_layer], ...]
+        spatial_layers = global_config.get("spatial_target_layers", [])  # [llm_layer, ...]
+        n_text_layers = global_config["text_config"]["num_hidden_layers"]
+        n_vision_layers = global_config["vision_config"]["num_hidden_layers"]
+        normalized_projector_map = []
+        if deepstack_map:
+            for deepstack_idx, (vision_layer, llm_layer) in enumerate(sorted(deepstack_map)):
+                if vision_layer < 0:
+                    vision_layer = n_vision_layers + vision_layer
+                if llm_layer < 0:
+                    llm_layer = n_text_layers + llm_layer
+                normalized_projector_map.append((vision_layer, llm_layer, "layerwise", deepstack_idx))
+        if spatial_layers:
+            spatial_vision_layer = global_config.get("spatial_vision_layer", -1)
+            if spatial_vision_layer < 0:
+                spatial_vision_layer = n_vision_layers + spatial_vision_layer
+            for spatial_idx, llm_layer in enumerate(spatial_layers):
+                normalized_projector_map.append((spatial_vision_layer, llm_layer, "spatial", spatial_idx))
+        return list(sorted(normalized_projector_map, key=(lambda entry: entry[1])))
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        normalized_projector_map = self.get_normalized_projector_map(self.global_config)
+        self._n_proj = len(normalized_projector_map)
+
+        self._tensor_prefix_map = {
+            f"model.{proj_type}_projectors.{type_idx}": proj_idx
+            for proj_idx, (_, _, proj_type, type_idx) in enumerate(normalized_projector_map)
+        }
+        self._vision_feature_layers = [vision_layer for vision_layer, _, _, _ in normalized_projector_map]
+        self._spatial_offsets = [
+            type_idx if proj_type == "spatial" else -1
+            for _, _, proj_type, type_idx in normalized_projector_map
+        ]
+
+    def set_gguf_parameters(self):
+        assert self.hparams_vision is not None
+        super().set_gguf_parameters()
+
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.GRANITE4_VISION)
+
+        # SigLIP encoder hparams
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-6))
+        self.gguf_writer.add_vision_use_gelu(True)
+
+        # Preprocessor
+        self.gguf_writer.add_vision_preproc_image_size(self.hparams.get("image_size", 384))
+
+        # QFormer projector config
+        ds_rate = self.global_config["downsample_rate"]
+        ds_parts = ds_rate.split("/")
+        assert len(ds_parts) == 2, f"Invalid 'downsample_rate' value: {ds_rate}"
+        query_side, window_side = [int(p) for p in ds_parts]
+        self.gguf_writer.add_vision_projector_query_side(query_side)
+        self.gguf_writer.add_vision_projector_window_side(window_side)
+
+        # Set vision feature layers
+        self.gguf_writer.add_vision_feature_layers(self._vision_feature_layers)
+
+        # Set the spatial offests per projector
+        self.gguf_writer.add_vision_spatial_offsets(self._spatial_offsets)
+
+        # Add flattened image grind pinpoints (resolution candidates internally)
+        if pinpoints := self.global_config.get("image_grid_pinpoints"):
+            # Flatten with h, w -> w, h inversion
+            pinpoints = [val for h, w in pinpoints for val in (w, h)]
+            self.gguf_writer.add_vision_image_grid_pinpoints(pinpoints)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, _ = item
+        if ("vision_model.head" in name or name.startswith("lm_head")):
+            return None
+        return super().filter_tensors(item)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+
+        # Detect projector tensors and bin them
+        projector_idx = None
+        for prefix, proj_idx in self._tensor_prefix_map.items():
+            if name.startswith(prefix):
+                projector_idx = proj_idx
+                break
+        if projector_idx is not None:
+            # If this projector tensor has a block id within the projector,
+            # alias the bid to projector_idx
+            #
+            # TODO: currently, none of the Granite 4 Vision models have
+            # projectors with multiple QFormer layers, so the `layer.{}` index
+            # is always 0. This allows us to simply map to a single `bid` that
+            # matches the projector index. If this changes, we'll need a
+            # convention that merges the two IDs.
+            id_matches = list(re.finditer(r"\.([0-9]+)\.", name))
+            all_ids = [int(m.group(1)) for m in id_matches]
+            assert len(all_ids) >= 1 and len(all_ids) <= 2, "Must have at least 1 and at most 2 ids in tensor names"
+            # If not layer id, just use the projector index
+            new_bid = projector_idx
+            if len(all_ids) == 1:
+                new_name = name[:id_matches[0].span(1)[0]] + str(new_bid) + name[id_matches[0].span(1)[1]:]
+            else: # len(all_ids) == 2
+                new_bid = projector_idx # + all_ids[1]
+                new_name = name[:id_matches[0].span(0)[0]] + name[id_matches[0].span(1)[1]:id_matches[1].span(1)[0]] + str(new_bid) + name[id_matches[1].span(1)[1]:]
+            yield from super().modify_tensors(data_torch, new_name, new_bid)
+            return
+        yield from super().modify_tensors(data_torch, name, bid)

conversion/mistral.py

@@ -105,8 +105,9 @@ class MistralModel(LlamaModel):
             gguf_writer.add_rope_scaling_yarn_log_mul(mscale_all_dim)
             gguf_writer.add_rope_scaling_orig_ctx_len(yarn_params["original_max_position_embeddings"])
 
-        if "llama_4_scaling" in hparams:
-            gguf_writer.add_attn_temperature_scale(hparams["llama_4_scaling"]["beta"])
+        llama_4_scaling = hparams.get("llama_4_scaling")
+        if llama_4_scaling is not None:
+            gguf_writer.add_attn_temperature_scale(llama_4_scaling["beta"])
 
 
 class MistralMoeModel(DeepseekV2Model):

convert_hf_to_gguf.py

@@ -238,7 +238,7 @@ def main() -> None:
             assert hparams.get("vision_encoder") is not None, "This model does not support multimodal"
             from conversion.pixtral import PixtralModel
             model_class = PixtralModel
-        elif "moe" in hparams:
+        elif hparams.get("moe") is not None:
             from conversion.mistral import MistralMoeModel
             model_class = MistralMoeModel
         else:

convert_lora_to_gguf.py

@@ -311,6 +311,10 @@ def parse_args() -> argparse.Namespace:
         "--base-model-id", type=str,
         help="the model ID of the base model, if it is not available locally or in the adapter config. If specified, it will ignore --base and load the base model config from the Hugging Face hub (Example: 'meta-llama/Llama-3.2-1B-Instruct')",
     )
+    parser.add_argument(
+        "--trust-remote-code", default=False, action="store_true",
+        help="trust remote code in the model",
+    )
     parser.add_argument(
         "lora_path", type=Path,
         help="directory containing Hugging Face PEFT LoRA config (adapter_model.json) and weights (adapter_model.safetensors or adapter_model.bin)",
@@ -319,11 +323,11 @@ def parse_args() -> argparse.Namespace:
     return parser.parse_args()
 
 
-def load_hparams_from_hf(hf_model_id: str) -> tuple[dict[str, Any], Path | None]:
+def load_hparams_from_hf(hf_model_id: str, trust_remote_code: bool) -> tuple[dict[str, Any], Path | None]:
     from huggingface_hub import try_to_load_from_cache
 
     # normally, adapter does not come with base model config, we need to load it from AutoConfig
-    config = AutoConfig.from_pretrained(hf_model_id)
+    config = AutoConfig.from_pretrained(hf_model_id, trust_remote_code=trust_remote_code)
     cache_dir = try_to_load_from_cache(hf_model_id, "config.json")
     cache_dir = Path(cache_dir).parent if isinstance(cache_dir, str) else None
 
@@ -372,13 +376,13 @@ if __name__ == '__main__':
     # load base model
     if base_model_id is not None:
         logger.info(f"Loading base model from Hugging Face: {base_model_id}")
-        hparams, dir_base_model = load_hparams_from_hf(base_model_id)
+        hparams, dir_base_model = load_hparams_from_hf(base_model_id, args.trust_remote_code)
     elif dir_base_model is None:
         if "base_model_name_or_path" in lparams:
             model_id = lparams["base_model_name_or_path"]
             logger.info(f"Loading base model from Hugging Face: {model_id}")
             try:
-                hparams, dir_base_model = load_hparams_from_hf(model_id)
+                hparams, dir_base_model = load_hparams_from_hf(model_id, args.trust_remote_code)
             except OSError as e:
                 logger.error(f"Failed to load base model config: {e}")
                 logger.error("Please try downloading the base model and add its path to --base")
@@ -393,7 +397,9 @@ if __name__ == '__main__':
 
     with torch.inference_mode():
         try:
-            model_class = get_model_class(hparams["architectures"][0])
+            model_arch = hparams.get("text_config", {}).get("architectures", hparams["architectures"])[0]
+            logger.info("Using model architecture: %s", model_arch)
+            model_class = get_model_class(model_arch)
         except NotImplementedError:
             logger.error(f"Model {hparams['architectures'][0]} is not supported")
             sys.exit(1)

ggml/CMakeLists.txt

@@ -4,8 +4,8 @@ project("ggml" C CXX ASM)
 
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 13)
-set(GGML_VERSION_PATCH 1)
+set(GGML_VERSION_MINOR 14)
+set(GGML_VERSION_PATCH 0)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")

ggml/include/ggml-rpc.h

@@ -8,10 +8,10 @@ extern "C" {
 
 #define RPC_PROTO_MAJOR_VERSION    4
 #define RPC_PROTO_MINOR_VERSION    0
-#define RPC_PROTO_PATCH_VERSION    0
+#define RPC_PROTO_PATCH_VERSION    1
 
 #ifdef  __cplusplus
-static_assert(GGML_OP_COUNT == 96, "GGML_OP_COUNT has changed - update RPC_PROTO_PATCH_VERSION");
+static_assert(GGML_OP_COUNT == 97, "GGML_OP_COUNT has changed - update RPC_PROTO_PATCH_VERSION");
 #endif
 
 #define GGML_RPC_MAX_SERVERS       16

ggml/include/ggml.h

@@ -535,6 +535,7 @@ extern "C" {
         GGML_OP_IM2COL,
         GGML_OP_IM2COL_BACK,
         GGML_OP_IM2COL_3D,
+        GGML_OP_COL2IM_1D,
         GGML_OP_CONV_2D,
         GGML_OP_CONV_3D,
         GGML_OP_CONV_2D_DW,
@@ -2007,6 +2008,16 @@ extern "C" {
         int                   d1, // dilation dimension 1
         bool                  is_2D);
 
+    // col2im_1d: scatter-add GEMM columns back to 1D signal
+    // a: [K*OC, T_in]  (columns from matmul, K = a->ne[0]/OC)
+    // result: [T_out, OC]  where T_out = (T_in - 1)*s0 + K - 2*p0
+    GGML_API struct ggml_tensor * ggml_col2im_1d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,   // columns [K*OC, T_in]
+        int                   s0,  // stride
+        int                   oc,  // output channels
+        int                   p0); // padding to crop from both sides
+
     GGML_API struct ggml_tensor * ggml_conv_1d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,   // convolution kernel
@@ -2542,10 +2553,16 @@ extern "C" {
     // TODO: add ggml_gated_delta_net_set_bcast() to be able to configure Q, K broadcast type: tiled vs interleaved [TAG_GGML_GDN_BCAST]
     // ref: https://github.com/ggml-org/llama.cpp/pull/19468#discussion_r2786394306
     //
-    // state is a 3D tensor of shape (S_v*S_v*H, K, n_seqs):
-    //   K == 1: output carries the final state only.
-    //   K  > 1: output carries K snapshot slots; the kernel writes the last min(n_tokens, K)
-    //   per-token snapshots into the trailing slots
+    // tensor shapes (S_k == S_v, H_v % H_k == 0):
+    //   q, k  : [S_k, H_k, n_tokens, n_seqs]
+    //   v     : [S_v, H_v, n_tokens, n_seqs]
+    //   g     : [1, H_v, n_tokens, n_seqs] (scalar gate) or [S_v, H_v, n_tokens, n_seqs] (KDA)
+    //   beta  : [1, H_v, n_tokens, n_seqs]
+    //   state : [S_v, S_v, H_v, n_seqs] -- initial recurrent state s0
+    //
+    // the output packs the attention scores [S_v, H_v, n_tokens, n_seqs] followed by K state
+    // snapshots, most-recent first (slot 0 = final state, slot s = state s tokens back). K == 1
+    // keeps only the final state; when n_tokens < K only slots 0..n_tokens-1 are written.
     GGML_API struct ggml_tensor * ggml_gated_delta_net(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,
@@ -2553,7 +2570,8 @@ extern "C" {
             struct ggml_tensor  * v,
             struct ggml_tensor  * g,
             struct ggml_tensor  * beta,
-            struct ggml_tensor  * state);
+            struct ggml_tensor  * state,
+            int64_t               K);
 
     // custom operators
 

ggml/src/ggml-backend-meta.cpp

@@ -776,8 +776,8 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
         GGML_ASSERT(src_ss[2].axis == GGML_BACKEND_SPLIT_AXIS_1);
         GGML_ASSERT(src_ss[3].axis == GGML_BACKEND_SPLIT_AXIS_1);
         GGML_ASSERT(src_ss[4].axis == GGML_BACKEND_SPLIT_AXIS_1);
-        // state shape is (S_v*S_v*H, K, n_seqs); the heads dim is nested inside axis 0,
-        // so a head-aligned split on the input cache reshapes to axis 0 here (not axis 2).
+        // state shape is [S_v, S_v, H_v, n_seqs] (s0 only); the heads dim is its own axis 2,
+        // so a head-aligned split on the input cache lands on axis 2 here.
         GGML_ASSERT(src_ss[5].axis == GGML_BACKEND_SPLIT_AXIS_2 || src_ss[5].axis == GGML_BACKEND_SPLIT_AXIS_1 || src_ss[5].axis == GGML_BACKEND_SPLIT_AXIS_0);
         return {GGML_BACKEND_SPLIT_AXIS_0, {0}, {1}, 1};
     };

ggml/src/ggml-cpu/ggml-cpu.c

@@ -1912,6 +1912,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_im2col_3d(params, tensor);
             } break;
+        case GGML_OP_COL2IM_1D:
+            {
+                ggml_compute_forward_col2im_1d(params, tensor);
+            } break;
         case GGML_OP_CONV_2D:
             {
                 ggml_compute_forward_conv_2d(params, tensor);
@@ -2343,6 +2347,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_CONV_2D:
         case GGML_OP_CONV_3D:
         case GGML_OP_CONV_2D_DW:
+        case GGML_OP_COL2IM_1D:
         case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_CONV_TRANSPOSE_2D:
             {
@@ -2943,7 +2948,7 @@ struct ggml_cplan ggml_graph_plan(
                 case GGML_OP_GATED_DELTA_NET:
                     {
                         const int64_t S_v = node->src[2]->ne[0];
-                        const int64_t K   = node->src[5]->ne[1];  // state is (D, K, n_seqs)
+                        const int64_t K   = ggml_get_op_params_i32(node, 0);
                         const int64_t per_thread = S_v + (K > 1 ? S_v * S_v : 0);
                         cur = per_thread * sizeof(float) * n_tasks;
                     } break;

ggml/src/ggml-cpu/ops.cpp

@@ -4008,12 +4008,12 @@ static void ggml_compute_forward_rms_norm_back_f32(
                 // dx := scale(dx, rrms)
                 float * dx = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
 
-                // dx[i00] = (x*(-sum_xdz/sum_eps) + dz) / sqrtf(mean_eps)
-                ggml_vec_cpy_f32  (ne00, dx, x);
-                // ggml_vec_scale_f32(ne00, dx, -mean_xdz/mean_eps);
-                ggml_vec_scale_f32(ne00, dx, (float)(-sum_xdz)/sum_eps);
-                ggml_vec_acc_f32  (ne00, dx, dz);
-                ggml_vec_scale_f32(ne00, dx, rrms);
+                // dx[i00] = (dz + x*(-sum_xdz/sum_eps)) * rrms
+                // note: https://github.com/ggml-org/ggml/issues/1491
+                const float scale_x = (float) (-sum_xdz) / sum_eps;
+                for (int64_t i00 = 0; i00 < ne00; i00++) {
+                    dx[i00] = (dz[i00] + x[i00] * scale_x) * rrms;
+                }
             }
         }
     }
@@ -6730,6 +6730,78 @@ static inline int64_t ggml_wrap_around(int64_t coord, int64_t size) {
     return (coord  + size) % size; // adding size avoids negative number weirdness
 }
 
+// ggml_compute_forward_col2im_1d
+//
+// Scatter-add columns [K*OC, T_in] -> signal [T_out, OC]
+// where T_out = (T_in - 1)*s + K - 2*p.  Gather approach: each output reads ceil(K/s) inputs.
+// Parallelized over the time axis so the split stays balanced whatever OC is.
+// Supports F32, F16, BF16 input/output (same type), F32 accumulator.
+
+template <typename elem_t>
+static void ggml_compute_forward_col2im_1d_impl(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src = dst->src[0];  // [K*OC, T_in]
+
+    GGML_ASSERT(ggml_is_contiguous(src));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
+    const int32_t OC = ((const int32_t *)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t *)(dst->op_params))[2];
+
+    const int64_t K_OC = src->ne[0];
+    const int64_t T_in = src->ne[1];
+    const int64_t K    = K_OC / OC;
+    const int64_t T_out = dst->ne[0];
+
+    const elem_t * col_data = (const elem_t *) src->data;
+    elem_t       * dst_data = (elem_t *) dst->data;
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    // Parallelize over the time axis: the split stays balanced whatever OC is,
+    // down to OC = 1 for mono audio, and threads read disjoint column bands
+    const int64_t dr = (T_out + nth - 1) / nth;
+    const int64_t it0 = dr * ith;
+    const int64_t it1 = it0 + dr < T_out ? it0 + dr : T_out;
+
+    for (int64_t oc = 0; oc < OC; oc++) {
+        for (int64_t t_out = it0; t_out < it1; t_out++) {
+            const int64_t t_abs = t_out + p0;  // absolute position in uncropped signal
+            // Gather: find all (t_in, k) where t_in * s + k == t_abs, 0 <= k < K
+            int64_t t_in_min = (t_abs - K + 1 + s0 - 1) / s0;  // ceil((t_abs-K+1)/s)
+            if (t_in_min < 0) t_in_min = 0;
+            int64_t t_in_max = t_abs / s0;
+            if (t_in_max >= T_in) t_in_max = T_in - 1;
+
+            float sum = 0.0f;
+            for (int64_t t_in = t_in_min; t_in <= t_in_max; t_in++) {
+                int64_t k = t_abs - t_in * s0;
+                if (k >= 0 && k < K) {
+                    // col layout: [K*OC, T_in], element (oc*K+k, t_in)
+                    sum += type_conversion_table<elem_t>::to_f32(col_data[(oc * K + k) + t_in * K_OC]);
+                }
+            }
+            // dst layout: [T_out, OC], element (t_out, oc)
+            dst_data[t_out + oc * T_out] = type_conversion_table<elem_t>::from_f32(sum);
+        }
+    }
+}
+
+void ggml_compute_forward_col2im_1d(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+    switch (dst->src[0]->type) {
+        case GGML_TYPE_F32:  ggml_compute_forward_col2im_1d_impl<float>      (params, dst); break;
+        case GGML_TYPE_F16:  ggml_compute_forward_col2im_1d_impl<ggml_fp16_t>(params, dst); break;
+        case GGML_TYPE_BF16: ggml_compute_forward_col2im_1d_impl<ggml_bf16_t>(params, dst); break;
+        default: GGML_ABORT("col2im_1d: unsupported type %d", dst->src[0]->type);
+    }
+}
+
 // ggml_compute_forward_conv_2d
 
 
@@ -10552,11 +10624,11 @@ static void ggml_compute_forward_gated_delta_net_one_chunk(
 
     const bool kda = (neg0 == S_v);
 
-    // state is 3D (S_v*S_v*H, K, n_seqs); K is the snapshot slot count.
-    const int64_t K = src_state->ne[1];
+    // K (snapshot slot count) is an op param; state holds s0 only [S_v, S_v, H, n_seqs].
+    const int64_t K = ggml_get_op_params_i32(dst, 0);
     GGML_ASSERT(K >= 1);
-    // per-seq stride in floats (slot 0 of seq s lives at state + s * seq_stride)
-    const int64_t state_seq_stride = src_state->nb[2] / sizeof(float);
+    // per-seq stride in floats (seq s starts at state + s * seq_stride)
+    const int64_t state_seq_stride = src_state->nb[3] / sizeof(float);
 
     const int64_t per_thread = S_v + (K > 1 ? S_v * S_v : 0);
     const int ith = params->ith;
@@ -10572,9 +10644,8 @@ static void ggml_compute_forward_gated_delta_net_one_chunk(
     float * attn_out_base  = (float *)dst->data;
     float * state_out_base = (float *)dst->data + attn_score_elems;
 
-    // snapshot slot mapping: target_slot = t - shift. When n_tokens < K only the last
-    // n_tokens slots are written; earlier slots are left untouched (caller-owned).
-    const int64_t shift = n_tokens - K;
+    // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+    // When n_tokens < K only slots 0..n_tokens-1 are written; older slots are caller-owned.
 
     const float * state_in_base = (const float *)src_state->data;
 
@@ -10602,7 +10673,7 @@ static void ggml_compute_forward_gated_delta_net_one_chunk(
             : state_out_base + (iv3 * H + iv1) * S_v * S_v;
 
         // copy input state into the working buffer and operate in-place
-        // state layout (D, K, n_seqs): slot 0 of seq iv3 starts at iv3 * state_seq_stride.
+        // state layout [S_v, S_v, H, n_seqs]: seq iv3 starts at iv3 * state_seq_stride.
         const float * s_in = state_in_base + iv3 * state_seq_stride + iv1 * S_v * S_v;
         memcpy(s_out, s_in, S_v * S_v * sizeof(float));
 
@@ -10655,7 +10726,7 @@ static void ggml_compute_forward_gated_delta_net_one_chunk(
             attn_data += S_v * H; // advance to next token
 
             if (K > 1) {
-                const int64_t target_slot = t - shift;
+                const int64_t target_slot = n_tokens - 1 - t;
                 if (target_slot >= 0 && target_slot < K) {
                     float * curr_state_o = state_out_base + target_slot * state_size_per_snap +
                                      (iv3 * H + iv1) * S_v * S_v;

ggml/src/ggml-cpu/ops.h

@@ -68,6 +68,7 @@ void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * p
 void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_im2col_3d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_col2im_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_3d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);

ggml/src/ggml-cuda/gated_delta_net.cu

@@ -39,9 +39,9 @@ gated_delta_net_cuda(const float * q,
     float *       attn_data        = dst;
     float *       state            = dst + attn_score_elems;
 
-    // input state layout (D, K, n_seqs) — seq stride is K * D = K * H * S_v * S_v.
+    // input state holds s0 only: [S_v, S_v, H, n_seqs] — seq stride is D = H * S_v * S_v.
     // output state layout (per-slot D * n_seqs) — same per-(seq,head) offset as before.
-    const int64_t state_in_offset      = sequence * K * H * S_v * S_v + h_idx * S_v * S_v;
+    const int64_t state_in_offset      = sequence * H * S_v * S_v + h_idx * S_v * S_v;
     const int64_t state_out_offset     = (sequence * H + h_idx) * S_v * S_v;
     state += state_out_offset;
     curr_state += state_in_offset + col * S_v;
@@ -143,12 +143,10 @@ gated_delta_net_cuda(const float * q,
         attn_data += S_v * H;
 
         if constexpr (keep_rs_t) {
-            // slot mapping: target_slot = t - shift. When n_tokens < K only the last n_tokens slots
-            // are written; earlier slots are left untouched (caller-owned).
-            const int shift = (int) n_tokens - K;
-
+            // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+            // When n_tokens < K only slots 0..n_tokens-1 are written; older slots are caller-owned.
             const int64_t state_size_per_token = S_v * S_v * H * n_seqs; // per-slot stride in output
-            const int target_slot = t - shift;
+            const int target_slot = (int) n_tokens - 1 - t;
             if (target_slot >= 0 && target_slot < K) {
                 float * curr_state = (dst + attn_score_elems) + target_slot * state_size_per_token + state_out_offset;
 #pragma unroll
@@ -286,8 +284,8 @@ void ggml_cuda_op_gated_delta_net(ggml_backend_cuda_context & ctx, ggml_tensor *
 
     cudaStream_t stream = ctx.stream();
 
-    // state is 3D (S_v*S_v*H, K, n_seqs); K is the snapshot slot count.
-    const int K = (int) src_state->ne[1];
+    // K (snapshot slot count) is an op param; state holds s0 only [S_v, S_v, H, n_seqs].
+    const int K = ggml_get_op_params_i32(dst, 0);
     const bool keep_rs = K > 1;
 
     if (kda) {

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -622,6 +622,18 @@ ggml_backend_cuda_context::~ggml_backend_cuda_context() {
 
 // cuda buffer
 
+struct ggml_backend_cuda_device_context {
+    int device;
+    std::string name;
+    std::string description;
+    std::string pci_bus_id;
+    int op_offload_min_batch_size;
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    std::mutex device_mutex;
+    int active_count = 0;
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+};
+
 struct ggml_backend_cuda_buffer_context {
     int device;
     void * dev_ptr = nullptr;
@@ -639,6 +651,13 @@ struct ggml_backend_cuda_buffer_context {
 
 static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) buffer->buft->device->context;
+    std::lock_guard<std::mutex> lock(dev_ctx->device_mutex);
+    dev_ctx->active_count--;
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+
     delete ctx;
 }
 
@@ -791,6 +810,12 @@ static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_bac
 
     ggml_backend_cuda_buffer_context * ctx = new ggml_backend_cuda_buffer_context(buft_ctx->device, dev_ptr);
 
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) buft->device->context;
+    std::lock_guard<std::mutex> lock(dev_ctx->device_mutex);
+    dev_ctx->active_count++;
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+
     return ggml_backend_buffer_init(buft, ggml_backend_cuda_buffer_interface, ctx, size);
 }
 
@@ -1490,6 +1515,12 @@ static bool ggml_backend_buft_is_cuda_host(ggml_backend_buffer_type_t buft) {
 }
 
 static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) buffer->buft->device->context;
+    std::lock_guard<std::mutex> lock(dev_ctx->device_mutex);
+    dev_ctx->active_count--;
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+
     CUDA_CHECK(cudaFreeHost(buffer->context));
 }
 
@@ -1498,6 +1529,8 @@ static void * ggml_cuda_host_malloc(size_t size) {
         return nullptr;
     }
 
+    ggml_cuda_set_device(0); // cudaMallocHost can create the implicit CUDA device context, make sure that this is consistently done on device 0.
+
     void * ptr = nullptr;
     cudaError_t err = cudaMallocHost((void **) &ptr, size);
     if (err != cudaSuccess) {
@@ -1523,6 +1556,12 @@ static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggm
     buffer->buft = buft;
     buffer->iface.free_buffer = ggml_backend_cuda_host_buffer_free_buffer;
 
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) buft->device->context;
+    std::lock_guard<std::mutex> lock(dev_ctx->device_mutex);
+    dev_ctx->active_count++;
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+
     return buffer;
 }
 
@@ -3140,6 +3179,12 @@ static const char * ggml_backend_cuda_get_name(ggml_backend_t backend) {
 static void ggml_backend_cuda_free(ggml_backend_t backend) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
 
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) backend->device->context;
+    std::lock_guard<std::mutex> lock(dev_ctx->device_mutex);
+    dev_ctx->active_count--;
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+
     delete cuda_ctx;
     delete backend;
 }
@@ -4871,14 +4916,6 @@ void ggml_backend_cuda_unregister_host_buffer(void * buffer) {
 
 // backend device
 
-struct ggml_backend_cuda_device_context {
-    int device;
-    std::string name;
-    std::string description;
-    std::string pci_bus_id;
-    int op_offload_min_batch_size;
-};
-
 static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
     ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
     return ctx->name.c_str();
@@ -4967,6 +5004,11 @@ static bool ggml_backend_cuda_get_available_uma_memory(long * available_memory_k
 
 static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
     ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    std::lock_guard<std::mutex> lock(ctx->device_mutex);
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+
     ggml_cuda_set_device(ctx->device);
     CUDA_CHECK(cudaMemGetInfo(free, total));
 
@@ -4993,6 +5035,13 @@ static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t *
     }
 #endif // defined(__linux__)
 
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    // If no backends or buffers are active, the cudaMemGetInfo call above lazily created a CUDA
+    // context that permanently consumes VRAM. Reset the device to free it.
+    if (ctx->active_count == 0) {
+        CUDA_CHECK(cudaDeviceReset());
+    }
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 }
 
 static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend_dev_t dev) {
@@ -5687,13 +5736,21 @@ ggml_backend_t ggml_backend_cuda_init(int device) {
         return nullptr;
     }
 
+    ggml_backend_dev_t dev = ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), device);
+
     ggml_backend_t cuda_backend = new ggml_backend {
         /* .guid    = */ ggml_backend_cuda_guid(),
         /* .iface   = */ ggml_backend_cuda_interface,
-        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), device),
+        /* .device  = */ dev,
         /* .context = */ ctx,
     };
 
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) dev->context;
+    std::lock_guard<std::mutex> lock(dev_ctx->device_mutex);
+    dev_ctx->active_count++;
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+
     return cuda_backend;
 }
 

ggml/src/ggml-cuda/mmvq.cu

@@ -411,7 +411,6 @@ static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_d
                 case GGML_TYPE_Q5_0:
                 case GGML_TYPE_Q5_1:
                 case GGML_TYPE_Q8_0:
-                case GGML_TYPE_Q4_K:
                     return 8;
                 case GGML_TYPE_Q6_K:
                     return 2;

ggml/src/ggml-cuda/ssm-scan.cu

@@ -67,6 +67,7 @@ __global__ void __launch_bounds__(splitD, 1)
     __shared__ CubTempStorage cub_temp_storage;
 
     BlockLoad(cub_temp_storage.load_temp).Load(A_block, regA);
+    __syncthreads();
     BlockLoad(cub_temp_storage.load_temp).Load(s0_block, regs0);
 #else
     const int stride_s0 = src0_nb2 / sizeof(float);
@@ -105,6 +106,7 @@ __global__ void __launch_bounds__(splitD, 1)
             regs0[n] = state;
         }
         y_block[i * stride_y + threadIdx.x] = sumf;
+        __syncthreads();
     }
 
 #ifdef USE_CUB
@@ -249,9 +251,8 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
         GGML_ASSERT(head_dim == 1);
         GGML_ASSERT(n_group == 1);
         const dim3 blocks(n_seq, (n_head + threads - 1) / threads, 1);
-        const int  smem_size = (threads * (d_state + 1) * 2) * sizeof(float);
         if (d_state == 16) {
-            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks, threads, smem_size, stream);
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks, threads, 0, stream);
             switch (n_tok)
             {
             case 1:

ggml/src/ggml-cuda/vendors/hip.h

@@ -219,9 +219,9 @@
 #define RDNA3
 #endif // defined(__GFX11__)
 
-#if defined(__gfx1150__) || defined(__gfx1151__)
+#if defined(__gfx1150__) || defined(__gfx1151__) || defined(__gfx1152__) || defined(__gfx1153__)
 #define RDNA3_5
-#endif // defined(__gfx1150__) || defined(__gfx1151__)
+#endif // defined(__gfx1150__) || defined(__gfx1151__) || defined(__gfx1152__) || defined(__gfx1153__)
 
 #if defined(RDNA3) && !defined(RDNA3_5)
 #define RDNA3_0

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -2538,7 +2538,7 @@ static bool ggml_hexagon_supported_gated_delta_net(const struct ggml_hexagon_ses
     const int64_t H        = v->ne[1];
     const int64_t n_tokens = v->ne[2];
     const int64_t n_seqs   = v->ne[3];
-    const int64_t K        = state->ne[1];
+    const int64_t K        = ggml_get_op_params_i32(op, 0);
 
     if (S_v <= 0 || S_v > 128 || H <= 0 || n_tokens <= 0 || n_seqs <= 0) {
         return false;
@@ -2551,7 +2551,8 @@ static bool ggml_hexagon_supported_gated_delta_net(const struct ggml_hexagon_ses
     if ((g->ne[0] != 1 && g->ne[0] != S_v) || beta->ne[0] != 1) {
         return false;
     }
-    if (ggml_nelements(state) != S_v * S_v * H * n_seqs * K) {
+    // state holds s0 only [S_v, S_v, H, n_seqs]; K is op param 0.
+    if (ggml_nelements(state) != S_v * S_v * H * n_seqs) {
         return false;
     }
     if (dst->ne[0] != S_v * H || dst->ne[1] != n_tokens * n_seqs + S_v * n_seqs * K) {

ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c

@@ -584,7 +584,7 @@ static void gated_delta_net_f32_pp_thread(unsigned int nth, unsigned int ith, vo
     const uint32_t H        = v->ne[1];
     const uint32_t n_tokens = v->ne[2];
     const uint32_t n_seqs   = v->ne[3];
-    const uint32_t K        = state->ne[1];
+    const uint32_t K        = octx->op_params[0];
 
     const uint32_t total_rows = H * n_seqs;
     if (ith >= total_rows) {
@@ -618,9 +618,8 @@ static void gated_delta_net_f32_pp_thread(unsigned int nth, unsigned int ith, vo
     struct fastdiv_values fd_rq3 = init_fastdiv_values(rq3);
     struct fastdiv_values fd_rk3 = init_fastdiv_values(rk3);
 
-    const uint64_t state_seq_stride = state->nb[2] / sizeof(float);
+    const uint64_t state_seq_stride = state->nb[3] / sizeof(float);
     const uint64_t state_size_per_snap = (uint64_t) S_v * S_v * H * n_seqs;
-    const int64_t shift = (int64_t) n_tokens - (int64_t) K;
 
     uint32_t ir_prefetch = ith;
     int spad_idx = 0;
@@ -630,7 +629,8 @@ static void gated_delta_net_f32_pp_thread(unsigned int nth, unsigned int ith, vo
         const uint32_t piv1 = fastmodulo(ir_prefetch, H, &fd_H);
         const uint32_t piv3 = fastdiv(ir_prefetch, &fd_H);
         const float * ps_in = state_in_base + (uint64_t) piv3 * state_seq_stride + (uint64_t) piv1 * S_v * S_v;
-        float * ps_out = state_out_base + (uint64_t) (K - 1) * state_size_per_snap + ((uint64_t) piv3 * H + piv1) * S_v * S_v;
+        // final state lands in snapshot slot 0 (most-recent-first ordering)
+        float * ps_out = state_out_base + ((uint64_t) piv3 * H + piv1) * S_v * S_v;
 
         // Push dummy write-back
         dma_queue_push(dma, dma_make_ptr(ps_out, s_work[spad_idx]),
@@ -661,7 +661,8 @@ static void gated_delta_net_f32_pp_thread(unsigned int nth, unsigned int ith, vo
         const uint32_t iq3 = fastdiv(iv3, &fd_rq3);
         const uint32_t ik3 = fastdiv(iv3, &fd_rk3);
 
-        float * s_out = state_out_base + (uint64_t) (K - 1) * state_size_per_snap + ((uint64_t) iv3 * H + iv1) * S_v * S_v;
+        // final state lands in snapshot slot 0 (most-recent-first ordering)
+        float * s_out = state_out_base + ((uint64_t) iv3 * H + iv1) * S_v * S_v;
 
         float * attn_data = dst_base + ((uint64_t) iv3 * n_tokens * H + iv1) * S_v;
 
@@ -792,7 +793,8 @@ static void gated_delta_net_f32_pp_thread(unsigned int nth, unsigned int ith, vo
             }
 
             if (K > 1) {
-                const int64_t target_slot = (int64_t) t - shift;
+                // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+                const int64_t target_slot = (int64_t) n_tokens - 1 - (int64_t) t;
                 if (target_slot >= 0 && target_slot < (int64_t) K) {
                     float * curr_state_o = state_out_base + (uint64_t) target_slot * state_size_per_snap + ((uint64_t) iv3 * H + iv1) * S_v * S_v;
                     if (curr_state_o != s_out) {
@@ -844,7 +846,6 @@ static void gated_delta_net_f32_tg_thread(unsigned int nth, unsigned int ith, vo
     const uint32_t S_v      = v->ne[0];
     const uint32_t H        = v->ne[1];
     const uint32_t n_seqs   = v->ne[3];
-    const uint32_t K        = state->ne[1];
 
     const uint32_t total_rows = H * n_seqs;
     if (ith >= total_rows) {
@@ -878,8 +879,7 @@ static void gated_delta_net_f32_tg_thread(unsigned int nth, unsigned int ith, vo
     struct fastdiv_values fd_rq3 = init_fastdiv_values(rq3);
     struct fastdiv_values fd_rk3 = init_fastdiv_values(rk3);
 
-    const uint64_t state_seq_stride = state->nb[2] / sizeof(float);
-    const uint64_t state_size_per_snap = (uint64_t) S_v * S_v * H * n_seqs;
+    const uint64_t state_seq_stride = state->nb[3] / sizeof(float);
 
     uint32_t ir_prefetch = ith;
     int spad_idx = 0;
@@ -889,7 +889,8 @@ static void gated_delta_net_f32_tg_thread(unsigned int nth, unsigned int ith, vo
         const uint32_t piv1 = fastmodulo(ir_prefetch, H, &fd_H);
         const uint32_t piv3 = fastdiv(ir_prefetch, &fd_H);
         const float * ps_in = state_in_base + (uint64_t) piv3 * state_seq_stride + (uint64_t) piv1 * S_v * S_v;
-        float * ps_out = state_out_base + (uint64_t) (K - 1) * state_size_per_snap + ((uint64_t) piv3 * H + piv1) * S_v * S_v;
+        // final state lands in snapshot slot 0 (most-recent-first ordering)
+        float * ps_out = state_out_base + ((uint64_t) piv3 * H + piv1) * S_v * S_v;
 
         // Push dummy write-back
         dma_queue_push(dma, dma_make_ptr(ps_out, s_work[spad_idx]),
@@ -920,7 +921,8 @@ static void gated_delta_net_f32_tg_thread(unsigned int nth, unsigned int ith, vo
         const uint32_t iq3 = fastdiv(iv3, &fd_rq3);
         const uint32_t ik3 = fastdiv(iv3, &fd_rk3);
 
-        float * s_out = state_out_base + (uint64_t) (K - 1) * state_size_per_snap + ((uint64_t) iv3 * H + iv1) * S_v * S_v;
+        // final state lands in snapshot slot 0 (most-recent-first ordering)
+        float * s_out = state_out_base + ((uint64_t) iv3 * H + iv1) * S_v * S_v;
 
         float * attn_data = dst_base + ((uint64_t) iv3 * H + iv1) * S_v;
 
@@ -1097,7 +1099,7 @@ int op_gated_delta_net(struct htp_ops_context * octx) {
     const uint32_t H        = v->ne[1];
     const uint32_t n_tokens = v->ne[2];
     const uint32_t n_seqs   = v->ne[3];
-    const uint32_t K        = state->ne[1];
+    const uint32_t K        = octx->op_params[0];
 
     if (S_v == 0 || S_v > HTP_GDN_MAX_SV || H == 0 || n_tokens == 0 || n_seqs == 0) {
         return HTP_STATUS_NO_SUPPORT;
@@ -1110,7 +1112,8 @@ int op_gated_delta_net(struct htp_ops_context * octx) {
         (n_seqs % q->ne[3]) != 0 || (n_seqs % k->ne[3]) != 0) {
         return HTP_STATUS_NO_SUPPORT;
     }
-    if (state->ne[0] * state->ne[2] * state->ne[3] != S_v * S_v * H * n_seqs) {
+    // state holds s0 only: [S_v, S_v, H, n_seqs]
+    if (state->ne[0] != S_v || state->ne[1] != S_v || state->ne[2] != H || state->ne[3] != n_seqs) {
         return HTP_STATUS_NO_SUPPORT;
     }
     if (dst->ne[0] != S_v * H || dst->ne[1] != n_tokens * n_seqs + S_v * n_seqs * K) {

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -590,8 +590,8 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_gated_delta_net(
     const int ne20 = op->src[2]->ne[0]; // S_v
     const int ne21 = op->src[2]->ne[1]; // H
     const int ne30 = op->src[3]->ne[0]; // G
-    // state is src[5], 3D (S_v*S_v*H, K, n_seqs); K is the snapshot slot count.
-    const int K = op->src[5]->ne[1];
+    // state is src[5], 4D [S_v, S_v, H_v, n_seqs] (s0 only); K is op param 0.
+    const int K = ggml_get_op_params_i32(op, 0);
 
     const int nsg = op->src[2]->ne[0]/32;
 
@@ -1738,10 +1738,14 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_im2col(ggml_meta
     GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
     GGML_ASSERT(op->type         == GGML_TYPE_F16 || op->type == GGML_TYPE_F32);
 
+    const bool is_2D = ((const int32_t *)(op->op_params))[6] == 1;
+    const int64_t KH = is_2D ? ne01 : 1;
+    const int64_t KW = ne00;
+
     char base[256];
     char name[256];
 
-    if (ne00*ne01 <= 1024) {
+    if (KH*KW <= 1024) {
         snprintf(base, 256, "kernel_im2col_%s", ggml_type_name(op->type));
     } else {
         snprintf(base, 256, "kernel_im2col_ext_%s", ggml_type_name(op->type));

ggml/src/ggml-metal/ggml-metal.metal

@@ -2599,9 +2599,9 @@ kernel void kernel_gated_delta_net_impl(
 
     const float scale = 1.0f / sqrt((float)S_v);
 
-    // input state layout (D, K, n_seqs): per-seq stride is K*H*D; we read slot 0.
+    // input state layout [S_v, S_v, H, n_seqs] (s0 only): per-seq stride is H*D.
     // state is stored transposed: M[i20][is] = S[is][i20], so row i20 is contiguous
-    const uint state_in_base = (i23*K*args.ne21 + i21)*S_v*S_v + i20*S_v;
+    const uint state_in_base = (i23*args.ne21 + i21)*S_v*S_v + i20*S_v;
     device const float * s_ptr = (device const float *) (s) + state_in_base;
 
     float ls[NSG];
@@ -2620,9 +2620,8 @@ kernel void kernel_gated_delta_net_impl(
     device const float * b_ptr = (device const float *) (b) + (i23*args.ne22*args.ne21 + i21);
     device const float * g_ptr = (device const float *) (g) + (i23*args.ne22*args.ne21 + i21)*G;
 
-    // snapshot slot mapping: target_slot = t - shift. When n_tokens < K, only the last
-    // n_tokens slots are written; earlier slots are left untouched (caller-owned).
-    const int shift = (int)args.ne22 - (int)K;
+    // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+    // When n_tokens < K, only slots 0..n_tokens-1 are written; older slots are caller-owned.
 
     // output state base offset: after attention scores
     const uint attn_size = args.ne22 * args.ne21 * S_v * args.ne23;
@@ -2680,7 +2679,7 @@ kernel void kernel_gated_delta_net_impl(
         g_ptr += args.ne21*G;
 
         if (K > 1) {
-            const int target_slot = (int)t - shift;
+            const int target_slot = (int)args.ne22 - 1 - (int)t;
             if (target_slot >= 0 && target_slot < (int)K) {
                 device float * dst_state = (device float *) (dst) + attn_size + (uint)target_slot * state_size_per_snap + state_out_base;
                 FOR_UNROLL (short j = 0; j < NSG; j++) {

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -558,7 +558,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_set_rows_f32_i64, kernel_set_rows_f32_i32, kernel_set_rows_f16_i64, kernel_set_rows_f16_i32;
     cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16;
     cl_kernel kernel_rope_multi_f32, kernel_rope_multi_f16, kernel_rope_vision_f32, kernel_rope_vision_f16;
-    cl_kernel kernel_cpy_f16_f16, kernel_cpy_f16_f32, kernel_cpy_f32_f16, kernel_cpy_f32_f32, kernel_cpy_i32_i32;
+    cl_kernel kernel_cpy_f16_f16, kernel_cpy_f16_f32, kernel_cpy_f32_f16, kernel_cpy_f32_f32, kernel_cpy_f32_f32_pack, kernel_cpy_i32_i32;
     cl_kernel kernel_mul_mat_f32_f32;
     cl_kernel kernel_mul_mat_f16_f16;
     cl_kernel kernel_mul_mat_f16_f32_1row;
@@ -639,7 +639,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_softplus_f16, kernel_softplus_f16_4, kernel_softplus_f16_nc;
     cl_kernel kernel_upscale;
     cl_kernel kernel_upscale_bilinear;
-    cl_kernel kernel_concat_f32;
+    cl_kernel kernel_concat_f32, kernel_concat_f32_pack;
     cl_kernel kernel_conv_2d_f16;
     cl_kernel kernel_conv_2d_f32;
     cl_kernel kernel_conv_2d_f16_f32;
@@ -1121,6 +1121,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
         CL_CHECK((backend_ctx->kernel_cpy_f16_f32 = clCreateKernel(prog, "kernel_cpy_f16_f32", &err), err));
         CL_CHECK((backend_ctx->kernel_cpy_f32_f16 = clCreateKernel(prog, "kernel_cpy_f32_f16", &err), err));
         CL_CHECK((backend_ctx->kernel_cpy_f32_f32 = clCreateKernel(prog, "kernel_cpy_f32_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_cpy_f32_f32_pack = clCreateKernel(prog, "kernel_cpy_f32_f32_pack", &err), err));
         CL_CHECK((backend_ctx->kernel_cpy_i32_i32 = clCreateKernel(prog, "kernel_cpy_i32_i32", &err), err));
         GGML_LOG_CONT(".");
     }
@@ -2615,6 +2616,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
         cl_program prog =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
         CL_CHECK((backend_ctx->kernel_concat_f32 = clCreateKernel(prog, "kernel_concat_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_concat_f32_pack = clCreateKernel(prog, "kernel_concat_f32_pack", &err), err));
         CL_CHECK(clReleaseProgram(prog));
         GGML_LOG_CONT(".");
     }
@@ -8552,7 +8554,14 @@ static void ggml_cl_get_rows(ggml_backend_t backend, const ggml_tensor * src0, c
         nth *= 2;
     }
 
-    size_t global_work_size[] = {(size_t)ne10*nth, (size_t)ne11, (size_t)ne12};
+    int nchunks = 1;
+    if (src0->type == GGML_TYPE_F32) {
+        const int chunk_target = nth * 4;
+        nchunks = (ne00 + chunk_target - 1) / chunk_target;
+        nchunks = MAX(1, MIN(nchunks, 64));
+    }
+
+    size_t global_work_size[] = {(size_t)ne10*nth*nchunks, (size_t)ne11, (size_t)ne12};
     size_t local_work_size[] = {(size_t)nth, 1, 1};
 
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
@@ -11128,7 +11137,9 @@ static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, con
 
     int nth = MIN(64, ne0);
 
-    cl_kernel kernel = backend_ctx->kernel_concat_f32;
+    const bool concat_pack = (dim == 0 && ne0 < 32);
+    cl_kernel kernel = concat_pack ? backend_ctx->kernel_concat_f32_pack
+                                   : backend_ctx->kernel_concat_f32;
 
     CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
     CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
@@ -11155,10 +11166,28 @@ static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, con
     CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &nb3));
     CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_int),   &dim));
 
-    size_t global_work_size[] = {(size_t)ne1*nth, (size_t)ne2, (size_t)ne3};
-    size_t local_work_size[] = {(size_t)nth, 1, 1};
+    if (concat_pack) {
+        // packed kernel needs the dst dims to unflatten its 1-D row index.
+        CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int), &ne1));
+        CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int), &ne2));
+        CL_CHECK(clSetKernelArg(kernel, 26, sizeof(int), &ne3));
 
-    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+        const int maxwg = (int)backend_ctx->get_kernel_workgroup_size(kernel);
+        const int base  = MIN(64, maxwg);
+        const int tpr   = MIN(ne0, base);                 // threads per row
+        const int rpw   = MAX(1, base / tpr);             // rows per workgroup
+        const int lsz   = tpr * rpw;
+        const int nrows = ne1*ne2*ne3;
+        const int nwg   = (nrows + rpw - 1) / rpw;
+        size_t global_work_size[] = {(size_t)nwg*lsz, 1, 1};
+        size_t local_work_size[]  = {(size_t)lsz, 1, 1};
+        backend_ctx->enqueue_ndrange_kernel(kernel, 1, global_work_size, local_work_size, dst);
+    } else {
+        size_t global_work_size[] = {(size_t)ne1*nth, (size_t)ne2, (size_t)ne3};
+        size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    }
 }
 
 static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
@@ -14536,7 +14565,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
             } else if (backend_ctx->gpu_family == ADRENO) {
                 nth0 = 64;
                 nth1 = 2;
-                ndst = 4;
+                ndst = 16;
             } else {
                 GGML_ASSERT(false && "TODO: Unknown GPU");
             }
@@ -16633,7 +16662,8 @@ static void ggml_cl_cpy(ggml_backend_t backend, const ggml_tensor * src0, const
                     kernel = backend_ctx->kernel_cpy_f32_f16;
                     break;
                 case GGML_TYPE_F32:
-                    kernel = backend_ctx->kernel_cpy_f32_f32;
+                    kernel = ne00 < 32 ? backend_ctx->kernel_cpy_f32_f32_pack
+                                       : backend_ctx->kernel_cpy_f32_f32;
                     break;
                 default:
                     GGML_ASSERT(false && "not implemented");
@@ -16685,12 +16715,27 @@ static void ggml_cl_cpy(ggml_backend_t backend, const ggml_tensor * src0, const
     CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb12));
     CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb13));
 
-    const int nth = MIN(64, ne00);
+    if (kernel == backend_ctx->kernel_cpy_f32_f32_pack) {
+        const int maxwg = (int)backend_ctx->get_kernel_workgroup_size(kernel);
+        const int base  = MIN(64, maxwg);
+        const int tpr   = MIN(ne00, base);                 // threads per row
+        const int rpw   = MAX(1, base / tpr);              // rows per workgroup
+        const int lsz   = tpr * rpw;                       // <= base <= maxwg
+        const int nrows = ne01*ne02*ne03;
+        const int nwg   = (nrows + rpw - 1) / rpw;
 
-    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
-    size_t local_work_size[] = {(size_t)nth, 1, 1};
+        size_t global_work_size[] = {(size_t)nwg*lsz, 1, 1};
+        size_t local_work_size[]  = {(size_t)lsz, 1, 1};
 
-    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, src1);
+        backend_ctx->enqueue_ndrange_kernel(kernel, 1, global_work_size, local_work_size, src1);
+    } else {
+        const int nth = MIN(64, ne00);
+
+        size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+        size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, src1);
+    }
 }
 
 static void ggml_cl_dup(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -17705,7 +17750,7 @@ static void ggml_cl_gated_delta_net(ggml_backend_t backend, ggml_tensor * dst) {
     const cl_uint H_v      = (cl_uint) src_v->ne[1];
     const cl_uint n_tokens = (cl_uint) src_v->ne[2];
     const cl_uint n_seqs   = (cl_uint) src_v->ne[3];
-    const cl_uint K        = (cl_uint) src_state->ne[1];
+    const cl_uint K        = (cl_uint) ggml_get_op_params_i32(dst, 0);
 
     int si;
     switch (S_v) {

ggml/src/ggml-opencl/kernels/concat.cl

@@ -49,3 +49,70 @@ kernel void kernel_concat_f32(
         *y = *x;
     }
 }
+
+kernel void kernel_concat_f32_pack(
+    global  const char * src0,
+    ulong                offset0,
+    global  const char * src1,
+    ulong                offset1,
+    global        char * dst,
+    ulong                offsetd,
+    int             ne00,
+    int             ne01,
+    int             ne02,
+    int             ne03,
+    ulong           nb00,
+    ulong           nb01,
+    ulong           nb02,
+    ulong           nb03,
+    ulong           nb10,
+    ulong           nb11,
+    ulong           nb12,
+    ulong           nb13,
+    int             ne0,
+    ulong           nb0,
+    ulong           nb1,
+    ulong           nb2,
+    ulong           nb3,
+    int             dim,
+    int             ne1,
+    int             ne2,
+    int             ne3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int lsz = get_local_size(0);
+    int tpr = min(ne0, lsz);          // threads per row
+    int rpw = lsz / tpr;              // rows per workgroup
+    int lid = get_local_id(0);
+    int row = get_group_id(0)*rpw + lid / tpr;
+    int lane = lid - (lid / tpr) * tpr;
+
+    int nrows = ne1*ne2*ne3;
+    if (row >= nrows) {
+        return;
+    }
+
+    int i1 = row % ne1;
+    int t  = row / ne1;
+    int i2 = t % ne2;
+    int i3 = t / ne2;
+
+    int o[4] = {0, 0, 0, 0};
+    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
+
+    for (int i0 = lane; i0 < ne0; i0 += tpr) {
+        global const float * x;
+        if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+            x = (global const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
+        } else {
+            x = (global const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
+        }
+
+        global float * y = (global float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+        *y = *x;
+    }
+}

ggml/src/ggml-opencl/kernels/cpy.cl

@@ -183,6 +183,65 @@ kernel void kernel_cpy_f32_f32(
     }
 }
 
+kernel void kernel_cpy_f32_f32_pack(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int lsz = get_local_size(0);
+    int tpr = min(ne00, lsz);          // threads per row
+    int rpw = lsz / tpr;               // rows per workgroup
+    int lid = get_local_id(0);
+    int row = get_group_id(0)*rpw + lid / tpr;
+    int lane = lid - (lid / tpr) * tpr;
+
+    int nrows = ne01*ne02*ne03;
+    if (row >= nrows) {
+        return;
+    }
+
+    int i01 = row % ne01;
+    int t   = row / ne01;
+    int i02 = t % ne02;
+    int i03 = t / ne02;
+
+    // linear index of the first element of this row, unflattened over dst dims
+    long n  = (long)row * ne00;
+    int i3  = (int)(n / ((long)ne2*ne1*ne0));
+    long rm = n - (long)i3*ne2*ne1*ne0;
+    int i2  = (int)(rm / ((long)ne1*ne0));
+    rm     -= (long)i2*ne1*ne0;
+    int i1  = (int)(rm / ne0);
+    int i0  = (int)(rm - (long)i1*ne0);
+
+    global float * dst_data = (global float *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = lane; i00 < ne00; i00 += tpr) {
+        global const float * src = (global float *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        dst_data[i00] = src[0];
+    }
+}
+
 kernel void kernel_cpy_i32_i32(
         global int * src0,
         ulong offset0,

ggml/src/ggml-opencl/kernels/gated_delta_net.cl

@@ -123,7 +123,8 @@ kernel void kernel_gated_delta_net(
     const uint iq3 = seq_id / rq3; // seq index for Q and K
 
     const uint state_size = S_V * S_V;
-    const uint state_base = (seq_id * K * H_v + head_id) * state_size;
+    // input state holds s0 only [S_v, S_v, H, n_seqs]: per-seq stride is H*D.
+    const uint state_base = (seq_id * H_v + head_id) * state_size;
     const uint q_off_base  = iq3 * sq3 + iq1 * sq1;
     const uint v_off_base  = seq_id * sv3 + head_id * sv1;
     const uint gb_off_base = seq_id * sb3 + head_id * sb1;
@@ -143,7 +144,8 @@ kernel void kernel_gated_delta_net(
         }
     }
 
-    const int shift = (int)n_tokens - (int)K;
+    // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+    // When n_tokens < K only slots 0..n_tokens-1 are written; older slots are caller-owned.
     uint attn_off = (seq_id * n_tokens * H_v + head_id) * S_V;
 
     for (uint t = 0; t < n_tokens; t++) {
@@ -219,7 +221,7 @@ kernel void kernel_gated_delta_net(
         attn_off += S_V * H_v;
 
         if (K > 1u) {
-            const int target_slot = (int)t - shift;
+            const int target_slot = (int)n_tokens - 1 - (int)t;
             if (target_slot >= 0 && target_slot < (int)K) {
                 #pragma unroll
                 for (uint cg = 0; cg < COLS_PER_LANE_GROUP; cg++) {

ggml/src/ggml-opencl/kernels/get_rows.cl

@@ -82,21 +82,27 @@ kernel void kernel_get_rows_f32(
     src1 = (global int*)((global char*)src1 + offset1);
     dst = (global float*)((global char*)dst + offsetd);
 
-    int i10 = get_group_id(0);
-    int i11 = get_group_id(1);
-    int i12 = get_group_id(2);
+    int nchunks = get_num_groups(0) / ne10;
+    int g       = get_group_id(0);
+    int i10     = g / nchunks;
+    int chunk   = g - i10 * nchunks;
+    int i11     = get_group_id(1);
+    int i12     = get_group_id(2);
 
     int r = ((global int *) ((global char *) src1 + i12*nb12 + i11*nb11 + i10*nb10))[0];
 
     int i02 = i11;
     int i03 = i12;
 
-    for (int ind = get_local_id(0); ind < ne00; ind += get_local_size(0)) {
-        if (ind >= ne00) {
-            return;
-        }
-        ((global float *) ((global char *) dst + i12*nb3 + i11*nb2 + i10*nb1))[ind] =
-            ((global float *) ((global char *) src0 + r*nb01 + i02*nb02 + i03*nb03))[ind];
+    global float * dst_row = (global float *) ((global char *) dst  + i12*nb3 + i11*nb2 + i10*nb1);
+    global float * src_row = (global float *) ((global char *) src0 + r*nb01 + i02*nb02 + i03*nb03);
+
+    int span  = (ne00 + nchunks - 1) / nchunks;
+    int start = chunk * span;
+    int end   = min(start + span, ne00);
+
+    for (int ind = start + get_local_id(0); ind < end; ind += get_local_size(0)) {
+        dst_row[ind] = src_row[ind];
     }
 }
 

ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl

@@ -33,13 +33,15 @@ inline float block_q_6_K_dot_y_flat(
     global uchar * blk_qh,
     global char  * blk_scales,
     global half  * blk_d,
-    global float * yy,
     int ib,
     int ip,
     int is,
-    int l0
+    int l0,
+    float4 y0,
+    float4 y1,
+    float4 y2,
+    float4 y3
 ) {
-    int y_offset   = 128*ip + l0;
     int q_offset_l =  64*ip + l0;
     int q_offset_h =  32*ip + l0;
 
@@ -48,36 +50,28 @@ inline float block_q_6_K_dot_y_flat(
     global uchar * qh = blk_qh     + ib*64 + q_offset_h;
     global char  * sc = blk_scales + ib*16 + is;
 
-    global float * y = yy + ib * QK_K + y_offset;
-
     float dall = blk_d[ib];
 
-    float  sumf = 0;
-    float4 sums = {0.f, 0.f, 0.f, 0.f};
+    // Vectorized loads: 3 uchar4 weight loads instead of 12 scalar byte reads.
+    // q_offset_l/h are 4-aligned, so these are aligned vector loads.
+    uchar4 q1v = vload4(0, q1);
+    uchar4 q2v = vload4(0, q2);
+    uchar4 qhv = vload4(0, qh);
 
-    sums.s0 += y[0+ 0] * ((float)((q1[0] & 0xF) | ((qh[0] & Q6_K_MASK1) << 4)) - 32.f);
-    sums.s1 += y[0+32] * ((float)((q2[0] & 0xF) | ((qh[0] & Q6_K_MASK2) << 2)) - 32.f);
-    sums.s2 += y[0+64] * ((float)((q1[0]  >> 4) | ((qh[0] & Q6_K_MASK3) << 0)) - 32.f);
-    sums.s3 += y[0+96] * ((float)((q2[0]  >> 4) | ((qh[0] & Q6_K_MASK4) >> 2)) - 32.f);
+    int4 q1i = convert_int4(q1v);
+    int4 q2i = convert_int4(q2v);
+    int4 qhi = convert_int4(qhv);
 
-    sums.s0 += y[1+ 0] * ((float)((q1[1] & 0xF) | ((qh[1] & Q6_K_MASK1) << 4)) - 32.f);
-    sums.s1 += y[1+32] * ((float)((q2[1] & 0xF) | ((qh[1] & Q6_K_MASK2) << 2)) - 32.f);
-    sums.s2 += y[1+64] * ((float)((q1[1]  >> 4) | ((qh[1] & Q6_K_MASK3) << 0)) - 32.f);
-    sums.s3 += y[1+96] * ((float)((q2[1]  >> 4) | ((qh[1] & Q6_K_MASK4) >> 2)) - 32.f);
+    // Reconstruct the four 6-bit weight groups (low/high nibble of ql OR'd with the
+    // matching 2-bit plane of qh), same arithmetic as the scalar version, then dot()
+    // against the cached activation lanes.
+    float4 w0 = convert_float4((q1i & 0xF) | ((qhi & Q6_K_MASK1) << 4)) - 32.f;
+    float4 w1 = convert_float4((q2i & 0xF) | ((qhi & Q6_K_MASK2) << 2)) - 32.f;
+    float4 w2 = convert_float4((q1i >> 4)  | ((qhi & Q6_K_MASK3)     )) - 32.f;
+    float4 w3 = convert_float4((q2i >> 4)  | ((qhi & Q6_K_MASK4) >> 2)) - 32.f;
 
-    sums.s0 += y[2+ 0] * ((float)((q1[2] & 0xF) | ((qh[2] & Q6_K_MASK1) << 4)) - 32.f);
-    sums.s1 += y[2+32] * ((float)((q2[2] & 0xF) | ((qh[2] & Q6_K_MASK2) << 2)) - 32.f);
-    sums.s2 += y[2+64] * ((float)((q1[2]  >> 4) | ((qh[2] & Q6_K_MASK3) << 0)) - 32.f);
-    sums.s3 += y[2+96] * ((float)((q2[2]  >> 4) | ((qh[2] & Q6_K_MASK4) >> 2)) - 32.f);
-
-    sums.s0 += y[3+ 0] * ((float)((q1[3] & 0xF) | ((qh[3] & Q6_K_MASK1) << 4)) - 32.f);
-    sums.s1 += y[3+32] * ((float)((q2[3] & 0xF) | ((qh[3] & Q6_K_MASK2) << 2)) - 32.f);
-    sums.s2 += y[3+64] * ((float)((q1[3]  >> 4) | ((qh[3] & Q6_K_MASK3) << 0)) - 32.f);
-    sums.s3 += y[3+96] * ((float)((q2[3]  >> 4) | ((qh[3] & Q6_K_MASK4) >> 2)) - 32.f);
-
-    sumf += dall * (sums.s0 * sc[0] + sums.s1 * sc[2] + sums.s2 * sc[4] + sums.s3 * sc[6]);
-
-    return sumf;
+    return dall * (dot(y0, w0) * sc[0] + dot(y1, w1) * sc[2] +
+                   dot(y2, w2) * sc[4] + dot(y3, w3) * sc[6]);
 }
 
 #undef N_DST
@@ -89,7 +83,7 @@ inline float block_q_6_K_dot_y_flat(
 #define N_SIMDGROUP 2
 #define N_SIMDWIDTH 16
 #elif defined (ADRENO_GPU)
-#define N_DST 4
+#define N_DST 16
 #define N_SIMDGROUP 2
 #define N_SIMDWIDTH 64
 #endif
@@ -146,49 +140,39 @@ kernel void kernel_mul_mv_q6_K_f32_flat(
     global half  * blk_d      = (global half  *) src0_d  + offset_src0_d;
     global float * yy         = (global float *) src1    + r1*ne10 + im*ne00*ne1;
 
-    int tid = get_sub_group_local_id()/BLOCK_STRIDE; // first block_stride groups have tid=0
-    int ix  = get_sub_group_local_id()%BLOCK_STRIDE; // first block is 0..block_stride-1
+    int tid = get_sub_group_local_id()%(N_SIMDWIDTH/BLOCK_STRIDE); // within-super-block part, 0..15
+    int ix  = get_sub_group_local_id()/(N_SIMDWIDTH/BLOCK_STRIDE); // super-block selector, 0..BLOCK_STRIDE-1
     int ip  = tid/8;   // first or second half of (super) block (0 or 1)
     int il  = tid%8;   // each half has 8 parts, one per scale
     int n   = 4;       // 4 scales at a time (and 4 sums)
     int l0  = n*il;    // offset into half-block, 0..28
     int is  = 8*ip + l0/16; // 0, 1, 8, 9
 
-    float4 sumf = 0;
+    float sumf[N_DST];
+    for (int row = 0; row < N_DST; row++) {
+        sumf[row] = 0.f;
+    }
 
     for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
-        if (first_row + 0 < ne01) {
-            sumf.s0 += block_q_6_K_dot_y_flat(blk_ql + 0*nb*128, blk_qh + 0*nb*64, blk_scales + 0*nb*16, blk_d + 0*nb, yy, ib, ip, is, l0);
-        }
-        if (first_row + 1 < ne01) {
-            sumf.s1 += block_q_6_K_dot_y_flat(blk_ql + 1*nb*128, blk_qh + 1*nb*64, blk_scales + 1*nb*16, blk_d + 1*nb, yy, ib, ip, is, l0);
-        }
-        if (first_row + 2 < ne01) {
-            sumf.s2 += block_q_6_K_dot_y_flat(blk_ql + 2*nb*128, blk_qh + 2*nb*64, blk_scales + 2*nb*16, blk_d + 2*nb, yy, ib, ip, is, l0);
-        }
-        if (first_row + 3 < ne01) {
-            sumf.s3 += block_q_6_K_dot_y_flat(blk_ql + 3*nb*128, blk_qh + 3*nb*64, blk_scales + 3*nb*16, blk_d + 3*nb, yy, ib, ip, is, l0);
+        global float * y = yy + ib * QK_K + 128*ip + l0;
+        float4 y0 = vload4(0, y +  0);
+        float4 y1 = vload4(0, y + 32);
+        float4 y2 = vload4(0, y + 64);
+        float4 y3 = vload4(0, y + 96);
+
+        for (int row = 0; row < N_DST; row++) {
+            if (first_row + row < ne01) {
+                sumf[row] += block_q_6_K_dot_y_flat(
+                    blk_ql + row*nb*128, blk_qh + row*nb*64, blk_scales + row*nb*16, blk_d + row*nb,
+                    ib, ip, is, l0, y0, y1, y2, y3);
+            }
         }
     }
 
-    float4 tot = (float4)(
-        sub_group_reduce_add(sumf.s0),
-        sub_group_reduce_add(sumf.s1),
-        sub_group_reduce_add(sumf.s2),
-        sub_group_reduce_add(sumf.s3)
-    );
-    if (get_sub_group_local_id() == 0) {
-        if (first_row + 0 < ne01) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
-        }
-        if (first_row + 1 < ne01) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
-        }
-        if (first_row + 2 < ne01) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
-        }
-        if (first_row + 3 < ne01) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+    for (int row = 0; row < N_DST; row++) {
+        float tot = sub_group_reduce_add(sumf[row]);
+        if (get_sub_group_local_id() == 0 && first_row + row < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
         }
     }
 }

ggml/src/ggml-sycl/gated_delta_net.cpp

@@ -44,9 +44,9 @@ void gated_delta_net_sycl(const float *     q,
     float *       attn_data        = dst;
     float *       state            = dst + attn_score_elems;
 
-    // input state layout (D, K, n_seqs) — seq stride is K * D = K * H * S_v * S_v.
+    // input state holds s0 only [S_v, S_v, H, n_seqs] — seq stride is D = H * S_v * S_v.
     // output state layout (per-slot D * n_seqs) — same per-(seq,head) offset as before.
-    const int64_t state_in_offset      = sequence * K * H * S_v * S_v + h_idx * S_v * S_v;
+    const int64_t state_in_offset      = sequence * H * S_v * S_v + h_idx * S_v * S_v;
     const int64_t state_out_offset     = (sequence * H + h_idx) * S_v * S_v;
     const int64_t state_size_per_token = S_v * S_v * H * n_seqs; // per-slot stride in output
     state += state_out_offset;
@@ -63,9 +63,8 @@ void gated_delta_net_sycl(const float *     q,
         s_shard[r]  = curr_state[i];
     }
 
-    // slot mapping: target_slot = t - shift. When n_tokens < K only the last n_tokens slots
-    // are written; earlier slots are left untouched (caller-owned).
-    const int shift = (int) n_tokens - K;
+    // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+    // When n_tokens < K only slots 0..n_tokens-1 are written; older slots are caller-owned.
 
     for (int t = 0; t < n_tokens; t++) {
         const float * q_t = q + iq3 * sq3 + t * sq2 + iq1 * sq1;
@@ -144,7 +143,7 @@ void gated_delta_net_sycl(const float *     q,
 
     // Write state back to global memory
         if constexpr (keep_rs_t) {
-            const int target_slot = t - shift;
+            const int target_slot = (int) n_tokens - 1 - t;
             if (target_slot >= 0 && target_slot < K) {
                 float * curr_state = (dst + attn_score_elems) + target_slot * state_size_per_token + state_out_offset;
 #pragma unroll
@@ -315,8 +314,8 @@ void ggml_sycl_op_gated_delta_net(ggml_backend_sycl_context & ctx, ggml_tensor *
 
     dpct::queue_ptr stream = ctx.stream();
 
-    // state is 3D (S_v*S_v*H, K, n_seqs); K is the snapshot slot count.
-    const int K = (int) src_state->ne[1];
+    // K (snapshot slot count) is an op param; state holds s0 only [S_v, S_v, H, n_seqs].
+    const int K = ggml_get_op_params_i32(dst, 0);
     const bool keep_rs = K > 1;
 
     if (kda) {

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

@@ -113,6 +113,21 @@ typedef struct VkPhysicalDeviceShaderBfloat16FeaturesKHR {
 } VkPhysicalDeviceShaderBfloat16FeaturesKHR;
 #endif
 
+#if !defined(VK_VALVE_shader_mixed_float_dot_product)
+#define VK_VALVE_shader_mixed_float_dot_product 1
+#define VK_VALVE_SHADER_MIXED_FLOAT_DOT_PRODUCT_SPEC_VERSION 1
+#define VK_VALVE_SHADER_MIXED_FLOAT_DOT_PRODUCT_EXTENSION_NAME "VK_VALVE_shader_mixed_float_dot_product"
+#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_MIXED_FLOAT_DOT_PRODUCT_FEATURES_VALVE ((VkStructureType)1000673000)
+typedef struct VkPhysicalDeviceShaderMixedFloatDotProductFeaturesVALVE {
+    VkStructureType    sType;
+    void*              pNext;
+    VkBool32           shaderMixedFloatDotProductFloat16AccFloat32;
+    VkBool32           shaderMixedFloatDotProductFloat16AccFloat16;
+    VkBool32           shaderMixedFloatDotProductBFloat16Acc;
+    VkBool32           shaderMixedFloatDotProductFloat8AccFloat32;
+} VkPhysicalDeviceShaderMixedFloatDotProductFeaturesVALVE;
+#endif
+
 #define ROUNDUP_POW2(M, N) (((M) + (N) - 1) & ~((N) - 1))
 #define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
 static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
@@ -705,6 +720,8 @@ struct vk_device_struct {
     bool coopmat2_bf16_support {};
     bool coopmat2_decode_vector;
 
+    bool dot2_f16 {};
+
     bool pipeline_executable_properties_support {};
 
     size_t idx;
@@ -1976,6 +1993,9 @@ struct ggml_backend_vk_context {
     // Cache most recent tensor that was converted into prealloc_y, and what pipeline it used to convert.
     vk_pipeline_struct * prealloc_y_last_pipeline_used {};
     const ggml_tensor * prealloc_y_last_tensor_used {};
+    // True when prealloc_y holds the padded fp16 layout used by the coopmat2 B decode-vector callback.
+    // If false, then it's contiguous.
+    bool prealloc_y_last_decode_vector_staging {};
 
     // Track which nodes have been used since the last sync, and whether they were written to
     std::vector<const ggml_tensor *> unsynced_nodes_written;
@@ -3374,7 +3394,9 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec
     switch (src0_type) {
     case GGML_TYPE_IQ1_S:
     case GGML_TYPE_IQ1_M:
-        lut_size = 2*2048 + 4*2048;
+        // Regular matmul uses the compact uint16_t IQ1 grid; the expanded
+        // uint32_t grid is only enabled for the q8_1/int-dot vector path.
+        lut_size = 2*2048;
         break;
     case GGML_TYPE_IQ2_XXS:
         lut_size = 8*256;
@@ -3652,9 +3674,10 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
         s_mmq_wg_denoms_k = { 32,  64,  1 };
 
         // spec constants and tile sizes for quant matmul_id
-        l_warptile_mmqid = { 256, 128, 128, 32, 1, device->subgroup_size };
-        m_warptile_mmqid = { 256, 128, 64, 32, 0, device->subgroup_size };
-        s_warptile_mmqid = { 256, 128, 64, 32, 0, device->subgroup_size };
+        const uint32_t mmqid_bk = device->coopmat2_decode_vector ? 64u : 32u;
+        l_warptile_mmqid = { 256, 128, 128, mmqid_bk, 1, device->subgroup_size };
+        m_warptile_mmqid = { 256, 128, 64,  mmqid_bk, 0, device->subgroup_size };
+        s_warptile_mmqid = { 256, 128, 64,  mmqid_bk, 0, device->subgroup_size };
         l_mmqid_wg_denoms = { 128, 128, 1 };
         m_mmqid_wg_denoms = { 128, 64, 1 };
         s_mmqid_wg_denoms = { 128, 64, 1 };
@@ -3916,8 +3939,13 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
             name = aligned ? "flash_attn_f32_f16_aligned" : "flash_attn_f32_f16";
         } else {
             if (device->fp16) {
-                if (f32acc) { spv_data = flash_attn_f32_f16_data;        spv_size = flash_attn_f32_f16_len; }
-                else        { spv_data = flash_attn_f32_f16_f16acc_data; spv_size = flash_attn_f32_f16_f16acc_len; }
+                if (device->dot2_f16) {
+                    if (f32acc) { spv_data = flash_attn_f32_f16_dot2_data;        spv_size = flash_attn_f32_f16_dot2_len; }
+                    else        { spv_data = flash_attn_f32_f16_dot2_f16acc_data; spv_size = flash_attn_f32_f16_dot2_f16acc_len; }
+                } else {
+                    if (f32acc) { spv_data = flash_attn_f32_f16_data;        spv_size = flash_attn_f32_f16_len; }
+                    else        { spv_data = flash_attn_f32_f16_f16acc_data; spv_size = flash_attn_f32_f16_f16acc_len; }
+                }
             } else {
                 spv_data = flash_attn_f32_f16_fp32_data;
                 spv_size = flash_attn_f32_f16_fp32_len;
@@ -4211,7 +4239,23 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
 #endif  // defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
     if (device->fp16) {
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
+        // Selects dot2 SPIR-V variant at runtime when device->dot2_f16 is true
 #define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
+        if (device->mul_mat ## ID ## _l[TYPE]) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+        if (device->mul_mat ## ID ## _m[TYPE]) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+        if (device->mul_mat ## ID ## _s[TYPE]) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+        if (device->mul_mat ## ID ## _l[TYPE]) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _len : NAMELC ## _aligned ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _data : NAMELC ## _aligned ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+        if (device->mul_mat ## ID ## _m[TYPE]) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _len : NAMELC ## _aligned ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _data : NAMELC ## _aligned ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+        if (device->mul_mat ## ID ## _s[TYPE]) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _len : NAMELC ## _aligned ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _data : NAMELC ## _aligned ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+
+        // bf16 scalar path promotes to f32, no dot2 variant
+#define CREATE_MM_NODOT2(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
         if (device->mul_mat ## ID ## _l[TYPE]) \
             ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
@@ -4246,7 +4290,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, , 0);
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, , 0);
 
-        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, , 0);
+        CREATE_MM_NODOT2(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, , 0);
 
         CREATE_MM2(GGML_TYPE_Q1_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q1_0], matmul_q1_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
         CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0], matmul_q4_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
@@ -4254,7 +4298,6 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
         CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0], matmul_q5_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
         CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1], matmul_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
         CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0], matmul_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
-
         CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K], matmul_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
         CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K], matmul_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
         CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K], matmul_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, , 0);
@@ -4294,8 +4337,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
             CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
             CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
             CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_subgroup_f16_f32, wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
-            CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
-
+            CREATE_MM_NODOT2(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
             CREATE_MM2(GGML_TYPE_Q1_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q1_0], matmul_id_subgroup_q1_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
             CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
             CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
@@ -4340,8 +4382,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
             CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
             CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
             CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
-            CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
-
+            CREATE_MM_NODOT2(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
             CREATE_MM2(GGML_TYPE_Q1_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q1_0], matmul_id_q1_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
             CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
             CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
@@ -4386,6 +4427,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
 #undef CREATE_MM2
 #undef CREATE_MMQ
 #undef CREATE_MM
+#undef CREATE_MM_NODOT2
     } else {
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
@@ -5084,6 +5126,14 @@ 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
+        int idx = 0;
+        for (uint32_t n : {64, 128, 256, 512}) {
+            const uint32_t block_size = std::min(device->subgroup_size, n);
+            ggml_vk_create_pipeline(device, device->pipeline_fwht_f32[idx], "fwht_shmem_f32", fwht_shmem_f32_len, fwht_shmem_f32_data, "main", 2, sizeof(vk_op_fwht_push_constants), {1, 1, 1}, { block_size, n }, 1);
+            ++idx;
+        }
     }
 
     const uint32_t cumsum_elem_per_thread = (device->vendor_id == VK_VENDOR_ID_AMD || device->vendor_id == VK_VENDOR_ID_INTEL) ? 2 : 4;
@@ -5441,6 +5491,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->integer_dot_product = false;
         device->shader_64b_indexing = false;
         bool bfloat16_support = false;
+        bool dot2_f16_support = false;
 
         for (const auto& properties : ext_props) {
             if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
@@ -5483,6 +5534,9 @@ static vk_device ggml_vk_get_device(size_t idx) {
                        !getenv("GGML_VK_DISABLE_BFLOAT16")) {
                 bfloat16_support = true;
 #endif
+            } else if (strcmp("VK_VALVE_shader_mixed_float_dot_product", properties.extensionName) == 0 &&
+                       !getenv("GGML_VK_DISABLE_DOT2")) {
+                dot2_f16_support = true;
             } else if (strcmp("VK_KHR_pipeline_executable_properties", properties.extensionName) == 0) {
                 pipeline_executable_properties_support = true;
             } else if (strcmp("VK_EXT_memory_priority", properties.extensionName) == 0 &&
@@ -5630,6 +5684,11 @@ static vk_device ggml_vk_get_device(size_t idx) {
 #endif
         device->subgroup_shuffle = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
                                    (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eShuffle);
+#ifdef __APPLE__
+        if (device->vendor_id == VK_VENDOR_ID_AMD) {
+            device->subgroup_shuffle = false;
+        }
+#endif
         device->subgroup_clustered = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
                                      (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eClustered);
 
@@ -5785,6 +5844,14 @@ static vk_device ggml_vk_get_device(size_t idx) {
             device_extensions.push_back("VK_KHR_shader_integer_dot_product");
         }
 
+        VkPhysicalDeviceShaderMixedFloatDotProductFeaturesVALVE dot2_features {};
+        dot2_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_MIXED_FLOAT_DOT_PRODUCT_FEATURES_VALVE;
+        if (dot2_f16_support) {
+            last_struct->pNext = (VkBaseOutStructure *)&dot2_features;
+            last_struct = (VkBaseOutStructure *)&dot2_features;
+            device_extensions.push_back("VK_VALVE_shader_mixed_float_dot_product");
+        }
+
         VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR pep_features {};
         pep_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR;
         if (pipeline_executable_properties_support) {
@@ -5819,6 +5886,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->bf16 = false;
 #endif
 
+        device->dot2_f16 = dot2_f16_support && dot2_features.shaderMixedFloatDotProductFloat16AccFloat32;
+
         device->pipeline_robustness = pl_robustness_features.pipelineRobustness;
 
         device->multi_add = vk12_props.shaderRoundingModeRTEFloat16 &&
@@ -6233,6 +6302,7 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     bool coopmat2_decode_vector_support = false;
     bool integer_dot_product = false;
     bool bfloat16_support = false;
+    bool dot2_f16_support = false;
 
     for (auto properties : ext_props) {
         if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) {
@@ -6262,6 +6332,9 @@ static void ggml_vk_print_gpu_info(size_t idx) {
                     !getenv("GGML_VK_DISABLE_BFLOAT16")) {
             bfloat16_support = true;
 #endif
+        } else if (strcmp("VK_VALVE_shader_mixed_float_dot_product", properties.extensionName) == 0 &&
+                    !getenv("GGML_VK_DISABLE_DOT2")) {
+            dot2_f16_support = true;
         }
     }
 
@@ -6336,6 +6409,15 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     }
 #endif
 
+#if defined(VK_NV_cooperative_matrix2)
+    VkPhysicalDeviceCooperativeMatrix2FeaturesNV coopmat2_features {};
+    coopmat2_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_2_FEATURES_NV;
+    if (coopmat2_support) {
+        last_struct->pNext = (VkBaseOutStructure *)&coopmat2_features;
+        last_struct = (VkBaseOutStructure *)&coopmat2_features;
+    }
+#endif
+
     VkPhysicalDeviceCooperativeMatrixDecodeVectorFeaturesNV coopmat2_decode_vector_features {};
     coopmat2_decode_vector_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_DECODE_VECTOR_FEATURES_NV;
     if (coopmat2_decode_vector_support) {
@@ -6343,6 +6425,13 @@ static void ggml_vk_print_gpu_info(size_t idx) {
         last_struct = (VkBaseOutStructure *)&coopmat2_decode_vector_features;
     }
 
+    VkPhysicalDeviceShaderMixedFloatDotProductFeaturesVALVE dot2_features {};
+    dot2_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_MIXED_FLOAT_DOT_PRODUCT_FEATURES_VALVE;
+    if (dot2_f16_support) {
+        last_struct->pNext = (VkBaseOutStructure *)&dot2_features;
+        last_struct = (VkBaseOutStructure *)&dot2_features;
+    }
+
     vkGetPhysicalDeviceFeatures2(physical_device, &device_features2);
 
     fp16 = fp16 && vk12_features.shaderFloat16;
@@ -6367,6 +6456,19 @@ static void ggml_vk_print_gpu_info(size_t idx) {
 #endif
                    && ggml_vk_khr_cooperative_matrix_support(props2.properties, driver_props, device_architecture);
 
+#if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+    coopmat2_support = coopmat2_support &&
+                       coopmat2_features.cooperativeMatrixWorkgroupScope &&
+                       coopmat2_features.cooperativeMatrixFlexibleDimensions &&
+                       coopmat2_features.cooperativeMatrixReductions &&
+                       coopmat2_features.cooperativeMatrixConversions &&
+                       coopmat2_features.cooperativeMatrixPerElementOperations &&
+                       coopmat2_features.cooperativeMatrixTensorAddressing &&
+                       coopmat2_features.cooperativeMatrixBlockLoads;
+#else
+    coopmat2_support = false;
+#endif
+
     coopmat2_decode_vector_support = coopmat2_decode_vector_support && coopmat2_decode_vector_features.cooperativeMatrixDecodeVector;
 #if !defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR_GLSLC_SUPPORT)
     coopmat2_decode_vector_support = false;
@@ -6376,9 +6478,12 @@ static void ggml_vk_print_gpu_info(size_t idx) {
                              : coopmat_support  ? "KHR_coopmat"
                              : "none";
 
+    bool dot2_f16 = dot2_f16_support && dot2_features.shaderMixedFloatDotProductFloat16AccFloat32;
+    const char *fp16_str = fp16 ? (dot2_f16 ? "dot2" : "1") : "0";
+
     std::string device_name = props2.properties.deviceName.data();
-    GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | bf16: %d | warp size: %zu | shared memory: %d | int dot: %d | matrix cores: %s\n",
-              idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, bf16, subgroup_size,
+    GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %s | bf16: %d | warp size: %zu | shared memory: %d | int dot: %d | matrix cores: %s\n",
+              idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16_str, bf16, subgroup_size,
               props2.properties.limits.maxComputeSharedMemorySize, integer_dot_product, matrix_cores.c_str());
 
     if (props2.properties.deviceType == vk::PhysicalDeviceType::eCpu) {
@@ -8075,6 +8180,40 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
     ggml_vk_sync_buffers(ctx, subctx);
 }
 
+// Copy/convert tensor into a caller-defined dense layout. Destination strides
+// are in output elements, not bytes.
+static void ggml_vk_cpy_to_strided(
+        ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline pipeline, const ggml_tensor * tensor,
+        const vk_subbuffer & in, const vk_subbuffer & out,
+        uint32_t nb10, uint32_t nb11, uint32_t nb12, uint32_t nb13) {
+    VK_LOG_DEBUG("ggml_vk_cpy_to_strided((" << tensor << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << "), ";
+    std::cerr << "dst_nb=(" << nb10 << ", " << nb11 << ", " << nb12 << ", " << nb13 << "), buffer in size=" << in.buffer->size << ", buffer out size=" << out.buffer->size << ")");
+    const int tensor_type_size = ggml_type_size(tensor->type);
+
+    const uint32_t ne = ggml_nelements(tensor);
+    std::array<uint32_t, 3> elements;
+
+    if (ne > 262144) {
+        elements = { 512, 512, CEIL_DIV(ne, 262144) };
+    } else if (ne > 512) {
+        elements = { 512, CEIL_DIV(ne, 512), 1 };
+    } else {
+        elements = { ne, 1, 1 };
+    }
+
+    vk_op_unary_push_constants pc = {
+        (uint32_t)ne,
+        (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3], (uint32_t)tensor->nb[0] / tensor_type_size, (uint32_t)tensor->nb[1] / tensor_type_size, (uint32_t)tensor->nb[2] / tensor_type_size, (uint32_t)tensor->nb[3] / tensor_type_size,
+        (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3], nb10, nb11, nb12, nb13,
+        0,
+        0.0f, 0.0f,
+        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    };
+    init_pushconst_fastdiv(pc);
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, pc, elements);
+    ggml_vk_sync_buffers(ctx, subctx);
+}
+
 static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type) {
     switch(type) {
         case GGML_TYPE_Q8_1:
@@ -8332,24 +8471,28 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     }
     if (y_non_contig) {
         if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() ||
-            ctx->prealloc_y_last_tensor_used != src1) {
+            ctx->prealloc_y_last_tensor_used != src1 ||
+            ctx->prealloc_y_last_decode_vector_staging) {
             if (ctx->prealloc_y_need_sync) {
                 ggml_vk_sync_buffers(ctx, subctx);
             }
             ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0));
             ctx->prealloc_y_last_pipeline_used = to_fp16_vk_1.get();
             ctx->prealloc_y_last_tensor_used = src1;
+            ctx->prealloc_y_last_decode_vector_staging = false;
         }
     }
     if (quantize_y) {
         if (ctx->prealloc_y_last_pipeline_used != to_q8_1.get() ||
-            ctx->prealloc_y_last_tensor_used != src1) {
+            ctx->prealloc_y_last_tensor_used != src1 ||
+            ctx->prealloc_y_last_decode_vector_staging) {
             if (ctx->prealloc_y_need_sync) {
                 ggml_vk_sync_buffers(ctx, subctx);
             }
             ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0), y_ne);
             ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
             ctx->prealloc_y_last_tensor_used = src1;
+            ctx->prealloc_y_last_decode_vector_staging = false;
         }
     }
 
@@ -8607,24 +8750,28 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     if (y_non_contig) {
         GGML_ASSERT(y_sz == ggml_type_size(src1->type) * y_ne);
         if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() ||
-            ctx->prealloc_y_last_tensor_used != src1) {
+            ctx->prealloc_y_last_tensor_used != src1 ||
+            ctx->prealloc_y_last_decode_vector_staging) {
             if (ctx->prealloc_y_need_sync) {
                 ggml_vk_sync_buffers(ctx, subctx);
             }
             ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, d_Qy, d_Y);
             ctx->prealloc_y_last_pipeline_used = to_fp16_vk_1.get();
             ctx->prealloc_y_last_tensor_used = src1;
+            ctx->prealloc_y_last_decode_vector_staging = false;
         }
     }
     if (quantize_y) {
         if (ctx->prealloc_y_last_pipeline_used != to_q8_1.get() ||
-            ctx->prealloc_y_last_tensor_used != src1) {
+            ctx->prealloc_y_last_tensor_used != src1 ||
+            ctx->prealloc_y_last_decode_vector_staging) {
             if (ctx->prealloc_y_need_sync) {
                 ggml_vk_sync_buffers(ctx, subctx);
             }
             ggml_vk_quantize_q8_1(ctx, subctx, d_Qy, d_Y, y_ne);
             ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
             ctx->prealloc_y_last_tensor_used = src1;
+            ctx->prealloc_y_last_decode_vector_staging = false;
         }
     }
 
@@ -9075,12 +9222,30 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     // Reformat and convert to fp16 if non-contiguous, or for coopmat2 for better perf
     const bool x_non_contig = (ctx->device->coopmat2 && src0->type == GGML_TYPE_F32) ||
                               !ggml_vk_dim01_contiguous(src0);
-    const bool y_non_contig = (ctx->device->coopmat2 && src1->type == GGML_TYPE_F32) ||
+    // If src0 is BF16, try to use a BF16 x BF16 multiply
+    ggml_type f16_type = src0->type == GGML_TYPE_BF16 ? GGML_TYPE_BF16 : GGML_TYPE_F16;
+#if defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR_GLSLC_SUPPORT)
+    // B must already be, or be convertible to, the matmul B type used by this path.
+    const bool y_decode_vector_supported = ctx->device->coopmat2_decode_vector &&
+                                           (f16_type != GGML_TYPE_BF16 || ctx->device->coopmat2_bf16_support) &&
+                                           (src1->type == GGML_TYPE_F32 || src1->type == f16_type);
+    // If B is copied to prealloc_y, we can choose a 4-element-aligned row stride.
+    const bool y_decode_vector_uses_prealloc = !ggml_vk_dim01_contiguous(src1) || src1->type != f16_type;
+    // Direct B reads are safe only if row starts and the original buffer offset are 4-element aligned.
+    const bool y_decode_vector_aligned =
+        (ne10 % 4 == 0) &&
+        (y_decode_vector_uses_prealloc || get_misalign_bytes(ctx, src1) % (4 * ggml_type_size(src1->type)) == 0);
+    // Stage B only when decode-vector is available and direct B reads would be misaligned.
+    const bool y_decode_vector_staging = y_decode_vector_supported && !y_decode_vector_aligned;
+#else
+    const bool y_decode_vector_staging = false;
+#endif
+    const bool y_non_contig = y_decode_vector_staging ||
+                              (ctx->device->coopmat2 && src1->type == GGML_TYPE_F32) ||
                               (src0->type == GGML_TYPE_BF16 && src1->type != GGML_TYPE_BF16) ||
                               !ggml_vk_dim01_contiguous(src1);
 
-    // If src0 is BF16, try to use a BF16 x BF16 multiply
-    ggml_type f16_type = src0->type == GGML_TYPE_BF16 ? GGML_TYPE_BF16 : GGML_TYPE_F16;
+    const uint32_t y_staged_row_stride = y_decode_vector_staging ? (uint32_t)ggml_vk_align_size(ne10, 4) : (uint32_t)ne10;
 
     const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
 
@@ -9119,11 +9284,11 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
     uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11;
     const uint64_t x_ne = ggml_nelements(src0);
-    const uint64_t y_ne = padded_n * ne10 * ne12 * ne13;
+    const uint64_t y_ne = (uint64_t)y_staged_row_stride * padded_n * ne12 * ne13;
     const uint64_t d_ne = ggml_nelements(dst);
 
     const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type);
-    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
+    const uint64_t qy_sz = ggml_type_size(src1->type) * ggml_nelements(src1) / ggml_blck_size(src1->type);
     const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne;
     const uint64_t y_sz = quantize_y ? (ggml_vk_align_size(y_ne, 128) * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
     const uint64_t ids_sz = nbi2;
@@ -9133,13 +9298,30 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     vk_pipeline to_fp16_vk_1 = nullptr;
     vk_pipeline to_q8_1 = nullptr;
 
+    auto make_y_staged_dst = [&]() {
+        ggml_tensor y_staged_dst = *src1;
+        y_staged_dst.type = f16_type;
+        y_staged_dst.nb[0] = ggml_type_size(f16_type);
+        y_staged_dst.nb[1] = y_staged_dst.nb[0] * y_staged_row_stride;
+        y_staged_dst.nb[2] = y_staged_dst.nb[1] * padded_n;
+        y_staged_dst.nb[3] = y_staged_dst.nb[2] * y_staged_dst.ne[2];
+        return y_staged_dst;
+    };
+
     if (x_non_contig) {
         to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, f16_type);
     } else {
         to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type);
     }
     if (y_non_contig) {
-        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, f16_type);
+        ggml_tensor y_staged_dst;
+        const ggml_tensor * y_staged_dst_ptr = nullptr;
+        if (y_decode_vector_staging) {
+            y_staged_dst = make_y_staged_dst();
+            y_staged_dst_ptr = &y_staged_dst;
+        }
+
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, y_staged_dst_ptr, f16_type);
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
@@ -9257,30 +9439,47 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     }
     if (y_non_contig) {
         if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() ||
-            ctx->prealloc_y_last_tensor_used != src1) {
+            ctx->prealloc_y_last_tensor_used != src1 ||
+            ctx->prealloc_y_last_decode_vector_staging != y_decode_vector_staging) {
             if (ctx->prealloc_y_need_sync) {
                 ggml_vk_sync_buffers(ctx, subctx);
             }
-            ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0));
+            if (y_decode_vector_staging) {
+                const ggml_tensor y_staged_dst = make_y_staged_dst();
+                const uint32_t y_staged_dst_type_size = ggml_type_size(y_staged_dst.type);
+                ggml_vk_cpy_to_strided(
+                    ctx, subctx, to_fp16_vk_1, src1,
+                    ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0),
+                    (uint32_t)(y_staged_dst.nb[0] / y_staged_dst_type_size),
+                    (uint32_t)(y_staged_dst.nb[1] / y_staged_dst_type_size),
+                    (uint32_t)(y_staged_dst.nb[2] / y_staged_dst_type_size),
+                    (uint32_t)(y_staged_dst.nb[3] / y_staged_dst_type_size));
+            } else {
+                ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0));
+            }
             ctx->prealloc_y_last_pipeline_used = to_fp16_vk_1.get();
             ctx->prealloc_y_last_tensor_used = src1;
+            ctx->prealloc_y_last_decode_vector_staging = y_decode_vector_staging;
         }
     }
     if (quantize_y) {
         if (ctx->prealloc_y_last_pipeline_used != to_q8_1.get() ||
-            ctx->prealloc_y_last_tensor_used != src1) {
+            ctx->prealloc_y_last_tensor_used != src1 ||
+            ctx->prealloc_y_last_decode_vector_staging) {
             if (ctx->prealloc_y_need_sync) {
                 ggml_vk_sync_buffers(ctx, subctx);
             }
             ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(ctx, d_Qy, qy_buf_offset), ggml_vk_subbuffer(ctx, d_Y, 0), y_ne);
             ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
             ctx->prealloc_y_last_tensor_used = src1;
+            ctx->prealloc_y_last_decode_vector_staging = false;
         }
     }
     ggml_vk_sync_buffers(ctx, subctx);
 
     uint32_t stride_batch_x = ne00*ne01;
-    uint32_t stride_batch_y = ne10*ne11;
+    uint32_t stride_b_y = y_decode_vector_staging ? y_staged_row_stride : ne10;
+    uint32_t stride_batch_y = y_decode_vector_staging ? y_staged_row_stride * padded_n : ne10*ne11;
 
     if (!ggml_vk_dim01_contiguous(src0) && !qx_needs_dequant) {
         stride_batch_x = src0->nb[0] / ggml_type_size(src0->type);
@@ -9295,7 +9494,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
         ctx, subctx, pipeline,
         { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz },
         { d_D, d_buf_offset, d_sz }, { d_ids, ids_buf_offset, ids_sz }, expert_count_buf,
-        ne01, ne21, ne10, ne10, ne10, ne01,
+        ne01, ne21, ne10, ne10, stride_b_y, ne01,
         stride_batch_x, stride_batch_y, ne20*ne21,
         n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11, padded_n
     );  // NOLINT
@@ -9453,24 +9652,28 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     if (y_non_contig) {
         GGML_ASSERT(y_sz == ggml_type_size(src1->type) * y_ne);
         if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() ||
-            ctx->prealloc_y_last_tensor_used != src1) {
+            ctx->prealloc_y_last_tensor_used != src1 ||
+            ctx->prealloc_y_last_decode_vector_staging) {
             if (ctx->prealloc_y_need_sync) {
                 ggml_vk_sync_buffers(ctx, subctx);
             }
             ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, d_Qy, d_Y);
             ctx->prealloc_y_last_pipeline_used = to_fp16_vk_1.get();
             ctx->prealloc_y_last_tensor_used = src1;
+            ctx->prealloc_y_last_decode_vector_staging = false;
         }
     }
     if (quantize_y) {
         if (ctx->prealloc_y_last_pipeline_used != to_q8_1.get() ||
-            ctx->prealloc_y_last_tensor_used != src1) {
+            ctx->prealloc_y_last_tensor_used != src1 ||
+            ctx->prealloc_y_last_decode_vector_staging) {
             if (ctx->prealloc_y_need_sync) {
                 ggml_vk_sync_buffers(ctx, subctx);
             }
             ggml_vk_quantize_q8_1(ctx, subctx, d_Qy, d_Y, y_ne);
             ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
             ctx->prealloc_y_last_tensor_used = src1;
+            ctx->prealloc_y_last_decode_vector_staging = false;
         }
     }
 
@@ -11325,7 +11528,6 @@ static void ggml_vk_gated_delta_net(ggml_backend_vk_context * ctx, vk_context& s
     const ggml_tensor * src_q     = dst->src[0];
     const ggml_tensor * src_v     = dst->src[2];
     const ggml_tensor * src_beta  = dst->src[4];
-    const ggml_tensor * src_state = dst->src[5];
 
     GGML_ASSERT(dst->buffer != nullptr);
 
@@ -11334,8 +11536,8 @@ static void ggml_vk_gated_delta_net(ggml_backend_vk_context * ctx, vk_context& s
     const uint32_t n_tokens = (uint32_t)src_v->ne[2];
     const uint32_t n_seqs   = (uint32_t)src_v->ne[3];
 
-    // state is 3D (S_v*S_v*H, K, n_seqs); K is the snapshot slot count.
-    const uint32_t K = (uint32_t)src_state->ne[1];
+    // K (snapshot slot count) is an op param; state holds s0 only [S_v, S_v, H, n_seqs].
+    const uint32_t K = (uint32_t)ggml_get_op_params_i32(dst, 0);
 
     const uint32_t s_off = S_v * H * n_tokens * n_seqs;
 
@@ -13695,7 +13897,9 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx, vk_contex
             ggml_vk_destroy_buffer(ctx->prealloc_y);
         }
         ctx->prealloc_y = ggml_vk_create_buffer_device(ctx->device, ctx->prealloc_size_y);
+        ctx->prealloc_y_last_pipeline_used = nullptr;
         ctx->prealloc_y_last_tensor_used = nullptr;
+        ctx->prealloc_y_last_decode_vector_staging = false;
     }
     if (ctx->prealloc_split_k == nullptr || (ctx->prealloc_size_split_k > 0 && ctx->prealloc_split_k->size < ctx->prealloc_size_split_k)) {
         VK_LOG_MEMORY("ggml_vk_preallocate_buffers(split_k_size: " << ctx->prealloc_size_split_k << ")");
@@ -14275,6 +14479,8 @@ static void ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
 static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
     VK_LOG_DEBUG("ggml_vk_graph_cleanup()");
     ctx->prealloc_y_last_pipeline_used = {};
+    ctx->prealloc_y_last_tensor_used = nullptr;
+    ctx->prealloc_y_last_decode_vector_staging = false;
 
     ctx->unsynced_nodes_written.clear();
     ctx->unsynced_nodes_read.clear();
@@ -14325,6 +14531,8 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) {
     ggml_vk_destroy_buffer(ctx->sync_staging);
 
     ctx->prealloc_y_last_pipeline_used = nullptr;
+    ctx->prealloc_y_last_tensor_used = nullptr;
+    ctx->prealloc_y_last_decode_vector_staging = false;
 
     ctx->prealloc_size_x = 0;
     ctx->prealloc_size_y = 0;
@@ -15504,6 +15712,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
 
     ctx->prealloc_y_last_pipeline_used = nullptr;
     ctx->prealloc_y_last_tensor_used = nullptr;
+    ctx->prealloc_y_last_decode_vector_staging = false;
 
     if (ctx->prealloc_size_add_rms_partials) {
         ggml_vk_preallocate_buffers(ctx, nullptr);
@@ -17744,7 +17953,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
             src_clone[4], src_clone[5], src_clone[6]);
         } else if (tensor->op == GGML_OP_GATED_DELTA_NET) {
             tensor_clone = ggml_gated_delta_net(ggml_ctx, src_clone[0], src_clone[1],
-            src_clone[2], src_clone[3], src_clone[4], src_clone[5]);
+            src_clone[2], src_clone[3], src_clone[4], src_clone[5],
+            ggml_get_op_params_i32(tensor, 0));
         } else if (tensor->op == GGML_OP_OPT_STEP_ADAMW) {
             src_clone[0]->flags = tensor->src[0]->flags;
             tensor_clone = ggml_opt_step_adamw(ggml_ctx, src_clone[0], src_clone[1],

ggml/src/ggml-vulkan/vulkan-shaders/dot_product_funcs.glsl

@@ -0,0 +1,27 @@
+#ifdef DOT2_F16
+#extension GL_EXT_spirv_intrinsics : require
+
+spirv_instruction(extensions = ["SPV_VALVE_mixed_float_dot_product"],
+                  capabilities = [6912], id = 6916)
+float v_dot2_f32_f16(f16vec2 a, f16vec2 b, float acc);
+
+ACC_TYPE dot_product(f16vec4 a, f16vec4 b, ACC_TYPE acc) {
+    return ACC_TYPE(v_dot2_f32_f16(a.zw, b.zw, v_dot2_f32_f16(a.xy, b.xy, float(acc))));
+}
+
+ACC_TYPE dot_product(f16vec2 a, f16vec2 b, ACC_TYPE acc) {
+    return ACC_TYPE(v_dot2_f32_f16(a, b, float(acc)));
+}
+
+#else
+
+ACC_TYPE dot_product(FLOAT_TYPEV4 a, FLOAT_TYPEV4 b, ACC_TYPE acc) {
+    return fma(ACC_TYPE(a.x), ACC_TYPE(b.x), fma(ACC_TYPE(a.y), ACC_TYPE(b.y),
+           fma(ACC_TYPE(a.z), ACC_TYPE(b.z), fma(ACC_TYPE(a.w), ACC_TYPE(b.w), acc))));
+}
+
+ACC_TYPE dot_product(FLOAT_TYPEV2 a, FLOAT_TYPEV2 b, ACC_TYPE acc) {
+    return fma(ACC_TYPE(a.x), ACC_TYPE(b.x), fma(ACC_TYPE(a.y), ACC_TYPE(b.y), acc));
+}
+
+#endif

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

@@ -21,6 +21,7 @@
 #extension GL_KHR_shader_subgroup_vote : enable
 
 #include "types.glsl"
+#include "dot_product_funcs.glsl"
 #include "flash_attn_base.glsl"
 #include "flash_attn_dequant.glsl"
 
@@ -318,7 +319,7 @@ void main() {
                         K_Tf = FLOAT_TYPEV4(data_kv4[k_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * k_stride / 4 + d * D_split + d_tid]);
                     }
                     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                        Sf[r][c] += dot(ACC_TYPEV4(Q_cache[r]), ACC_TYPEV4(K_Tf));
+                        Sf[r][c] = dot_product(Q_cache[r], K_Tf, Sf[r][c]);
                     }
                 }
             }
@@ -341,7 +342,7 @@ void main() {
                         K_Tf = FLOAT_TYPEV4(data_kv4[k_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * k_stride / 4 + d * D_split + d_tid]);
                     }
                     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                        Sf[r][c] += dot(ACC_TYPEV4(Qf[tile_row(r) * qf_stride + d * D_split + d_tid]), ACC_TYPEV4(K_Tf));
+                        Sf[r][c] = dot_product(Qf[tile_row(r) * qf_stride + d * D_split + d_tid], K_Tf, Sf[r][c]);
                     }
                 }
             }

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

@@ -1,14 +1,16 @@
 #version 450
 
 #extension GL_EXT_control_flow_attributes : require
+#ifndef FWHT_SHMEM
 #extension GL_KHR_shader_subgroup_basic : enable
 #extension GL_KHR_shader_subgroup_shuffle : enable
+#endif
+
+layout(constant_id = 0) const uint BLOCK_SIZE = 32;
+layout(constant_id = 1) const uint N = 128;
 
 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;
@@ -20,35 +22,72 @@ layout(push_constant) uniform parameter
 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;
+const uint EL_W = N / BLOCK_SIZE;
+
+#ifdef FWHT_SHMEM
+shared float shmem[4 * N];
+#endif
 
 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) {
+#ifdef FWHT_SHMEM
+    const uint tid = gl_LocalInvocationID.x;
+    const uint shmem_base = gl_LocalInvocationID.y * N;
+    const uint row_id = gl_LocalInvocationID.y;
+#else
+    const uint tid = gl_SubgroupInvocationID;
+    const uint row_id = gl_SubgroupID;
+#endif
+
+    for (uint base_row = gl_WorkGroupID.x * gl_WorkGroupSize.y;
+            base_row < n_rows;
+            base_row += gl_NumWorkGroups.x * gl_WorkGroupSize.y) {
+        const uint row = base_row + row_id;
         const uint row_offset = row * N;
 
+#ifndef FWHT_SHMEM
+        if (row >= n_rows) {
+            continue;
+        }
+#endif
+
         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;
+            reg[i] = row < n_rows ? data_a[src_offset + row_offset + i * BLOCK_SIZE + tid] * scale : 0.0;
         }
 
+#ifdef FWHT_SHMEM
         [[unroll]]
-        for (uint h = 1; h < WARP_SIZE; h <<= 1) {
+        for (uint h = 1; h < BLOCK_SIZE; h <<= 1) {
+            [[unroll]]
+            for (uint i = 0; i < EL_W; ++i) {
+                shmem[shmem_base + i * BLOCK_SIZE + tid] = reg[i];
+            }
+            barrier();
+            [[unroll]]
+            for (uint j = 0; j < EL_W; ++j) {
+                const float val = reg[j];
+                const float other = shmem[shmem_base + j * BLOCK_SIZE + (tid ^ h)];
+                reg[j] = (tid & h) == 0 ? val + other : other - val;
+            }
+            barrier();
+        }
+#else
+        [[unroll]]
+        for (uint h = 1; h < BLOCK_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;
+                reg[j] = (tid & h) == 0 ? val + val2 : val2 - val;
             }
         }
+#endif
 
         [[unroll]]
-        for (uint h = WARP_SIZE; h < N; h <<= 1) {
-            const uint step = h / WARP_SIZE;
+        for (uint h = BLOCK_SIZE; h < N; h <<= 1) {
+            const uint step = h / BLOCK_SIZE;
             [[unroll]]
             for (uint j = 0; j < EL_W; j += 2 * step) {
                 [[unroll]]
@@ -61,9 +100,16 @@ void main() {
             }
         }
 
-        [[unroll]]
-        for (uint i = 0; i < EL_W; ++i) {
-            data_d[dst_offset + row_offset + i * WARP_SIZE + lane] = reg[i];
+#ifdef FWHT_SHMEM
+        if (row < n_rows) {
+#endif
+            [[unroll]]
+            for (uint i = 0; i < EL_W; ++i) {
+                data_d[dst_offset + row_offset + i * BLOCK_SIZE + tid] = reg[i];
+            }
+#ifdef FWHT_SHMEM
         }
+        barrier();
+#endif
     }
 }

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

@@ -102,8 +102,8 @@ void main() {
     const uint iq3 = seq_id / rq3;
 
     const uint state_size = S_V * S_V;
-    // input state layout (D, K, n_seqs): per-seq stride is K*H*D; we read slot 0.
-    const uint state_in_base       = (seq_id * K * H + head_id) * state_size;
+    // input state holds s0 only [S_v, S_v, H, n_seqs]: per-seq stride is H*D.
+    const uint state_in_base       = (seq_id * H + head_id) * state_size;
     // output state layout per slot: same per-(seq,head) offset as the single-slot case.
     const uint state_out_base      = (seq_id * H + head_id) * state_size;
     const uint state_size_per_snap = state_size * H * n_seqs;
@@ -113,9 +113,8 @@ void main() {
         s_shard[r] = FLOAT_TYPE(data_state[state_in_base + col * S_V + r * LANES_PER_COLUMN + lane]);
     }
 
-    // snapshot slot mapping: target_slot = t - shift. When n_tokens < K, only the last
-    // n_tokens slots are written; earlier slots are left untouched (caller-owned).
-    const int shift = int(n_tokens) - int(K);
+    // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+    // When n_tokens < K, only slots 0..n_tokens-1 are written; older slots are caller-owned.
 
     uint attn_off = (seq_id * n_tokens * H + head_id) * S_V;
 
@@ -172,7 +171,7 @@ void main() {
         attn_off += S_V * H;
 
         if (K > 1u) {
-            const int target_slot = int(t) - shift;
+            const int target_slot = int(n_tokens) - 1 - int(t);
             if (target_slot >= 0 && target_slot < int(K)) {
                 const uint slot_base = s_off + uint(target_slot) * state_size_per_snap + state_out_base;
                 [[unroll]] for (uint r = 0; r < ROWS_PER_LANE; r++) {

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

@@ -4,6 +4,7 @@
 #extension GL_EXT_integer_dot_product : require
 
 #define MMQ
+#define NEEDS_IQ1S_GRID_GPU
 #define B_TYPE block_q8_1_x4
 
 #include "mul_mat_vec_base.glsl"

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

@@ -29,6 +29,7 @@
 #endif
 
 #include "types.glsl"
+#include "dot_product_funcs.glsl"
 
 #ifndef LOAD_VEC_A
 #define LOAD_VEC_A 1
@@ -329,15 +330,8 @@ void main() {
                         [[unroll]] for (uint cr = 0; cr < TM / 2; cr++) {
                             // [WNITER][TN][WMITER][TM / 2] -> [wsic][cc][wsir][cr]
                             const uint sums_idx = (wsic * TN + cc) * WMITER * (TM / 2) + wsir * (TM / 2) + cr;
-                        #if defined(DATA_A_F32) || defined(DATA_A_F16)
-                            sums[sums_idx].x = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].y), ACC_TYPE(cache_b.y),
-                                               fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].z), ACC_TYPE(cache_b.z), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].w), ACC_TYPE(cache_b.w), sums[sums_idx].x))));
-                            sums[sums_idx].y = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].y), ACC_TYPE(cache_b.y),
-                                               fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].z), ACC_TYPE(cache_b.z), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].w), ACC_TYPE(cache_b.w), sums[sums_idx].y))));
-                        #else
-                            sums[sums_idx].x = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].y), ACC_TYPE(cache_b.y), sums[sums_idx].x));
-                            sums[sums_idx].y = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].y), ACC_TYPE(cache_b.y), sums[sums_idx].y));
-                        #endif
+                            sums[sums_idx].x = dot_product(cache_a[wsir * TM + 2 * cr    ], cache_b, sums[sums_idx].x);
+                            sums[sums_idx].y = dot_product(cache_a[wsir * TM + 2 * cr + 1], cache_b, sums[sums_idx].y);
                         }
                     }
                 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp

@@ -11,6 +11,9 @@
 #extension GL_KHR_memory_scope_semantics : enable
 #extension GL_KHR_cooperative_matrix : enable
 #extension GL_NV_cooperative_matrix2 : enable
+#ifdef GGML_VULKAN_COOPMAT2_DECODE_VECTOR
+#extension GL_NV_cooperative_matrix_decode_vector : enable
+#endif
 #extension GL_EXT_buffer_reference : enable
 #extension GL_KHR_shader_subgroup_ballot : enable
 #extension GL_KHR_shader_subgroup_vote : enable
@@ -69,10 +72,13 @@ layout (push_constant) uniform parameter
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+#if defined(MUL_MAT_ID) && defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR)
+layout (binding = 1) readonly buffer B4 {B_TYPEV4 data_b_v4[];};
+#endif
 
 #if QUANT_K > 1
 #include "dequant_funcs_cm2.glsl"
-#if defined(dequantFuncA_v) && defined(GL_NV_cooperative_matrix_decode_vector)
+#if defined(dequantFuncA_v) && defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR)
 #define DECODEFUNCA , dequantFuncA, dequantFuncA_v
 #else
 #define DECODEFUNCA , dequantFuncA
@@ -113,11 +119,33 @@ B_TYPE decodeFuncB(const in decodeBufB bl, const in uint blockCoords[2], const i
     const uint row_i = blockCoords[0];
 
     const u16vec4 row_idx = row_ids[row_i];
-    B_TYPE ret = data_b[row_idx.y * p.batch_stride_b + row_idx.x * p.stride_b + blockCoords[1]];
+#if defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR)
+    // The decode-vector path gives B a K-dimension tensor-layout block size of BK.
+    const uint k = blockCoords[1] * BK + coordInBlock[1];
+#else
+    const uint k = blockCoords[1];
+#endif
+    B_TYPE ret = data_b[row_idx.y * p.batch_stride_b + row_idx.x * p.stride_b + k];
 
     return ret;
 }
 
+#if defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR)
+B_TYPEV4 decodeFuncB_v(const in decodeBufB bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const uint row_i = blockCoords[0];
+
+    const u16vec4 row_idx = row_ids[row_i];
+    const uint k = blockCoords[1] * BK + coordInBlock[1];
+    const uint base = row_idx.y * p.batch_stride_b + row_idx.x * p.stride_b + k;
+
+    return data_b_v4[base >> 2];
+}
+#define DECODEFUNCB , decodeFuncB, decodeFuncB_v
+#else
+#define DECODEFUNCB , decodeFuncB
+#endif
+
 D_TYPE perElemOpD(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t ir, const in uint32_t ic)
 {
     uint dr = ir * BM + r;
@@ -287,6 +315,9 @@ void main() {
     tensorLayoutA = setTensorLayoutBlockSizeNV(tensorLayoutA, 1, QUANT_K);
     tensorLayoutAClamp = setTensorLayoutBlockSizeNV(tensorLayoutAClamp, 1, QUANT_K);
 #endif
+#if defined(MUL_MAT_ID) && defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR)
+    tensorLayoutB = setTensorLayoutBlockSizeNV(tensorLayoutB, 1, BK);
+#endif
 
     // Use end_k rather than p.K as the dimension because that's what
     // we need to bound check against when using split_k.
@@ -499,7 +530,7 @@ void main() {
                     coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
-                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);
+                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover4, block_k, BK), tensorViewTranspose DECODEFUNCB);
 
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 } else {
@@ -507,7 +538,7 @@ void main() {
                     coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
-                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);
+                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover4, block_k, BK), tensorViewTranspose DECODEFUNCB);
 
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 }
@@ -543,7 +574,7 @@ void main() {
                     coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
-                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
+                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover2, block_k, BK), tensorViewTranspose DECODEFUNCB);
 
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 } else {
@@ -551,7 +582,7 @@ void main() {
                     coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
-                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
+                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover2, block_k, BK), tensorViewTranspose DECODEFUNCB);
 
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 }
@@ -588,7 +619,7 @@ void main() {
 
                 coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
 #ifdef MUL_MAT_ID
-                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BN, block_k, BK), tensorViewTranspose DECODEFUNCB);
 #else
                 coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
 #endif
@@ -600,7 +631,7 @@ void main() {
 
                 coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
 #ifdef MUL_MAT_ID
-                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BN, block_k, BK), tensorViewTranspose DECODEFUNCB);
 #else
                 coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
 #endif

ggml/src/ggml-vulkan/vulkan-shaders/types.glsl

@@ -598,9 +598,10 @@ const uint[1024] iq1s_grid_const = {
     0x55dd55df, 0x55d555d7, 0x5503550c, 0x557f5501, 0x5577557d, 0x55405575, 0x555d555f, 0x55555557
 };
 
+#if defined(NEEDS_IQ1S_GRID_GPU)
 // Same content as iq1s_grid_const except each 2-bit value is expanded to 4-bit
 // and has 1 added to it (allows packed values to be extracted with & 0x0F0F0F0F
-// and 0xF0F0F0F0).
+// and 0xF0F0F0F0). This is only used by the q8_1/int-dot vector path.
 const uint32_t[2048] iq1s_grid_gpu_const = {
     0x00000000, 0x00000002, 0x00000101, 0x00000200, 0x00000202, 0x00010001, 0x00010101, 0x00020000,
     0x00020002, 0x00020200, 0x00020202, 0x01000101, 0x01010001, 0x01010100, 0x01010102, 0x01020101,
@@ -859,9 +860,12 @@ const uint32_t[2048] iq1s_grid_gpu_const = {
     0x20222020, 0x20222022, 0x20222220, 0x20222222, 0x21212021, 0x21212120, 0x21212122, 0x22202020,
     0x22202022, 0x22202220, 0x22202222, 0x22212121, 0x22222020, 0x22222022, 0x22222220, 0x22222222,
 };
+#endif
 
 shared uint16_t iq1s_grid[2048];
+#if defined(NEEDS_IQ1S_GRID_GPU)
 shared uint32_t iq1s_grid_gpu[2048];
+#endif
 
 #define NEEDS_INIT_IQ_SHMEM
 void init_iq_shmem(uvec3 wgsize)
@@ -875,12 +879,14 @@ void init_iq_shmem(uvec3 wgsize)
             iq1s_grid[2*idx+1] = g.y;
         }
     }
+#if defined(NEEDS_IQ1S_GRID_GPU)
     [[unroll]] for (uint i = 0; i < iq1s_grid_gpu_const.length(); i += wgsize.x) {
         uint idx = i + gl_LocalInvocationIndex.x;
         if (iq1s_grid_gpu_const.length() % wgsize.x == 0 || idx < iq1s_grid_gpu_const.length()) {
             iq1s_grid_gpu[idx] = iq1s_grid_gpu_const[idx];
         }
     }
+#endif
     barrier();
 }
 #endif

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

@@ -336,7 +336,8 @@ void string_to_spv_func(std::string name, std::string in_path, std::string out_p
     // disable spirv-opt for coopmat shaders for https://github.com/ggml-org/llama.cpp/issues/10734
     // disable spirv-opt for bf16 shaders for https://github.com/ggml-org/llama.cpp/issues/15344
     // disable spirv-opt for rope shaders for https://github.com/ggml-org/llama.cpp/issues/16860
-    if (!coopmat && name.find("bf16") == std::string::npos && name.find("rope") == std::string::npos) {
+    // disable spirv-opt for dot2 shaders (spirv-opt doesn't recognize SPV_VALVE_mixed_float_dot_product capability)
+    if (!coopmat && name.find("bf16") == std::string::npos && name.find("rope") == std::string::npos && name.find("_dot2") == std::string::npos) {
         cmd.push_back("-O");
     }
 
@@ -427,10 +428,11 @@ void string_to_spv(std::string name, const std::string& source, const std::map<s
     generate_dep_file = false;
 }
 
-void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool coopmat2, bool f16acc) {
+void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool coopmat2, bool f16acc, bool dot2 = false) {
     std::string load_vec = coopmat2 ? "1" : fp16 ? "8" : "4";
     std::string aligned_b_type_f32 = coopmat2 ? "float" : fp16 ? "mat2x4" : "vec4";
     std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4";
+    std::string dot2_sfx = dot2 ? "_dot2" : "";
 
     std::map<std::string, std::string> base_dict;
     std::string shader_name = "matmul";
@@ -457,6 +459,15 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
     if (coopmat) {
         base_dict["COOPMAT"] = "1";
     }
+#if defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR_GLSLC_SUPPORT)
+    if (coopmat2) {
+        base_dict["GGML_VULKAN_COOPMAT2_DECODE_VECTOR"] = "1";
+    }
+#endif
+
+    if (dot2) {
+        base_dict["DOT2_F16"] = "1";
+    }
 
     const std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp";
 
@@ -523,11 +534,11 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
     };
 
     // Shaders with f16 B_TYPE
-    string_to_spv(shader_name + "_f32_f16",         source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F32", "1"},                                                     {"B_TYPE", "float16_t"},        {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
-    string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f32_f16" + dot2_sfx,              source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F32", "1"},                                                     {"B_TYPE", "float16_t"},        {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f32_f16" + dot2_sfx + "_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
 
-    string_to_spv(shader_name + "_f16",             source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F16", "1"},                                                     {"B_TYPE", "float16_t"},        {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
-    string_to_spv(shader_name + "_f16_aligned",     source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f16" + dot2_sfx,              source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F16", "1"},                                                     {"B_TYPE", "float16_t"},            {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f16" + dot2_sfx + "_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16},     {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
 
     // bf16
     {
@@ -548,8 +559,10 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
         if (!(coopmat || coopmat2))
 #endif
         {
-            string_to_spv(shader_name + "_bf16",         source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"},                             {"B_TYPE", coopmat2 ? "bfloat16_t" : "uint16_t"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}}),                   fp16, coopmat, coopmat2, f16acc);
-            string_to_spv(shader_name + "_bf16_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"},  {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            if (!dot2) {
+                string_to_spv(shader_name + "_bf16",         source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"},                             {"B_TYPE", coopmat2 ? "bfloat16_t" : "uint16_t"}, {"B_TYPEV4", "bf16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}}),                   fp16, coopmat, coopmat2, f16acc);
+                string_to_spv(shader_name + "_bf16_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"},  {"B_TYPEV4", "bf16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            }
         }
     }
 
@@ -579,18 +592,18 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
 
         // don't generate f32 variants for coopmat2
         if (!coopmat2) {
-            string_to_spv(shader_name + "_" + tname + "_f32",         source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float"},            {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
-            string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f32" + dot2_sfx,              source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float"},            {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f32" + dot2_sfx + "_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
         }
 
         if (tname != "f16" && tname != "f32") {
-            string_to_spv(shader_name + "_" + tname + "_f16",         source_name,  merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float16_t"},        {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
-            string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name,  merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f16" + dot2_sfx,              source_name,  merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float16_t"},        {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f16" + dot2_sfx + "_aligned", source_name,  merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
         }
 
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
-        // Integer dot mmq performs better with f32 accumulators
-        if (!f16acc && !coopmat && !coopmat2 && (is_legacy_quant(tname) || is_k_quant(tname) || tname == "mxfp4")) {
+        // Integer dot mmq performs better with f32 accumulators (different shader, skip for dot2)
+        if (!f16acc && !coopmat && !coopmat2 && !dot2 && (is_legacy_quant(tname) || is_k_quant(tname) || tname == "mxfp4")) {
             string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
         }
 #endif
@@ -608,6 +621,10 @@ void process_shaders() {
         matmul_shaders(true, matmul_id_type, false, false, false);
         matmul_shaders(true, matmul_id_type, false, false, true);
 
+        // dot2 variants (scalar fp16 only)
+        matmul_shaders(true, matmul_id_type, false, false, false, true);
+        matmul_shaders(true, matmul_id_type, false, false, true,  true);
+
         if (matmul_id_type != MatMulIdType::DEFAULT) {
 #if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
             // Coopmat, fp32acc and fp16acc
@@ -655,6 +672,12 @@ void process_shaders() {
 
             string_to_spv("flash_attn_f32_f16", "flash_attn.comp",
                 merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}}), fp16, false, false, f16acc);
+
+            if (fp16) {
+                string_to_spv("flash_attn_f32_f16_dot2", "flash_attn.comp",
+                    merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"DOT2_F16", "1"}}), fp16, false, false, f16acc);
+            }
+
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
             string_to_spv("flash_attn_f32_f16", "flash_attn.comp",
                 merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"MMQ", "1"}, {"FA_MMQ_MIXED", "1"}}), fp16, false, false, f16acc, "_int8");
@@ -957,6 +980,7 @@ 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("fwht_shmem_f32", "fwht.comp", {{"FWHT_SHMEM", "1"}});
     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"}}));

ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp

@@ -448,15 +448,19 @@ struct ggml_webgpu_upscale_pipeline_key_hash {
 /** Concat **/
 
 struct ggml_webgpu_concat_pipeline_key {
-    int type;
+    int  type;
+    bool src_overlap;
 
-    bool operator==(const ggml_webgpu_concat_pipeline_key & other) const { return type == other.type; }
+    bool operator==(const ggml_webgpu_concat_pipeline_key & other) const {
+        return type == other.type && src_overlap == other.src_overlap;
+    }
 };
 
 struct ggml_webgpu_concat_pipeline_key_hash {
     size_t operator()(const ggml_webgpu_concat_pipeline_key & key) const {
         size_t seed = 0;
         ggml_webgpu_hash_combine(seed, key.type);
+        ggml_webgpu_hash_combine(seed, key.src_overlap);
         return seed;
     }
 };
@@ -640,7 +644,8 @@ inline size_t ggml_webgpu_flash_attn_tensor_offset(const ggml_tensor * tensor) {
 
 inline bool ggml_webgpu_flash_attn_float_vec4_aligned(const ggml_tensor * K, size_t storage_offset_alignment) {
     const uint32_t offset_elems =
-        (uint32_t) ((ggml_webgpu_flash_attn_tensor_offset(K) & (storage_offset_alignment - 1)) / ggml_type_size(K->type));
+        (uint32_t) ((ggml_webgpu_flash_attn_tensor_offset(K) & (storage_offset_alignment - 1)) /
+                    ggml_type_size(K->type));
     return offset_elems % GGML_WEBGPU_FLASH_ATTN_TILE_KV_VEC_WIDTH == 0u;
 }
 
@@ -651,8 +656,10 @@ inline bool ggml_webgpu_flash_attn_float_vec4_aligned(const ggml_tensor * K,
            ggml_webgpu_flash_attn_float_vec4_aligned(V, storage_offset_alignment);
 }
 
-inline bool ggml_webgpu_flash_attn_kv_direct(
-    const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, uint32_t kv_direct_align) {
+inline bool ggml_webgpu_flash_attn_kv_direct(const ggml_tensor * Q,
+                                             const ggml_tensor * K,
+                                             const ggml_tensor * V,
+                                             uint32_t            kv_direct_align) {
     return K->type == GGML_TYPE_F16 && V->type == GGML_TYPE_F16 && (Q->ne[0] % kv_direct_align == 0) &&
            (K->ne[1] % GGML_WEBGPU_KV_SEQ_PAD == 0);
 }
@@ -667,10 +674,10 @@ inline ggml_webgpu_flash_attn_common_pipeline_key ggml_webgpu_flash_attn_make_co
     key.dst_type                                   = context.dst->type;
     key.head_dim_qk                                = (uint32_t) context.src0->ne[0];
     key.head_dim_v                                 = (uint32_t) context.src2->ne[0];
-    key.kv_direct          = ggml_webgpu_flash_attn_kv_direct(context.src0, context.src1, context.src2, kv_direct_align);
-    key.kv_overlap         = ggml_webgpu_tensor_overlap(context.src1, context.src2);
-    key.has_mask           = context.src3 != nullptr;
-    key.has_sinks          = context.src4 != nullptr;
+    key.kv_direct  = ggml_webgpu_flash_attn_kv_direct(context.src0, context.src1, context.src2, kv_direct_align);
+    key.kv_overlap = ggml_webgpu_tensor_overlap(context.src1, context.src2);
+    key.has_mask   = context.src3 != nullptr;
+    key.has_sinks  = context.src4 != nullptr;
     key.uses_logit_softcap = ggml_get_op_params_f32(context.dst, 2) != 0.0f;
     return key;
 }
@@ -1723,7 +1730,7 @@ class ggml_webgpu_shader_lib {
         key.type                              = context.dst->type;
         key.d_state                           = (int) context.src0->ne[0];
         key.xbc_overlap                       = ggml_webgpu_tensor_overlap(context.src1, context.src4) &&
-                          ggml_webgpu_tensor_overlap(context.src1, context.src5);
+                                                ggml_webgpu_tensor_overlap(context.src1, context.src5);
 
         auto it = ssm_scan_pipelines.find(key);
         if (it != ssm_scan_pipelines.end()) {
@@ -2634,6 +2641,7 @@ class ggml_webgpu_shader_lib {
     webgpu_pipeline get_concat_pipeline(const ggml_webgpu_shader_lib_context & context) {
         ggml_webgpu_concat_pipeline_key key = {};
         key.type                            = context.dst->type;
+        key.src_overlap                     = ggml_webgpu_tensor_overlap(context.src0, context.src1);
 
         auto it = concat_pipelines.find(key);
         if (it != concat_pipelines.end()) {
@@ -2656,11 +2664,17 @@ class ggml_webgpu_shader_lib {
                 GGML_ABORT("Unsupported type for concat shader");
         }
 
+        if (key.src_overlap) {
+            defines.push_back("SRC_OVERLAP");
+            variant += "_src_overlap";
+        }
+
         defines.push_back(std::string("WG_SIZE=") + std::to_string(context.max_wg_size));
 
         auto processed           = preprocessor.preprocess(wgsl_concat, defines);
-        auto decisions           = std::make_shared<ggml_webgpu_generic_shader_decisions>();
+        auto decisions           = std::make_shared<ggml_webgpu_binary_shader_decisions>();
         decisions->wg_size       = context.max_wg_size;
+        decisions->src_overlap   = key.src_overlap;
         webgpu_pipeline pipeline = ggml_webgpu_create_pipeline(device, processed, variant);
         pipeline.context         = decisions;
         concat_pipelines[key]    = pipeline;

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -621,10 +621,11 @@ static void ggml_backend_webgpu_buffer_memset(webgpu_global_context & ctx,
                                               uint32_t                value,
                                               size_t                  offset,
                                               size_t                  size) {
-    std::vector<uint32_t>             params       = { (uint32_t) offset, (uint32_t) size, value };
-    std::vector<wgpu::BindGroupEntry> entries      = { ggml_webgpu_make_bind_group_entry(0, buf, 0, buf.GetSize()) };
-    size_t                            bytes_per_wg = ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup * ctx->capabilities.memset_bytes_per_thread;
-    uint32_t                          wg_x         = CEIL_DIV(size + 3, bytes_per_wg);
+    std::vector<uint32_t>             params  = { (uint32_t) offset, (uint32_t) size, value };
+    std::vector<wgpu::BindGroupEntry> entries = { ggml_webgpu_make_bind_group_entry(0, buf, 0, buf.GetSize()) };
+    size_t                            bytes_per_wg =
+        ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup * ctx->capabilities.memset_bytes_per_thread;
+    uint32_t wg_x = CEIL_DIV(size + 3, bytes_per_wg);
 
     ctx->queue.WriteBuffer(ctx->memset_params_buf, 0, params.data(), params.size() * sizeof(uint32_t));
 
@@ -1244,7 +1245,7 @@ static webgpu_encoded_op ggml_webgpu_gated_delta_net(webgpu_context & ctx,
     const uint32_t h        = (uint32_t) src2->ne[1];
     const uint32_t n_tokens = (uint32_t) src2->ne[2];
     const uint32_t n_seqs   = (uint32_t) src2->ne[3];
-    const uint32_t K        = (uint32_t) src5->ne[1];
+    const uint32_t K        = (uint32_t) ggml_get_op_params_i32(dst, 0);
     const float    scale    = 1.0f / sqrtf((float) s_v);
     uint32_t       scale_u32;
     memcpy(&scale_u32, &scale, sizeof(scale_u32));
@@ -1362,7 +1363,7 @@ static webgpu_encoded_op ggml_webgpu_get_rows(webgpu_context & ctx,
     shader_lib_ctx.src0                           = src;
     shader_lib_ctx.src1                           = nullptr;
     shader_lib_ctx.dst                            = dst;
-    shader_lib_ctx.max_wg_size                    = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup;
+    shader_lib_ctx.max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup;
 
     webgpu_pipeline pipeline  = ctx->shader_lib->get_get_rows_pipeline(shader_lib_ctx);
     auto *          decisions = static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context.get());
@@ -2169,8 +2170,10 @@ static webgpu_encoded_op ggml_webgpu_unary_op(webgpu_context & ctx, ggml_tensor
         entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 1, dst));
     }
 
-    uint32_t wg_x = CEIL_DIV(ne, decisions->wg_size);
-    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
+    uint32_t wg_x, wg_y;
+    uint32_t total_wg = CEIL_DIV(ggml_nelements(dst), decisions->wg_size);
+    compute_2d_workgroups(total_wg, ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension, wg_x, wg_y);
+    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x, wg_y);
 }
 
 static webgpu_encoded_op ggml_webgpu_binary_op(webgpu_context & ctx,
@@ -2244,8 +2247,10 @@ static webgpu_encoded_op ggml_webgpu_binary_op(webgpu_context & ctx,
         }
     }
 
-    uint32_t wg_x = CEIL_DIV(ne, decisions->wg_size);
-    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
+    uint32_t wg_x, wg_y;
+    uint32_t total_wg = CEIL_DIV(ggml_nelements(dst), decisions->wg_size);
+    compute_2d_workgroups(total_wg, ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension, wg_x, wg_y);
+    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x, wg_y);
 }
 
 static webgpu_encoded_op ggml_webgpu_add_id(webgpu_context & ctx,
@@ -2305,33 +2310,6 @@ static webgpu_encoded_op ggml_webgpu_concat(webgpu_context & ctx,
     uint32_t ne  = (uint32_t) ggml_nelements(dst);
     uint32_t dim = (uint32_t) dst->op_params[0];
 
-    std::vector<uint32_t> params = {
-        ne,
-        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type)),
-        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type)),
-        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
-        (uint32_t) (src0->nb[0] / ggml_type_size(src0->type)),
-        (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),
-        (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),
-        (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),
-        (uint32_t) (src1->nb[0] / ggml_type_size(src1->type)),
-        (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),
-        (uint32_t) (src1->nb[2] / ggml_type_size(src1->type)),
-        (uint32_t) (src1->nb[3] / ggml_type_size(src1->type)),
-        (uint32_t) dst->ne[0],
-        (uint32_t) dst->ne[1],
-        (uint32_t) dst->ne[2],
-        (uint32_t) dst->ne[3],
-        dim,
-        (uint32_t) src0->ne[dim]
-    };
-
-    std::vector<wgpu::BindGroupEntry> entries = {
-        ggml_webgpu_make_tensor_bind_group_entry(ctx, 0, src0),
-        ggml_webgpu_make_tensor_bind_group_entry(ctx, 1, src1),
-        ggml_webgpu_make_tensor_bind_group_entry(ctx, 2, dst),
-    };
-
     ggml_webgpu_shader_lib_context shader_lib_ctx = {};
     shader_lib_ctx.src0                           = src0;
     shader_lib_ctx.src1                           = src1;
@@ -2339,8 +2317,52 @@ static webgpu_encoded_op ggml_webgpu_concat(webgpu_context & ctx,
     shader_lib_ctx.max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup;
 
     webgpu_pipeline pipeline  = ctx->shader_lib->get_concat_pipeline(shader_lib_ctx);
-    auto *          decisions = static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context.get());
-    uint32_t        wg_x      = CEIL_DIV(ne, decisions->wg_size);
+    auto *          decisions = static_cast<ggml_webgpu_binary_shader_decisions *>(pipeline.context.get());
+
+    uint32_t offset_src0   = (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type));
+    uint32_t offset_src1   = (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type));
+    size_t   merged_offset = 0;
+    size_t   merged_size   = 0;
+    if (decisions->src_overlap) {
+        const ggml_webgpu_merged_binding_range merged_range =
+            ggml_webgpu_tensor_merged_binding_range(ctx, { src0, src1 });
+        merged_offset = merged_range.offset;
+        merged_size   = merged_range.size;
+        offset_src0   = ggml_webgpu_tensor_merged_element_offset(src0, merged_range);
+        offset_src1   = ggml_webgpu_tensor_merged_element_offset(src1, merged_range);
+    }
+
+    std::vector<uint32_t> params = { ne,
+                                     offset_src0,
+                                     offset_src1,
+                                     (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
+                                     (uint32_t) (src0->nb[0] / ggml_type_size(src0->type)),
+                                     (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),
+                                     (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),
+                                     (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),
+                                     (uint32_t) (src1->nb[0] / ggml_type_size(src1->type)),
+                                     (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),
+                                     (uint32_t) (src1->nb[2] / ggml_type_size(src1->type)),
+                                     (uint32_t) (src1->nb[3] / ggml_type_size(src1->type)),
+                                     (uint32_t) dst->ne[0],
+                                     (uint32_t) dst->ne[1],
+                                     (uint32_t) dst->ne[2],
+                                     (uint32_t) dst->ne[3],
+                                     dim,
+                                     (uint32_t) src0->ne[dim] };
+
+    std::vector<wgpu::BindGroupEntry> entries = {};
+    if (decisions->src_overlap) {
+        entries.push_back(
+            ggml_webgpu_make_bind_group_entry(0, ggml_webgpu_tensor_buf(src0), merged_offset, merged_size));
+        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 1, dst));
+    } else {
+        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 0, src0));
+        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 1, src1));
+        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 2, dst));
+    }
+
+    uint32_t wg_x = CEIL_DIV(ne, decisions->wg_size);
     return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
 }
 
@@ -2673,8 +2695,10 @@ static webgpu_encoded_op ggml_webgpu_scale(webgpu_context & ctx, ggml_tensor * s
         entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 1, dst));
     }
 
-    uint32_t wg_x = CEIL_DIV(ggml_nelements(dst), decisions->wg_size);
-    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
+    uint32_t wg_x, wg_y;
+    uint32_t total_wg = CEIL_DIV(ggml_nelements(dst), decisions->wg_size);
+    compute_2d_workgroups(total_wg, ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension, wg_x, wg_y);
+    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x, wg_y);
 }
 
 static webgpu_encoded_op ggml_webgpu_soft_max(webgpu_context & ctx,
@@ -3751,7 +3775,8 @@ static ggml_guid_t ggml_backend_webgpu_guid(void) {
 
 static void ggml_webgpu_init_memset_pipeline(webgpu_global_context & ctx) {
     // we use the maximum workgroup size for the memset pipeline
-    size_t max_threads = ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup * ctx->capabilities.limits.maxComputeWorkgroupsPerDimension;
+    size_t max_threads = ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup *
+                         ctx->capabilities.limits.maxComputeWorkgroupsPerDimension;
     // Size the bytes_per_thread so that the largest buffer size can be handled
     ctx->capabilities.memset_bytes_per_thread =
         CEIL_DIV(ctx->capabilities.limits.maxStorageBufferBindingSize, max_threads);
@@ -4228,9 +4253,9 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
                 const uint32_t q_tile =
                     use_subgroup_matrix ? capabilities.sg_mat_m : GGML_WEBGPU_FLASH_ATTN_TILE_Q_TILE;
                 const uint32_t kv_granularity = use_subgroup_matrix ? capabilities.sg_mat_n : 1u;
-                const bool     kv_direct = use_subgroup_matrix ?
-                                               ggml_webgpu_flash_attn_kv_direct(src0, src1, src2, capabilities.sg_mat_k) :
-                                               false;
+                const bool kv_direct = use_subgroup_matrix ?
+                                           ggml_webgpu_flash_attn_kv_direct(src0, src1, src2, capabilities.sg_mat_k) :
+                                           false;
                 const uint32_t max_kv_tile = ggml_webgpu_flash_attn_max_kv_tile(
                     capabilities.limits.maxComputeWorkgroupStorageSize, q_tile, kv_granularity, (uint32_t) src0->ne[0],
                     (uint32_t) src2->ne[0], op->src[3] != nullptr, kv_direct);

ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl

@@ -130,10 +130,13 @@ fn update(dst_i: u32, src0_i: u32, src1_i: u32) {
 }
 
 @compute @workgroup_size(WG_SIZE)
-fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-    if (gid.x < params.ne) {
-        let src0_i = params.offset_src0 + src0_index(gid.x);
-        let src1_i = params.offset_src1 + src1_index(gid.x);
-        update(params.offset_dst + gid.x, src0_i, src1_i);
+fn main(@builtin(global_invocation_id) gid: vec3<u32>,
+    @builtin(num_workgroups)       num_wg:  vec3<u32>) {
+    let threads_per_group = u32(WG_SIZE);
+    let i = gid.x + (num_wg.x * threads_per_group) * gid.y;
+    if (i < params.ne) {
+        let src0_i = params.offset_src0 + src0_index(i);
+        let src1_i = params.offset_src1 + src1_index(i);
+        update(params.offset_dst + i, src0_i, src1_i);
     }
 }

ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl

@@ -31,6 +31,16 @@ struct Params {
 #define DataType i32
 #endif
 
+#ifdef SRC_OVERLAP
+@group(0) @binding(0)
+var<storage, read_write> merged_src: array<DataType>;
+
+@group(0) @binding(1)
+var<storage, read_write> dst: array<DataType>;
+
+@group(0) @binding(2)
+var<uniform> params: Params;
+#else
 @group(0) @binding(0)
 var<storage, read_write> src0: array<DataType>;
 
@@ -42,7 +52,7 @@ var<storage, read_write> dst: array<DataType>;
 
 @group(0) @binding(3)
 var<uniform> params: Params;
-
+#endif
 @compute @workgroup_size(WG_SIZE)
 fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
 
@@ -62,14 +72,22 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
                              ni[1] * params.stride_src0_1 +
                              ni[2] * params.stride_src0_2 +
                              ni[3] * params.stride_src0_3;
+#ifdef SRC_OVERLAP
+            dst[params.offset_dst + gid.x] = merged_src[params.offset_src0 + src_i];
+#else
             dst[params.offset_dst + gid.x] = src0[params.offset_src0 + src_i];
+#endif
         } else {
             ni[params.dim] -= params.src0_nedim;
             let src_i = ni[0] * params.stride_src1_0 +
                              ni[1] * params.stride_src1_1 +
                              ni[2] * params.stride_src1_2 +
                              ni[3] * params.stride_src1_3;
+#ifdef SRC_OVERLAP
+            dst[params.offset_dst + gid.x] = merged_src[params.offset_src1 + src_i];
+#else
             dst[params.offset_dst + gid.x] = src1[params.offset_src1 + src_i];
+#endif
         }
     }
 }

ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl

@@ -63,10 +63,10 @@ fn main(
     let iq3 = seq_id / params.rq3;
 
     let state_size = S_V * S_V;
-    let state_in_base = (seq_id * params.K * params.h + head_id) * state_size;
+    // input state holds s0 only [S_v, S_v, H, n_seqs]: per-seq stride is H*D.
+    let state_in_base = (seq_id * params.h + head_id) * state_size;
     let state_out_base = (seq_id * params.h + head_id) * state_size;
     let state_size_per_snap = state_size * params.h * params.n_seqs;
-    let shift = i32(params.n_tokens) - i32(params.K);
 
     var state: array<f32, S_V>;
     for (var i = 0u; i < S_V; i++) {
@@ -128,7 +128,8 @@ fn main(
         attn_off += S_V * params.h;
 
         if (params.K > 1u) {
-            let target_slot = i32(t) - shift;
+            // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+            let target_slot = i32(params.n_tokens) - 1 - i32(t);
             if (target_slot >= 0 && target_slot < i32(params.K)) {
                 let slot_base = params.s_off + u32(target_slot) * state_size_per_snap + state_out_base;
                 for (var i = 0u; i < S_V; i++) {

ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl

@@ -98,72 +98,50 @@ fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u3
 }
 #endif // INIT_SRC0_SHMEM_Q1_0
 
-#ifdef INIT_SRC0_SHMEM_Q4_0
+#if defined(INIT_SRC0_SHMEM_Q4_0) || defined(INIT_SRC0_SHMEM_Q4_1) || defined(INIT_SRC0_SHMEM_Q5_0) || defined(INIT_SRC0_SHMEM_Q5_1) || defined(INIT_SRC0_SHMEM_Q8_0) || defined(INIT_SRC0_SHMEM_Q8_1) || defined(INIT_SRC0_SHMEM_MXFP4)
 const BLOCK_SIZE = 32u;
-const BLOCK_SIZE_BYTES = 18u;
 // the number of blocks per k-tile. Note that this currently only works if TILE_K is a multiple of BLOCK_SIZE, which may need to be rethought for larger quantized types.
 override BLOCKS_K = TILE_K/BLOCK_SIZE;
 const NQ = 16u;
+#if defined(INIT_SRC0_SHMEM_Q8_0) || defined(INIT_SRC0_SHMEM_Q8_1)
+const BYTES_PER_THREAD = 16u; // NQ(16) weights use 16 bytes of q
+#else
 const BYTES_PER_THREAD = 8u; // NQ(16) weights use 8 bytes of q
+#endif
 const BYTES_PER_INNER_LOOP = 4u; // == sizeof(q_packed)
 
 fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
     for (var i = thread_id * NQ; i < TILE_SRC0_SHMEM; i += TOTAL_WORKGROUP_SIZE * NQ) {
-        let blck_idx = i / BLOCK_SIZE;
+        let block_idx = i / BLOCK_SIZE;
         let block_offset = (i % BLOCK_SIZE) / NQ;
-        let shmem_idx = blck_idx * BLOCK_SIZE + block_offset * BYTES_PER_THREAD;
+        let shmem_idx = block_idx * BLOCK_SIZE + block_offset * BYTES_PER_THREAD;
 
-        let tile_m = blck_idx / BLOCKS_K;
+        let tile_m = block_idx / BLOCKS_K;
         let global_m = offset_m + tile_m;
-        let block_k = blck_idx % BLOCKS_K;
+        let block_k = block_idx % BLOCKS_K;
         let global_block_k = k_outer / BLOCK_SIZE + block_k;
 
         if (global_m < params.m && global_block_k < params.k / BLOCK_SIZE) {
             let src0_idx = batch_offset + global_m * params.stride_01 + global_block_k;
-            let block_byte_base = src0_idx * BLOCK_SIZE_BYTES;
+
+#ifdef INIT_SRC0_SHMEM_Q4_0
+            let block_byte_base = src0_idx * 18u; // BLOCK_SIZE_BYTES = 18u;
             let d = load_f16_at_src0(block_byte_base);
 
-            // store NQ(16) weights
+            // load NQ(16) weights
             for (var j = 0u; j < BYTES_PER_THREAD / BYTES_PER_INNER_LOOP; j += 1) {
-
                 let q_byte_offset = block_byte_base + 2u + block_offset * BYTES_PER_THREAD + j * BYTES_PER_INNER_LOOP;
                 let q_packed = load_u32_at_src0(q_byte_offset);
                 dequant_q4_0_packed_to_shmem(q_packed, d, shmem_idx + j * BYTES_PER_INNER_LOOP);
             }
-        }
-    }
-}
-#endif // INIT_SRC0_SHMEM_Q4_0
+#elif INIT_SRC0_SHMEM_Q4_1
+            let block_byte_base = src0_idx * 20u; // BLOCK_SIZE_BYTES = 20u;
+            let dm = unpack2x16float(load_u32_at_src0_aligned(block_byte_base));
+            let d = f16(dm[0]);
+            let m = f16(dm[1]);
 
-#ifdef INIT_SRC0_SHMEM_Q4_1
-const BLOCK_SIZE = 32u;
-const BLOCK_SIZE_BYTES = 20u;
-// the number of blocks per k-tile. Note that this currently only works if TILE_K is a multiple of BLOCK_SIZE, which may need to be rethought for larger quantized types.
-override BLOCKS_K = TILE_K/BLOCK_SIZE;
-const NQ = 16u;
-const BYTES_PER_THREAD = 8u; // NQ(16) weights use 8 bytes of q
-const BYTES_PER_INNER_LOOP = 4u; // == sizeof(q_packed)
-
-fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
-    for (var i = thread_id * NQ; i < TILE_SRC0_SHMEM; i += TOTAL_WORKGROUP_SIZE * NQ) {
-        let blck_idx = i / BLOCK_SIZE;
-        let block_offset = (i % BLOCK_SIZE) / NQ;
-        let shmem_idx = blck_idx * BLOCK_SIZE + block_offset * BYTES_PER_THREAD;
-
-        let tile_m = blck_idx / BLOCKS_K;
-        let global_m = offset_m + tile_m;
-        let block_k = blck_idx % BLOCKS_K;
-        let global_block_k = k_outer / BLOCK_SIZE + block_k;
-
-        if (global_m < params.m && global_block_k < params.k / BLOCK_SIZE) {
-            let src0_idx = batch_offset + global_m * params.stride_01 + global_block_k;
-            let block_byte_base = src0_idx * BLOCK_SIZE_BYTES;
-            let d = load_f16_at_src0(block_byte_base);
-            let m = load_f16_at_src0(block_byte_base + 2u);
-
-            // store NQ(16) weights
+            // load NQ(16) weights
             for (var j = 0u; j < BYTES_PER_THREAD / BYTES_PER_INNER_LOOP; j += 1) {
-
                 let q_byte_offset = block_byte_base + 4u + block_offset * BYTES_PER_THREAD + j * BYTES_PER_INNER_LOOP;
                 let q_packed = load_u32_at_src0(q_byte_offset);
 
@@ -175,41 +153,13 @@ fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u3
                     shmem[shmem_idx + j * BYTES_PER_INNER_LOOP + k + 16u] = q_hi;
                 }
             }
-        }
-    }
-}
-#endif // INIT_SRC0_SHMEM_Q4_1
-
-#ifdef INIT_SRC0_SHMEM_Q5_0
-const BLOCK_SIZE = 32u;
-const BLOCK_SIZE_BYTES = 22u;
-// the number of blocks per k-tile. Note that this currently only works if TILE_K is a multiple of BLOCK_SIZE, which may need to be rethought for larger quantized types.
-// tile_k is defined as 32u, so blocks_k ends up being 1 always
-override BLOCKS_K = TILE_K / BLOCK_SIZE;
-const NQ = 16u;
-const BYTES_PER_THREAD = 8u; // NQ(16) weights use 8 bytes of q
-const BYTES_PER_INNER_LOOP = 4u; // == sizeof(q_packed)
-
-fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
-
-    for (var i = thread_id * NQ; i < TILE_SRC0_SHMEM; i += TOTAL_WORKGROUP_SIZE * NQ) {
-        let blck_idx    = i / BLOCK_SIZE;
-        let block_offset = (i % BLOCK_SIZE) / NQ;
-        let shmem_idx   = blck_idx * BLOCK_SIZE + block_offset * BYTES_PER_THREAD;
-
-        let tile_m   = blck_idx / BLOCKS_K;
-        let global_m = offset_m + tile_m;
-        let block_k  = blck_idx % BLOCKS_K;
-        let global_block_k = k_outer / BLOCK_SIZE + block_k;
-
-        if (global_m < params.m && global_block_k < params.k / BLOCK_SIZE) {
-            let src0_idx  = batch_offset + global_m * params.stride_01 + global_block_k;
-            let block_byte_base = src0_idx * BLOCK_SIZE_BYTES;
+#elif INIT_SRC0_SHMEM_Q5_0
+            let block_byte_base = src0_idx * 22u; // BLOCK_SIZE_BYTES = 22u;
 
             let d  = load_f16_at_src0(block_byte_base);
             let qh_packed = load_u32_at_src0(block_byte_base + 2u);
 
-            // store NQ(16) weights
+            // load NQ(16) weights
             for (var j = 0u; j < BYTES_PER_THREAD / BYTES_PER_INNER_LOOP; j += 1) {
                 let q_byte_offset = block_byte_base + 6u + block_offset * BYTES_PER_THREAD + j * BYTES_PER_INNER_LOOP;
                 let q_packed = load_u32_at_src0(q_byte_offset);
@@ -226,44 +176,18 @@ fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u3
                     shmem[shmem_idx + j * BYTES_PER_INNER_LOOP + k + 16u] = q_hi;
                 }
             }
-        }
-    }
-}
-#endif // INIT_SRC0_SHMEM_Q5_0
+#elif INIT_SRC0_SHMEM_Q5_1
+            let block_byte_base = src0_idx * 24u; // BLOCK_SIZE_BYTES = 24u;
 
-#ifdef INIT_SRC0_SHMEM_Q5_1
-const BLOCK_SIZE = 32u;
-const BLOCK_SIZE_BYTES = 24u;
-// the number of blocks per k-tile. Note that this currently only works if TILE_K is a multiple of BLOCK_SIZE, which may need to be rethought for larger quantized types.
-override BLOCKS_K = TILE_K / BLOCK_SIZE;
-const NQ = 16u;
-const BYTES_PER_THREAD = 8u; // NQ(16) weights use 8 bytes of q
-const BYTES_PER_INNER_LOOP = 4u; // == sizeof(q_packed)
+            let dm = unpack2x16float(load_u32_at_src0_aligned(block_byte_base));
+            let d  = f16(dm[0]);
+            let m = f16(dm[1]);
+            let qh_packed = load_u32_at_src0_aligned(block_byte_base + 4u);
 
-fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
-
-    for (var i = thread_id * NQ; i < TILE_SRC0_SHMEM; i += TOTAL_WORKGROUP_SIZE * NQ) {
-        let blck_idx    = i / BLOCK_SIZE;
-        let block_offset = (i % BLOCK_SIZE) / NQ;
-        let shmem_idx   = blck_idx * BLOCK_SIZE + block_offset * BYTES_PER_THREAD;
-
-        let tile_m   = blck_idx / BLOCKS_K;
-        let global_m = offset_m + tile_m;
-        let block_k  = blck_idx % BLOCKS_K;
-        let global_block_k = k_outer / BLOCK_SIZE + block_k;
-
-        if (global_m < params.m && global_block_k < params.k / BLOCK_SIZE) {
-            let src0_idx  = batch_offset + global_m * params.stride_01 + global_block_k;
-            let block_byte_base = src0_idx * BLOCK_SIZE_BYTES;
-
-            let d  = load_f16_at_src0(block_byte_base);
-            let m = load_f16_at_src0(block_byte_base + 2u);
-            let qh_packed = load_u32_at_src0(block_byte_base + 4u);
-
-            // store NQ(16) weights
+            // load NQ(16) weights
             for (var j = 0u; j < BYTES_PER_THREAD / BYTES_PER_INNER_LOOP; j += 1) {
                 let q_byte_offset = block_byte_base + 8u + block_offset * BYTES_PER_THREAD + j * BYTES_PER_INNER_LOOP;
-                let q_packed = load_u32_at_src0(q_byte_offset);
+                let q_packed = load_u32_at_src0_aligned(q_byte_offset);
 
                 for (var k = 0u; k < BYTES_PER_INNER_LOOP; k++) {
                     let q_byte = get_byte(q_packed, k);
@@ -277,461 +201,306 @@ fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u3
                     shmem[shmem_idx + j * BYTES_PER_INNER_LOOP + k + 16u] = q_hi;
                 }
             }
-        }
-    }
-}
-#endif // INIT_SRC0_SHMEM_Q5_1
-
-#ifdef INIT_SRC0_SHMEM_Q8_0
-const BLOCK_SIZE = 32u;
-const BLOCK_SIZE_BYTES = 34u;
-// the number of blocks per k-tile. Note that this currently only works if TILE_K is a multiple of BLOCK_SIZE, which may need to be rethought for larger quantized types.
-override BLOCKS_K = TILE_K/BLOCK_SIZE;
-const NQ = 16u;
-const BYTES_PER_THREAD = 16u; // NQ(16) weights use 16 bytes of q
-const BYTES_PER_INNER_LOOP = 4u; // == sizeof(q_packed)
-
-fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
-    for (var i = thread_id * NQ; i < TILE_SRC0_SHMEM; i += TOTAL_WORKGROUP_SIZE * NQ) {
-        let blck_idx = i / BLOCK_SIZE;
-        let block_offset = (i % BLOCK_SIZE) / NQ;
-        let shmem_idx = blck_idx * BLOCK_SIZE + block_offset * BYTES_PER_THREAD;
-
-        let tile_m = blck_idx / BLOCKS_K;
-        let global_m = offset_m + tile_m;
-        let block_k = blck_idx % BLOCKS_K;
-        let global_block_k = k_outer / BLOCK_SIZE + block_k;
-
-        if (global_m < params.m && global_block_k < params.k / BLOCK_SIZE) {
-            let src0_idx = batch_offset + global_m * params.stride_01 + global_block_k;
-            let block_byte_base = src0_idx * BLOCK_SIZE_BYTES;
+#elif INIT_SRC0_SHMEM_Q8_0
+            let block_byte_base = src0_idx * 34u; // BLOCK_SIZE_BYTES = 34u;
             let d = load_f16_at_src0(block_byte_base);
 
-            // store NQ(16) weights
+            // load NQ(16) weights
             for (var j = 0u; j < BYTES_PER_THREAD / BYTES_PER_INNER_LOOP; j += 1) {
                 let q_byte_offset = block_byte_base + 2u + block_offset * BYTES_PER_THREAD + j * BYTES_PER_INNER_LOOP;
                 let q_packed = load_u32_at_src0(q_byte_offset);
                 dequant_q8_0_packed_to_shmem(q_packed, d, shmem_idx + j * BYTES_PER_INNER_LOOP);
             }
-        }
-    }
-}
-#endif // INIT_SRC0_SHMEM_Q8_0
+#elif INIT_SRC0_SHMEM_Q8_1
+            let block_byte_base = src0_idx * 36u; // BLOCK_SIZE_BYTES = 36u;
+            let dm = unpack2x16float(load_u32_at_src0_aligned(block_byte_base));
+            let d = f16(dm[0]);
+            let m = f16(dm[1]);
 
-#ifdef INIT_SRC0_SHMEM_Q8_1
-const BLOCK_SIZE = 32u;
-const BLOCK_SIZE_BYTES = 36u;
-// the number of blocks per k-tile. Note that this currently only works if TILE_K is a multiple of BLOCK_SIZE, which may need to be rethought for larger quantized types.
-override BLOCKS_K = TILE_K/BLOCK_SIZE;
-const NQ = 16u;
-const BYTES_PER_THREAD = 16u; // NQ(16) weights use 16 bytes of q
-const BYTES_PER_INNER_LOOP = 4u; // == sizeof(q_packed)
-
-fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
-    for (var i = thread_id * NQ; i < TILE_SRC0_SHMEM; i += TOTAL_WORKGROUP_SIZE * NQ) {
-        let blck_idx = i / BLOCK_SIZE;
-        let block_offset = (i % BLOCK_SIZE) / NQ;
-        let shmem_idx = blck_idx * BLOCK_SIZE + block_offset * BYTES_PER_THREAD;
-
-        let tile_m = blck_idx / BLOCKS_K;
-        let global_m = offset_m + tile_m;
-        let block_k = blck_idx % BLOCKS_K;
-        let global_block_k = k_outer / BLOCK_SIZE + block_k;
-
-        if (global_m < params.m && global_block_k < params.k / BLOCK_SIZE) {
-            let src0_idx = batch_offset + global_m * params.stride_01 + global_block_k;
-            let block_byte_base = src0_idx * BLOCK_SIZE_BYTES;
-            let d = load_f16_at_src0(block_byte_base);
-            let m = load_f16_at_src0(block_byte_base + 2u);
-
-            // store NQ(16) weights
+            // load NQ(16) weights
             for (var j = 0u; j < BYTES_PER_THREAD / BYTES_PER_INNER_LOOP; j += 1) {
                 let q_byte_offset = block_byte_base + 4u + block_offset * BYTES_PER_THREAD + j * BYTES_PER_INNER_LOOP;
                 let q_packed = load_u32_at_src0(q_byte_offset);
                 for (var k = 0u; k < BYTES_PER_INNER_LOOP; k++) {
                     let q_byte = get_byte_i32(q_packed, k);
-
                     let q_val = f16(q_byte) * d + m;
                     shmem[shmem_idx + j * BYTES_PER_INNER_LOOP + k] = q_val;
                 }
             }
+#elif INIT_SRC0_SHMEM_MXFP4
+            let block_byte_base = src0_idx * 17u;
+            let eu8 = get_byte(load_u32_at_src0_aligned(block_byte_base), block_byte_base & 3u);
+            let e = ldexp(1.0, i32(eu8) - 128);
+
+            // load NQ(16) weights
+            for (var j = 0u; j < BYTES_PER_THREAD / BYTES_PER_INNER_LOOP; j += 1) {
+                let q_byte_offset = block_byte_base + 1u + block_offset * BYTES_PER_THREAD + j * BYTES_PER_INNER_LOOP;
+                let q_packed = load_u32_at_src0(q_byte_offset);
+                for (var k = 0u; k < BYTES_PER_INNER_LOOP; k++) {
+                    let q_byte = get_byte(q_packed, k);
+                    let q_hi = f32(kvalues_mxfp4[(q_byte >> 4) & 0xF]) * e;
+                    let q_lo = f32(kvalues_mxfp4[q_byte & 0xF]) * e;
+                    shmem[shmem_idx + j * BYTES_PER_INNER_LOOP + k] = f16(q_lo);
+                    shmem[shmem_idx + j * BYTES_PER_INNER_LOOP + k + 16u] = f16(q_hi);
+                }
+            }
+#endif
         }
     }
 }
-#endif // INIT_SRC0_SHMEM_Q8_1
+#endif
+
+// k-quants
+#if defined(INIT_SRC0_SHMEM_Q2_K) || defined(INIT_SRC0_SHMEM_Q3_K) || defined(INIT_SRC0_SHMEM_Q4_K) || defined(INIT_SRC0_SHMEM_Q5_K) || defined(INIT_SRC0_SHMEM_Q6_K)
+const BLOCK_SIZE = 256u;
+const NQ = 4u;
+
+fn store_shmem_kquants(val: vec4<f16>, idx: u32) {
+    shmem[idx] = val.x;
+    shmem[idx + 1] = val.y;
+    shmem[idx + 2] = val.z;
+    shmem[idx + 3] = val.w;
+}
+
+fn load_byte_at_src0_aligned(byte_offset: u32) -> u32 {
+    return get_byte(load_u32_at_src0_aligned(byte_offset), byte_offset % 4u);
+}
+
+fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
+    for (var elem_idx = thread_id * NQ; elem_idx < TILE_SRC0_SHMEM; elem_idx += TOTAL_WORKGROUP_SIZE * NQ) {
+        let tile_m = elem_idx / TILE_K;
+        let tile_k = elem_idx % TILE_K;
+
+        let global_m = offset_m + tile_m;
+        let global_k = k_outer + tile_k;
+
+        if (global_m >= params.m || global_k >= params.k) {
+            store_shmem_kquants(vec4<f16>(f16(0.0), f16(0.0), f16(0.0), f16(0.0)), elem_idx);
+            continue;
+        }
+
+        let block_k    = global_k / BLOCK_SIZE;
+        let k_in_block = global_k % BLOCK_SIZE; // k_in_block % 4 == 0;
+
+        let src0_idx = batch_offset + global_m * params.stride_01 + block_k;
 
 #ifdef INIT_SRC0_SHMEM_Q2_K
-const BLOCK_SIZE = 256u;
-const BLOCK_SIZE_BYTES = 84u;
+        let block_byte_base  = src0_idx * 84u; // BLOCK_SIZE_BYTES =  84u;
+        let scales_byte_base = block_byte_base;
+        let qs_byte_base     = block_byte_base + 16u;
+        let dm_byte_base     = block_byte_base + 80u;
 
-fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
-    // Use standard thread layout instead of lane/row_group
-    for (var elem_idx = thread_id; elem_idx < TILE_SRC0_SHMEM; elem_idx += TOTAL_WORKGROUP_SIZE) {
-        let tile_m = elem_idx / TILE_K;
-        let tile_k = elem_idx % TILE_K;
+        let d_packed = unpack2x16float(load_u32_at_src0_aligned(dm_byte_base));
+        let d        = f16(d_packed[0]);
+        let dmin     = f16(d_packed[1]);
 
-        let global_m = offset_m + tile_m;
-        let global_k = k_outer + tile_k;
+        let chunk        = k_in_block / 128u;
+        let pos_in_chunk = k_in_block % 32u;
+        let sub_block    = k_in_block / 16u;
+        let shift_phase  = (k_in_block % 128u) / 32u;
 
-        if (global_m >= params.m || global_k >= params.k) {
-            shmem[elem_idx] = f16(0.0);
-            continue;
-        }
+        // whole 2 bits (4 elems)
+        let qs_word = load_u32_at_src0_aligned(qs_byte_base + 32u * chunk + 1u * pos_in_chunk);
+        let qs_vec4 = vec4<f16>(
+            f16((qs_word >> (2u * shift_phase +  0u)) & 0x3u),
+            f16((qs_word >> (2u * shift_phase +  8u)) & 0x3u),
+            f16((qs_word >> (2u * shift_phase + 16u)) & 0x3u),
+            f16((qs_word >> (2u * shift_phase + 24u)) & 0x3u),
+        );
 
-        let block_k = global_k / BLOCK_SIZE;
-        let k_in_block = global_k % BLOCK_SIZE;
+        let scale = load_byte_at_src0_aligned(scales_byte_base + sub_block);
 
-        let src0_idx = batch_offset + global_m * params.stride_01 + block_k;
-        let block_byte_base = src0_idx * BLOCK_SIZE_BYTES;
+        let dl = d * f16(scale & 0xFu);
+        let ml = dmin * f16(scale >> 4u);
 
-        let d = load_f16_at_src0(block_byte_base + 80u);
-        let dmin = load_f16_at_src0(block_byte_base + 82u);
+        store_shmem_kquants(qs_vec4 * dl - ml, elem_idx);
+#elif INIT_SRC0_SHMEM_Q3_K
+        let block_byte_base  = src0_idx * 110u; // BLOCK_SIZE_BYTES = 110u;
+        let hmask_byte_base  = block_byte_base +  0u;
+        let qs_byte_base     = block_byte_base + 32u;
+        let scales_byte_base = block_byte_base + 96u;
 
-        // Decode the element at position k_in_block
-        let block_of_32 = k_in_block / 32u;
-        let pos_in_32 = k_in_block % 32u;
+        let d_all = load_f16_at_src0(block_byte_base + 108u);
 
-        let q_b_idx = (block_of_32 / 4u) * 32u;
-        let shift = (block_of_32 % 4u) * 2u;
-        let k = (pos_in_32 / 16u) * 16u;
-        let l = pos_in_32 % 16u;
+        let chunk        = k_in_block / 128u;
+        let pos_in_chunk = k_in_block % 32u;
+        let sub_block    = k_in_block / 16u;
+        let shift_phase  = (k_in_block % 128u) / 32u;
 
-        let is = k_in_block / 16u;
+        let hmask_block       = pos_in_chunk;
+        let hmask_shift_phase = k_in_block / 32u;
 
-        let sc_packed = load_u32_at_src0(block_byte_base + 4u * (is / 4u));
-        let sc = get_byte(sc_packed, is % 4u);
+        // low 2 bits (4 elems)
+        let q_lo2_word = load_u32_at_src0(qs_byte_base + 32u * chunk + 1u * hmask_block);
+        let q_lo2_vec4 = vec4<f16>(
+            f16((q_lo2_word >> (2u * shift_phase +  0u)) & 3u),
+            f16((q_lo2_word >> (2u * shift_phase +  8u)) & 3u),
+            f16((q_lo2_word >> (2u * shift_phase + 16u)) & 3u),
+            f16((q_lo2_word >> (2u * shift_phase + 24u)) & 3u)
+        );
 
-        let dl = d * f16(sc & 0xFu);
-        let ml = dmin * f16(sc >> 4u);
+        // high 1 bit (4 elems)
+        let q_hi1_word = load_u32_at_src0(hmask_byte_base + pos_in_chunk);
+        let q_hi1_vec4 = vec4<f16>(
+            f16(select(4.0, 0.0, ((q_hi1_word >> (1u * hmask_shift_phase +  0u)) & 1u) == 1u)),
+            f16(select(4.0, 0.0, ((q_hi1_word >> (1u * hmask_shift_phase +  8u)) & 1u) == 1u)),
+            f16(select(4.0, 0.0, ((q_hi1_word >> (1u * hmask_shift_phase + 16u)) & 1u) == 1u)),
+            f16(select(4.0, 0.0, ((q_hi1_word >> (1u * hmask_shift_phase + 24u)) & 1u) == 1u))
+        );
 
-        let q_idx = q_b_idx + k + l;
-        let q_packed = load_u32_at_src0(block_byte_base + 16u + 4u * (q_idx / 4u));
-        let q_byte = get_byte(q_packed, q_idx % 4u);
-        let qs_val = (q_byte >> shift) & 3u;
+        let q_vec4 = q_lo2_vec4 - q_hi1_vec4;
 
-        let q_val = f16(qs_val) * dl - ml;
-        shmem[elem_idx] = q_val;
-    }
-}
-#endif // INIT_SRC0_SHMEM_Q2_K
+        let scale_low4 = (load_byte_at_src0_aligned(scales_byte_base + (sub_block % 8u)) >> (4u * (sub_block / 8u))) & 0xFu;
+        let scale_hi2  = (load_byte_at_src0_aligned(scales_byte_base + 8u + (sub_block % 4u)) >> (2u * (sub_block / 4u))) & 3u;
+        let dl         = d_all * (f16((scale_hi2 << 4u) | scale_low4) - 32.0);
 
-#ifdef INIT_SRC0_SHMEM_Q3_K
-const BLOCK_SIZE = 256u;
-const BLOCK_SIZE_BYTES = 110u;
+        store_shmem_kquants(dl * q_vec4, elem_idx);
+#elif INIT_SRC0_SHMEM_Q4_K
+        let block_byte_base = src0_idx * 144u; // BLOCK_SIZE_BYTES = 144u;
+        let dm_byte_base    = block_byte_base +  0u;
+        let scale_byte_base = block_byte_base +  4u;
+        let qs_byte_base    = block_byte_base + 16u;
 
-fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
-    for (var elem_idx = thread_id; elem_idx < TILE_SRC0_SHMEM; elem_idx += TOTAL_WORKGROUP_SIZE) {
-        let tile_m = elem_idx / TILE_K;
-        let tile_k = elem_idx % TILE_K;
+        let dm   = unpack2x16float(load_u32_at_src0_aligned(dm_byte_base));
+        let d    = f16(dm[0]);
+        let dmin = f16(dm[1]);
 
-        let global_m = offset_m + tile_m;
-        let global_k = k_outer + tile_k;
+        let chunk        = k_in_block / 64u;
+        let pos_in_chunk = (k_in_block % 64u) % 32u;
+        let sub_block    = k_in_block / 32u;
+        let shift_phase  = sub_block & 1u;
 
-        if (global_m >= params.m || global_k >= params.k) {
-            shmem[elem_idx] = f16(0.0);
-            continue;
-        }
-
-        let block_k = global_k / BLOCK_SIZE;
-        let k_in_block = global_k % BLOCK_SIZE;
-
-        let src0_idx = batch_offset + global_m * params.stride_01 + block_k;
-        let block_byte_base = src0_idx * BLOCK_SIZE_BYTES;
-
-        let d = load_f16_at_src0(block_byte_base + 108u);
-
-        // Load and unpack scales
-        let kmask1: u32 = 0x03030303u;
-        let kmask2: u32 = 0x0f0f0f0fu;
-
-        var scale_vals: array<u32, 4>;
-        for (var i: u32 = 0u; i < 4u; i++) {
-            scale_vals[i] = load_u32_at_src0(block_byte_base + 96u + 4u * i);
-        }
-
-        var tmp: u32 = scale_vals[2];
-        scale_vals[2] = ((scale_vals[0] >> 4u) & kmask2) | (((tmp >> 4u) & kmask1) << 4u);
-        scale_vals[3] = ((scale_vals[1] >> 4u) & kmask2) | (((tmp >> 6u) & kmask1) << 4u);
-        scale_vals[0] = (scale_vals[0] & kmask2) | ((tmp & kmask1) << 4u);
-        scale_vals[1] = (scale_vals[1] & kmask2) | (((tmp >> 2u) & kmask1) << 4u);
-
-        // Load hmask and qs arrays
-        var hmask_vals: array<u32, 8>;
-        for (var i: u32 = 0u; i < 8u; i++) {
-            hmask_vals[i] = load_u32_at_src0(block_byte_base + 4u * i);
-        }
-
-        var qs_vals: array<u32, 16>;
-        for (var i: u32 = 0u; i < 16u; i++) {
-            qs_vals[i] = load_u32_at_src0(block_byte_base + 32u + 4u * i);
-        }
-
-        let half = k_in_block / 128u;           // 0 or 1
-        let pos_in_half = k_in_block % 128u;    // 0-127
-        let shift_group = pos_in_half / 32u;    // 0-3
-        let pos_in_32 = pos_in_half % 32u;      // 0-31
-        let k_group = pos_in_32 / 16u;          // 0 or 1
-        let l = pos_in_32 % 16u;                // 0-15
-
-        let q_b_idx = half * 32u;               // 0 or 32
-        let shift = shift_group * 2u;           // 0, 2, 4, 6
-        let k = k_group * 16u;                  // 0 or 16
-        let is = k_in_block / 16u;              // 0-15
-
-        // m increments every 32 elements across entire 256 element block
-        let m_shift = k_in_block / 32u;         // 0-7
-        let m: u32 = 1u << m_shift;             // 1,2,4,8,16,32,64,128
-
-        let sc = get_byte(scale_vals[is / 4u], is % 4u);
-        let dl = d * (f16(sc) - 32.0);
-
-        let q_idx = q_b_idx + k + l;
-        let hm_idx = k + l;
-
-        let q_byte = get_byte(qs_vals[q_idx / 4u], q_idx % 4u);
-        let hmask_byte = get_byte(hmask_vals[hm_idx / 4u], hm_idx % 4u);
-
-        let hm = select(4.0, 0.0, (hmask_byte & m) != 0);
-        let qs_val = (q_byte >> shift) & 3u;
-
-        let q_val = (f16(qs_val) - f16(hm)) * dl;
-        shmem[elem_idx] = q_val;
-    }
-}
-
-#endif // INIT_SRC0_SHMEM_Q3_K
-
-#ifdef INIT_SRC0_SHMEM_Q4_K
-const BLOCK_SIZE = 256u;
-const BLOCK_SIZE_BYTES = 144u;
-
-fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
-    for (var elem_idx = thread_id; elem_idx < TILE_SRC0_SHMEM; elem_idx += TOTAL_WORKGROUP_SIZE) {
-        let tile_m = elem_idx / TILE_K;
-        let tile_k = elem_idx % TILE_K;
-
-        let global_m = offset_m + tile_m;
-        let global_k = k_outer + tile_k;
-
-        if (global_m >= params.m || global_k >= params.k) {
-            shmem[elem_idx] = f16(0.0);
-            continue;
-        }
-
-        let block_k = global_k / BLOCK_SIZE;
-        let k_in_block = global_k % BLOCK_SIZE;
-
-        let src0_idx = batch_offset + global_m * params.stride_01 + block_k;
-        let block_byte_base = src0_idx * BLOCK_SIZE_BYTES;
-
-        let d = load_f16_at_src0(block_byte_base);
-        let dmin = load_f16_at_src0(block_byte_base + 2u);
-
-        // Map k_in_block to loop structure:
-        // Outer loop over 64-element groups (alternating q_b_idx)
-        // Inner loop over 2 shifts per group
-        let group_of_64 = k_in_block / 64u;  // 0-3 (maps to q_b_idx)
-        let pos_in_64 = k_in_block % 64u;    // 0-63
-        let shift_group = pos_in_64 / 32u;   // 0 or 1
-        let l = pos_in_64 % 32u;             // 0-31
-
-        let q_b_idx = group_of_64 * 32u;     // 0, 32, 64, 96
-        let shift = shift_group * 4u;        // 0 or 4
-        let is = k_in_block / 32u;           // 0-7
+        // whole 4 bits (4 elems)
+        let qs_word = load_u32_at_src0_aligned(qs_byte_base + 32u * chunk + 1u * pos_in_chunk);
+        let qs_vec4 = vec4<f16>(
+            f16((qs_word >> (4u * shift_phase +  0u)) & 0xFu),
+            f16((qs_word >> (4u * shift_phase +  8u)) & 0xFu),
+            f16((qs_word >> (4u * shift_phase + 16u)) & 0xFu),
+            f16((qs_word >> (4u * shift_phase + 24u)) & 0xFu)
+        );
 
         var sc: u32;
         var mn: u32;
 
-        let scale_base = block_byte_base + 4u;
-
-        if (is < 4u) {
-            let sc_byte = get_byte(load_u32_at_src0(scale_base), is % 4u);
-            let min_byte = get_byte(load_u32_at_src0(scale_base + 4), is % 4u);
-            sc = sc_byte & 63u;
-            mn = min_byte & 63u;
+        if (sub_block < 4u) {
+            let sc_byte  = get_byte(load_u32_at_src0_aligned(scale_byte_base), sub_block % 4u);
+            let min_byte = get_byte(load_u32_at_src0_aligned(scale_byte_base + 4), sub_block % 4u);
+            sc           = sc_byte & 63u;
+            mn           = min_byte & 63u;
         } else {
-            let sc_min_lo = get_byte(load_u32_at_src0(scale_base + 8), (is + 4u) % 4u);
-            let sc_hi = get_byte(load_u32_at_src0(scale_base), (is - 4u) % 4u);
-            let min_hi = get_byte(load_u32_at_src0(scale_base + 4), is % 4u);
-
-            sc = (sc_min_lo & 0xFu) | ((sc_hi >> 6u) << 4u);
-            mn = (sc_min_lo >> 4u) | ((min_hi >> 6u) << 4u);
+            let sc_min_lo = get_byte(load_u32_at_src0_aligned(scale_byte_base + 8), (sub_block + 4u) % 4u);
+            let sc_hi     = get_byte(load_u32_at_src0_aligned(scale_byte_base), (sub_block - 4u) % 4u);
+            let min_hi    = get_byte(load_u32_at_src0_aligned(scale_byte_base + 4), sub_block % 4u);
+            sc            = (sc_min_lo & 0xFu) | ((sc_hi >> 6u) << 4u);
+            mn            = (sc_min_lo >> 4u) | ((min_hi >> 6u) << 4u);
         }
 
         let dl = d * f16(sc);
         let ml = dmin * f16(mn);
 
-        let q_idx = q_b_idx + l;
-        let q_packed = load_u32_at_src0(block_byte_base + 16u + 4u * (q_idx / 4u));
+        store_shmem_kquants(dl * qs_vec4 - vec4(ml, ml, ml, ml), elem_idx);
+#elif INIT_SRC0_SHMEM_Q5_K
+        let block_byte_base = src0_idx * 176u; // BLOCK_SIZE_BYTES = 176u;
+        let dm_byte_base    = block_byte_base +  0u;
+        let scale_byte_base = block_byte_base +  4u;
+        let qh_byte_base    = block_byte_base + 16u;
+        let qs_byte_base    = block_byte_base + 48u;
 
-        let q_byte = get_byte(q_packed, q_idx % 4u);
-        let qs_val = (q_byte >> shift) & 0xFu;
+        let dm   = unpack2x16float(load_u32_at_src0_aligned(dm_byte_base));
+        let d    = f16(dm[0]);
+        let dmin = f16(dm[1]);
 
-        let q_val = f16(qs_val) * dl - ml;
-        shmem[elem_idx] = q_val;
-    }
-}
-#endif // INIT_SRC0_SHMEM_Q4_K
+        let chunk        = k_in_block / 64u;
+        let pos_in_chunk = (k_in_block % 64u) % 32u;
+        let sub_block    = k_in_block / 32u;
+        let shift_phase  = sub_block & 1u;
 
-#ifdef INIT_SRC0_SHMEM_Q5_K
-const BLOCK_SIZE = 256u;
-const BLOCK_SIZE_BYTES = 176u;
+        let qh_block       = k_in_block % 32u;
+        let qh_shift_phase = sub_block;
 
-fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
-    for (var elem_idx = thread_id; elem_idx < TILE_SRC0_SHMEM; elem_idx += TOTAL_WORKGROUP_SIZE) {
-        let tile_m = elem_idx / TILE_K;
-        let tile_k = elem_idx % TILE_K;
+        // low 4 bits (4 elems)
+        let qs_word     = load_u32_at_src0_aligned(qs_byte_base + 32u * chunk + 1u * pos_in_chunk);
+        let qs_lo4_vec4 = vec4<f16>(
+            f16((qs_word >> (4u * shift_phase +  0u)) & 0xFu),
+            f16((qs_word >> (4u * shift_phase +  8u)) & 0xFu),
+            f16((qs_word >> (4u * shift_phase + 16u)) & 0xFu),
+            f16((qs_word >> (4u * shift_phase + 24u)) & 0xFu)
+        );
 
-        let global_m = offset_m + tile_m;
-        let global_k = k_outer + tile_k;
-
-        if (global_m >= params.m || global_k >= params.k) {
-            shmem[elem_idx] = f16(0.0);
-            continue;
-        }
-
-        let block_k = global_k / BLOCK_SIZE;
-        let k_in_block = global_k % BLOCK_SIZE;
-
-        let src0_idx = batch_offset + global_m * params.stride_01 + block_k;
-        let block_byte_base = src0_idx * BLOCK_SIZE_BYTES;
-
-        let d = load_f16_at_src0(block_byte_base);
-        let dmin = load_f16_at_src0(block_byte_base + 2u);
-
-
-        // The original loop processes elements in groups of 64
-        // Each group of 64: q_b_idx cycles through [0,32,64,96], shift cycles [0,4]
-        // But u increments EVERY 32 elements (after each l loop)
-        let group_of_64 = k_in_block / 64u;  // 0-3
-        let pos_in_64 = k_in_block % 64u;    // 0-63
-        let shift_group = pos_in_64 / 32u;   // 0 or 1
-        let l = pos_in_64 % 32u;             // 0-31
-
-        let q_b_idx = group_of_64 * 32u;     // 0, 32, 64, 96
-        let shift = shift_group * 4u;        // 0 or 4
-        let is = k_in_block / 32u;           // 0-7
-
-        // u increments every 32 elements (0->1, 1->2, 2->4, 3->8, 4->16, 5->32, 6->64, 7->128)
-        let u_shift = k_in_block / 32u;      // 0-7
-        let u: u32 = 1u << u_shift;
+        // high 1 bit (4 elems)
+        let qh_word = load_u32_at_src0_aligned(qh_byte_base + qh_block);
+        let qh_vec4 = vec4<f16>(
+            f16(select(0.0, 16.0, ((qh_word >> (1u * qh_shift_phase +  0u)) & 1u) == 1u)),
+            f16(select(0.0, 16.0, ((qh_word >> (1u * qh_shift_phase +  8u)) & 1u) == 1u)),
+            f16(select(0.0, 16.0, ((qh_word >> (1u * qh_shift_phase + 16u)) & 1u) == 1u)),
+            f16(select(0.0, 16.0, ((qh_word >> (1u * qh_shift_phase + 24u)) & 1u) == 1u))
+        );
 
         var sc: u32;
         var mn: u32;
 
-        let scale_base = block_byte_base + 4u;
-
-        if (is < 4u) {
-            let sc_byte = get_byte(load_u32_at_src0(scale_base), is % 4u);
-            let min_byte = get_byte(load_u32_at_src0(scale_base + 4), is % 4u);
-            sc = sc_byte & 63u;
-            mn = min_byte & 63u;
+        if (sub_block < 4u) {
+            let sc_byte  = get_byte(load_u32_at_src0_aligned(scale_byte_base), sub_block % 4u);
+            let min_byte = get_byte(load_u32_at_src0_aligned(scale_byte_base + 4), sub_block % 4u);
+            sc           = sc_byte & 63u;
+            mn           = min_byte & 63u;
         } else {
-            let sc_min_lo = get_byte(load_u32_at_src0(scale_base + 8), (is + 4u) % 4u);
-            let sc_hi = get_byte(load_u32_at_src0(scale_base), (is - 4u) % 4u);
-            let min_hi = get_byte(load_u32_at_src0(scale_base + 4), is % 4u);
-
-            sc = (sc_min_lo & 0xFu) | ((sc_hi >> 6u) << 4u);
-            mn = (sc_min_lo >> 4u) | ((min_hi >> 6u) << 4u);
+            let sc_min_lo = get_byte(load_u32_at_src0_aligned(scale_byte_base + 8), (sub_block + 4u) % 4u);
+            let sc_hi     = get_byte(load_u32_at_src0_aligned(scale_byte_base), (sub_block - 4u) % 4u);
+            let min_hi    = get_byte(load_u32_at_src0_aligned(scale_byte_base + 4), sub_block % 4u);
+            sc            = (sc_min_lo & 0xFu) | ((sc_hi >> 6u) << 4u);
+            mn            = (sc_min_lo >> 4u) | ((min_hi >> 6u) << 4u);
         }
 
         let dl = d * f16(sc);
         let ml = dmin * f16(mn);
 
-        let q_idx = q_b_idx + l;
-        let q_packed = load_u32_at_src0(block_byte_base + 48u + 4u * (q_idx / 4u));
+        store_shmem_kquants((qh_vec4 + qs_lo4_vec4) * dl - vec4<f16>(ml, ml, ml, ml), elem_idx);
+#elif INIT_SRC0_SHMEM_Q6_K
+        let block_byte_base  = src0_idx * 210u; // BLOCK_SIZE_BYTES = 210u;
+        let ql_byte_base     = block_byte_base;
+        let qh_byte_base     = block_byte_base + 128u;
+        let scales_byte_base = block_byte_base + 192u;
+        let d_byte_base      = block_byte_base + 208u;
 
-        let q_byte = get_byte(q_packed, q_idx % 4u);
+        let d = load_f16_at_src0(d_byte_base);
 
-        let qh_packed = load_u32_at_src0(block_byte_base + 16u + 4u * (l / 4u));
+        let chunk           = k_in_block / 128u;
+        let ql_pos_in_chunk = (k_in_block % 128u) % 64u;
+        let qh_pos_in_chunk = (k_in_block % 128u) % 32u;
+        let sub_block       = k_in_block / 16u;
+        let ql_shift_phase  = (k_in_block % 128u) / 64u;
+        let qh_shift_phase  = (k_in_block % 128u) / 32u;
 
-        let qh_byte = get_byte(qh_packed, l % 4u);
+        // low 4 bits (4 elems)
+        let ql_word     = load_u32_at_src0(ql_byte_base + 64u * chunk + 1u * ql_pos_in_chunk);
+        let ql_lo4_vec4 = vec4<u32>(
+            (ql_word >> (4u * ql_shift_phase +  0u)) & 0xFu,
+            (ql_word >> (4u * ql_shift_phase +  8u)) & 0xFu,
+            (ql_word >> (4u * ql_shift_phase + 16u)) & 0xFu,
+            (ql_word >> (4u * ql_shift_phase + 24u)) & 0xFu
+        );
 
-        let qs_val = (q_byte >> shift) & 0xFu;
-        let qh_val = select(0.0, 16.0, (qh_byte & u) != 0);
+        // hi 2 bits (4 elems)
+        let qh_word     = load_u32_at_src0(qh_byte_base + 32u * chunk + 1u * qh_pos_in_chunk);
+        let qh_hi2_vec4 = vec4<u32>(
+            ((qh_word >> (2u * qh_shift_phase +  0u)) & 0x3u) << 4u,
+            ((qh_word >> (2u * qh_shift_phase +  8u)) & 0x3u) << 4u,
+            ((qh_word >> (2u * qh_shift_phase + 16u)) & 0x3u) << 4u,
+            ((qh_word >> (2u * qh_shift_phase + 24u)) & 0x3u) << 4u,
+        );
 
-        let q_val = (f16(qs_val) + f16(qh_val)) * dl - ml;
-        shmem[elem_idx] = q_val;
+        let q_vec4 = vec4<f16>(qh_hi2_vec4 | ql_lo4_vec4) - vec4<f16>(32.0, 32.0, 32.0, 32.0);
+
+        let scale_byte = scales_byte_base + 1u * sub_block;
+        let scale_word = load_u32_at_src0_aligned(scale_byte);
+        let scale      = get_byte_i32(scale_word, scale_byte & 3u);
+
+        store_shmem_kquants(d * q_vec4 * f16(scale), elem_idx);
+#endif
     }
 }
-
-#endif // INIT_SRC0_SHMEM_Q5_K
-
-#ifdef INIT_SRC0_SHMEM_Q6_K
-const BLOCK_SIZE = 256u;
-const BLOCK_SIZE_BYTES = 210u;
-
-fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
-    for (var elem_idx = thread_id; elem_idx < TILE_SRC0_SHMEM; elem_idx += TOTAL_WORKGROUP_SIZE) {
-        let tile_m = elem_idx / TILE_K;
-        let tile_k = elem_idx % TILE_K;
-
-        let global_m = offset_m + tile_m;
-        let global_k = k_outer + tile_k;
-
-        if (global_m >= params.m || global_k >= params.k) {
-            shmem[elem_idx] = f16(0.0);
-            continue;
-        }
-
-        let block_k = global_k / BLOCK_SIZE;
-        let k_in_block = global_k % BLOCK_SIZE;
-
-        let src0_idx = batch_offset + global_m * params.stride_01 + block_k;
-        let block_byte_base = src0_idx * BLOCK_SIZE_BYTES;
-
-        let half = k_in_block / 128u;
-        let pos_in_half = k_in_block % 128u;
-        let quarter = pos_in_half / 32u;
-        let l = pos_in_half % 32u;
-
-        let ql_b_idx = half * 64u;
-        let qh_b_idx = half * 32u;
-        let sc_b_idx = half * 8u;
-
-        // Load only ql13 word needed
-        let ql13_flat = ql_b_idx + l;
-        let ql13 = load_u32_at_src0(block_byte_base + ql13_flat);
-        let ql13_b = get_byte(ql13, 0u);
-
-        // Load only ql24 word needed
-        let ql24_flat = ql_b_idx + l + 32u;
-        let ql24 = load_u32_at_src0(block_byte_base + ql24_flat);
-        let ql24_b = get_byte(ql24, 0u);
-
-        // Load only qh word needed
-        let qh_flat = qh_b_idx + l;
-        let qh = load_u32_at_src0(block_byte_base + 128u + qh_flat);
-        let qh_b = get_byte(qh, 0u);
-
-        let q1 = f16((ql13_b & 0xFu) | ((qh_b & 3u) << 4u)) - f16(32.0);
-        let q2 = f16((ql24_b & 0xFu) | (((qh_b >> 2u) & 3u) << 4u)) - f16(32.0);
-        let q3 = f16((ql13_b >> 4u) | (((qh_b >> 4u) & 3u) << 4u)) - f16(32.0);
-        let q4 = f16((ql24_b >> 4u) | (((qh_b >> 6u) & 3u) << 4u)) - f16(32.0);
-
-        // Load only the scale word needed
-        let is = l / 16u;
-        let sc_idx = sc_b_idx + is + quarter * 2u;
-        let sc = load_u32_at_src0(block_byte_base + 192u + sc_idx);
-        let sc_val = get_byte_i32(sc, 0u);
-
-        let d = load_f16_at_src0(block_byte_base + 208u);
-
-        var q_val: f16;
-        if (quarter == 0u) {
-            q_val = q1;
-        } else if (quarter == 1u) {
-            q_val = q2;
-        } else if (quarter == 2u) {
-            q_val = q3;
-        } else {
-            q_val = q4;
-        }
-
-        shmem[elem_idx] = d * f16(sc_val) * q_val;
-    }
-}
-#endif // INIT_SRC0_SHMEM_Q6_K
+#endif // k-quants
 
 #ifdef INIT_SRC0_SHMEM_IQ4_NL
 const BLOCK_SIZE = 32u;
@@ -1155,48 +924,3 @@ fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u3
     }
 }
 #endif // INIT_SRC0_SHMEM_IQ3_S
-
-#ifdef INIT_SRC0_SHMEM_MXFP4
-const BLOCK_SIZE = 32u;
-const BLOCK_SIZE_BYTES = 17u;
-// the number of blocks per k-tile. Note that this currently only works if TILE_K is a multiple of BLOCK_SIZE, which may need to be rethought for larger quantized types.
-override BLOCKS_K = TILE_K/BLOCK_SIZE;
-const NQ = 16u;
-const BYTES_PER_THREAD = 8u; // NQ(16) weights uses 8 bytes of q
-const BYTES_PER_INNER_LOOP = 4u; // == sizeof(q_packed)
-
-fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
-    for (var i = thread_id * NQ; i < TILE_SRC0_SHMEM; i += TOTAL_WORKGROUP_SIZE * NQ) {
-        let blck_idx = i / BLOCK_SIZE;
-        let block_offset = (i % BLOCK_SIZE) / NQ;
-        let shmem_idx = blck_idx * BLOCK_SIZE + block_offset * BYTES_PER_THREAD;
-
-        let tile_m = blck_idx / BLOCKS_K;
-        let global_m = offset_m + tile_m;
-        let block_k = blck_idx % BLOCKS_K;
-        let global_block_k = k_outer / BLOCK_SIZE + block_k;
-
-        if (global_m < params.m && global_block_k < params.k / BLOCK_SIZE) {
-            let src0_idx = batch_offset + global_m * params.stride_01 + global_block_k;
-            let block_byte_base = src0_idx * BLOCK_SIZE_BYTES;
-            let eu8 = get_byte(load_u32_at_src0(block_byte_base), 0);
-            let e = ldexp(1.0, i32(eu8) - 128);
-
-            // store NQ(16) weights
-            for (var j = 0u; j < BYTES_PER_THREAD / BYTES_PER_INNER_LOOP; j += 1) {
-
-                let q_byte_offset = block_byte_base + 1u + block_offset * BYTES_PER_THREAD + j * BYTES_PER_INNER_LOOP;
-                let q_packed = load_u32_at_src0(q_byte_offset);
-
-                for (var k = 0u; k < BYTES_PER_INNER_LOOP; k++) {
-                    let q_byte = get_byte(q_packed, k);
-                    let q_hi = f32(kvalues_mxfp4[(q_byte >> 4) & 0xF]) * e;
-                    let q_lo = f32(kvalues_mxfp4[q_byte & 0xF]) * e;
-                    shmem[shmem_idx + j * BYTES_PER_INNER_LOOP + k] = f16(q_lo);
-                    shmem[shmem_idx + j * BYTES_PER_INNER_LOOP + k + 16u] = f16(q_hi);
-                }
-            }
-        }
-    }
-}
-#endif // INIT_SRC0_SHMEM_MXFP4

ggml/src/ggml-webgpu/wgsl-shaders/scale.wgsl

@@ -43,12 +43,14 @@ struct Params {
 var<storage, read_write> src: array<f32>;
 
 @compute @workgroup_size(WG_SIZE)
-fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-    if (gid.x >= params.ne) {
+fn main(
+    @builtin(global_invocation_id) gid: vec3<u32>,
+    @builtin(num_workgroups)       num_wg:  vec3<u32>) {
+    let threads_per_group = u32(WG_SIZE);
+    var i = gid.x + (num_wg.x * threads_per_group) * gid.y;
+    if (i >= params.ne) {
         return;
     }
-
-    var i = gid.x;
     let i3 = i / (params.ne2 * params.ne1 * params.ne0);
     i = i % (params.ne2 * params.ne1 * params.ne0);
     let i2 = i / (params.ne1 * params.ne0);

ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl

@@ -66,11 +66,14 @@ fn erf_approx(x: TYPE) -> TYPE {
 }
 
 @compute @workgroup_size(WG_SIZE)
-fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-    if (gid.x >= params.ne) {
+fn main(@builtin(global_invocation_id) gid: vec3<u32>,
+    @builtin(num_workgroups)       num_wg:  vec3<u32>) {
+    let threads_per_group = u32(WG_SIZE);
+    let flat_i = gid.x + (num_wg.x * threads_per_group) * gid.y;
+    if (flat_i >= params.ne) {
         return;
     }
-    var i = gid.x;
+    var i = flat_i;
     let ne2 = params.ne2;
 #ifdef DIAG
     let ne1 = params.ne0;
@@ -205,6 +208,6 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
 #ifdef INPLACE
     src[params.offset_src + src_idx] = res;
 #else
-    dst[params.offset_dst + gid.x] = res;
+    dst[params.offset_dst + flat_i] = res;
 #endif
 }

ggml/src/ggml.c

@@ -1031,6 +1031,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "IM2COL",
     "IM2COL_BACK",
     "IM2COL_3D",
+    "COL2IM_1D",
     "CONV_2D",
     "CONV_3D",
     "CONV_2D_DW",
@@ -1080,7 +1081,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "GLU",
 };
 
-static_assert(GGML_OP_COUNT == 96, "GGML_OP_COUNT != 96");
+static_assert(GGML_OP_COUNT == 97, "GGML_OP_COUNT != 97");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1141,6 +1142,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "im2col(x)",
     "im2col_back(x)",
     "im2col_3d(x)",
+    "col2im_1d(x)",
     "conv_2d(x)",
     "conv_3d(x)",
     "conv_2d_dw(x)",
@@ -1190,7 +1192,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "glu(x)",
 };
 
-static_assert(GGML_OP_COUNT == 96, "GGML_OP_COUNT != 96");
+static_assert(GGML_OP_COUNT == 97, "GGML_OP_COUNT != 97");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -4541,6 +4543,41 @@ struct ggml_tensor * ggml_conv_1d_dw_ph(
     return ggml_conv_1d_dw(ctx, a, b, s0, a->ne[0] / 2, d0);
 }
 
+// ggml_col2im_1d
+
+struct ggml_tensor * ggml_col2im_1d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   s0,
+        int                   oc,
+        int                   p0) {
+    GGML_ASSERT(ggml_is_matrix(a));
+    GGML_ASSERT(ggml_is_contiguous(a));
+    GGML_ASSERT(a->type == GGML_TYPE_F32 || a->type == GGML_TYPE_F16 || a->type == GGML_TYPE_BF16);
+    GGML_ASSERT(s0 > 0);
+    GGML_ASSERT(oc > 0);
+    GGML_ASSERT(p0 >= 0);
+
+    const int64_t K_OC = a->ne[0];
+    const int64_t T_in = a->ne[1];
+    const int64_t K = K_OC / oc;
+    const int64_t T_out = (T_in - 1) * s0 + K - 2 * p0;
+
+    GGML_ASSERT(K_OC == K * oc);  // a->ne[0] must be a whole number of oc blocks
+    GGML_ASSERT(K > 0 && T_out > 0);
+
+    const int64_t ne[4] = { T_out, oc, 1, 1 };
+    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, 2, ne);
+
+    int32_t params[] = { s0, (int32_t)oc, (int32_t)p0 };
+    ggml_set_op_params(result, params, sizeof(params));
+
+    result->op     = GGML_OP_COL2IM_1D;
+    result->src[0] = a;
+
+    return result;
+}
+
 // ggml_conv_transpose_1d
 
 static int64_t ggml_calc_conv_transpose_1d_output_size(int64_t ins, int64_t ks, int s, int p, int d) {
@@ -6186,7 +6223,8 @@ struct ggml_tensor * ggml_gated_delta_net(
         struct ggml_tensor  * v,
         struct ggml_tensor  * g,
         struct ggml_tensor  * beta,
-        struct ggml_tensor  * state) {
+        struct ggml_tensor  * state,
+        int64_t               K) {
     GGML_ASSERT(ggml_is_contiguous_rows(q));
     GGML_ASSERT(ggml_is_contiguous_rows(k));
     GGML_ASSERT(ggml_is_contiguous_rows(v));
@@ -6210,15 +6248,18 @@ struct ggml_tensor * ggml_gated_delta_net(
     GGML_ASSERT(g->ne[0] == 1 || g->ne[0] == S_v);
     GGML_ASSERT(beta->ne[0] == 1);
 
-    // state is a 3D tensor (S_v*S_v*H, K, n_seqs). K is the snapshot slot count.
-    GGML_ASSERT(state->ne[0] == S_v * S_v * H);
-    GGML_ASSERT(state->ne[2] == n_seqs);
-    GGML_ASSERT(state->ne[3] == 1);
-    const int64_t K = state->ne[1];
+    // state holds the initial state s0 only: [S_v, S_v, H, n_seqs]. K (snapshot slot count) is an op param.
+    GGML_ASSERT(state->ne[0] == S_v);
+    GGML_ASSERT(state->ne[1] == S_v);
+    GGML_ASSERT(state->ne[2] == H);
+    GGML_ASSERT(state->ne[3] == n_seqs);
+    GGML_ASSERT(K >= 1);
     const int64_t state_rows = K * S_v * n_seqs;
     const int64_t ne[4] = { S_v * H, n_tokens * n_seqs + state_rows, 1, 1 };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
+    ggml_set_op_params_i32(result, 0, (int32_t) K);
+
     result->op     = GGML_OP_GATED_DELTA_NET;
     result->src[0] = q;
     result->src[1] = k;

gguf-py/gguf/constants.py

@@ -128,6 +128,7 @@ class Keys:
         MOE_LATENT_SIZE                   = "{arch}.moe_latent_size"
         NEXTN_PREDICT_LAYERS              = "{arch}.nextn_predict_layers"
         NUM_DEEPSTACK_LAYERS              = "{arch}.n_deepstack_layers"
+        DEEPSTACK_MAPPING                 = "{arch}.deepstack_mapping"
         POOLING_TYPE                      = "{arch}.pooling_type"
         LOGIT_SCALE                       = "{arch}.logit_scale"
         DECODER_START_TOKEN_ID            = "{arch}.decoder_start_token_id"
@@ -271,7 +272,8 @@ class Keys:
         CHAT_TEMPLATE_N      = "tokenizer.chat_template.{name}"
         CHAT_TEMPLATES       = "tokenizer.chat_templates"
         # Normalizer constants
-        NORMALIZER_LOWERCASE = "tokenizer.ggml.normalizer.lowercase"
+        NORMALIZER_LOWERCASE     = "tokenizer.ggml.normalizer.lowercase"
+        NORMALIZER_STRIP_ACCENTS = "tokenizer.ggml.normalizer.strip_accents"
         # FIM/Infill special tokens constants
         FIM_PRE_ID           = "tokenizer.ggml.fim_pre_token_id"
         FIM_SUF_ID           = "tokenizer.ggml.fim_suf_token_id"
@@ -325,6 +327,8 @@ class Keys:
         WA_PATTERN_MODE       = "clip.vision.wa_pattern_mode"  # used by mimovl, per-layer -1/0/1
         IS_DEEPSTACK_LAYERS   = "clip.vision.is_deepstack_layers"
         WINDOW_SIZE           = "clip.vision.window_size"
+        FEATURE_LAYERS        = "clip.vision.feature_layer" # Granite4 Vision
+        IMAGE_GRID_PINPOINTS  = "clip.vision.image_grid_pinpoints" # Granite4 Vision
 
         class Attention:
             HEAD_COUNT      = "clip.vision.attention.head_count"
@@ -333,6 +337,9 @@ class Keys:
 
         class Projector:
             SCALE_FACTOR    = "clip.vision.projector.scale_factor"
+            QUERY_SIDE      = "clip.vision.projector.query_side"
+            WINDOW_SIDE     = "clip.vision.projector.window_side"
+            SPATIAL_OFFSETS = "clip.vision.projector.spatial_offsets"
 
         class SAM:
             BLOCK_COUNT         = "clip.vision.sam.block_count"
@@ -434,6 +441,7 @@ class MODEL_ARCH(IntEnum):
     GEMMA3           = auto()
     GEMMA3N          = auto()
     GEMMA4           = auto()
+    GEMMA4_ASSISTANT = auto()
     GEMMA_EMBEDDING  = auto()
     STARCODER2       = auto()
     RWKV6            = auto()
@@ -531,6 +539,8 @@ class VISION_PROJECTOR_TYPE(IntEnum):
 class MODEL_TENSOR(IntEnum):
     TOKEN_EMBD           = auto()
     TOKEN_EMBD_NORM      = auto()
+    MASKED_EMBD_CENTROIDS= auto()
+    MASKED_EMBD_ORDERING = auto()
     TOKEN_TYPES          = auto()
     POS_EMBD             = auto()
     OUTPUT               = auto()
@@ -821,6 +831,31 @@ class MODEL_TENSOR(IntEnum):
     V_RESMPL_QUERY_768   = auto() # Deepseek-OCR-2
     V_RESMPL_QUERY_1024  = auto() # Deepseek-OCR-2
 
+    # qformer projector (vision) - Granite4 Vision
+    V_QF_PROJ_QUERY      = auto()
+    V_QF_PROJ_NORM       = auto()
+    V_QF_PROJ_LINEAR     = auto()
+    V_QF_SELF_ATTN_Q     = auto()
+    V_QF_SELF_ATTN_K     = auto()
+    V_QF_SELF_ATTN_V     = auto()
+    V_QF_SELF_ATTN_O     = auto()
+    V_QF_SELF_ATTN_NORM  = auto()
+    V_QF_CROSS_ATTN_Q    = auto()
+    V_QF_CROSS_ATTN_K    = auto()
+    V_QF_CROSS_ATTN_V    = auto()
+    V_QF_CROSS_ATTN_O    = auto()
+    V_QF_CROSS_ATTN_NORM = auto()
+    V_QF_FFN_UP          = auto()
+    V_QF_FFN_DOWN        = auto()
+    V_QF_FFN_NORM        = auto()
+    V_PROJ_NORM          = auto()
+    # multi-projector (bid => projector id) - Granite4 vision
+    V_MULTI_PROJ_IMG_POS   = auto()
+    V_MULTI_PROJ_QUERY     = auto()
+    V_MULTI_PROJ_NORM      = auto()
+    V_MULTI_PROJ_LINEAR    = auto()
+    V_MULTI_PROJ_POST_NORM = auto()
+
     # audio (mtmd)
     A_ENC_EMBD_POS        = auto()
     A_ENC_EMBD_NORM       = auto()
@@ -866,6 +901,8 @@ class MODEL_TENSOR(IntEnum):
     A_PER_DIM_K_SCALE     = auto() # gemma4
     A_PER_DIM_SCALE       = auto() # gemma4
     # nextn/mtp
+    NEXTN_PROJ_PRE       = auto()
+    NEXTN_PROJ_POST      = auto()
     NEXTN_EH_PROJ        = auto()
     NEXTN_EMBED_TOKENS   = auto()
     NEXTN_ENORM          = auto()
@@ -885,7 +922,7 @@ class MODEL_TENSOR(IntEnum):
     A_CTC_OUT              = auto()
     A_CTC_OUT_MID          = auto()
     A_ENC_ATTN_REL_POS_EMB = auto()
-    # qformer projector
+    # audio qformer projector
     A_QF_PROJ_QUERY        = auto()
     A_QF_PROJ_NORM         = auto()
     A_QF_PROJ_LINEAR       = auto()
@@ -955,6 +992,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.GEMMA3:           "gemma3",
     MODEL_ARCH.GEMMA3N:          "gemma3n",
     MODEL_ARCH.GEMMA4:           "gemma4",
+    MODEL_ARCH.GEMMA4_ASSISTANT: "gemma4-assistant",
     MODEL_ARCH.GEMMA_EMBEDDING:  "gemma-embedding",
     MODEL_ARCH.STARCODER2:       "starcoder2",
     MODEL_ARCH.RWKV6:            "rwkv6",
@@ -1052,6 +1090,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.TOKEN_EMBD:                "token_embd",
     MODEL_TENSOR.TOKEN_EMBD_NORM:           "token_embd_norm",
     MODEL_TENSOR.TOKEN_TYPES:               "token_types",
+    MODEL_TENSOR.MASKED_EMBD_CENTROIDS:     "masked_embd_centroids",
+    MODEL_TENSOR.MASKED_EMBD_ORDERING:      "masked_embd_ordering",
     MODEL_TENSOR.POS_EMBD:                  "position_embd",
     MODEL_TENSOR.OUTPUT_NORM:               "output_norm",
     MODEL_TENSOR.OUTPUT:                    "output",
@@ -1337,10 +1377,33 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_SAM_NECK:                "v.sam.neck.{bid}",
     MODEL_TENSOR.V_SAM_NET_2:               "v.sam.net_2",
     MODEL_TENSOR.V_SAM_NET_3:               "v.sam.net_3",
-    MODEL_TENSOR.V_ENC_EMBD_IMGNL:          "v.image_newline", # Deepseek-OCR
+    MODEL_TENSOR.V_ENC_EMBD_IMGNL:          "v.image_newline", # Deepseek-OCR, Granite4Vision
     MODEL_TENSOR.V_ENC_EMBD_VSEP:           "v.view_seperator", # Deepseek-OCR
     MODEL_TENSOR.V_RESMPL_QUERY_768:        "v.resample_query_768", # Deepseek-OCR-2 qwen2
     MODEL_TENSOR.V_RESMPL_QUERY_1024:       "v.resample_query_1024", # Deepseek-OCR-2 qwen2
+    # Granite4 Vision
+    # qformer layers (bid => proj_id)
+    # NOTE: Names align with A_QF_*
+    MODEL_TENSOR.V_QF_SELF_ATTN_Q:          "v.proj_blk.{bid}.self_attn_q",
+    MODEL_TENSOR.V_QF_SELF_ATTN_K:          "v.proj_blk.{bid}.self_attn_k",
+    MODEL_TENSOR.V_QF_SELF_ATTN_V:          "v.proj_blk.{bid}.self_attn_v",
+    MODEL_TENSOR.V_QF_SELF_ATTN_O:          "v.proj_blk.{bid}.self_attn_out",
+    MODEL_TENSOR.V_QF_SELF_ATTN_NORM:       "v.proj_blk.{bid}.self_attn_norm",
+    MODEL_TENSOR.V_QF_CROSS_ATTN_Q:         "v.proj_blk.{bid}.cross_attn_q",
+    MODEL_TENSOR.V_QF_CROSS_ATTN_K:         "v.proj_blk.{bid}.cross_attn_k",
+    MODEL_TENSOR.V_QF_CROSS_ATTN_V:         "v.proj_blk.{bid}.cross_attn_v",
+    MODEL_TENSOR.V_QF_CROSS_ATTN_O:         "v.proj_blk.{bid}.cross_attn_out",
+    MODEL_TENSOR.V_QF_CROSS_ATTN_NORM:      "v.proj_blk.{bid}.cross_attn_norm",
+    MODEL_TENSOR.V_QF_FFN_UP:               "v.proj_blk.{bid}.ffn_up",
+    MODEL_TENSOR.V_QF_FFN_DOWN:             "v.proj_blk.{bid}.ffn_down",
+    MODEL_TENSOR.V_QF_FFN_NORM:             "v.proj_blk.{bid}.ffn_norm",
+    # multi-projector (bid => projector ID)
+    MODEL_TENSOR.V_MULTI_PROJ_IMG_POS:   "v.proj_blk.{bid}.img_pos",
+    MODEL_TENSOR.V_MULTI_PROJ_QUERY:     "v.proj_blk.{bid}.query",
+    MODEL_TENSOR.V_MULTI_PROJ_NORM:      "v.proj_blk.{bid}.norm",
+    MODEL_TENSOR.V_MULTI_PROJ_LINEAR:    "v.proj_blk.{bid}.linear",
+    MODEL_TENSOR.V_MULTI_PROJ_POST_NORM: "v.proj_blk.{bid}.post_norm",
+
     # audio (mtmd)
     # note: all audio tensor names must use prefix "a." or "mm.a."
     MODEL_TENSOR.A_ENC_EMBD_POS:            "a.position_embd",
@@ -1417,6 +1480,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.A_QF_FFN_DOWN:             "a.proj_blk.{bid}.ffn_down",
     MODEL_TENSOR.A_QF_FFN_NORM:             "a.proj_blk.{bid}.ffn_norm",
     # NextN/MTP
+    MODEL_TENSOR.NEXTN_PROJ_PRE:            "nextn.pre_projection",
+    MODEL_TENSOR.NEXTN_PROJ_POST:           "nextn.post_projection",
     MODEL_TENSOR.NEXTN_EH_PROJ:             "blk.{bid}.nextn.eh_proj",
     MODEL_TENSOR.NEXTN_EMBED_TOKENS:        "blk.{bid}.nextn.embed_tokens",
     MODEL_TENSOR.NEXTN_ENORM:               "blk.{bid}.nextn.enorm",
@@ -1522,6 +1587,29 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_SAM_NET_3,
         MODEL_TENSOR.V_RESMPL_QUERY_768,
         MODEL_TENSOR.V_RESMPL_QUERY_1024,
+        MODEL_TENSOR.V_PROJ_NORM,
+        MODEL_TENSOR.V_QF_PROJ_QUERY,
+        MODEL_TENSOR.V_QF_PROJ_NORM,
+        MODEL_TENSOR.V_QF_PROJ_LINEAR,
+        MODEL_TENSOR.V_QF_SELF_ATTN_Q,
+        MODEL_TENSOR.V_QF_SELF_ATTN_K,
+        MODEL_TENSOR.V_QF_SELF_ATTN_V,
+        MODEL_TENSOR.V_QF_SELF_ATTN_O,
+        MODEL_TENSOR.V_QF_SELF_ATTN_NORM,
+        MODEL_TENSOR.V_QF_CROSS_ATTN_Q,
+        MODEL_TENSOR.V_QF_CROSS_ATTN_K,
+        MODEL_TENSOR.V_QF_CROSS_ATTN_V,
+        MODEL_TENSOR.V_QF_CROSS_ATTN_O,
+        MODEL_TENSOR.V_QF_CROSS_ATTN_NORM,
+        MODEL_TENSOR.V_QF_FFN_UP,
+        MODEL_TENSOR.V_QF_FFN_DOWN,
+        MODEL_TENSOR.V_QF_FFN_NORM,
+        MODEL_TENSOR.V_QF_PROJ_NORM,
+        MODEL_TENSOR.V_MULTI_PROJ_IMG_POS,
+        MODEL_TENSOR.V_MULTI_PROJ_QUERY,
+        MODEL_TENSOR.V_MULTI_PROJ_LINEAR,
+        MODEL_TENSOR.V_MULTI_PROJ_NORM,
+        MODEL_TENSOR.V_MULTI_PROJ_POST_NORM,
         # audio
         MODEL_TENSOR.A_ENC_EMBD_POS,
         MODEL_TENSOR.A_ENC_EMBD_NORM,
@@ -2500,6 +2588,26 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.PER_LAYER_PROJ_NORM,
         MODEL_TENSOR.PER_LAYER_POST_NORM,
     ],
+    MODEL_ARCH.GEMMA4_ASSISTANT: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.MASKED_EMBD_CENTROIDS,
+        MODEL_TENSOR.MASKED_EMBD_ORDERING,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.NEXTN_PROJ_PRE,
+        MODEL_TENSOR.NEXTN_PROJ_POST,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.FFN_PRE_NORM,
+        MODEL_TENSOR.FFN_POST_NORM,
+        MODEL_TENSOR.LAYER_OUT_SCALE,
+    ],
     MODEL_ARCH.GEMMA_EMBEDDING: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT,
@@ -4388,6 +4496,7 @@ class VisionProjectorType:
     MINICPMV4_6    = "minicpmv4_6"
     GRANITE_SPEECH = "granite_speech"  # audio
     MIMOVL         = "mimovl"
+    GRANITE4_VISION = "granite4_vision"
 
 
 # Items here are (block size, type size)

gguf-py/gguf/gguf_writer.py

@@ -959,8 +959,13 @@ class GGUFWriter:
         self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value)
 
     def add_num_deepstack_layers(self, count: int) -> None:
+        """Add scalar deepstack layer count (qwen3vl format)"""
         self.add_uint32(Keys.LLM.NUM_DEEPSTACK_LAYERS.format(arch=self.arch), count)
 
+    def add_deepstack_mapping(self, layers: Sequence[int]) -> None:
+        """Add per-layer deepstack projector indices (Granite4 Vision format)"""
+        self.add_array(Keys.LLM.DEEPSTACK_MAPPING.format(arch=self.arch), list(layers))
+
     def add_rope_dimension_count(self, count: int) -> None:
         self.add_uint32(Keys.Rope.DIMENSION_COUNT.format(arch=self.arch), count)
 
@@ -1119,6 +1124,9 @@ class GGUFWriter:
     def add_normalizer_lowercase(self, value: bool) -> None:
         self.add_bool(Keys.Tokenizer.NORMALIZER_LOWERCASE, value)
 
+    def add_normalizer_strip_accents(self, value: bool) -> None:
+        self.add_bool(Keys.Tokenizer.NORMALIZER_STRIP_ACCENTS, value)
+
     def add_eot_token_id(self, id: int) -> None:
         self.add_uint32(Keys.Tokenizer.EOT_ID, id)
 
@@ -1184,6 +1192,15 @@ class GGUFWriter:
     def add_vision_preproc_image_size(self, value: int) -> None:
         self.add_uint32(Keys.ClipVision.PREPROC_IMAGE_SIZE, value)
 
+    def add_vision_projector_query_side(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.Projector.QUERY_SIDE, value)
+
+    def add_vision_projector_window_side(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.Projector.WINDOW_SIDE, value)
+
+    def add_vision_spatial_offsets(self, layers: Sequence[int]) -> None:
+        self.add_array(Keys.ClipVision.Projector.SPATIAL_OFFSETS, layers)
+
     def add_vision_image_mean(self, values: Sequence[float]) -> None:
         self.add_array(Keys.ClipVision.IMAGE_MEAN, values)
 
@@ -1240,6 +1257,12 @@ class GGUFWriter:
     def add_vision_window_size(self, value: int) -> None:
         self.add_uint32(Keys.ClipVision.WINDOW_SIZE, value)
 
+    def add_vision_feature_layers(self, layers: Sequence[int]) -> None:
+        self.add_array(Keys.ClipVision.FEATURE_LAYERS, layers)
+
+    def add_vision_image_grid_pinpoints(self, layers: Sequence[Sequence[int]]) -> None:
+        self.add_array(Keys.ClipVision.IMAGE_GRID_PINPOINTS, layers)
+
     def add_vision_sam_layers_count(self, value: int) -> None:
         self.add_uint32(Keys.ClipVision.SAM.BLOCK_COUNT, value)
 

gguf-py/gguf/tensor_mapping.py

@@ -37,6 +37,14 @@ class TensorNameMap:
             "model.embed",                               # talkie
         ),
 
+        # Masked embeddings
+        MODEL_TENSOR.MASKED_EMBD_CENTROIDS: (
+            "masked_embedding.centroids",                # gemma-4 E2B/E4B assistants
+        ),
+        MODEL_TENSOR.MASKED_EMBD_ORDERING: (
+            "masked_embedding.token_ordering",           # gemma-4 E2B/E4B assistants
+        ),
+
         # Token type embeddings
         MODEL_TENSOR.TOKEN_TYPES: (
             "embeddings.token_type_embeddings",  # bert nomic-bert
@@ -1408,6 +1416,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.V_ENC_EMBD_PATCH: (
+            "model.vision_tower.vision_model.embeddings.patch_embedding", # Granite4Vision
             "vision_tower.vision_model.embeddings.patch_embedding",
             "model.vision_tower.embeddings.patch_embedding", # minicpmv4_6
             "model.vision_tower.embeddings.patch_embeddings.projection", # Intern-S1
@@ -1439,6 +1448,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.V_ENC_EMBD_POS: (
+            "model.vision_tower.vision_model.embeddings.position_embedding", # Granite4Vision
             "vision_tower.vision_model.embeddings.position_embedding",
             "model.vision_tower.embeddings.position_embedding", # minicpmv4_6
             "model.vision_tower.embeddings.position_embeddings", # Intern-S1
@@ -1456,8 +1466,9 @@ class TensorNameMap:
             "model.vision_embedder.pos_embedding", # gemma4 unified
         ),
 
+        # TODO: I think these should all be moved to mapping_cfg?
         MODEL_TENSOR.V_ENC_EMBD_IMGNL: (
-            "model.image_newline",  # Deepseek-OCR
+            "model.image_newline",  # Deepseek-OCR, Granite4Vision
             "vit.perceive.image_newline", # HunyuanVL
         ),
 
@@ -1477,6 +1488,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_Q: (
+            "model.vision_tower.vision_model.encoder.layers.{bid}.self_attn.q_proj", # Granite4Vision
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.q_proj",
             "model.vision_tower.encoder.layers.{bid}.self_attn.q_proj", # minicpmv4_6
             "model.vision_tower.encoder.layer.{bid}.attention.q_proj", # Intern-S1
@@ -1502,6 +1514,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_K: (
+            "model.vision_tower.vision_model.encoder.layers.{bid}.self_attn.k_proj", # Granite4Vision
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.k_proj",
             "model.vision_tower.encoder.layers.{bid}.self_attn.k_proj", # minicpmv4_6
             "model.vision_tower.encoder.layer.{bid}.attention.k_proj", # Intern-S1
@@ -1527,6 +1540,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_V: (
+            "model.vision_tower.vision_model.encoder.layers.{bid}.self_attn.v_proj", # Granite4Vision
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.v_proj",
             "model.vision_tower.encoder.layers.{bid}.self_attn.v_proj", # minicpmv4_6
             "model.vision_tower.encoder.layer.{bid}.attention.v_proj", # Intern-S1
@@ -1545,6 +1559,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.V_ENC_INPUT_NORM: (
+            "model.vision_tower.vision_model.encoder.layers.{bid}.layer_norm1", # Granite4Vision
             "vision_tower.vision_model.encoder.layers.{bid}.layer_norm1",
             "model.vision_tower.encoder.layers.{bid}.layer_norm1", # minicpmv4_6
             "vision_tower.vision_model.encoder.layers.{bid}.norm1", # InternVL
@@ -1567,6 +1582,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_O: (
+            "model.vision_tower.vision_model.encoder.layers.{bid}.self_attn.out_proj", # Granite4Vision
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.out_proj",
             "model.vision_tower.encoder.layers.{bid}.self_attn.out_proj", # minicpmv4_6
             "vision_tower.vision_model.encoder.layers.{bid}.attn.proj", # InternVL
@@ -1595,6 +1611,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.V_ENC_POST_ATTN_NORM: (
+            "model.vision_tower.vision_model.encoder.layers.{bid}.layer_norm2", # Granite4Vision
             "vision_tower.vision_model.encoder.layers.{bid}.layer_norm2",
             "model.vision_tower.encoder.layers.{bid}.layer_norm2", # minicpmv4_6
             "vision_tower.vision_model.encoder.layers.{bid}.norm2", # InternVL
@@ -1618,6 +1635,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.V_ENC_FFN_UP: (
+            "model.vision_tower.vision_model.encoder.layers.{bid}.mlp.fc1", # Granite4Vision
             "vision_tower.vision_model.encoder.layers.{bid}.mlp.fc1",
             "model.vision_tower.encoder.layers.{bid}.mlp.fc1", # minicpmv4_6
             "model.vision_tower.encoder.layer.{bid}.mlp.fc1", # Intern-S1
@@ -1649,6 +1667,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.V_ENC_FFN_DOWN: (
+            "model.vision_tower.vision_model.encoder.layers.{bid}.mlp.fc2", # Granite4Vision
             "vision_tower.vision_model.encoder.layers.{bid}.mlp.fc2",
             "model.vision_tower.encoder.layers.{bid}.mlp.fc2", # minicpmv4_6
             "model.vision_tower.encoder.layer.{bid}.mlp.fc2", # Intern-S1
@@ -1706,6 +1725,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.V_POST_NORM: (
+            "model.vision_tower.vision_model.post_layernorm", # Granite4Vision
             "vision_tower.vision_model.post_layernorm",
             "model.vision_tower.post_layernorm", # minicpmv4_6
             "model.vision_model.post_layernorm", # SmolVLM
@@ -1952,6 +1972,82 @@ class TensorNameMap:
             "model.vision_tower.std_scale", # gemma4
         ),
 
+        # For these tensors, bid => projector ID
+        MODEL_TENSOR.V_MULTI_PROJ_IMG_POS: (
+            "model.layerwise_projectors.{bid}.image_positions", # Granite4 Vision
+            "model.spatial_projectors.{bid}.image_positions",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_MULTI_PROJ_QUERY: (
+            "model.layerwise_projectors.{bid}.query", # Granite4 Vision
+            "model.spatial_projectors.{bid}.query",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_MULTI_PROJ_LINEAR: (
+            "model.layerwise_projectors.{bid}.out_linear", # Granite4 Vision
+            "model.spatial_projectors.{bid}.out_linear",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_MULTI_PROJ_NORM: (
+            "model.layerwise_projectors.{bid}.norm", # Granite4 Vision
+            "model.spatial_projectors.{bid}.norm",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_MULTI_PROJ_POST_NORM: (
+            "model.layerwise_projectors.{bid}.qformer.layernorm", # Granite4 Vision
+            "model.spatial_projectors.{bid}.qformer.layernorm",   # Granite4 Vision
+        ),
+
+        # For these tensors, bid => proj-id
+        MODEL_TENSOR.V_QF_SELF_ATTN_Q: (
+            "model.layerwise_projectors.qformer.encoder.layer.{bid}.attention.attention.query", # Granite4 Vision
+            "model.spatial_projectors.qformer.encoder.layer.{bid}.attention.attention.query",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_QF_SELF_ATTN_K: (
+            "model.layerwise_projectors.qformer.encoder.layer.{bid}.attention.attention.key", # Granite4 Vision
+            "model.spatial_projectors.qformer.encoder.layer.{bid}.attention.attention.key",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_QF_SELF_ATTN_V: (
+            "model.layerwise_projectors.qformer.encoder.layer.{bid}.attention.attention.value", # Granite4 Vision
+            "model.spatial_projectors.qformer.encoder.layer.{bid}.attention.attention.value",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_QF_SELF_ATTN_O: (
+            "model.layerwise_projectors.qformer.encoder.layer.{bid}.attention.output.dense", # Granite4 Vision
+            "model.spatial_projectors.qformer.encoder.layer.{bid}.attention.output.dense",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_QF_SELF_ATTN_NORM: (
+            "model.layerwise_projectors.qformer.encoder.layer.{bid}.attention.output.LayerNorm", # Granite4 Vision
+            "model.spatial_projectors.qformer.encoder.layer.{bid}.attention.output.LayerNorm",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_QF_CROSS_ATTN_Q: (
+            "model.layerwise_projectors.qformer.encoder.layer.{bid}.crossattention.attention.query", # Granite4 Vision
+            "model.spatial_projectors.qformer.encoder.layer.{bid}.crossattention.attention.query",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_QF_CROSS_ATTN_K: (
+            "model.layerwise_projectors.qformer.encoder.layer.{bid}.crossattention.attention.key", # Granite4 Vision
+            "model.spatial_projectors.qformer.encoder.layer.{bid}.crossattention.attention.key",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_QF_CROSS_ATTN_V: (
+            "model.layerwise_projectors.qformer.encoder.layer.{bid}.crossattention.attention.value", # Granite4 Vision
+            "model.spatial_projectors.qformer.encoder.layer.{bid}.crossattention.attention.value",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_QF_CROSS_ATTN_O: (
+            "model.layerwise_projectors.qformer.encoder.layer.{bid}.crossattention.output.dense", # Granite4 Vision
+            "model.spatial_projectors.qformer.encoder.layer.{bid}.crossattention.output.dense",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_QF_CROSS_ATTN_NORM: (
+            "model.layerwise_projectors.qformer.encoder.layer.{bid}.crossattention.output.LayerNorm", # Granite4 Vision
+            "model.spatial_projectors.qformer.encoder.layer.{bid}.crossattention.output.LayerNorm",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_QF_FFN_UP: (
+            "model.layerwise_projectors.qformer.encoder.layer.{bid}.intermediate_query.dense", # Granite4 Vision
+            "model.spatial_projectors.qformer.encoder.layer.{bid}.intermediate_query.dense",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_QF_FFN_DOWN: (
+            "model.layerwise_projectors.qformer.encoder.layer.{bid}.output_query.dense", # Granite4 Vision
+            "model.spatial_projectors.qformer.encoder.layer.{bid}.output_query.dense",   # Granite4 Vision
+        ),
+        MODEL_TENSOR.V_QF_FFN_NORM: (
+            "model.layerwise_projectors.qformer.encoder.layer.{bid}.output_query.LayerNorm", # Granite4 Vision
+            "model.spatial_projectors.qformer.encoder.layer.{bid}.output_query.LayerNorm",   # Granite4 Vision
+        ),
+
         # audio (mtmd)
 
         MODEL_TENSOR.A_ENC_EMBD_POS: (
@@ -2279,6 +2375,14 @@ class TensorNameMap:
         ),
 
         # NextN/MTP tensors
+        MODEL_TENSOR.NEXTN_PROJ_PRE: (
+            "pre_projection",
+        ),
+
+        MODEL_TENSOR.NEXTN_PROJ_POST: (
+            "post_projection",
+        ),
+
         MODEL_TENSOR.NEXTN_EH_PROJ: (
             "model.layers.{bid}.eh_proj",
         ),

gguf-py/gguf/vocab.py

@@ -53,6 +53,7 @@ class SpecialVocab:
     special_token_ids: dict[str, int]
     chat_template: str | Sequence[Mapping[str, str]] | None
     normalizer_lowercase: bool | None
+    normalizer_strip_accents: bool | None
 
     def __init__(
         self, path: str | os.PathLike[str], load_merges: bool = False,
@@ -66,6 +67,7 @@ class SpecialVocab:
         self.merges = []
         self.chat_template = None
         self.normalizer_lowercase = None
+        self.normalizer_strip_accents = None
         if special_token_types is not None:
             self.special_token_types = special_token_types
         else:
@@ -108,6 +110,10 @@ class SpecialVocab:
             if not quiet:
                 logger.info(f'Setting normalizer_lowercase to {self.normalizer_lowercase}')
             gw.add_normalizer_lowercase(self.normalizer_lowercase)
+        if self.normalizer_strip_accents is not None:
+            if not quiet:
+                logger.info(f'Setting normalizer_strip_accents to {self.normalizer_strip_accents}')
+            gw.add_normalizer_strip_accents(self.normalizer_strip_accents)
 
     def _load(self, path: Path) -> None:
         self._try_load_from_tokenizer_json(path)
@@ -155,17 +161,21 @@ class SpecialVocab:
     def _parse_normalizer(self, normalizer: dict) -> None:
         # ref: https://huggingface.co/docs/tokenizers/api/normalizers
         #
-        # Detects lowercase normalization in three possible formats:
-        # 1. Standalone: {"type": "Lowercase"}
-        # 2. BertNormalizer attribute: {"type": "BertNormalizer", "lowercase": true, ...}
-        # 3. Nested in Sequence: {"type": "Sequence", "normalizers": [...]}
+        # Extracts normalizer flags from three possible formats:
+        # 1. Standalone:           {"type": "Lowercase"}
+        # 2. BertNormalizer attrs: {"type": "BertNormalizer", ...}
+        # 3. Nested in Sequence:   {"type": "Sequence", "normalizers": [...]}
 
         normalizer_type = normalizer.get('type')
         if normalizer_type == 'Lowercase':
             self.normalizer_lowercase = True
+        elif normalizer_type == 'StripAccents':
+            self.normalizer_strip_accents = True
         elif normalizer_type == 'BertNormalizer':
             if 'lowercase' in normalizer:
                 self.normalizer_lowercase = normalizer['lowercase']
+            if 'strip_accents' in normalizer:
+                self.normalizer_strip_accents = normalizer['strip_accents']
         elif normalizer_type == 'Sequence':
             for norm in normalizer.get('normalizers', []):
                 self._parse_normalizer(norm)
@@ -246,6 +256,11 @@ class SpecialVocab:
                             if special_first := tmpl_single[0].get('SpecialToken', {}).get('id'):
                                 if not tokenizer_config:
                                     special_bos = special_first
+                                elif special_first not in (special_bos, special_cls):
+                                    if not special_bos:
+                                        tokenizer_config['bos_token'] = special_bos = special_first
+                                    if not special_cls:
+                                        tokenizer_config['cls_token'] = special_cls = special_first
                                 self.add_special_token['bos'] = True if special_first in (special_bos, special_cls) else False
                                 if special_first not in (special_bos, special_cls):
                                     logger.warning(f'Unknown leading special token {special_first!r} in TemplateProcessing<single>')

include/llama.h

@@ -388,6 +388,10 @@ extern "C" {
         // note: the samplers must be sampler chains (i.e. use llama_sampler_chain_init)
         struct llama_sampler_seq_config * samplers;
         size_t                            n_samplers;
+
+        // a source/target/parent context
+        // can be utilized in various ways, for example by sharing results or llama_memory between 2 contexts
+        struct llama_context * ctx_other;
     };
 
     struct llama_model_tensor_override {

models/templates/LFM2.5-8B-A1B.jinja

@@ -0,0 +1,115 @@
+{{- bos_token -}}
+{%- set preserve_thinking = preserve_thinking | default(false) -%}
+
+{%- macro format_arg_value(arg_value) -%}
+    {%- if arg_value is string -%}
+        {{- "'" + arg_value + "'" -}}
+    {%- elif arg_value is mapping -%}
+        {{- arg_value | tojson -}}
+    {%- else -%}
+        {{- arg_value | string -}}
+    {%- endif -%}
+{%- endmacro -%}
+
+{%- macro parse_content(content) -%}
+    {%- if content is string -%}
+        {{- content -}}
+    {%- else -%}
+        {%- set _ns = namespace(result="") -%}
+        {%- for item in content -%}
+            {%- if item["type"] == "image" -%}
+                {%- set _ns.result = _ns.result + "<image>" -%}
+            {%- elif item["type"] == "text" -%}
+                {%- set _ns.result = _ns.result + item["text"] -%}
+            {%- else -%}
+                {%- set _ns.result = _ns.result + item | tojson -%}
+            {%- endif -%}
+        {%- endfor -%}
+        {{- _ns.result -}}
+    {%- endif -%}
+{%- endmacro -%}
+
+{%- macro render_tool_calls(tool_calls) -%}
+    {%- set tool_calls_ns = namespace(tool_calls=[]) -%}
+    {%- for tool_call in tool_calls -%}
+        {%- set func_name = tool_call["function"]["name"] -%}
+        {%- set func_args = tool_call["function"]["arguments"] -%}
+        {%- set args_ns = namespace(arg_strings=[]) -%}
+        {%- for arg_name, arg_value in func_args.items() -%}
+            {%- set args_ns.arg_strings = args_ns.arg_strings + [arg_name + "=" + format_arg_value(arg_value)] -%}
+        {%- endfor -%}
+        {%- set tool_calls_ns.tool_calls = tool_calls_ns.tool_calls + [func_name + "(" + (args_ns.arg_strings | join(", ")) + ")"] -%}
+    {%- endfor -%}
+    {{- "<|tool_call_start|>[" + (tool_calls_ns.tool_calls | join(", ")) + "]<|tool_call_end|>" -}}
+{%- endmacro -%}
+
+{%- set ns = namespace(system_prompt="", last_user_index=-1) -%}
+{%- if messages[0]["role"] == "system" -%}
+    {%- if messages[0].get("content") -%}
+        {%- set ns.system_prompt = parse_content(messages[0]["content"]) -%}
+    {%- endif -%}
+    {%- set messages = messages[1:] -%}
+{%- endif -%}
+{%- if tools -%}
+    {%- set ns.system_prompt = ns.system_prompt + ("\n" if ns.system_prompt else "") + "List of tools: [" -%}
+    {%- for tool in tools -%}
+        {%- if tool is not string -%}
+            {%- set tool = tool | tojson -%}
+        {%- endif -%}
+        {%- set ns.system_prompt = ns.system_prompt + tool -%}
+        {%- if not loop.last -%}
+            {%- set ns.system_prompt = ns.system_prompt + ", " -%}
+        {%- endif -%}
+    {%- endfor -%}
+    {%- set ns.system_prompt = ns.system_prompt + "]" -%}
+{%- endif -%}
+{%- if ns.system_prompt -%}
+    {{- "<|im_start|>system\n" + ns.system_prompt + "<|im_end|>\n" -}}
+{%- endif -%}
+{%- for message in messages -%}
+    {%- if message["role"] == "user" -%}
+        {%- set ns.last_user_index = loop.index0 -%}
+    {%- endif -%}
+{%- endfor -%}
+{%- for message in messages -%}
+    {{- "<|im_start|>" + message.role + "\n" -}}
+    {%- if message.role == "assistant" -%}
+        {%- generation -%}
+        {%- if message.thinking is defined and (preserve_thinking or loop.index0 > ns.last_user_index) -%}
+            {{- "<think>" + message.thinking + "</think>" -}}
+        {%- endif -%}
+        {%- set _cfm_tag = "CONTINUE_FINAL_MESSAGE_TAG " -%}
+        {%- set _has_cfm = false -%}
+        {%- if message.content is defined -%}
+            {%- set content = parse_content(message.content) -%}
+            {%- if not (preserve_thinking or loop.index0 > ns.last_user_index) -%}
+                {%- if "</think>" in content -%}
+                    {%- set content = content.split("</think>")[-1] | trim -%}
+                {%- endif -%}
+            {%- endif -%}
+            {%- if message.tool_calls is defined and content.endswith(_cfm_tag) -%}
+                {%- set _has_cfm = true -%}
+                {%- set _trunc_len = (content | length) - (_cfm_tag | length) -%}
+                {{- content[:_trunc_len] -}}
+            {%- else -%}
+                {{- content -}}
+            {%- endif -%}
+        {%- endif -%}
+        {%- if message.tool_calls is defined -%}
+            {{- render_tool_calls(message.tool_calls) -}}
+        {%- endif -%}
+        {%- if _has_cfm -%}
+            {{- _cfm_tag -}}
+        {%- endif -%}
+        {{- "<|im_end|>\n" -}}
+        {%- endgeneration -%}
+    {%- else %}
+        {%- if message.get("content") -%}
+            {{- parse_content(message["content"]) -}}
+        {%- endif -%}
+        {{- "<|im_end|>\n" -}}
+    {%- endif %}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+    {{- "<|im_start|>assistant\n" -}}
+{%- endif -%}

scripts/sync-ggml.last

@@ -1 +1 @@
-1e33fed33e87c43aa4c4078e2a9c239d4c1f1bd3
+7142aa6bf9fcaeec0fef8d80fcd90afe4268adf1

src/llama-arch.cpp

@@ -57,6 +57,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_GEMMA3,           "gemma3"           },
     { LLM_ARCH_GEMMA3N,          "gemma3n"          },
     { LLM_ARCH_GEMMA4,           "gemma4"           },
+    { LLM_ARCH_GEMMA4_ASSISTANT, "gemma4-assistant" },
     { LLM_ARCH_GEMMA_EMBEDDING,  "gemma-embedding"  },
     { LLM_ARCH_STARCODER2,       "starcoder2"       },
     { LLM_ARCH_MAMBA,            "mamba"            },
@@ -196,6 +197,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_MOE_LATENT_SIZE,                   "%s.moe_latent_size"                   },
     { LLM_KV_NEXTN_PREDICT_LAYERS,              "%s.nextn_predict_layers"              },
     { LLM_KV_NUM_DEEPSTACK_LAYERS,              "%s.n_deepstack_layers"                },
+    { LLM_KV_DEEPSTACK_MAPPING,                 "%s.deepstack_mapping"                 },
     { LLM_KV_HIDDEN_ACT,                        "%s.hidden_activation"                 },
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
@@ -297,39 +299,40 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_DENSE_3_FEAT_IN,        "%s.dense_3_feat_in"   },
     { LLM_KV_DENSE_3_FEAT_OUT,       "%s.dense_3_feat_out"  },
 
-    { LLM_KV_TOKENIZER_MODEL,                "tokenizer.ggml.model"                    },
-    { LLM_KV_TOKENIZER_PRE,                  "tokenizer.ggml.pre"                      },
-    { LLM_KV_TOKENIZER_LIST,                 "tokenizer.ggml.tokens"                   },
-    { LLM_KV_TOKENIZER_TOKEN_TYPE,           "tokenizer.ggml.token_type"               },
-    { LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,     "tokenizer.ggml.token_type_count"         },
-    { LLM_KV_TOKENIZER_SCORES,               "tokenizer.ggml.scores"                   },
-    { LLM_KV_TOKENIZER_MERGES,               "tokenizer.ggml.merges"                   },
-    { LLM_KV_TOKENIZER_BOS_ID,               "tokenizer.ggml.bos_token_id"             },
-    { LLM_KV_TOKENIZER_EOS_ID,               "tokenizer.ggml.eos_token_id"             },
-    { LLM_KV_TOKENIZER_EOT_ID,               "tokenizer.ggml.eot_token_id"             },
-    { LLM_KV_TOKENIZER_EOM_ID,               "tokenizer.ggml.eom_token_id"             },
-    { LLM_KV_TOKENIZER_UNK_ID,               "tokenizer.ggml.unknown_token_id"         },
-    { LLM_KV_TOKENIZER_SEP_ID,               "tokenizer.ggml.seperator_token_id"       },
-    { LLM_KV_TOKENIZER_PAD_ID,               "tokenizer.ggml.padding_token_id"         },
-    { LLM_KV_TOKENIZER_CLS_ID,               "tokenizer.ggml.cls_token_id"             },
-    { LLM_KV_TOKENIZER_MASK_ID,              "tokenizer.ggml.mask_token_id"            },
-    { LLM_KV_TOKENIZER_ADD_BOS,              "tokenizer.ggml.add_bos_token"            },
-    { LLM_KV_TOKENIZER_ADD_EOS,              "tokenizer.ggml.add_eos_token"            },
-    { LLM_KV_TOKENIZER_ADD_SEP,              "tokenizer.ggml.add_sep_token"            },
-    { LLM_KV_TOKENIZER_ADD_PREFIX,           "tokenizer.ggml.add_space_prefix"         },
-    { LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,      "tokenizer.ggml.remove_extra_whitespaces" },
-    { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, "tokenizer.ggml.precompiled_charsmap"     },
-    { LLM_KV_TOKENIZER_HF_JSON,              "tokenizer.huggingface.json"              },
-    { LLM_KV_TOKENIZER_RWKV,                 "tokenizer.rwkv.world"                    },
-    { LLM_KV_TOKENIZER_CHAT_TEMPLATE,        "tokenizer.chat_template"                 },
-    { LLM_KV_TOKENIZER_NORMALIZER_LOWERCASE, "tokenizer.ggml.normalizer.lowercase"     },
-    { LLM_KV_TOKENIZER_FIM_PRE_ID,           "tokenizer.ggml.fim_pre_token_id"         },
-    { LLM_KV_TOKENIZER_FIM_SUF_ID,           "tokenizer.ggml.fim_suf_token_id"         },
-    { LLM_KV_TOKENIZER_FIM_MID_ID,           "tokenizer.ggml.fim_mid_token_id"         },
-    { LLM_KV_TOKENIZER_FIM_PAD_ID,           "tokenizer.ggml.fim_pad_token_id"         },
-    { LLM_KV_TOKENIZER_FIM_REP_ID,           "tokenizer.ggml.fim_rep_token_id"         },
-    { LLM_KV_TOKENIZER_FIM_SEP_ID,           "tokenizer.ggml.fim_sep_token_id"         },
-    { LLM_KV_TOKENIZER_SUPPRESS_TOKENS,      "tokenizer.ggml.suppress_tokens"          },
+    { LLM_KV_TOKENIZER_MODEL,                    "tokenizer.ggml.model"                    },
+    { LLM_KV_TOKENIZER_PRE,                      "tokenizer.ggml.pre"                      },
+    { LLM_KV_TOKENIZER_LIST,                     "tokenizer.ggml.tokens"                   },
+    { LLM_KV_TOKENIZER_TOKEN_TYPE,               "tokenizer.ggml.token_type"               },
+    { LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,         "tokenizer.ggml.token_type_count"         },
+    { LLM_KV_TOKENIZER_SCORES,                   "tokenizer.ggml.scores"                   },
+    { LLM_KV_TOKENIZER_MERGES,                   "tokenizer.ggml.merges"                   },
+    { LLM_KV_TOKENIZER_BOS_ID,                   "tokenizer.ggml.bos_token_id"             },
+    { LLM_KV_TOKENIZER_EOS_ID,                   "tokenizer.ggml.eos_token_id"             },
+    { LLM_KV_TOKENIZER_EOT_ID,                   "tokenizer.ggml.eot_token_id"             },
+    { LLM_KV_TOKENIZER_EOM_ID,                   "tokenizer.ggml.eom_token_id"             },
+    { LLM_KV_TOKENIZER_UNK_ID,                   "tokenizer.ggml.unknown_token_id"         },
+    { LLM_KV_TOKENIZER_SEP_ID,                   "tokenizer.ggml.seperator_token_id"       },
+    { LLM_KV_TOKENIZER_PAD_ID,                   "tokenizer.ggml.padding_token_id"         },
+    { LLM_KV_TOKENIZER_CLS_ID,                   "tokenizer.ggml.cls_token_id"             },
+    { LLM_KV_TOKENIZER_MASK_ID,                  "tokenizer.ggml.mask_token_id"            },
+    { LLM_KV_TOKENIZER_ADD_BOS,                  "tokenizer.ggml.add_bos_token"            },
+    { LLM_KV_TOKENIZER_ADD_EOS,                  "tokenizer.ggml.add_eos_token"            },
+    { LLM_KV_TOKENIZER_ADD_SEP,                  "tokenizer.ggml.add_sep_token"            },
+    { LLM_KV_TOKENIZER_ADD_PREFIX,               "tokenizer.ggml.add_space_prefix"         },
+    { LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,          "tokenizer.ggml.remove_extra_whitespaces" },
+    { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,     "tokenizer.ggml.precompiled_charsmap"     },
+    { LLM_KV_TOKENIZER_HF_JSON,                  "tokenizer.huggingface.json"              },
+    { LLM_KV_TOKENIZER_RWKV,                     "tokenizer.rwkv.world"                    },
+    { LLM_KV_TOKENIZER_CHAT_TEMPLATE,            "tokenizer.chat_template"                 },
+    { LLM_KV_TOKENIZER_NORMALIZER_LOWERCASE,     "tokenizer.ggml.normalizer.lowercase"     },
+    { LLM_KV_TOKENIZER_NORMALIZER_STRIP_ACCENTS, "tokenizer.ggml.normalizer.strip_accents" },
+    { LLM_KV_TOKENIZER_FIM_PRE_ID,               "tokenizer.ggml.fim_pre_token_id"         },
+    { LLM_KV_TOKENIZER_FIM_SUF_ID,               "tokenizer.ggml.fim_suf_token_id"         },
+    { LLM_KV_TOKENIZER_FIM_MID_ID,               "tokenizer.ggml.fim_mid_token_id"         },
+    { LLM_KV_TOKENIZER_FIM_PAD_ID,               "tokenizer.ggml.fim_pad_token_id"         },
+    { LLM_KV_TOKENIZER_FIM_REP_ID,               "tokenizer.ggml.fim_rep_token_id"         },
+    { LLM_KV_TOKENIZER_FIM_SEP_ID,               "tokenizer.ggml.fim_sep_token_id"         },
+    { LLM_KV_TOKENIZER_SUPPRESS_TOKENS,          "tokenizer.ggml.suppress_tokens"          },
 
     { LLM_KV_ADAPTER_TYPE,                    "adapter.type"               },
     { LLM_KV_ADAPTER_LORA_ALPHA,              "adapter.lora.alpha"         },
@@ -452,6 +455,8 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
     { LLM_TENSOR_FFN_NORM_EXPS,                          "blk.%d.ffn_norm_exps" },
     { LLM_TENSOR_ATTN_K_B,                               "blk.%d.attn_k_b" },
     { LLM_TENSOR_ATTN_V_B,                               "blk.%d.attn_v_b" },
+    { LLM_TENSOR_NEXTN_PROJ_PRE,                         "nextn.pre_projection" },
+    { LLM_TENSOR_NEXTN_PROJ_POST,                        "nextn.post_projection" },
     { LLM_TENSOR_NEXTN_EH_PROJ,                          "blk.%d.nextn.eh_proj" },
     { LLM_TENSOR_NEXTN_EMBED_TOKENS,                     "blk.%d.nextn.embed_tokens" },
     { LLM_TENSOR_NEXTN_ENORM,                            "blk.%d.nextn.enorm" },
@@ -555,6 +560,8 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
     { LLM_TENSOR_INDEXER_PROJ,                           "blk.%d.indexer.proj" },
     { LLM_TENSOR_INDEXER_ATTN_K,                         "blk.%d.indexer.attn_k" },
     { LLM_TENSOR_INDEXER_ATTN_Q_B,                       "blk.%d.indexer.attn_q_b" },
+    { LLM_TENSOR_MASKED_EMBD_CENTROIDS,                  "masked_embd_centroids" },
+    { LLM_TENSOR_MASKED_EMBD_ORDERING,                   "masked_embd_ordering" },
 };
 
 // declare information about the model weight tensors:
@@ -764,6 +771,8 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_INDEXER_PROJ,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_INDEXER_ATTN_K,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_INDEXER_ATTN_Q_B,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_NEXTN_PROJ_PRE,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_NEXTN_PROJ_POST,            {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
     // NextN/MTP tensors are stored per-block (blk.%d.nextn.*) even though only the
     // last nextn_predict_layers blocks carry them. Classify as LAYER_REPEATING so
     // the model loader doesn't fault on the block index.
@@ -777,6 +786,8 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     // latent projections feed ggml_mul_mat, the buft probe must use MUL_MAT to keep them on GPU
     {LLM_TENSOR_FFN_LATENT_DOWN,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_FFN_LATENT_UP,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_MASKED_EMBD_CENTROIDS,      {LLM_TENSOR_LAYER_INPUT,     GGML_OP_NONE}},
+    {LLM_TENSOR_MASKED_EMBD_ORDERING,       {LLM_TENSOR_LAYER_INPUT,     GGML_OP_NONE}},
 };
 
 LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}

src/llama-arch.h

@@ -61,6 +61,7 @@ enum llm_arch {
     LLM_ARCH_GEMMA3,
     LLM_ARCH_GEMMA3N,
     LLM_ARCH_GEMMA4,
+    LLM_ARCH_GEMMA4_ASSISTANT,
     LLM_ARCH_GEMMA_EMBEDDING,
     LLM_ARCH_STARCODER2,
     LLM_ARCH_MAMBA,
@@ -200,6 +201,7 @@ enum llm_kv {
     LLM_KV_MOE_LATENT_SIZE,
     LLM_KV_NEXTN_PREDICT_LAYERS,
     LLM_KV_NUM_DEEPSTACK_LAYERS,
+    LLM_KV_DEEPSTACK_MAPPING,
     LLM_KV_HIDDEN_ACT,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
@@ -312,6 +314,7 @@ enum llm_kv {
     LLM_KV_TOKENIZER_RWKV,
     LLM_KV_TOKENIZER_CHAT_TEMPLATE,
     LLM_KV_TOKENIZER_NORMALIZER_LOWERCASE,
+    LLM_KV_TOKENIZER_NORMALIZER_STRIP_ACCENTS,
     LLM_KV_TOKENIZER_FIM_PRE_ID,
     LLM_KV_TOKENIZER_FIM_SUF_ID,
     LLM_KV_TOKENIZER_FIM_MID_ID,
@@ -556,14 +559,19 @@ enum llm_tensor {
     LLM_TENSOR_INDEXER_PROJ,
     LLM_TENSOR_INDEXER_ATTN_K,
     LLM_TENSOR_INDEXER_ATTN_Q_B,
+    LLM_TENSOR_NEXTN_PROJ_PRE,
+    LLM_TENSOR_NEXTN_PROJ_POST,
     LLM_TENSOR_NEXTN_EH_PROJ,
     LLM_TENSOR_NEXTN_EMBED_TOKENS,
     LLM_TENSOR_NEXTN_ENORM,
     LLM_TENSOR_NEXTN_HNORM,
     LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
     LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
+    LLM_TENSOR_MASKED_EMBD_CENTROIDS,
+    LLM_TENSOR_MASKED_EMBD_ORDERING,
 };
 
+
 enum llm_tensor_layer {
     LLM_TENSOR_LAYER_INPUT,
     LLM_TENSOR_LAYER_REPEATING,

src/llama-context.cpp

@@ -69,9 +69,10 @@ llama_context::llama_context(
     cparams.embeddings_nextn_masked = false;
     cparams.offload_kqv             = params.offload_kqv;
     cparams.no_perf                 = params.no_perf;
-    cparams.pooling_type            = params.pooling_type;
     cparams.warmup                  = false;
 
+    cparams.ctx_type     = params.ctx_type;
+    cparams.pooling_type = params.pooling_type;
 
     cparams.n_ctx            = params.n_ctx           == 0    ? hparams.n_ctx_train           : params.n_ctx;
     cparams.rope_freq_base   = params.rope_freq_base  == 0.0f ? hparams.rope_freq_base_train  : params.rope_freq_base;
@@ -84,7 +85,17 @@ llama_context::llama_context(
     cparams.cb_eval           = params.cb_eval;
     cparams.cb_eval_user_data = params.cb_eval_user_data;
 
-    cparams.ctx_type          = params.ctx_type;
+    cparams.ctx_other = nullptr;
+
+    // TODO: more generic
+    if (model.arch == LLM_ARCH_GEMMA4_ASSISTANT) {
+        if (params.ctx_other == nullptr) {
+            // TODO: change from runtime_error to llama_exception to avoid printing error message
+            throw std::runtime_error("Gemma4Assistant requires ctx_other to be set (this is normal during memory fitting)");
+        }
+
+        cparams.ctx_other = params.ctx_other;
+    }
 
     // Initialize backend samplers here so they are part of the sampling graph
     // before the reserve passes run later in this function. This avoids a later
@@ -300,10 +311,11 @@ llama_context::llama_context(
     // init the memory module
     if (!hparams.vocab_only) {
         llama_memory_params params_mem = {
-            /*.type_k   =*/ params.type_k,
-            /*.type_v   =*/ params.type_v,
-            /*.swa_full =*/ params.swa_full,
-            /*.ctx_type= */ cparams.ctx_type,
+            /*.type_k    =*/ params.type_k,
+            /*.type_v    =*/ params.type_v,
+            /*.swa_full  =*/ params.swa_full,
+            /*.ctx_type  =*/ cparams.ctx_type,
+            /*.mem_other =*/ llama_get_memory(cparams.ctx_other),
         };
 
         memory.reset(model.create_memory(params_mem, cparams));
@@ -341,7 +353,7 @@ llama_context::llama_context(
         // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
         bool pipeline_parallel =
             model.n_devices() > 1 &&
-            model.n_gpu_layers() > model.hparams.n_layer() &&
+            model.n_gpu_layers() > model.hparams.n_layer_all &&
             model.split_mode() == LLAMA_SPLIT_MODE_LAYER &&
             cparams.offload_kqv &&
             !model.has_tensor_overrides();
@@ -904,7 +916,7 @@ float * llama_context::get_embeddings_nextn_ith(int32_t i) {
             throw std::runtime_error("no nextn embeddings");
         }
 
-        const uint32_t n_embd = model.hparams.n_embd;
+        const uint32_t n_embd = model.hparams.n_embd_out();
 
         if (!cparams.embeddings_nextn_masked) {
             // unmasked: nextn rows are stored densely, indexed by raw token position.
@@ -1473,7 +1485,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
         ggml_backend_t backend_h = ggml_backend_sched_get_tensor_backend(sched.get(), t_h_nextn);
         GGML_ASSERT(backend_h != nullptr);
 
-        const uint32_t n_embd = hparams.n_embd;
+        const uint32_t n_embd = hparams.n_embd_out();
         GGML_ASSERT(n_tokens*n_embd <= (int64_t) embd_nextn.size);
         ggml_backend_tensor_get_async(backend_h, t_h_nextn, embd_nextn.data, 0, n_tokens*n_embd*sizeof(float));
     }
@@ -1924,7 +1936,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
                 ggml_backend_t backend_h = ggml_backend_sched_get_tensor_backend(sched.get(), t_h_nextn);
                 GGML_ASSERT(backend_h != nullptr);
 
-                const uint32_t n_embd  = hparams.n_embd;
+                const uint32_t n_embd  = hparams.n_embd_out();
                 float * embd_nextn_out = embd_nextn.data + offset*n_embd;
 
                 GGML_ASSERT((offset + n_rows)*n_embd <= (int64_t) embd_nextn.size);
@@ -2017,7 +2029,6 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
 
     const auto n_batch    = cparams.n_batch;
     const auto n_vocab    = vocab.n_tokens();
-    const auto n_embd     = hparams.n_embd;
     const auto n_embd_out = hparams.n_embd_out();
 
     bool has_logits     = true;
@@ -2036,12 +2047,12 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
 
     logits.size     = has_logits     ? n_vocab*n_outputs_max     : 0;
     embd.size       = has_embd       ? n_embd_out*n_outputs_max  : 0;
-    embd_nextn.size = has_embd_nextn ? n_embd*n_outputs_max      : 0;
+    embd_nextn.size = has_embd_nextn ? n_embd_out*n_outputs_max  : 0;
 
     if (has_embd_nextn && !cparams.embeddings_nextn_masked) {
         // unmasked: nextn row exists for every token in the batch, not just
         // those flagged via batch.logits[i] -> size by token count instead.
-        embd_nextn.size = (size_t) n_embd * n_batch;
+        embd_nextn.size = (size_t) n_embd_out * n_batch;
     }
 
     // Allocate backend sampling output buffers if there are backend samplers configured.
@@ -2351,7 +2362,7 @@ llm_graph_cb llama_context::graph_get_cb() const {
 
         // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
         // FIXME: fix in ggml_backend_sched
-        const bool full_offload = model.n_gpu_layers() > model.hparams.n_layer();
+        const bool full_offload = model.n_gpu_layers() > model.hparams.n_layer_all;
         if (ubatch.n_tokens < 32 || full_offload) {
             if (il != -1 && strcmp(name, "norm") == 0) {
                 const auto & dev_layer = model.dev_layer(il);
@@ -3375,6 +3386,7 @@ llama_context_params llama_context_default_params() {
         /*.kv_unified                  =*/ false,
         /*.sampler                     =*/ nullptr,
         /*.n_sampler                   =*/ 0,
+        /*.ctx_other                   =*/ nullptr,
     };
 
     return result;
@@ -3454,7 +3466,6 @@ llama_context * llama_init_from_model(
         return nullptr;
     }
 
-
     try {
         auto * ctx = new llama_context(*model, params);
         return ctx;
@@ -3593,6 +3604,14 @@ void llama_set_embeddings_nextn(llama_context * ctx, bool value, bool masked) {
     ctx->set_embeddings_nextn(value, masked);
 }
 
+llama_memory_t llama_get_memory(const struct llama_context * ctx) {
+    if (!ctx) {
+        return nullptr;
+    }
+
+    return ctx->get_memory();
+}
+
 float * llama_get_embeddings_nextn(llama_context * ctx) {
     ctx->synchronize();
 
@@ -3656,7 +3675,7 @@ struct ggml_cgraph * llama_graph_reserve(
         uint32_t n_tokens,
         uint32_t n_seqs,
         uint32_t n_outputs) {
-    auto * memory = ctx->get_memory();
+    auto memory = ctx->get_memory();
     llama_memory_context_ptr mctx;
     if (memory) {
         mctx = memory->init_full();
@@ -3696,10 +3715,6 @@ int32_t llama_set_adapter_cvec(
 // memory
 //
 
-llama_memory_t llama_get_memory(const struct llama_context * ctx) {
-    return ctx->get_memory();
-}
-
 void llama_memory_clear(llama_memory_t mem, bool data) {
     if (!mem) {
         return;
@@ -4010,3 +4025,7 @@ void llama_opt_epoch(
 llama_memory_breakdown llama_get_memory_breakdown(const struct llama_context * ctx) {
     return ctx->memory_breakdown();
 }
+
+llama_context * llama_get_ctx_other(struct llama_context * ctx) {
+    return ctx->get_cparams().ctx_other;
+}

src/llama-context.h

@@ -6,6 +6,7 @@
 #include "llama-graph.h"
 #include "llama-adapter.h"
 #include "llama-impl.h"
+#include "llama-memory.h"
 
 #include "ggml-cpp.h"
 #include "ggml-opt.h"
@@ -273,7 +274,7 @@ private:
 
     llama_cross cross; // TODO: tmp for handling cross-attention - need something better probably
 
-    std::unique_ptr<llama_memory_i> memory;
+    llama_memory_ptr memory;
 
     // decode output (2-dimensional array: [n_outputs][n_vocab])
     buffer_view<float> logits = {nullptr, 0};

src/llama-cparams.h

@@ -49,4 +49,6 @@ struct llama_cparams {
 
     ggml_backend_sched_eval_callback cb_eval;
     void * cb_eval_user_data;
+
+    llama_context * ctx_other;
 };

src/llama-ext.h

@@ -100,3 +100,5 @@ LLAMA_API float * llama_get_embeddings_nextn(struct llama_context * ctx);
 
 // LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);
 LLAMA_API float * llama_get_embeddings_nextn_ith(struct llama_context * ctx, int32_t i);
+
+LLAMA_API llama_context * llama_get_ctx_other(struct llama_context * ctx);

src/llama-graph.cpp

@@ -397,7 +397,7 @@ static void print_mask(const T * data, int64_t n_tokens, int64_t n_kv, int64_t n
         case LLAMA_SWA_TYPE_SYMMETRIC: swa_type_str = "LLAMA_SWA_TYPE_SYMMETRIC"; break;
     };
 
-    LLAMA_LOG_DEBUG("%s: n_swa : %d, n_kv: %d, swq_type: %s\n", __func__, (int)n_swa, (int)n_kv, swa_type_str);
+    LLAMA_LOG_DEBUG("%s: n_swa : %d, n_kv: %d, swa_type: %s\n", __func__, (int)n_swa, (int)n_kv, swa_type_str);
     LLAMA_LOG_DEBUG("%s: '0' = can attend, '∞' = masked\n", __func__);
     LLAMA_LOG_DEBUG("%s: Rows = query tokens, Columns = key/value tokens\n\n", __func__);
 
@@ -565,7 +565,10 @@ void llm_graph_input_attn_kv_iswa::set_input(const llama_ubatch * ubatch) {
     if (self_k_idxs && self_k_idxs->buffer) {
         mctx->get_base()->set_input_k_idxs(self_k_idxs, ubatch);
         mctx->get_base()->set_input_v_idxs(self_v_idxs, ubatch);
+    }
 
+    // the kq mask guards on its own buffer: shared cells leave idxs unbacked while the mask stays live
+    if (self_kq_mask && self_kq_mask->buffer) {
         mctx->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
     }
 
@@ -573,7 +576,9 @@ void llm_graph_input_attn_kv_iswa::set_input(const llama_ubatch * ubatch) {
     if (self_k_idxs_swa && self_k_idxs_swa->buffer) {
         mctx->get_swa()->set_input_k_idxs(self_k_idxs_swa, ubatch);
         mctx->get_swa()->set_input_v_idxs(self_v_idxs_swa, ubatch);
+    }
 
+    if (self_kq_mask_swa && self_kq_mask_swa->buffer) {
         mctx->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
     }
 
@@ -605,7 +610,9 @@ bool llm_graph_input_attn_kv_iswa::can_reuse(const llm_graph_params & params) {
     if (self_k_idxs && self_k_idxs->buffer) {
         res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
       //res &= self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
+    }
 
+    if (self_kq_mask && self_kq_mask->buffer) {
         res &= can_reuse_kq_mask(self_kq_mask, mctx->get_base(), params.ubatch, params.cparams);
     }
 
@@ -613,7 +620,9 @@ bool llm_graph_input_attn_kv_iswa::can_reuse(const llm_graph_params & params) {
     if (self_k_idxs_swa && self_k_idxs_swa->buffer) {
         res &= self_k_idxs_swa->ne[0] == params.ubatch.n_tokens;
       //res &= self_v_idxs_swa->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
+    }
 
+    if (self_kq_mask_swa && self_kq_mask_swa->buffer) {
         res &= can_reuse_kq_mask(self_kq_mask_swa, mctx->get_swa(), params.ubatch, params.cparams);
     }
 
@@ -756,7 +765,9 @@ void llm_graph_input_mem_hybrid_iswa::set_input(const llama_ubatch * ubatch) {
     if (inp_attn->self_k_idxs && inp_attn->self_k_idxs->buffer) {
         attn_ctx->get_base()->set_input_k_idxs(inp_attn->self_k_idxs, ubatch);
         attn_ctx->get_base()->set_input_v_idxs(inp_attn->self_v_idxs, ubatch);
+    }
 
+    if (inp_attn->self_kq_mask && inp_attn->self_kq_mask->buffer) {
         attn_ctx->get_base()->set_input_kq_mask(inp_attn->self_kq_mask, ubatch, cparams.causal_attn);
     }
 
@@ -764,7 +775,9 @@ void llm_graph_input_mem_hybrid_iswa::set_input(const llama_ubatch * ubatch) {
     if (inp_attn->self_k_idxs_swa && inp_attn->self_k_idxs_swa->buffer) {
         attn_ctx->get_swa()->set_input_k_idxs(inp_attn->self_k_idxs_swa, ubatch);
         attn_ctx->get_swa()->set_input_v_idxs(inp_attn->self_v_idxs_swa, ubatch);
+    }
 
+    if (inp_attn->self_kq_mask_swa && inp_attn->self_kq_mask_swa->buffer) {
         attn_ctx->get_swa()->set_input_kq_mask(inp_attn->self_kq_mask_swa, ubatch, cparams.causal_attn);
     }
 
@@ -810,18 +823,18 @@ bool llm_graph_input_mem_hybrid_iswa::can_reuse(const llm_graph_params & params)
     if (inp_attn->self_k_idxs && inp_attn->self_k_idxs->buffer) {
         res &= inp_attn->self_k_idxs->ne[0] == params.ubatch.n_tokens;
       //res &= inp_attn->self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
-
-        res &= can_reuse_kq_mask(inp_attn->self_kq_mask, attn_ctx->get_base(), params.ubatch, params.cparams);
     }
 
+    res &= can_reuse_kq_mask(inp_attn->self_kq_mask, attn_ctx->get_base(), params.ubatch, params.cparams);
+
     // swa tensors may not be allocated if there are no SWA attention layers
     if (inp_attn->self_k_idxs_swa && inp_attn->self_k_idxs_swa->buffer) {
         res &= inp_attn->self_k_idxs_swa->ne[0] == params.ubatch.n_tokens;
       //res &= inp_attn->self_v_idxs_swa->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
-
-        res &= can_reuse_kq_mask(inp_attn->self_kq_mask_swa, attn_ctx->get_swa(), params.ubatch, params.cparams);
     }
 
+    res &= can_reuse_kq_mask(inp_attn->self_kq_mask_swa, attn_ctx->get_swa(), params.ubatch, params.cparams);
+
     res &= inp_rs->s_copy->ne[0] == mctx->get_recr()->get_n_rs();
 
     res &= inp_rs->s_copy_main->ne[0]  == params.ubatch.n_seqs;
@@ -1006,6 +1019,7 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     ubatch           (params.ubatch),
     n_embd           (hparams.n_embd),
     n_layer          (hparams.n_layer()),
+    n_layer_nextn    (hparams.n_layer_nextn),
     n_rot            (hparams.n_rot()),
     n_ctx            (cparams.n_ctx),
     n_head           (hparams.n_head()),
@@ -1859,7 +1873,12 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
     res->t_inp_embd = cur;
 
     // For Granite architecture
-    if (hparams.f_embedding_scale != 0.0f) {
+    // NOTE: For deepstack models, only apply scale to token inputs (ie text-only input).
+    //  Raw embeddings are assumed to be multimodal inputs that should not be scaled.
+    if (hparams.f_embedding_scale != 0.0f && (ubatch.token || hparams.n_deepstack_layers == 0)) {
+        if (!ggml_is_contiguous(cur)) {
+            cur = ggml_cont(ctx0, cur);
+        }
         cur = ggml_scale(ctx0, cur, hparams.f_embedding_scale);
     }
 

src/llama-graph.h

@@ -784,6 +784,7 @@ struct llm_graph_context {
 
     const int64_t n_embd;
     const int64_t n_layer;
+    const int64_t n_layer_nextn;
     const int64_t n_rot;
     const int64_t n_ctx;       // user-specified context size (can be different from n_ctx_train)
     const int64_t n_head;

src/llama-hparams.cpp

@@ -91,6 +91,10 @@ uint32_t llama_hparams::n_rot(uint32_t il) const {
 }
 
 uint32_t llama_hparams::n_embd_inp() const {
+    if (n_embd_inp_impl > 0) {
+        return n_embd_inp_impl;
+    }
+
     uint32_t n_embd_inp = n_embd;
 
     if (n_deepstack_layers > 0) {

src/llama-hparams.h

@@ -185,6 +185,9 @@ struct llama_hparams {
     // for Classifiers
     uint32_t n_cls_out = 1;
 
+    // input embedding dimension (0 = use n_embd)
+    uint32_t n_embd_inp_impl = 0;
+
     // output embedding dimension (0 = use n_embd)
     uint32_t n_embd_out_impl = 0;
 
@@ -219,8 +222,19 @@ struct llama_hparams {
     uint32_t indexer_top_k     = 0;
 
     // qwen3vl deepstack
+    // When parsed from GGUF, this implies the first N layers consume the first
+    // N deepstack embeddings. Use deepstack_mapping_arr if you need a more
+    // complex mapping. If using deepstack_mapping_arr, also make sure to set
+    // n_deepstack_layers to the number of unique deepstack layers so that
+    // n_embd_imp is accurate (see granite.cpp).
+    // TODO: can be expressed via the `new n_embd_inp_impl` and remove this param
     uint32_t n_deepstack_layers = 0;
 
+    // deepstack layer array (Granite4 Vision)
+    // -1  => no deepstack
+    // >=0 => input embedding index for deepstack injection
+    std::array<int32_t, LLAMA_MAX_LAYERS> deepstack_mapping_arr;
+
     // gemma4 per-layer embedding
     uint32_t n_embd_per_layer = 0;
 

src/llama-kv-cache-dsa.cpp

@@ -32,7 +32,7 @@ llama_kv_cache_dsa::llama_kv_cache_dsa(
     kv_mla = std::make_unique<llama_kv_cache>(
             model, model.hparams, type_k, type_v,
             v_trans, offload, unified, kv_size, n_seq_max, n_pad,
-            n_swa, swa_type, filter, reuse);
+            n_swa, swa_type, nullptr, filter, reuse, nullptr);
 
     // we use llama_kv_cache for caching indexer keys
     // by hand-tweaking some hparams we fool it to create
@@ -49,7 +49,7 @@ llama_kv_cache_dsa::llama_kv_cache_dsa(
     kv_lid = std::make_unique<llama_kv_cache>(
             model, hparams_lid, type_k, type_v,
             v_trans, offload, unified, kv_size, n_seq_max, n_pad,
-            n_swa, swa_type, filter, reuse);
+            n_swa, swa_type, nullptr, filter, reuse, nullptr);
 }
 
 void llama_kv_cache_dsa::clear(bool data) {

src/llama-kv-cache-iswa.cpp

@@ -23,8 +23,10 @@ llama_kv_cache_iswa::llama_kv_cache_iswa(
                  uint32_t   n_seq_max,
                  uint32_t   n_ubatch,
                  uint32_t   n_pad,
+           llama_memory_t   mem_other,
     const layer_filter_cb & filter,
-    const  layer_reuse_cb & reuse) : hparams(model.hparams), unified(unified) {
+    const  layer_reuse_cb & reuse,
+    const  layer_share_cb & share) : hparams(model.hparams), unified(unified) {
 
     // chain filters
     const layer_filter_cb filter_base = [&](int32_t il) {
@@ -59,17 +61,27 @@ llama_kv_cache_iswa::llama_kv_cache_iswa(
 
     LLAMA_LOG_INFO("%s: creating non-SWA KV cache, size = %u cells\n", __func__, size_base);
 
+    llama_memory_t mem_other_base = nullptr;
+    if (mem_other) {
+        mem_other_base = static_cast<llama_kv_cache_iswa *>(mem_other)->get_base();
+    }
+
+    llama_memory_t mem_other_swa = nullptr;
+    if (mem_other) {
+        mem_other_swa = static_cast<llama_kv_cache_iswa *>(mem_other)->get_swa();
+    }
+
     kv_base = std::make_unique<llama_kv_cache>(
             model, hparams, type_k, type_v,
             v_trans, offload, unified, size_base, n_seq_max, n_pad,
-            0, LLAMA_SWA_TYPE_NONE, filter_base, reuse);
+            0, LLAMA_SWA_TYPE_NONE, mem_other_base, filter_base, reuse, share);
 
     LLAMA_LOG_INFO("%s: creating     SWA KV cache, size = %u cells\n", __func__, size_swa);
 
     kv_swa = std::make_unique<llama_kv_cache>(
             model, hparams, type_k, type_v,
             v_trans, offload, unified, size_swa, n_seq_max, n_pad,
-            hparams.n_swa, hparams.swa_type, filter_swa, reuse);
+            hparams.n_swa, hparams.swa_type, mem_other_swa, filter_swa, reuse, share);
 }
 
 void llama_kv_cache_iswa::clear(bool data) {

src/llama-kv-cache-iswa.h

@@ -25,8 +25,10 @@ public:
                      uint32_t   n_seq_max,
                      uint32_t   n_ubatch,
                      uint32_t   n_pad,
+               llama_memory_t   mem_other,
         const layer_filter_cb & filter,
-        const  layer_reuse_cb & reuse);
+        const  layer_reuse_cb & reuse,
+        const  layer_share_cb & share);
 
     ~llama_kv_cache_iswa() = default;
 

src/llama-kv-cache.cpp

@@ -90,10 +90,26 @@ llama_kv_cache::llama_kv_cache(
                  uint32_t   n_pad,
                  uint32_t   n_swa,
            llama_swa_type   swa_type,
+           llama_memory_t   mem_other,
     const layer_filter_cb & filter,
-    const  layer_reuse_cb & reuse) :
+    const  layer_reuse_cb & reuse,
+    const  layer_share_cb & share) :
     model(model), hparams(hparams), v_trans(v_trans),
-    n_seq_max(n_seq_max), n_stream(unified ? 1 : n_seq_max), n_pad(n_pad), n_swa(n_swa), swa_type(swa_type) {
+    n_seq_max(n_seq_max), n_stream(unified ? 1 : n_seq_max), n_pad(n_pad), n_swa(n_swa), swa_type(swa_type),
+    other(static_cast<llama_kv_cache *>(mem_other)),
+    v_cells_impl(other ? other->v_cells_impl : std::make_shared<llama_kv_cells_vec>()),
+    v_cells(*v_cells_impl) {
+
+    // shared cells view the source cache's K/V tensors, so the cell count
+    // follows the source allocation: a fitted target can be smaller than the
+    // draft default and oversized views would overflow the source tensors
+    if (other) {
+        const uint32_t size_other = other->get_size();
+        if (kv_size != size_other) {
+            LLAMA_LOG_WARN("%s: kv_size = %u overridden to %u to match the shared source cache\n", __func__, kv_size, size_other);
+            kv_size = size_other;
+        }
+    }
 
     GGML_ASSERT(kv_size % n_pad == 0);
 
@@ -171,6 +187,24 @@ llama_kv_cache::llama_kv_cache(
             continue;
         }
 
+        if (share && other) {
+            const int32_t il_share = share(il);
+
+            if (il_share >= 0) {
+                const auto & layer_share = other->layers[other->map_layer_ids[il_share]];
+
+                LLAMA_LOG_WARN("%s: layer %3d: sharing with layer %d. k = %p, v = %p\n", __func__, il, il_share,
+                        layer_share.k->data, layer_share.v->data);
+
+                map_layer_ids[il] = layers.size();
+
+                layers.push_back(layer_share);
+                layers.back().il = il;
+
+                continue;
+            }
+        }
+
         if (n_embd_head_k_all == 0) {
             n_embd_head_k_all = (int32_t) hparams.n_embd_head_k(il);
         } else if (n_embd_head_k_all > 0 && n_embd_head_k_all != (int32_t) hparams.n_embd_head_k(il)) {
@@ -282,29 +316,38 @@ llama_kv_cache::llama_kv_cache(
                 ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f));
     }
 
-    const char * LLAMA_ATTN_ROT_DISABLE = getenv("LLAMA_ATTN_ROT_DISABLE");
-    const bool attn_rot_disable = LLAMA_ATTN_ROT_DISABLE ? atoi(LLAMA_ATTN_ROT_DISABLE) : false;
-    if (attn_rot_disable) {
-        LLAMA_LOG_WARN("%s: attention rotation force disabled (LLAMA_ATTN_ROT_DISABLE)\n", __func__);
+    // TODO: refactor [TAG_KV_CACHE_SHARE_CELLS]
+    if (other) {
+        n_embd_head_k_all = other->n_embd_head_k_all;
+        n_embd_head_v_all = other->n_embd_head_v_all;
+
+        attn_rot_k = other->attn_rot_k;
+        attn_rot_v = other->attn_rot_v;
+    } else {
+        const char * LLAMA_ATTN_ROT_DISABLE = getenv("LLAMA_ATTN_ROT_DISABLE");
+        const bool attn_rot_disable = LLAMA_ATTN_ROT_DISABLE ? atoi(LLAMA_ATTN_ROT_DISABLE) : false;
+        if (attn_rot_disable) {
+            LLAMA_LOG_WARN("%s: attention rotation force disabled (LLAMA_ATTN_ROT_DISABLE)\n", __func__);
+        }
+
+        attn_rot_k =
+            !attn_rot_disable &&
+            n_embd_head_k_all > 0 &&
+            ggml_is_quantized(type_k) &&
+            hparams.n_embd_head_k() % 64 == 0;
+
+        // always create Hadamard rotation tensors for DeepSeek V3.2 DSA lightning indexer
+        if (model.arch == LLM_ARCH_DEEPSEEK32 && hparams.n_embd_head_k_full == hparams.indexer_head_size) {
+            attn_rot_k = true;
+        }
+
+        attn_rot_v =
+            !attn_rot_disable &&
+            n_embd_head_v_all > 0 &&
+            ggml_is_quantized(type_v) &&
+            hparams.n_embd_head_v() % 64 == 0;
     }
 
-    attn_rot_k =
-        !attn_rot_disable &&
-        n_embd_head_k_all > 0 &&
-        ggml_is_quantized(type_k) &&
-        hparams.n_embd_head_k() % 64 == 0;
-
-    // always create Hadamard rotation tensors for DeepSeek V3.2 DSA lightning indexer
-    if (model.arch == LLM_ARCH_DEEPSEEK32 && hparams.n_embd_head_k_full == hparams.indexer_head_size) {
-        attn_rot_k = true;
-    }
-
-    attn_rot_v =
-        !attn_rot_disable &&
-        n_embd_head_v_all > 0 &&
-        ggml_is_quantized(type_v) &&
-        hparams.n_embd_head_v() % 64 == 0;
-
     LLAMA_LOG_INFO("%s: attn_rot_k = %d, n_embd_head_k_all = %d\n", __func__, attn_rot_k, n_embd_head_k_all);
     LLAMA_LOG_INFO("%s: attn_rot_v = %d, n_embd_head_k_all = %d\n", __func__, attn_rot_v, n_embd_head_v_all);
 
@@ -347,6 +390,11 @@ void llama_kv_cache::clear(bool data) {
 }
 
 bool llama_kv_cache::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+    // TODO: refactor [TAG_KV_CACHE_SHARE_CELLS]
+    if (other) {
+        return true;
+    }
+
     GGML_ASSERT(seq_id == -1 || (seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()));
 
     if (p0 < 0) {
@@ -410,6 +458,11 @@ bool llama_kv_cache::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
 }
 
 void llama_kv_cache::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+    // TODO: refactor [TAG_KV_CACHE_SHARE_CELLS]
+    if (other) {
+        return;
+    }
+
     GGML_ASSERT(seq_id_src >= 0 && (size_t) seq_id_src < seq_to_stream.size());
     GGML_ASSERT(seq_id_dst >= 0 && (size_t) seq_id_dst < seq_to_stream.size());
 
@@ -497,6 +550,11 @@ void llama_kv_cache::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, ll
 }
 
 void llama_kv_cache::seq_keep(llama_seq_id seq_id) {
+    // TODO: refactor [TAG_KV_CACHE_SHARE_CELLS]
+    if (other) {
+        return;
+    }
+
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
 
     auto & cells = v_cells[seq_to_stream[seq_id]];
@@ -519,6 +577,11 @@ void llama_kv_cache::seq_keep(llama_seq_id seq_id) {
 }
 
 void llama_kv_cache::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
+    // TODO: refactor [TAG_KV_CACHE_SHARE_CELLS]
+    if (other) {
+        return;
+    }
+
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
     GGML_ASSERT(hparams.n_pos_per_embd() == 1 && "seq_add() is only supported for n_pos_per_embd() == 1");
 
@@ -564,6 +627,11 @@ void llama_kv_cache::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, ll
 }
 
 void llama_kv_cache::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+    // TODO: refactor [TAG_KV_CACHE_SHARE_CELLS]
+    if (other) {
+        return;
+    }
+
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
     GGML_ASSERT(hparams.n_pos_per_embd() == 1 && "seq_div() is only supported for n_pos_per_embd() == 1");
 
@@ -598,6 +666,11 @@ void llama_kv_cache::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, in
 }
 
 llama_pos llama_kv_cache::seq_pos_min(llama_seq_id seq_id) const {
+    // TODO: refactor [TAG_KV_CACHE_SHARE_CELLS]
+    if (other) {
+        return other->seq_pos_min(seq_id);
+    }
+
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
 
     const auto & cells = v_cells[seq_to_stream[seq_id]];
@@ -606,6 +679,11 @@ llama_pos llama_kv_cache::seq_pos_min(llama_seq_id seq_id) const {
 }
 
 llama_pos llama_kv_cache::seq_pos_max(llama_seq_id seq_id) const {
+    // TODO: refactor [TAG_KV_CACHE_SHARE_CELLS]
+    if (other) {
+        return other->seq_pos_max(seq_id);
+    }
+
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
 
     const auto & cells = v_cells[seq_to_stream[seq_id]];
@@ -746,6 +824,11 @@ llama_kv_cache::slot_info_vec_t llama_kv_cache::prepare(const std::vector<llama_
 }
 
 bool llama_kv_cache::update(llama_context * lctx, bool do_shift, const stream_copy_info & sc_info) {
+    // TODO: refactor [TAG_KV_CACHE_SHARE_CELLS]
+    if (other) {
+        return true;
+    }
+
     bool updated = false;
 
     auto * sched = lctx->get_sched();
@@ -1021,6 +1104,11 @@ llama_kv_cache::slot_info llama_kv_cache::find_slot(const llama_ubatch & ubatch,
 }
 
 void llama_kv_cache::apply_ubatch(const slot_info & sinfo, const llama_ubatch & ubatch) {
+    // TODO: refactor [TAG_KV_CACHE_SHARE_CELLS]
+    if (other) {
+        return;
+    }
+
     // keep track of the max sequence position that we would overwrite with this ubatch
     // for non-SWA cache, this would be always empty
     llama_seq_id seq_pos_max_rm[LLAMA_MAX_SEQ];
@@ -1815,6 +1903,9 @@ void llm_graph_input_k_shift::set_input(const llama_ubatch * ubatch) {
 }
 
 ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_context * lctx) const {
+    // TODO: refactor [TAG_KV_CACHE_SHARE_CELLS]
+    GGML_ASSERT(!other);
+
     auto * ctx = res->get_ctx();
     auto * gf  = res->get_gf();
 
@@ -1860,6 +1951,11 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
 }
 
 void llama_kv_cache::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    // TODO: refactor [TAG_KV_CACHE_SHARE_CELLS]
+    if (other) {
+        return;
+    }
+
     GGML_UNUSED(flags);
 
     io.write(&n_stream, sizeof(n_stream));
@@ -1925,6 +2021,11 @@ void llama_kv_cache::state_write(llama_io_write_i & io, llama_seq_id seq_id, lla
 }
 
 void llama_kv_cache::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    // TODO: refactor [TAG_KV_CACHE_SHARE_CELLS]
+    if (other) {
+        return;
+    }
+
     GGML_UNUSED(flags);
 
     GGML_ASSERT(seq_id == -1 || (seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()));

src/llama-kv-cache.h

@@ -98,7 +98,7 @@ public:
     //       likely through `struct llama_memory_params`
     llama_kv_cache(
             const llama_model & model,
-            const llama_hparams & hparams,
+          const llama_hparams & hparams,
                     ggml_type   type_k,
                     ggml_type   type_v,
                          bool   v_trans,
@@ -109,8 +109,10 @@ public:
                      uint32_t   n_pad,
                      uint32_t   n_swa,
                llama_swa_type   swa_type,
+               llama_memory_t   mem_other,
         const layer_filter_cb & filter,
-        const  layer_reuse_cb & reuse);
+        const  layer_reuse_cb & reuse,
+        const  layer_share_cb & share);
 
     ~llama_kv_cache() = default;
 
@@ -264,7 +266,12 @@ private:
     // note: this is not part of the KV state and it's only used to speed-up the find_slot() method
     std::vector<uint32_t> v_heads;
 
-    std::vector<llama_kv_cells> v_cells;
+    // TODO: temporary until we refactor to be able to share the same cells between 2 kv caches [TAG_KV_CACHE_SHARE_CELLS]
+    llama_kv_cache * other;
+
+    std::shared_ptr<llama_kv_cells_vec> v_cells_impl;
+
+    llama_kv_cells_vec & v_cells;
 
     // maps from a sequence id to a stream id
     std::vector<uint32_t> seq_to_stream;

src/llama-kv-cells.h

@@ -531,3 +531,5 @@ private:
         }
     }
 };
+
+using llama_kv_cells_vec = std::vector<llama_kv_cells>;

src/llama-memory-hybrid-iswa.cpp

@@ -43,9 +43,11 @@ llama_memory_hybrid_iswa::llama_memory_hybrid_iswa(
         n_seq_max,
         n_ubatch,
         n_pad,
+        nullptr,
         filter_attn == nullptr ?
             [&](int32_t il) { return !hparams.is_recr(il); }
             : filter_attn,
+        nullptr,
         nullptr
     )),
     mem_recr(new llama_memory_recurrent(

src/llama-memory-hybrid.cpp

@@ -44,9 +44,11 @@ llama_memory_hybrid::llama_memory_hybrid(
         n_pad,
         n_swa,
         swa_type,
+        nullptr,
         filter_attn == nullptr ?
             [&](int32_t il) { return !hparams.is_recr(il); }
             : filter_attn,
+        nullptr,
         nullptr
     )),
     mem_recr(new llama_memory_recurrent(

src/llama-memory.h

@@ -23,6 +23,8 @@ struct llama_memory_params {
     bool swa_full;
 
     llama_context_type ctx_type;
+
+    llama_memory_t mem_other;
 };
 
 enum llama_memory_status {
@@ -76,6 +78,8 @@ struct llama_memory_i {
     // return negative value to indicate that the layer il should not reuse memory
     using layer_reuse_cb = std::function<int32_t(int32_t il)>;
 
+    using layer_share_cb = std::function<int32_t(int32_t il)>;
+
     virtual ~llama_memory_i() = default;
 
     // split the input batch into a set of ubatches and verify that they can fit into the cache

src/llama-model-loader.cpp

@@ -393,6 +393,7 @@ namespace GGUFMeta {
     }
 
     template bool llama_model_loader::get_arr<std::vector<std::string>>(enum llm_kv kid, std::vector<std::string> & result, bool required);
+    template bool llama_model_loader::get_arr<std::array<int32_t, 512>>(enum llm_kv kid, std::array<int32_t, 512> & result, bool required);
 
     template<typename T>
     bool llama_model_loader::get_key(const std::string & key, T & result, bool required) {

src/llama-model-saver.cpp

@@ -229,6 +229,7 @@ void llama_model_saver::add_kv_from_model() {
     add_kv(LLM_KV_MOE_EVERY_N_LAYERS,                hparams.moe_every_n_layers);
     add_kv(LLM_KV_NEXTN_PREDICT_LAYERS,              hparams.n_layer_nextn);
     add_kv(LLM_KV_NUM_DEEPSTACK_LAYERS,              hparams.n_deepstack_layers);
+    add_kv(LLM_KV_DEEPSTACK_MAPPING,                 hparams.deepstack_mapping_arr);
     add_kv(LLM_KV_POOLING_TYPE,                      uint32_t(hparams.pooling_type));
     add_kv(LLM_KV_LOGIT_SCALE,                       hparams.f_logit_scale);
     add_kv(LLM_KV_DECODER_START_TOKEN_ID,            hparams.dec_start_token_id);

src/llama-model.cpp

@@ -139,6 +139,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
             return new llama_model_gemma3n(params);
         case LLM_ARCH_GEMMA4:
             return new llama_model_gemma4(params);
+        case LLM_ARCH_GEMMA4_ASSISTANT:
+            return new llama_model_gemma4_assistant(params);
         case LLM_ARCH_GEMMA_EMBEDDING:
             return new llama_model_gemma_embedding(params);
         case LLM_ARCH_STARCODER2:
@@ -553,10 +555,12 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
     };
 
     auto get_split_granularity = [&](int64_t blck_size, uint32_t il, const std::vector<std::pair<int64_t, uint32_t>> & segments) -> std::vector<int64_t> {
+        // for better performance it may make sense to round up blck_size to a higher power of 2 so that more efficient kernels can be used
         if (hparams.is_recr(il)) {
             // linear attention
-            const int64_t head_dim  = hparams.ssm_d_state;
-            const int64_t granularity_qkv = std::lcm(blck_size, head_dim);
+            const int64_t head_dim        = hparams.ssm_d_state;
+            const int64_t blck_size_perf  = std::lcm(blck_size, 128);
+            const int64_t granularity_qkv = std::lcm(blck_size_perf, head_dim);
             if (std::regex_match(tensor_name, pattern_qkv_weight) || std::regex_match(tensor_name, pattern_attn_gate_weight) ||
                     std::regex_match(tensor_name, pattern_ssm_conv1d) || std::regex_match(tensor_name, pattern_ssm_out_weight)) {
                 return std::vector<int64_t>(segments.size(), granularity_qkv);
@@ -578,17 +582,24 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
             // regular attention
             const uint32_t n_gqa    = hparams.n_gqa(il);
             const uint32_t n_embd_q = n_gqa * hparams.n_embd_head_k(il);
-            if (std::regex_match(tensor_name, pattern_attn_sinks)) {
-                GGML_ASSERT(segments.size() == 1);
-                return {std::lcm(n_embd_q, blck_size)/n_embd_q * n_gqa};
+
+            // to handle head sizes like 80, only increase granularity while it doesn't cause underutilization
+            int64_t blck_size_perf = blck_size;
+            while (blck_size_perf < 128 && blck_size_perf*ud->n_devices < n_embd_q) {
+                blck_size_perf *= 2;
             }
 
-            const int64_t granularity_q = std::lcm(n_embd_q, blck_size);
+            if (std::regex_match(tensor_name, pattern_attn_sinks)) {
+                GGML_ASSERT(segments.size() == 1);
+                return {std::lcm(n_embd_q, blck_size_perf)/n_embd_q * n_gqa};
+            }
+
+            const int64_t granularity_q = std::lcm(n_embd_q, blck_size_perf);
             if (std::regex_match(tensor_name, pattern_q_weight) || std::regex_match(tensor_name, pattern_q_bias)) {
                 GGML_ASSERT(segments.size() == 1);
                 // some models have Q gate tensors, for those cases the granularity needs to be doubled:
                 if (ud->model->arch == LLM_ARCH_QWEN3NEXT || ud->model->arch == LLM_ARCH_QWEN35 || ud->model->arch == LLM_ARCH_QWEN35MOE) {
-                    return {std::lcm(2*n_embd_q, blck_size)};
+                    return {std::lcm(2*n_embd_q, blck_size_perf)};
                 }
                 return {granularity_q};
             }
@@ -613,8 +624,9 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
         // FFN
         if (std::regex_match(tensor_name, pattern_ffn_up_gate_weight) || std::regex_match(tensor_name, pattern_ffn_up_gate_bias) ||
                 std::regex_match(tensor_name, pattern_ffn_gate_up_weight) || std::regex_match(tensor_name, pattern_ffn_down_weight)) {
+            const int64_t blck_size_perf = std::lcm(blck_size, 128);
             GGML_ASSERT(segments.size() == 1);
-            return {blck_size};
+            return {blck_size_perf};
         }
 
         // everything else
@@ -627,7 +639,6 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
     tensor_config tc = get_tensor_config();
     split_state.axis = tc.axis;
     if (split_state.axis >= 0 && split_state.axis < GGML_MAX_DIMS) {
-        const int64_t ne_full = tensor->ne[split_state.axis];
         const int64_t blck_size = ggml_blck_size(tc.tensor_axis_0->type);
         const float * tensor_split = ud->model->tensor_split();
         std::vector<float> tensor_split_scan;
@@ -644,7 +655,6 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
             const int64_t  ne_s = segments[is].first;
             const uint32_t nr_s = segments[is].second;
             const int64_t  g_s  = granularity[is];
-            GGML_ASSERT(ne_full % g_s == 0);
             int64_t low = 0;
             size_t j = 0;
             for (; j < ud->n_devices - 1; j++) {
@@ -1092,6 +1102,9 @@ void llama_model_base::load_hparams(llama_model_loader & ml) {
     ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH,  hparams.n_ff_arr,   hparams.n_layer(), false);
     ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer(), false);
 
+    // Populate deepstack_mapping_arr - initialized to -1 (no deepstack)
+    std::fill(hparams.deepstack_mapping_arr.begin(), hparams.deepstack_mapping_arr.end(), -1);
+
     // n_head_kv is optional, default to n_head
     hparams.n_head_kv_arr = hparams.n_head_arr;
 
@@ -1194,7 +1207,7 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
     const auto & use_mlock    = params.use_mlock;
     const auto & tensor_split = params.tensor_split;
 
-    const int n_layer = hparams.n_layer_all;
+    const int n_layer_all = hparams.n_layer_all;
     const int n_gpu_layers = this->n_gpu_layers();
 
     const bool use_mmap_buffer = true;
@@ -1251,10 +1264,10 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
         splits[i] /= split_sum;
     }
 
-    const int i_gpu_start = std::max(n_layer + 1 - n_gpu_layers, 0);
-    const int act_gpu_layers = devices.empty() ? 0 : std::min(n_gpu_layers, n_layer + 1);
+    const int i_gpu_start = std::max(n_layer_all + 1 - n_gpu_layers, 0);
+    const int act_gpu_layers = devices.empty() ? 0 : std::min(n_gpu_layers, n_layer_all + 1);
     auto get_layer_buft_list = [&](int il) -> llama_model::impl::layer_dev {
-        const bool is_swa = il < n_layer && hparams.is_swa(il);
+        const bool is_swa = il < n_layer_all && hparams.is_swa(il);
         if (il < i_gpu_start || (il - i_gpu_start) >= act_gpu_layers) {
             LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s, is_swa = %d\n", il, ggml_backend_dev_name(cpu_dev), is_swa);
             return {cpu_dev, &pimpl->cpu_buft_list};
@@ -1270,13 +1283,13 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
     pimpl->dev_input = { cpu_dev, &pimpl->cpu_buft_list };
 
     // assign the repeating layers to the devices according to the splits
-    pimpl->dev_layer.resize(n_layer);
-    for (int il = 0; il < n_layer; ++il) {
+    pimpl->dev_layer.resize(n_layer_all);
+    for (int il = 0; il < n_layer_all; ++il) {
         pimpl->dev_layer[il] = get_layer_buft_list(il);
     }
 
     // assign the output layer
-    pimpl->dev_output = get_layer_buft_list(n_layer);
+    pimpl->dev_output = get_layer_buft_list(n_layer_all);
 
     const auto TENSOR_NOT_REQUIRED = llama_model_loader::TENSOR_NOT_REQUIRED;
 
@@ -1292,14 +1305,14 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
             throw std::runtime_error("model has expert layers but no expert layers are used");
         }
 
-        layers.resize(n_layer);
+        layers.resize(n_layer_all);
 
         // call the per-model loading function
         load_arch_tensors(ml);
 
         // generic pass: load optional per-tensor/per-expert ".scale" tensors (e.g. NVFP4 scale2)
         // this avoids having to add scale loading to every architecture
-        for (int i = 0; i < n_layer; ++i) {
+        for (int i = 0; i < n_layer_all; ++i) {
             auto & layer = layers[i];
 
             // attention weight scales (per-tensor, shape {1})
@@ -1557,7 +1570,7 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
     }
 
     if (llama_supports_gpu_offload()) {
-        const int n_gpu = std::min(n_gpu_layers, n_layer);
+        const int n_gpu = std::min(n_gpu_layers, n_layer_all);
 
         int n_repeating = n_gpu;
         if (n_repeating > 0) {
@@ -1566,8 +1579,8 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
         }
         LLAMA_LOG_INFO("%s: offloading %d repeating layers to GPU\n", __func__, n_repeating);
 
-        const int max_backend_supported_layers = n_layer + 1;
-        const int max_offloadable_layers       = n_layer + 1;
+        const int max_backend_supported_layers = n_layer_all + 1;
+        const int max_offloadable_layers       = n_layer_all + 1;
 
         LLAMA_LOG_INFO("%s: offloaded %d/%d layers to GPU\n", __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers);
     }
@@ -1670,10 +1683,10 @@ uint64_t llama_model::n_elements() const {
 void llama_model::print_info() const {
     const std::string rope_scaling_type = llama_rope_scaling_type_name(hparams.rope_scaling_type_train);
 
-    auto print_f = [](const std::function<uint32_t(uint32_t)> & f, uint32_t n) {
+    auto print_f = [](const std::function<int32_t(uint32_t)> & f, uint32_t n) {
         bool is_var = false;
 
-        std::vector<uint32_t> v;
+        std::vector<int32_t> v;
         for (uint32_t i = 0; i < n; ++i) {
             v.push_back(f(i));
             if (v[i] != v[0]) {
@@ -1706,19 +1719,21 @@ void llama_model::print_info() const {
 
     if (!hparams.vocab_only) {
         LLAMA_LOG_INFO("%s: n_ctx_train           = %u\n",     __func__, hparams.n_ctx_train);
-        LLAMA_LOG_INFO("%s: n_embd                = %u\n",     __func__, hparams.n_embd);
         LLAMA_LOG_INFO("%s: n_embd_inp            = %u\n",     __func__, hparams.n_embd_inp());
+        LLAMA_LOG_INFO("%s: n_embd                = %u\n",     __func__, hparams.n_embd);
+        LLAMA_LOG_INFO("%s: n_embd_out            = %u\n",     __func__, hparams.n_embd_out());
         LLAMA_LOG_INFO("%s: n_layer               = %u\n",     __func__, hparams.n_layer());
-        LLAMA_LOG_INFO("%s: n_head                = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_head(il);    }, hparams.n_layer()).c_str());
-        LLAMA_LOG_INFO("%s: n_head_kv             = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer()).c_str());
+        LLAMA_LOG_INFO("%s: n_layer_all           = %u\n",     __func__, hparams.n_layer_all);
+        LLAMA_LOG_INFO("%s: n_head                = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_head(il);    }, hparams.n_layer_all).c_str());
+        LLAMA_LOG_INFO("%s: n_head_kv             = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer_all).c_str());
         LLAMA_LOG_INFO("%s: n_rot                 = %u\n",     __func__, hparams.n_rot_full);
         LLAMA_LOG_INFO("%s: n_swa                 = %u\n",     __func__, hparams.n_swa);
         LLAMA_LOG_INFO("%s: is_swa_any            = %u\n",     __func__, hparams.is_swa_any());
         LLAMA_LOG_INFO("%s: n_embd_head_k         = %u\n",     __func__, hparams.n_embd_head_k_full);
         LLAMA_LOG_INFO("%s: n_embd_head_v         = %u\n",     __func__, hparams.n_embd_head_v_full);
-        LLAMA_LOG_INFO("%s: n_gqa                 = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_gqa(il);        }, hparams.n_layer()).c_str());
-        LLAMA_LOG_INFO("%s: n_embd_k_gqa          = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_embd_k_gqa(il); }, hparams.n_layer()).c_str());
-        LLAMA_LOG_INFO("%s: n_embd_v_gqa          = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_embd_v_gqa(il); }, hparams.n_layer()).c_str());
+        LLAMA_LOG_INFO("%s: n_gqa                 = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_gqa(il);        }, hparams.n_layer_all).c_str());
+        LLAMA_LOG_INFO("%s: n_embd_k_gqa          = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_embd_k_gqa(il); }, hparams.n_layer_all).c_str());
+        LLAMA_LOG_INFO("%s: n_embd_v_gqa          = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_embd_v_gqa(il); }, hparams.n_layer_all).c_str());
         LLAMA_LOG_INFO("%s: f_norm_eps            = %.1e\n",   __func__, hparams.f_norm_eps);
         LLAMA_LOG_INFO("%s: f_norm_rms_eps        = %.1e\n",   __func__, hparams.f_norm_rms_eps);
         LLAMA_LOG_INFO("%s: f_clamp_kqv           = %.1e\n",   __func__, hparams.f_clamp_kqv);
@@ -1726,7 +1741,7 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: f_logit_scale         = %.1e\n",   __func__, hparams.f_logit_scale);
         LLAMA_LOG_INFO("%s: f_attn_scale          = %.1e\n",   __func__, hparams.f_attention_scale);
         LLAMA_LOG_INFO("%s: f_attn_value_scale    = %.4f\n",   __func__, hparams.f_attn_value_scale);
-        LLAMA_LOG_INFO("%s: n_ff                  = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_ff(il); }, hparams.n_layer()).c_str());
+        LLAMA_LOG_INFO("%s: n_ff                  = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_ff(il); }, hparams.n_layer_all).c_str());
         LLAMA_LOG_INFO("%s: n_expert              = %u\n",     __func__, hparams.n_expert);
         LLAMA_LOG_INFO("%s: n_expert_used         = %u\n",     __func__, hparams.n_expert_used);
         LLAMA_LOG_INFO("%s: n_expert_groups       = %d\n",     __func__, hparams.n_expert_groups);
@@ -1747,6 +1762,14 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_ctx_orig_yarn       = %u\n",     __func__, hparams.n_ctx_orig_yarn);
         LLAMA_LOG_INFO("%s: rope_yarn_log_mul     = %.4f\n",   __func__, hparams.rope_yarn_log_mul);
         LLAMA_LOG_INFO("%s: rope_finetuned        = %s\n",     __func__, hparams.rope_finetuned ? "yes" : "unknown");
+        if (arch == LLM_ARCH_GRANITE &&
+            std::any_of(hparams.deepstack_mapping_arr.begin(),
+                        hparams.deepstack_mapping_arr.end(),
+                        [](const auto & entry) { return entry >= 0; })) {
+            LLAMA_LOG_INFO("%s: deepstack_mapping_arr = %s\n", __func__,
+                           print_f([&](uint32_t il) { return hparams.deepstack_mapping_arr[il]; },
+                           hparams.n_layer_all).c_str());
+        }
         // MRoPE (Multi-axis Rotary Position Embedding) sections
         if (const auto & s = hparams.rope_sections; s[0] || s[1] || s[2] || s[3]) {
             LLAMA_LOG_INFO("%s: mrope sections        = [%d, %d, %d, %d]\n", __func__, s[0], s[1], s[2], s[3]);
@@ -2094,8 +2117,9 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                             /* filter_recr       */ std::move(filter_recr));
                     }
                 } else {
-                    llama_memory_i::layer_reuse_cb reuse = nullptr;
                     llama_kv_cache::layer_filter_cb filter = nullptr;
+                    llama_memory_i::layer_reuse_cb reuse = nullptr;
+                    llama_kv_cache::layer_share_cb share = nullptr;
 
                     if (arch == LLM_ARCH_GEMMA3N || arch == LLM_ARCH_GEMMA4) {
                         reuse = [&](uint32_t il) {
@@ -2124,20 +2148,53 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                     if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
                         GGML_ASSERT(hparams.is_swa_any());
 
-                        res = new llama_kv_cache_iswa(
-                                *this,
-                                params.type_k,
-                                params.type_v,
-                                !cparams.flash_attn,
-                                cparams.offload_kqv,
-                                params.swa_full,
-                                cparams.kv_unified,
-                                cparams.n_ctx_seq,
-                                cparams.n_seq_max,
-                                cparams.n_ubatch,
-                                1,
-                                filter,
-                                reuse);
+                        if (arch == LLM_ARCH_GEMMA4_ASSISTANT) {
+                            llama_memory_t mem_other = llama_get_memory(cparams.ctx_other);
+
+                            share = [&](int32_t il) {
+                                const llama_model * model_other = llama_get_model(cparams.ctx_other);
+
+                                if (hparams.is_swa(il)) {
+                                    return llama_model_n_layer(model_other) - 2;
+                                }
+
+                                return llama_model_n_layer(model_other) - 1;
+                            };
+
+                            res = new llama_kv_cache_iswa(
+                                    *this,
+                                    params.type_k,
+                                    params.type_v,
+                                    !cparams.flash_attn,
+                                    cparams.offload_kqv,
+                                    params.swa_full,
+                                    cparams.kv_unified,
+                                    cparams.n_ctx_seq,
+                                    cparams.n_seq_max,
+                                    cparams.n_ubatch,
+                                    1,
+                                    mem_other,
+                                    filter,
+                                    reuse,
+                                    share);
+                        } else {
+                            res = new llama_kv_cache_iswa(
+                                    *this,
+                                    params.type_k,
+                                    params.type_v,
+                                    !cparams.flash_attn,
+                                    cparams.offload_kqv,
+                                    params.swa_full,
+                                    cparams.kv_unified,
+                                    cparams.n_ctx_seq,
+                                    cparams.n_seq_max,
+                                    cparams.n_ubatch,
+                                    1,
+                                    nullptr,
+                                    filter,
+                                    reuse,
+                                    share);
+                        }
                     } else {
                         GGML_ASSERT(!hparams.is_swa_any());
 
@@ -2154,7 +2211,9 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                                 1,
                                 hparams.n_swa,
                                 hparams.swa_type,
+                                nullptr,
                                 filter,
+                                nullptr,
                                 nullptr);
                     }
                 }
@@ -2387,6 +2446,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_GEMMA3:
         case LLM_ARCH_GEMMA3N:
         case LLM_ARCH_GEMMA4:
+        case LLM_ARCH_GEMMA4_ASSISTANT:
         case LLM_ARCH_GEMMA_EMBEDDING:
         case LLM_ARCH_STARCODER2:
         case LLM_ARCH_OPENELM:

src/llama-model.h

@@ -548,6 +548,10 @@ struct llama_model {
     struct ggml_tensor * output_s    = nullptr;
     struct ggml_tensor * output_in_s = nullptr;
 
+    // NextN/MTP model-level projections
+    struct ggml_tensor * nextn_proj_pre  = nullptr;
+    struct ggml_tensor * nextn_proj_post = nullptr;
+
     // classifier
     struct ggml_tensor * cls       = nullptr;
     struct ggml_tensor * cls_b     = nullptr;
@@ -702,6 +706,7 @@ const char * llm_type_name(llm_type type);
 #define LLAMA_LOAD_LOCALS \
     const int     n_layer        = hparams.n_layer();        GGML_UNUSED(n_layer); \
     const int     n_layer_all    = hparams.n_layer_all;      GGML_UNUSED(n_layer_all); \
+    const int     n_layer_nextn  = hparams.n_layer_nextn;    GGML_UNUSED(n_layer_nextn); \
     const int64_t n_head         = hparams.n_head();         GGML_UNUSED(n_head); \
     const int64_t n_head_kv      = hparams.n_head_kv();      GGML_UNUSED(n_head_kv); \
     const int64_t n_embd         = hparams.n_embd;           GGML_UNUSED(n_embd); \

src/llama-vocab.cpp

@@ -764,7 +764,7 @@ struct llm_tokenizer_wpm_session {
 
     void tokenize(const std::string & text, std::vector<llama_token> & output) {
         // normalize and split by whitespace
-        std::vector<std::string> words = preprocess(text, vocab.get_normalizer_lowercase());
+        std::vector<std::string> words = preprocess(text, vocab.get_normalizer_opts());
         // bos token prepended already
 
         // find the longest tokens that form the words
@@ -809,11 +809,14 @@ struct llm_tokenizer_wpm_session {
     }
 
     // TODO: reduce string copies by using cpts_offs array
-    static std::vector<std::string> preprocess(const std::string & text, bool lowercase)  {
-        const std::vector<uint32_t> cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text));
+    static std::vector<std::string> preprocess(const std::string & text, const llama_vocab::normalizer_options & normalizer_opts)  {
+        std::vector<uint32_t> cpts = unicode_cpts_from_utf8(text);
+        if (normalizer_opts.strip_accents) {
+            cpts = unicode_cpts_normalize_nfd(cpts);
+        }
         std::vector<std::string> words(1, "");
 
-        for (const uint32_t cpt : cpts_nfd) {
+        for (const uint32_t cpt : cpts) {
             const auto flags = unicode_cpt_flags_from_cpt(cpt);
 
             if (flags.is_whitespace) {
@@ -828,7 +831,11 @@ struct llm_tokenizer_wpm_session {
                 continue;
             }
 
-            const std::string s = unicode_cpt_to_utf8(lowercase ? unicode_tolower(cpt) : cpt);
+            if (normalizer_opts.strip_accents && flags.is_accent_mark) {
+                continue;
+            }
+
+            const std::string s = unicode_cpt_to_utf8(normalizer_opts.lowercase ? unicode_tolower(cpt) : cpt);
             if (flags.is_punctuation || ( cpt < 0x7F && flags.is_symbol ) || is_chinese_char(cpt)) {
                 if (words.back().size()) {  // finish previous word if any
                     words.emplace_back();
@@ -1692,7 +1699,7 @@ struct llm_tokenizer_whitespace_session : llm_tokenizer_bpe_session {
     llm_tokenizer_whitespace_session(const llama_vocab & vocab, const llm_tokenizer_bpe & tokenizer) : llm_tokenizer_bpe_session{vocab, tokenizer}, vocab{vocab} {}
 
     void tokenize(const std::string & text, std::vector<llama_token> & output) override {
-        const bool lowercase = vocab.get_normalizer_lowercase();
+        const bool lowercase = vocab.get_normalizer_opts().lowercase;
 
         std::string segment;
         auto flush = [&]() {
@@ -1797,7 +1804,9 @@ struct llama_vocab::impl {
     bool remove_extra_whitespaces   = false;
     bool escape_whitespaces         = true;
     bool treat_whitespace_as_suffix = false;
-    bool normalizer_lowercase       = true; // Lowercase normalizer (tokenizer.json)
+
+    // BertNormalizer options
+    llama_vocab::normalizer_options normalizer_opts;
 
     std::unordered_map<std::string, llama_token> token_to_id;
     std::vector<token_data>                      id_to_token;
@@ -2172,7 +2181,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             } else if (
                     tokenizer_pre == "whitespace") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_WHITESPACE;
-                normalizer_lowercase = false;
+                normalizer_opts.lowercase = false;
             } else if (
                     tokenizer_pre == "refact") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_REFACT;
@@ -2532,8 +2541,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             }
         }
 
-        // Lowercase normalizer flag (consulted by WPM / whitespace BPE)
-        ml.get_key(LLM_KV_TOKENIZER_NORMALIZER_LOWERCASE, normalizer_lowercase, false);
+        // BertNormalizer options
+        ml.get_key(LLM_KV_TOKENIZER_NORMALIZER_LOWERCASE,     normalizer_opts.lowercase,     false);
+        normalizer_opts.strip_accents = normalizer_opts.lowercase;
+        ml.get_key(LLM_KV_TOKENIZER_NORMALIZER_STRIP_ACCENTS, normalizer_opts.strip_accents, false);
 
         // suppress tokens
         {
@@ -3969,8 +3980,8 @@ bool llama_vocab::get_treat_whitespace_as_suffix() const {
     return pimpl->treat_whitespace_as_suffix;
 }
 
-bool llama_vocab::get_normalizer_lowercase() const {
-    return pimpl->normalizer_lowercase;
+const llama_vocab::normalizer_options & llama_vocab::get_normalizer_opts() const {
+    return pimpl->normalizer_opts;
 }
 
 const std::vector<llama_token> & llama_vocab::get_suppress_tokens() const {