server : fix incorrect usage of llama_get_embeddings() (#14225)

* server : fix incorrect usage of llama_get_embeddings()

ggml-ci

* cont : fix the fix

ggml-ci

.devops/intel.Dockerfile

@@ -49,19 +49,23 @@ COPY --from=build /app/full /app
 
 WORKDIR /app
 
-RUN apt-get update \
-    && apt-get install -y \
-    git \
-    python3 \
-    python3-pip \
-    && pip install --upgrade pip setuptools wheel \
-    && pip install -r requirements.txt \
-    && apt autoremove -y \
-    && apt clean -y \
-    && rm -rf /tmp/* /var/tmp/* \
-    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
-    && find /var/cache -type f -delete
+RUN apt-get update && \
+    apt-get install -y \
+        git \
+        python3 \
+        python3-pip \
+        python3-venv && \
+    python3 -m venv /opt/venv && \
+    . /opt/venv/bin/activate && \
+    pip install --upgrade pip setuptools wheel && \
+    pip install -r requirements.txt && \
+    apt autoremove -y && \
+    apt clean -y && \
+    rm -rf /tmp/* /var/tmp/* && \
+    find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete && \
+    find /var/cache -type f -delete
 
+ENV PATH="/opt/venv/bin:$PATH"
 
 ENTRYPOINT ["/app/tools.sh"]
 

.github/workflows/build.yml

@@ -306,6 +306,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
+          export GGML_VK_VISIBLE_DEVICES=0
           # This is using llvmpipe and runs slower than other backends
           ctest -L main --verbose --timeout 3600
 
@@ -687,12 +688,12 @@ jobs:
     strategy:
       matrix:
         include:
-          - build: 'cpu-x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF'
+          - build: 'cpu-x64 (static)'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
           - build: 'openblas-x64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
+            defines: '-DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
           - build: 'llvm-arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
           - build: 'llvm-arm64-opencl-adreno'
@@ -777,6 +778,7 @@ jobs:
           cmake -S . -B build ${{ matrix.defines }} `
             -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
+          cp $env:CURL_PATH/bin/libcurl-*.dll build/bin/Release
 
       - name: Add libopenblas.dll
         id: add_libopenblas_dll

CMakeLists.txt

@@ -89,6 +89,14 @@ option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/common.cmake)
 
+if (NOT DEFINED LLAMA_BUILD_NUMBER)
+    set(LLAMA_BUILD_NUMBER        ${BUILD_NUMBER})
+endif()
+if (NOT DEFINED LLAMA_BUILD_COMMIT)
+    set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
+endif()
+set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
+
 # override ggml options
 set(GGML_ALL_WARNINGS   ${LLAMA_ALL_WARNINGS})
 set(GGML_FATAL_WARNINGS ${LLAMA_FATAL_WARNINGS})
@@ -155,6 +163,8 @@ if (LLAMA_USE_SYSTEM_GGML)
 endif()
 
 if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML)
+    set(GGML_BUILD_NUMBER ${LLAMA_BUILD_NUMBER})
+    set(GGML_BUILD_COMMIT ${LLAMA_BUILD_COMMIT})
     add_subdirectory(ggml)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
@@ -204,10 +214,6 @@ endif()
 include(GNUInstallDirs)
 include(CMakePackageConfigHelpers)
 
-set(LLAMA_BUILD_NUMBER        ${BUILD_NUMBER})
-set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
-set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
-
 set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header  files")
 set(LLAMA_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
 set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")

Makefile

@@ -367,7 +367,7 @@ ifdef LLAMA_SERVER_SSL
 endif
 
 ifndef GGML_NO_CPU_AARCH64
-	MK_CPPFLAGS += -DGGML_USE_CPU_AARCH64
+	MK_CPPFLAGS += -DGGML_USE_CPU_REPACK
 endif
 
 # warnings
@@ -970,7 +970,7 @@ OBJ_GGML = \
 	$(DIR_GGML)/src/ggml-threading.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu_cpp.o \
-	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-aarch64.o \
+	$(DIR_GGML)/src/ggml-cpu/repack.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-hbm.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-quants.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-traits.o \

README.md

@@ -6,7 +6,7 @@
 [![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
 [![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
 
-[Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggml-org/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
+[Roadmap](https://github.com/users/ggerganov/projects/7) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
@@ -18,7 +18,6 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 ## Hot topics
 
 - 🔥 Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
-- **GGML developer experience survey (organized and reviewed by NVIDIA):** [link](https://forms.gle/Gasw3cRgyhNEnrwK9)
 - A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141](https://github.com/ggml-org/llama.cpp/pull/13141)), `libllava` will be deprecated
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639

ci/run.sh

@@ -39,7 +39,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=OFF"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"

common/CMakeLists.txt

@@ -7,8 +7,8 @@ llama_add_compile_flags()
 # Build info header
 #
 
-if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
-    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
+if(EXISTS "${PROJECT_SOURCE_DIR}/.git")
+    set(GIT_DIR "${PROJECT_SOURCE_DIR}/.git")
 
     # Is git submodule
     if(NOT IS_DIRECTORY "${GIT_DIR}")
@@ -18,36 +18,26 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
         if (SLASH_POS EQUAL 0)
             set(GIT_DIR "${REAL_GIT_DIR}")
         else()
-            set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../${REAL_GIT_DIR}")
+            set(GIT_DIR "${PROJECT_SOURCE_DIR}/${REAL_GIT_DIR}")
         endif()
     endif()
 
     if(EXISTS "${GIT_DIR}/index")
-        set(GIT_INDEX "${GIT_DIR}/index")
+        # For build-info.cpp below
+        set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS "${GIT_DIR}/index")
     else()
         message(WARNING "Git index not found in git repository.")
-        set(GIT_INDEX "")
     endif()
 else()
     message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
-    set(GIT_INDEX "")
 endif()
 
-# Add a custom command to rebuild build-info.cpp when .git/index changes
-add_custom_command(
-    OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp"
-    COMMENT "Generating build details from Git"
-    COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
-            -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
-            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-            -DCMAKE_SYSTEM_NAME=${CMAKE_SYSTEM_NAME} -DCMAKE_SYSTEM_PROCESSOR=${CMAKE_SYSTEM_PROCESSOR}
-            -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
-    WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
-    DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
-    VERBATIM
-)
+set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in")
+set(OUTPUT_FILE   "${CMAKE_CURRENT_BINARY_DIR}/build-info.cpp")
+configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
+
 set(TARGET build_info)
-add_library(${TARGET} OBJECT build-info.cpp)
+add_library(${TARGET} OBJECT ${OUTPUT_FILE})
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()

common/arg.cpp

@@ -988,10 +988,6 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         params.tensor_buft_overrides.push_back({nullptr, nullptr});
     }
 
-    if (params.reranking && params.embedding) {
-        throw std::invalid_argument("error: either --embedding or --reranking can be specified, but not both");
-    }
-
     if (!params.chat_template.empty() && !common_chat_verify_template(params.chat_template, params.use_jinja)) {
         throw std::runtime_error(string_format(
             "error: the supplied chat template is not supported: %s%s\n",
@@ -2747,9 +2743,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_EMBEDDINGS"));
     add_opt(common_arg(
         {"--reranking", "--rerank"},
-        string_format("enable reranking endpoint on server (default: %s)", params.reranking ? "enabled" : "disabled"),
+        string_format("enable reranking endpoint on server (default: %s)", "disabled"),
         [](common_params & params) {
-            params.reranking = true;
+            params.embedding = true;
+            params.pooling_type = LLAMA_POOLING_TYPE_RANK;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_RERANKING"));
     add_opt(common_arg(

common/build-info.cpp.in

@@ -1,4 +1,4 @@
-int LLAMA_BUILD_NUMBER = @BUILD_NUMBER@;
-char const *LLAMA_COMMIT = "@BUILD_COMMIT@";
+int LLAMA_BUILD_NUMBER = @LLAMA_BUILD_NUMBER@;
+char const *LLAMA_COMMIT = "@LLAMA_BUILD_COMMIT@";
 char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
 char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";

common/chat-parser.cpp

@@ -49,6 +49,7 @@ bool common_chat_msg_parser::add_tool_call(const std::string & name, const std::
 
     // LOG_DBG("Tool call arguments:\n\traw: %s\n\tresult: %s\n", arguments.c_str(), tool_call.arguments.c_str());
     result_.tool_calls.emplace_back(tool_call);
+
     return true;
 }
 bool common_chat_msg_parser::add_tool_call(const json & tool_call) {
@@ -378,3 +379,7 @@ std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parse
         /* .is_partial = */ found_healing_marker,
     };
 }
+
+void common_chat_msg_parser::clear_tools() {
+    result_.tool_calls.clear();
+}

common/chat-parser.h

@@ -115,4 +115,6 @@ class common_chat_msg_parser {
         const std::vector<std::vector<std::string>> & args_paths = {},
         const std::vector<std::vector<std::string>> & content_paths = {}
     );
+
+    void clear_tools();
 };

common/chat.cpp

@@ -1921,7 +1921,9 @@ common_chat_msg common_chat_parse(const std::string & input, bool is_partial, co
     } catch (const common_chat_msg_partial_exception & ex) {
         LOG_DBG("Partial parse: %s\n", ex.what());
         if (!is_partial) {
-            throw std::runtime_error(ex.what());
+            builder.clear_tools();
+            builder.move_to(0);
+            common_chat_parse_content_only(builder);
         }
     }
     auto msg = builder.result();

common/cmake/build-info-gen-cpp.cmake

@@ -1,24 +0,0 @@
-include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
-
-set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp.in")
-set(OUTPUT_FILE   "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp")
-
-# Only write the build info if it changed
-if(EXISTS ${OUTPUT_FILE})
-    file(READ ${OUTPUT_FILE} CONTENTS)
-    string(REGEX MATCH "LLAMA_COMMIT = \"([^\"]*)\";" _ ${CONTENTS})
-    set(OLD_COMMIT ${CMAKE_MATCH_1})
-    string(REGEX MATCH "LLAMA_COMPILER = \"([^\"]*)\";" _ ${CONTENTS})
-    set(OLD_COMPILER ${CMAKE_MATCH_1})
-    string(REGEX MATCH "LLAMA_BUILD_TARGET = \"([^\"]*)\";" _ ${CONTENTS})
-    set(OLD_TARGET ${CMAKE_MATCH_1})
-    if (
-        NOT OLD_COMMIT   STREQUAL BUILD_COMMIT   OR
-        NOT OLD_COMPILER STREQUAL BUILD_COMPILER OR
-        NOT OLD_TARGET   STREQUAL BUILD_TARGET
-    )
-        configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
-    endif()
-else()
-    configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
-endif()

common/common.cpp

@@ -466,7 +466,7 @@ size_t string_find_partial_stop(const std::string_view & str, const std::string_
 
 std::string regex_escape(const std::string & s) {
     static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
-    return std::regex_replace(s, special_chars, "\\$0");
+    return std::regex_replace(s, special_chars, "\\$&");
 }
 
 std::string string_join(const std::vector<std::string> & values, const std::string & separator) {
@@ -767,6 +767,9 @@ bool fs_validate_filename(const std::string & filename) {
     return true;
 }
 
+#include <iostream>
+
+
 // returns true if successful, false otherwise
 bool fs_create_directory_with_parents(const std::string & path) {
 #ifdef _WIN32
@@ -784,9 +787,16 @@ bool fs_create_directory_with_parents(const std::string & path) {
     // process path from front to back, procedurally creating directories
     while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
         const std::wstring subpath = wpath.substr(0, pos_slash);
-        const wchar_t * test = subpath.c_str();
 
-        const bool success = CreateDirectoryW(test, NULL);
+        pos_slash += 1;
+
+        // skip the drive letter, in some systems it can return an access denied error
+        if (subpath.length() == 2 && subpath[1] == ':') {
+            continue;
+        }
+
+        const bool success = CreateDirectoryW(subpath.c_str(), NULL);
+
         if (!success) {
             const DWORD error = GetLastError();
 
@@ -800,8 +810,6 @@ bool fs_create_directory_with_parents(const std::string & path) {
                 return false;
             }
         }
-
-        pos_slash += 1;
     }
 
     return true;
@@ -897,34 +905,6 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
-    if (params.reranking) {
-        bool ok = true;
-
-        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
-        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
-
-        if (!has_eos && !has_sep) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
-            ok = false;
-        } else if (!has_eos) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
-        } else if (!has_sep) {
-            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        if (!ok) {
-            llama_model_free(model);
-
-            return iparams;
-        }
-    }
-
     auto cparams = common_context_params_to_llama(params);
 
     llama_context * lctx = llama_init_from_model(model, cparams);
@@ -966,6 +946,35 @@ struct common_init_result common_init_from_params(common_params & params) {
         }
     }
 
+    if (llama_pooling_type(lctx) == LLAMA_POOLING_TYPE_RANK) {
+        bool ok = true;
+
+        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
+        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
+
+        if (!has_eos && !has_sep) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
+            ok = false;
+        } else if (!has_eos) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
+        } else if (!has_sep) {
+            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (!ok) {
+            llama_free(lctx);
+            llama_model_free(model);
+
+            return iparams;
+        }
+    }
+
     // load and optionally apply lora adapters
     for (auto & la : params.lora_adapters) {
         llama_adapter_lora_ptr lora;
@@ -1143,11 +1152,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.op_offload        = !params.no_op_offload;
     cparams.swa_full          = params.swa_full;
 
-    if (params.reranking) {
-        cparams.embeddings    = true;
-        cparams.pooling_type  = LLAMA_POOLING_TYPE_RANK;
-    }
-
     cparams.type_k = params.cache_type_k;
     cparams.type_v = params.cache_type_v;
 

common/common.h

@@ -355,7 +355,6 @@ struct common_params {
     int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
     std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
     std::string embd_sep   = "\n";  // separator of embeddings
-    bool reranking         = false; // enable reranking support on server
 
     // server params
     int32_t port           = 8080;         // server listens on this network port

convert_hf_to_gguf.py

@@ -519,7 +519,7 @@ class TextModel(ModelBase):
     def set_gguf_parameters(self):
         self.gguf_writer.add_block_count(self.block_count)
 
-        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions"], optional=True)) is not None:
+        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions", "max_length"], optional=True)) is not None:
             self.gguf_writer.add_context_length(n_ctx)
             logger.info(f"gguf: context length = {n_ctx}")
 
@@ -556,8 +556,11 @@ class TextModel(ModelBase):
             logger.info(f"gguf: experts used count = {n_experts_used}")
 
         if (head_dim := self.hparams.get("head_dim")) is not None:
-            self.gguf_writer.add_key_length(head_dim)
-            self.gguf_writer.add_value_length(head_dim)
+            # Workaround for incorrect AutoConfig value for DeepSeekV3 (is set correctly in DeepSeekV2Model class)
+            # https://github.com/huggingface/transformers/blob/19224c3642705c5b6988c9f5f4251f83323d05ae/src/transformers/models/deepseek_v3/configuration_deepseek_v3.py#L210
+            if self.hparams.get("model_type") != "deepseek_v3":
+                self.gguf_writer.add_key_length(head_dim)
+                self.gguf_writer.add_value_length(head_dim)
 
         self.gguf_writer.add_file_type(self.ftype)
         logger.info(f"gguf: file type = {self.ftype}")
@@ -2017,6 +2020,20 @@ class LlamaModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("ArceeForCausalLM")
+class ArceeModel(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.ARCEE
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self._try_set_pooling_type()
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
+
+
 @ModelBase.register(
     "LlavaForConditionalGeneration", # pixtral
     "Mistral3ForConditionalGeneration", # mistral small 3.1
@@ -4059,6 +4076,34 @@ class NomicBertModel(BertModel):
         raise ValueError(f"unknown tokenizer: {toktyp}")
 
 
+@ModelBase.register("NeoBERT", "NeoBERTLMHead", "NeoBERTForSequenceClassification")
+class NeoBert(BertModel):
+    model_arch = gguf.MODEL_ARCH.NEO_BERT
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # NeoBERT uses 2/3 of the intermediate size as feed forward length
+        self.gguf_writer.add_feed_forward_length(int(2 * self.hparams["intermediate_size"] / 3))
+        self.gguf_writer.add_rope_freq_base(10000.0)  # default value for NeoBERT
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+
+        f_rms_eps = self.hparams.get("norm_eps", 1e-6)  # default value for NeoBERT
+        self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
+        logger.info(f"gguf: rms norm epsilon = {f_rms_eps}")
+
+        self.gguf_writer.add_pooling_type(gguf.PoolingType.CLS) # https://huggingface.co/chandar-lab/NeoBERT#how-to-use
+
+    def modify_tensors(self, data_torch, name, bid):
+        if name.startswith("decoder."):
+            return []
+
+        if name.startswith("model."):
+            name = name[6:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
 class XLMRobertaModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT
@@ -4798,25 +4843,6 @@ class OlmoeModel(TextModel):
 class JinaBertV2Model(BertModel):
     model_arch = gguf.MODEL_ARCH.JINA_BERT_V2
 
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.intermediate_size = self.hparams["intermediate_size"]
-
-    def get_tensors(self):
-        for name, data in super().get_tensors():
-            if 'gated_layer' in name:
-                d1 = data[:self.intermediate_size, :]
-                name1 = name.replace('gated_layers', 'gated_layers_w')
-                name1 = name1.replace('up_gated_layer', 'gated_layers_v')
-                d2 = data[self.intermediate_size:, :]
-                name2 = name.replace('gated_layers', 'gated_layers_v')
-                name2 = name2.replace('up_gated_layer', 'gated_layers_w')
-                yield name1, d1
-                yield name2, d2
-                continue
-
-            yield name, data
-
     def set_vocab(self):
         tokenizer_class = 'BertTokenizer'
         with open(self.dir_model / "tokenizer_config.json", "r", encoding="utf-8") as f:
@@ -4832,14 +4858,6 @@ class JinaBertV2Model(BertModel):
         self.gguf_writer.add_add_bos_token(True)
         self.gguf_writer.add_add_eos_token(True)
 
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # if name starts with "bert.", remove the prefix
-        # e.g. https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
-        if name.startswith("bert."):
-            name = name[5:]
-
-        return super().modify_tensors(data_torch, name, bid)
-
 
 @ModelBase.register("OpenELMForCausalLM")
 class OpenELMModel(TextModel):
@@ -5286,6 +5304,34 @@ class DeepseekV2Model(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("Dots1ForCausalLM")
+class Dots1Model(Qwen2MoeModel):
+    model_arch = gguf.MODEL_ARCH.DOTS1
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.hparams["num_experts"] = self.hparams["n_routed_experts"]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_leading_dense_block_count(self.hparams["first_k_dense_replace"])
+        self.gguf_writer.add_expert_shared_count(self.hparams["n_shared_experts"])
+        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
+        self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
+
+        if self.hparams["scoring_func"] == "noaux_tc":
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+        else:
+            raise ValueError(f"Unsupported scoring_func value: {self.hparams['scoring_func']}")
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+        if "shared_experts" in name:
+            return [(self.map_tensor_name(name), data_torch)]
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("PLMForCausalLM")
 class PLMModel(TextModel):
     model_arch = gguf.MODEL_ARCH.PLM

docs/function-calling.md

@@ -11,7 +11,7 @@ Function calling is supported for all models (see https://github.com/ggml-org/ll
   - Llama 3.1 / 3.3 (including builtin tools support - tool names for `wolfram_alpha`, `web_search` / `brave_search`, `code_interpreter`), Llama 3.2
   - Functionary v3.1 / v3.2
   - Hermes 2/3, Qwen 2.5
-  - Qwen 2.5 Coder (WIP: https://github.com/ggml-org/llama.cpp/pull/12034)
+  - Qwen 2.5 Coder
   - Mistral Nemo
   - Firefunction v2
   - Command R7B

docs/multimodal.md

@@ -107,3 +107,7 @@ NOTE: some models may require large context window, for example: `-c 8192`
 (tool_name) -hf ggml-org/Qwen2.5-Omni-3B-GGUF
 (tool_name) -hf ggml-org/Qwen2.5-Omni-7B-GGUF
 ```
+
+## Finding more models:
+
+GGUF models on Huggingface with vision capabilities can be found here: https://huggingface.co/models?pipeline_tag=image-text-to-text&sort=trending&search=gguf

examples/gritlm/gritlm.cpp

@@ -41,12 +41,11 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
 
         // add input to batch (this increments n_tokens)
         for (int32_t j = 0; j < n_toks; j++) {
-            common_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
+            common_batch_add(batch, inputs[j], j, { 0 }, true);
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
         llama_memory_clear(llama_get_memory(ctx), true);
-        llama_set_embeddings(ctx, true);
         llama_set_causal_attn(ctx, false);
 
         // run model
@@ -103,7 +102,6 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
     llama_token eos_token = llama_vocab_eos(vocab);
 
     llama_memory_clear(llama_get_memory(ctx), true);
-    llama_set_embeddings(ctx, false);
     llama_set_causal_attn(ctx, true);
 
     llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
@@ -166,6 +164,8 @@ int main(int argc, char * argv[]) {
     llama_model_params mparams = common_model_params_to_llama(params);
     llama_context_params cparams = common_context_params_to_llama(params);
 
+    cparams.embeddings = true;
+
     llama_backend_init();
 
     llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
@@ -213,6 +213,8 @@ int main(int argc, char * argv[]) {
         std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[1].c_str(), cosine_sim_q1_d1);
     }
 
+    llama_set_embeddings(ctx, false);
+
     // ### Generation ###
     // GritLM models are not finetuned with system prompts, as you can just include system-like instructions together with your user instruction
     {

ggml/CMakeLists.txt

@@ -105,7 +105,7 @@ message(DEBUG "GGML_NATIVE_DEFAULT : ${GGML_NATIVE_DEFAULT}")
 message(DEBUG "INS_ENB             : ${INS_ENB}")
 
 option(GGML_CPU_HBM          "ggml: use memkind for CPU HBM" OFF)
-option(GGML_CPU_AARCH64      "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
+option(GGML_CPU_REPACK       "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
 option(GGML_CPU_KLEIDIAI     "ggml: use KleidiAI optimized kernels if applicable" OFF)
 option(GGML_SSE42            "ggml: enable SSE 4.2"          ${INS_ENB})
 option(GGML_AVX              "ggml: enable AVX"              ${INS_ENB})
@@ -172,6 +172,7 @@ option(GGML_HIP                             "ggml: use HIP"
 option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
 option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
+option(GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12   "ggml: enable rocWMMA FlashAttention on GFX12"    OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)

ggml/cmake/common.cmake

@@ -36,8 +36,7 @@ function(ggml_get_system_arch)
             (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
             CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64|amd64)$"))
         set(GGML_SYSTEM_ARCH "x86" PARENT_SCOPE)
-    elseif ("${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "ppc64le " OR
-            "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ")
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc|power")
         set(GGML_SYSTEM_ARCH "PowerPC" PARENT_SCOPE)
     elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
         set(GGML_SYSTEM_ARCH "loongarch64"  PARENT_SCOPE)

ggml/src/CMakeLists.txt

@@ -212,6 +212,7 @@ endif()
 
 add_library(ggml
             ggml-backend-reg.cpp)
+add_library(ggml::ggml ALIAS ggml)
 
 target_link_libraries(ggml PUBLIC ggml-base)
 
@@ -269,17 +270,23 @@ endfunction()
 function(ggml_add_cpu_backend_variant tag_name)
     set(GGML_CPU_TAG_NAME ${tag_name})
     # other: OPENMP LLAMAFILE CPU_HBM
-    foreach (feat NATIVE
-                  SSE42
-                  AVX AVX2 BMI2 AVX_VNNI FMA F16C
-                  AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
-                  AMX_TILE AMX_INT8 AMX_BF16)
-        set(GGML_${feat} OFF)
-    endforeach()
+    if (GGML_SYSTEM_ARCH STREQUAL "x86")
+        foreach (feat NATIVE
+                      SSE42
+                      AVX AVX2 BMI2 AVX_VNNI FMA F16C
+                      AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
+                      AMX_TILE AMX_INT8 AMX_BF16)
+            set(GGML_${feat} OFF)
+        endforeach()
 
-    foreach (feat ${ARGN})
-        set(GGML_${feat} ON)
-    endforeach()
+        foreach (feat ${ARGN})
+            set(GGML_${feat} ON)
+        endforeach()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "ARM")
+        foreach (feat ${ARGN})
+            set(GGML_INTERNAL_${feat} ON)
+        endforeach()
+    endif()
 
     ggml_add_cpu_backend_variant_impl(${tag_name})
 endfunction()
@@ -289,6 +296,8 @@ ggml_add_backend(CPU)
 if (GGML_CPU_ALL_VARIANTS)
     if (NOT GGML_BACKEND_DL)
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL")
+    elseif (GGML_CPU_ARM_ARCH)
+        message(FATAL_ERROR "Cannot use both GGML_CPU_ARM_ARCH and GGML_CPU_ALL_VARIANTS")
     endif()
     if (GGML_SYSTEM_ARCH STREQUAL "x86")
         ggml_add_cpu_backend_variant(x64)
@@ -302,8 +311,30 @@ if (GGML_CPU_ALL_VARIANTS)
             # MSVC doesn't support AMX
             ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
         endif()
+    elseif(GGML_SYSTEM_ARCH STREQUAL "ARM")
+        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+            # Many of these features are optional so we build versions with popular
+            # combinations and name the backends based on the version they were
+            # first released with
+            ggml_add_cpu_backend_variant(armv8.0_1)
+            ggml_add_cpu_backend_variant(armv8.2_1    DOTPROD)
+            ggml_add_cpu_backend_variant(armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
+            ggml_add_cpu_backend_variant(armv8.2_3    DOTPROD FP16_VECTOR_ARITHMETIC SVE)
+            ggml_add_cpu_backend_variant(armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8)
+            ggml_add_cpu_backend_variant(armv8.6_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2)
+            ggml_add_cpu_backend_variant(armv9.2_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SME)
+            ggml_add_cpu_backend_variant(armv9.2_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2 SME)
+        elseif (CMAKE_SYSTEM_NAME MATCHES "Android")
+            # Android-specific backends with SoC-compatible feature sets
+            ggml_add_cpu_backend_variant(android_armv8.0_1)
+            ggml_add_cpu_backend_variant(android_armv8.2_1    DOTPROD)
+            ggml_add_cpu_backend_variant(android_armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
+            ggml_add_cpu_backend_variant(android_armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC MATMUL_INT8)
+        else()
+            message(FATAL_ERROR "Unsupported ARM target OS: ${CMAKE_SYSTEM_NAME}")
+        endif()
     else()
-        message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported on ${GGML_SYSTEM_ARCH}")
+        message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported with ${GGML_SYSTEM_ARCH} on ${CMAKE_SYSTEM_NAME}")
     endif()
 elseif (GGML_CPU)
     ggml_add_cpu_backend_variant_impl("")

ggml/src/ggml-common.h

@@ -1074,6 +1074,10 @@ GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
     0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
 GGML_TABLE_END()
 
+GGML_TABLE_BEGIN(int8_t, kvalues_iq4nl, 16)
+    -127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113,
+GGML_TABLE_END()
+
 #define NGRID_IQ1S 2048
 #define IQ1S_DELTA 0.125f
 #define IQ1M_DELTA 0.125f

ggml/src/ggml-cpu/CMakeLists.txt

@@ -1,3 +1,17 @@
+function(ggml_add_cpu_backend_features cpu_name arch)
+    # The feature detection code is compiled as a separate target so that
+    # it can be built without the architecture flags
+    # Since multiple variants of the CPU backend may be included in the same
+    # build, using set_source_files_properties() to set the arch flags is not possible
+    set(GGML_CPU_FEATS_NAME ${cpu_name}-feats)
+    add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/arch/${arch}/cpu-feats.cpp)
+    target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
+    target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARGN})
+    target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
+    set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_link_libraries(${cpu_name} PRIVATE ${GGML_CPU_FEATS_NAME})
+endfunction()
+
 function(ggml_add_cpu_backend_variant_impl tag_name)
     if (tag_name)
         set(GGML_CPU_NAME ggml-cpu-${tag_name})
@@ -10,14 +24,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     list (APPEND GGML_CPU_SOURCES
         ggml-cpu/ggml-cpu.c
         ggml-cpu/ggml-cpu.cpp
-        ggml-cpu/ggml-cpu-aarch64.cpp
-        ggml-cpu/ggml-cpu-aarch64.h
-        ggml-cpu/ggml-cpu-hbm.cpp
-        ggml-cpu/ggml-cpu-hbm.h
-        ggml-cpu/ggml-cpu-quants.c
-        ggml-cpu/ggml-cpu-quants.h
-        ggml-cpu/ggml-cpu-traits.cpp
-        ggml-cpu/ggml-cpu-traits.h
+        ggml-cpu/repack.cpp
+        ggml-cpu/repack.h
+        ggml-cpu/hbm.cpp
+        ggml-cpu/hbm.h
+        ggml-cpu/quants.c
+        ggml-cpu/quants.h
+        ggml-cpu/traits.cpp
+        ggml-cpu/traits.h
         ggml-cpu/amx/amx.cpp
         ggml-cpu/amx/amx.h
         ggml-cpu/amx/mmq.cpp
@@ -84,6 +98,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
     if (GGML_SYSTEM_ARCH STREQUAL "ARM")
         message(STATUS "ARM detected")
+        list(APPEND GGML_CPU_SOURCES
+            ggml-cpu/arch/arm/quants.c
+            ggml-cpu/arch/arm/repack.cpp
+            )
+
         if (MSVC AND NOT CMAKE_C_COMPILER_ID STREQUAL "Clang")
             message(FATAL_ERROR "MSVC is not supported for ARM, use clang")
         else()
@@ -138,6 +157,46 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             else()
                 if (GGML_CPU_ARM_ARCH)
                     list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
+                elseif(GGML_CPU_ALL_VARIANTS)
+                    # Begin with the lowest baseline
+                    set(ARM_MCPU "armv8-a")
+                    set(ARCH_TAGS "")
+                    set(ARCH_DEFINITIONS "")
+
+                    # When a feature is selected, bump the MCPU to the first
+                    # version that supported it
+                    if (GGML_INTERNAL_DOTPROD)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+dotprod")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_DOTPROD)
+                    endif()
+                    if (GGML_INTERNAL_FP16_VECTOR_ARITHMETIC)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+fp16")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_FP16_VECTOR_ARITHMETIC)
+                    endif()
+                    if (GGML_INTERNAL_SVE)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+sve")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_SVE)
+                    endif()
+                    if (GGML_INTERNAL_MATMUL_INT8)
+                        set(ARM_MCPU "armv8.6-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+i8mm")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_MATMUL_INT8)
+                    endif()
+                    if (GGML_INTERNAL_SVE2)
+                        set(ARM_MCPU "armv8.6-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+sve2")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_SVE2)
+                    endif()
+                    if (GGML_INTERNAL_SME)
+                        set(ARM_MCPU "armv9.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+sme")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_SME)
+                    endif()
+                    list(APPEND ARCH_FLAGS "-march=${ARM_MCPU}${ARCH_TAGS}")
+                    ggml_add_cpu_backend_features(${GGML_CPU_NAME} arm ${ARCH_DEFINITIONS})
                 endif()
             endif()
 
@@ -167,6 +226,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "x86")
         message(STATUS "x86 detected")
+        list(APPEND GGML_CPU_SOURCES
+            ggml-cpu/arch/x86/quants.c
+            ggml-cpu/arch/x86/repack.cpp
+            )
+
         if (MSVC)
             # instruction set detection for MSVC only
             if (GGML_NATIVE)
@@ -296,21 +360,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 # the feature check relies on ARCH_DEFINITIONS, but it is not set with GGML_NATIVE
                 message(FATAL_ERROR "GGML_NATIVE is not compatible with GGML_BACKEND_DL, consider using GGML_CPU_ALL_VARIANTS")
             endif()
-
-            # The feature detection code is compiled as a separate target so that
-            # it can be built without the architecture flags
-            # Since multiple variants of the CPU backend may be included in the same
-            # build, using set_source_files_properties() to set the arch flags is not possible
-            set(GGML_CPU_FEATS_NAME ${GGML_CPU_NAME}-feats)
-            add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/cpu-feats-x86.cpp)
-            target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
-            target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARCH_DEFINITIONS})
-            target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
-            set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-            target_link_libraries(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_FEATS_NAME})
+            ggml_add_cpu_backend_features(${GGML_CPU_NAME} x86 ${ARCH_DEFINITIONS})
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
         message(STATUS "PowerPC detected")
+        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/powerpc/quants.c)
         if (GGML_NATIVE)
             if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
                 file(READ "/proc/cpuinfo" POWER10_M)
@@ -338,6 +392,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "loongarch64")
         message(STATUS "loongarch64 detected")
+        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/loongarch/quants.c)
+
         list(APPEND ARCH_FLAGS -march=loongarch64)
         if (GGML_LASX)
             list(APPEND ARCH_FLAGS -mlasx)
@@ -347,6 +403,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
         message(STATUS "riscv64 detected")
+        list(APPEND GGML_CPU_SOURCES
+            ggml-cpu/arch/riscv/quants.c
+            ggml-cpu/arch/riscv/repack.cpp
+            )
         if (GGML_RVV)
             if (GGML_XTHEADVECTOR)
                 list(APPEND ARCH_FLAGS -march=rv64gc_xtheadvector -mabi=lp64d)
@@ -358,6 +418,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         message(STATUS "s390x detected")
+        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/s390/quants.c)
         file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
         string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
 
@@ -381,12 +442,16 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         if (GGML_VXE)
             list(APPEND ARCH_FLAGS -mvx -mzvector)
         endif()
+    elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
+        message(STATUS "Wasm detected")
+        list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)
     else()
-        message(STATUS "Unknown architecture")
+        message(WARNING "Unknown CPU architecture. Falling back to generic implementations.")
+        list(APPEND ARCH_FLAGS -DGGML_CPU_GENERIC)
     endif()
 
-    if (GGML_CPU_AARCH64)
-        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_AARCH64)
+    if (GGML_CPU_REPACK)
+        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_REPACK)
     endif()
 
     if (GGML_CPU_KLEIDIAI)

ggml/src/ggml-cpu/amx/amx.cpp

@@ -5,7 +5,7 @@
 #include "ggml-backend.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #if defined(__gnu_linux__)
 #include <sys/syscall.h>

ggml/src/ggml-cpu/amx/mmq.cpp

@@ -8,7 +8,7 @@
 #include "mmq.h"
 #include "ggml-impl.h"
 #include "ggml-cpu-impl.h"
-#include "ggml-cpu-quants.h"
+#include "quants.h"
 #include "ggml-quants.h"
 #include <algorithm>
 #include <type_traits>

ggml/src/ggml-cpu/apple-fallback.h

@@ -0,0 +1,88 @@
+#pragma once
+
+// Solve alias issue for Apple targets (currently PowerPC, x86, and ARM64).
+// Mach-O has a weak alias equivalent but no practical compiler support can
+// be found, so we need to do it manually.
+// ref: https://stackoverflow.com/questions/42757744
+//
+// This file is a complement to native implementations in the `arch` folder.
+// A kernel in quants.c or repack.cpp is either:
+// - implemented in the `arch` folder, or
+// - defined in this file to remove the `_generic` suffix
+
+#if defined(GGML_CPU_GENERIC)
+// quants.c
+#define quantize_row_q8_0_generic quantize_row_q8_0
+#define quantize_row_q8_1_generic quantize_row_q8_1
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
+#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
+#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
+#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
+#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
+#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
+#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
+#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
+#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
+#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__aarch64__) || defined(__arm__)
+// repack.cpp
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#elif defined(__x86_64__) || defined(__i386__)
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__POWERPC__)
+// ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679
+// quants.c
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#endif

ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp

@@ -0,0 +1,94 @@
+#include "ggml-backend-impl.h"
+
+#if defined(__aarch64__)
+
+#if defined(__linux__)
+#include <sys/auxv.h>
+#elif defined(__APPLE__)
+#include <sys/sysctl.h>
+#endif
+
+#if !defined(HWCAP2_I8MM)
+#define HWCAP2_I8MM (1 << 13)
+#endif
+
+#if !defined(HWCAP2_SME)
+#define HWCAP2_SME (1 << 23)
+#endif
+
+struct aarch64_features {
+    // has_neon not needed, aarch64 has NEON guaranteed
+    bool has_dotprod     = false;
+    bool has_fp16_va     = false;
+    bool has_sve         = false;
+    bool has_sve2        = false;
+    bool has_i8mm        = false;
+    bool has_sme         = false;
+
+    aarch64_features() {
+#if defined(__linux__)
+        uint32_t hwcap = getauxval(AT_HWCAP);
+        uint32_t hwcap2 = getauxval(AT_HWCAP2);
+
+        has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
+        has_fp16_va = !!(hwcap & HWCAP_FPHP);
+        has_sve     = !!(hwcap & HWCAP_SVE);
+        has_sve2    = !!(hwcap2 & HWCAP2_SVE2);
+        has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
+        has_sme     = !!(hwcap2 & HWCAP2_SME);
+#elif defined(__APPLE__)
+        int oldp = 0;
+        size_t size = sizeof(oldp);
+
+        if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) == 0) {
+            has_dotprod = static_cast<bool>(oldp);
+        }
+
+        if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) == 0) {
+            has_i8mm = static_cast<bool>(oldp);
+        }
+
+        if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) == 0) {
+            has_sme = static_cast<bool>(oldp);
+        }
+
+        // Apple apparently does not implement SVE yet
+#endif
+    }
+};
+
+static int ggml_backend_cpu_aarch64_score() {
+    int score = 1;
+    aarch64_features af;
+
+#ifdef GGML_USE_DOTPROD
+    if (!af.has_dotprod) { return 0; }
+    score += 1<<1;
+#endif
+#ifdef GGML_USE_FP16_VECTOR_ARITHMETIC
+    if (!af.has_fp16_va) { return 0; }
+    score += 1<<2;
+#endif
+#ifdef GGML_USE_SVE
+    if (!af.has_sve) { return 0; }
+    score += 1<<3;
+#endif
+#ifdef GGML_USE_MATMUL_INT8
+    if (!af.has_i8mm) { return 0; }
+    score += 1<<4;
+#endif
+#ifdef GGML_USE_SVE2
+    if (!af.has_sve2) { return 0; }
+    score += 1<<5;
+#endif
+#ifdef GGML_USE_SME
+    if (!af.has_sme) { return 0; }
+    score += 1<<6;
+#endif
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_aarch64_score)
+
+# endif // defined(__aarch64__)

ggml/src/ggml-cpu/arch/riscv/repack.cpp

@@ -0,0 +1,396 @@
+#define GGML_COMMON_IMPL_CPP
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+#include "ggml-backend-impl.h"
+
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-cpu-impl.h"
+#include "traits.h"
+
+#include <cmath>
+#include <cstring>
+#include <cassert>
+#include <cstdlib> // for qsort
+#include <cstdio>  // for GGML_ASSERT
+
+#define GGML_CPU_CLANG_WORKAROUND
+#include "../../repack.h"
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+#endif
+
+#define UNUSED GGML_UNUSED
+
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined __riscv_v
+    if (__riscv_vlenb() >= QK4_0) {
+        const size_t vl = QK4_0;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+            vfloat32m1_t sumf = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+            for (int l = 0; l < nb; l++) {
+                const int64_t a0 = *(const int64_t *)&a_ptr[l].qs[0];
+                const int64_t a1 = *(const int64_t *)&a_ptr[l].qs[8];
+                const int64_t a2 = *(const int64_t *)&a_ptr[l].qs[16];
+                const int64_t a3 = *(const int64_t *)&a_ptr[l].qs[24];
+                __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment constraints
+                const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a0, vl / 4));
+                const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a1, vl / 4));
+                const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a2, vl / 4));
+                const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a3, vl / 4));
+
+                const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
+                const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
+                const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
+                const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
+                const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
+                const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
+                const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
+
+                const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_hi_m));
+                const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                // vector version needs Zvfhmin extension
+                const float a_scale = GGML_FP16_TO_FP32(a_ptr[l].d);
+                const float b_scales[8] = {
+                    GGML_FP16_TO_FP32(b_ptr[l].d[0]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[1]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[2]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[3]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[4]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[5]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[6]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[7])
+                };
+                const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
+                const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
+                sumf = __riscv_vfmacc_vv_f32m1(sumf, tmp1, b_scales_vec, vl / 4);
+            }
+            __riscv_vse32_v_f32m1(s + x * ncols_interleaved, sumf, vl / 4);
+        }
+        return;
+    }
+
+#endif
+    {
+        float sumf[8];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        sumi = 0;
+                        for (int i = 0; i < blocklen; ++i) {
+                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                        }
+                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined __riscv_v
+    if (__riscv_vlenb() >= QK4_0) {
+        const size_t vl = QK4_0;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+                vfloat32m1_t sumf0 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                vfloat32m1_t sumf1 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                vfloat32m1_t sumf2 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                vfloat32m1_t sumf3 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                for (int l = 0; l < nb; l++) {
+                    const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
+                    const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
+                    const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
+                    const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
+                    const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
+                    const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
+                    const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
+
+                    // vector version needs Zvfhmin extension
+                    const float a_scales[4] = {
+                        GGML_FP16_TO_FP32(a_ptr[l].d[0]),
+                        GGML_FP16_TO_FP32(a_ptr[l].d[1]),
+                        GGML_FP16_TO_FP32(a_ptr[l].d[2]),
+                        GGML_FP16_TO_FP32(a_ptr[l].d[3])
+                    };
+                    const float b_scales[8] = {
+                        GGML_FP16_TO_FP32(b_ptr[l].d[0]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[1]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[2]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[3]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[4]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[5]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[6]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[7])
+                    };
+                    const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
+
+                    const int64_t A0 = *(const int64_t *)&a_ptr[l].qs[0];
+                    const int64_t A4 = *(const int64_t *)&a_ptr[l].qs[32];
+                    const int64_t A8 = *(const int64_t *)&a_ptr[l].qs[64];
+                    const int64_t Ac = *(const int64_t *)&a_ptr[l].qs[96];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l0;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A0, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A4, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A8, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ac, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l0 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l0));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[0], vl / 4);
+                        sumf0 = __riscv_vfmacc_vv_f32m1(sumf0, tmp1, b_scales_vec, vl / 4);
+                    }
+
+                    const int64_t A1 = *(const int64_t *)&a_ptr[l].qs[8];
+                    const int64_t A5 = *(const int64_t *)&a_ptr[l].qs[40];
+                    const int64_t A9 = *(const int64_t *)&a_ptr[l].qs[72];
+                    const int64_t Ad = *(const int64_t *)&a_ptr[l].qs[104];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l1;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A1, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A5, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A9, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ad, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l1 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l1));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[1], vl / 4);
+                        sumf1 = __riscv_vfmacc_vv_f32m1(sumf1, tmp1, b_scales_vec, vl / 4);
+                    }
+
+                    const int64_t A2 = *(const int64_t *)&a_ptr[l].qs[16];
+                    const int64_t A6 = *(const int64_t *)&a_ptr[l].qs[48];
+                    const int64_t Aa = *(const int64_t *)&a_ptr[l].qs[80];
+                    const int64_t Ae = *(const int64_t *)&a_ptr[l].qs[112];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l2;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A2, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A6, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Aa, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ae, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l2 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l2));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[2], vl / 4);
+                        sumf2 = __riscv_vfmacc_vv_f32m1(sumf2, tmp1, b_scales_vec, vl / 4);
+                    }
+
+                    const int64_t A3 = *(const int64_t *)&a_ptr[l].qs[24];
+                    const int64_t A7 = *(const int64_t *)&a_ptr[l].qs[56];
+                    const int64_t Ab = *(const int64_t *)&a_ptr[l].qs[88];
+                    const int64_t Af = *(const int64_t *)&a_ptr[l].qs[120];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l3;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A3, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A7, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ab, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Af, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l3 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l3));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[3], vl / 4);
+                        sumf3 = __riscv_vfmacc_vv_f32m1(sumf3, tmp1, b_scales_vec, vl / 4);
+                    }
+                }
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 0) * bs + x * ncols_interleaved], sumf0, vl / 4);
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 1) * bs + x * ncols_interleaved], sumf1, vl / 4);
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 2) * bs + x * ncols_interleaved], sumf2, vl / 4);
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 3) * bs + x * ncols_interleaved], sumf3, vl / 4);
+            }
+        }
+
+        return;
+    }
+
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}

ggml/src/ggml-cpu/common.h

@@ -1,7 +1,7 @@
 #pragma once
 
 #include "ggml.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-impl.h"
 

ggml/src/ggml-cpu/ggml-cpu-aarch64.h

@@ -1,8 +0,0 @@
-#pragma once
-
-#include "ggml-cpu-traits.h"
-#include "ggml.h"
-
-// GGML internal header
-
-ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -503,6 +503,34 @@ static __m256 __lasx_xvreplfr2vr_s(const float val) {
 // TODO: move to ggml-threading
 void ggml_barrier(struct ggml_threadpool * tp);
 
+void ggml_threadpool_chunk_set(struct ggml_threadpool * tp, int value);
+int  ggml_threadpool_chunk_add(struct ggml_threadpool * tp, int value);
+
 #ifdef __cplusplus
 }
 #endif
+
+#define GGML_DO_PRAGMA_(x) _Pragma (#x)
+#define GGML_DO_PRAGMA(x) GGML_DO_PRAGMA_(x)
+#if defined(GGML_CPU_GENERIC) || defined(__HIPCC__) || defined(__APPLE__)
+// Note for Apple targets:
+// - clang: aliases are not supported on darwin
+// - all native kernels need to be implemented in both x86 and arm files
+// - on iOS, tvOS, and visionOS, if cmake cannot determine the target architecture, all `_generic` names are replaced by defines
+# define GGML_WEAK_ALIAS(name, alias)
+#elif defined(__GNUC__)
+// GCC/Clang on *nix
+# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(weak name = alias) // NOLINT
+#elif defined(_MSC_VER) && defined(_WIN64)
+// MSVC
+// Note: C name mangling varies across different calling conventions
+// see https://learn.microsoft.com/en-us/cpp/build/reference/decorated-names?view=msvc-170
+# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(comment(linker, "/alternatename:" #name "=" #alias))
+#elif defined(_MSC_VER) && defined(WIN32)
+// ref: https://github.com/ggml-org/whisper.cpp/pull/3239#issuecomment-2958224591
+# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(comment(linker, "/alternatename:_" #name "=_" #alias))
+#else
+# error "Unsupported compiler for GGML_WEAK_ALIAS"
+#endif
+
+#define GGML_CPU_NATIVE_IMPL(name) GGML_WEAK_ALIAS(name, name ## _generic)

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

@@ -3,11 +3,11 @@
 
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
-#include "ggml-cpu-quants.h"
+#include "quants.h"
 #include "ggml-threading.h"
 #include "unary-ops.h"
 #include "binary-ops.h"
@@ -559,6 +559,14 @@ void ggml_barrier(struct ggml_threadpool * tp) {
 #endif
 }
 
+void ggml_threadpool_chunk_set(struct ggml_threadpool * tp, int value) {
+    atomic_store_explicit(&tp->current_chunk, value, memory_order_relaxed);
+}
+
+int ggml_threadpool_chunk_add(struct ggml_threadpool * tp, int value) {
+    return atomic_fetch_add_explicit(&tp->current_chunk, value, memory_order_relaxed);
+}
+
 #if defined(__gnu_linux__)
 static cpu_set_t ggml_get_numa_affinity(void) {
     cpu_set_t cpuset;

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -1,8 +1,8 @@
 #include "ggml-backend.h"
 #include "ggml-backend-impl.h"
 #include "ggml-cpu.h"
-#include "ggml-cpu-aarch64.h"
-#include "ggml-cpu-traits.h"
+#include "repack.h"
+#include "traits.h"
 #include "ggml-impl.h"
 #include "amx/amx.h"
 
@@ -11,7 +11,7 @@
 #include <vector>
 
 #ifdef GGML_USE_CPU_HBM
-#    include "ggml-cpu-hbm.h"
+#    include "hbm.h"
 #endif
 
 #ifdef GGML_USE_CPU_KLEIDIAI
@@ -51,9 +51,9 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
         }
 #endif
 
-#ifdef GGML_USE_CPU_AARCH64
-        if (ggml_backend_cpu_aarch64_buffer_type()) {
-            bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
+#ifdef GGML_USE_CPU_REPACK
+        if (ggml_backend_cpu_repack_buffer_type()) {
+            bufts.push_back(ggml_backend_cpu_repack_buffer_type());
         }
 #endif
 
@@ -596,8 +596,8 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
     #ifdef GGML_USE_CPU_KLEIDIAI
         features.push_back({ "KLEIDIAI", "1" });
     #endif
-    #ifdef GGML_USE_CPU_AARCH64
-        features.push_back({ "AARCH64_REPACK", "1" });
+    #ifdef GGML_USE_CPU_REPACK
+        features.push_back({ "REPACK", "1" });
     #endif
 
         features.push_back({ nullptr, nullptr });

ggml/src/ggml-cpu/hbm.cpp

@@ -5,7 +5,7 @@
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
 
-#include "ggml-cpu-hbm.h"
+#include "hbm.h"
 
 // buffer type HBM
 

ggml/src/ggml-cpu/kleidiai/kleidiai.cpp

@@ -26,7 +26,7 @@
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
 #include "ggml-threading.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #include "kernels.h"
 

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -53,7 +53,6 @@
 #include "ggml-cpu-impl.h"
 #include "ggml-quants.h"
 
-#include <atomic>
 #include <array>
 #include <type_traits>
 
@@ -394,8 +393,6 @@ class tinyBLAS {
 
     template <int RM, int RN, int BM>
     NOINLINE void gemm(int64_t m, int64_t n, int64_t BN) {
-        static std::atomic<int64_t> current_chunk;
-
         GGML_ASSERT(m % (RM * BM) == 0);
         const int64_t ytiles = m / (RM * BM);
         const int64_t xtiles = (n + RN -1) / RN;
@@ -410,7 +407,7 @@ class tinyBLAS {
         if (params->ith == 0) {
             GGML_ASSERT( jj_BN * SIZE_BN + (NB_BN - jj_BN) * (SIZE_BN - 1) == xtiles);
             // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
-            std::atomic_store_explicit(&current_chunk, (int64_t)params->nth, std::memory_order_relaxed);
+            ggml_threadpool_chunk_set(params->threadpool, params->nth);
         }
 
         ggml_barrier(params->threadpool);
@@ -439,8 +436,7 @@ class tinyBLAS {
                 GGML_ASSERT(jj == jj2);
             }
 
-            // next step.
-            job = std::atomic_fetch_add_explicit(&current_chunk, (int64_t)1, std::memory_order_relaxed);
+            job = ggml_threadpool_chunk_add(params->threadpool, 1);
         }
 
         ggml_barrier(params->threadpool);

ggml/src/ggml-cpu/quants.h

@@ -58,6 +58,32 @@ void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const
 void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
+// Generic implementation
+void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy,  size_t by, int nrc);
+void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
 #ifdef __cplusplus
 }
 #endif

ggml/src/ggml-cpu/repack.h

@@ -0,0 +1,103 @@
+#pragma once
+
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+
+#include "traits.h"
+#include "ggml.h"
+
+// GGML internal header
+
+ggml_backend_buffer_type_t ggml_backend_cpu_repack_buffer_type(void);
+
+template <int K> constexpr int QK_0() {
+    if constexpr (K == 4) {
+        return QK4_0;
+    }
+    if constexpr (K == 8) {
+        return QK8_0;
+    }
+    return -1;
+}
+
+template <int K, int N> struct block {
+    ggml_half d[N];                         // deltas for N qK_0 blocks
+    int8_t    qs[(QK_0<K>() * N * K) / 8];  // quants for N qK_0 blocks
+};
+
+// control size
+static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding");
+static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding");
+static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding");
+static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding");
+
+using block_q4_0x4 = block<4, 4>;
+using block_q4_0x8 = block<4, 8>;
+using block_q8_0x4 = block<8, 4>;
+using block_q8_0x8 = block<8, 8>;
+
+struct block_q4_Kx8 {
+    ggml_half d[8];      // super-block scale for quantized scales
+    ggml_half dmin[8];   // super-block scale for quantized mins
+    uint8_t scales[96];  // scales and mins, quantized with 6 bits
+    uint8_t qs[1024];    // 4--bit quants
+};
+
+static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
+
+struct block_q8_Kx4 {
+    float d[4];              // delta
+    int8_t qs[QK_K * 4];     // quants
+    int16_t bsums[QK_K / 4]; // sum of quants in groups of 16
+};
+
+static_assert(sizeof(block_q8_Kx4) == sizeof(float) * 4 + QK_K * 4 + (QK_K / 4) * sizeof(int16_t), "wrong q8_K block size/padding");
+
+struct block_iq4_nlx4 {
+    ggml_half d[4];            // deltas for 4 iq4_nl blocks
+    uint8_t   qs[QK4_NL * 2];  // nibbles / quants for 4 iq4_nl blocks
+};
+
+static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+// Workaround for clang:
+// clang++ complains: ``error: call to 'ggml_gemm_q4_0_4x4_q8_0' is ambiguous''
+// repro: https://godbolt.org/z/oKdeWKonM (ICE), https://godbolt.org/z/1szq6P36v (ambiguous call)
+#if defined(GGML_CPU_CLANG_WORKAROUND) || defined(__APPLE__) || !(defined(__GNUC__) && defined(__clang__)) || defined(__HIPCC__)
+void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+#endif // !defined(__clang__)
+
+// Native implementations
+void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif

ggml/src/ggml-cpu/traits.cpp

@@ -1,4 +1,4 @@
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"

ggml/src/ggml-cuda/common.cuh

@@ -207,9 +207,9 @@ typedef float2 dfloat2;
 #define FP16_MMA_AVAILABLE
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 
-#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
+#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 #define FP16_MMA_AVAILABLE
-#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define NEW_MMA_AVAILABLE
@@ -262,11 +262,11 @@ static bool cp_async_available(const int cc) {
 }
 
 static constexpr __device__ int ggml_cuda_get_physical_warp_size() {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
-    return __AMDGCN_WAVEFRONT_SIZE;
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
+    return 64;
 #else
     return 32;
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
 }
 
 [[noreturn]]
@@ -466,9 +466,6 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }
 
-// TODO: move to ggml-common.h
-static constexpr __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
 typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
 
 static __device__ __forceinline__ float get_alibi_slope(

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

@@ -10,6 +10,8 @@ __global__ void __launch_bounds__(splitD, 2)
                  float * __restrict__ dst, const int64_t L) {
     GGML_UNUSED(src1_nb0);
     GGML_UNUSED(src2_nb0);
+
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     const int bidx = blockIdx.x;  // split along B
     const int bidy = blockIdx.y;  // split along D
     const int tid  = threadIdx.x;
@@ -44,16 +46,16 @@ __global__ void __launch_bounds__(splitD, 2)
     if (N == 16) {
 #pragma unroll
         for (size_t i = 0; i < splitD / 4; i += 2) {
-            float value = A_block[(wid * warpSize + i) * stride_A + wtid];
+            float value = A_block[(wid * warp_size + i) * stride_A + wtid];
             // todo: bank conflict
             // I am always confused with how to use the swizzling method to solve
             // bank conflit. Hoping somebody can tell me.
-            smem_A[(wid * warpSize + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
+            smem_A[(wid * warp_size + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
         }
 #pragma unroll
         for (size_t i = 0; i < splitD / 4; i += 2) {
-            float value = s0_block[(wid * warpSize + i) * stride_s0 + wtid];
-            smem_s0[(wid * warpSize + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
+            float value = s0_block[(wid * warp_size + i) * stride_s0 + wtid];
+            smem_s0[(wid * warp_size + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
         }
     }
 

ggml/src/ggml-hip/CMakeLists.txt

@@ -113,6 +113,10 @@ if (GGML_HIP_ROCWMMA_FATTN)
     add_compile_definitions(GGML_HIP_ROCWMMA_FATTN)
 endif()
 
+if (GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12 OR ${hip_VERSION} VERSION_GREATER_EQUAL 7.0)
+    add_compile_definitions(GGML_HIP_ROCWMMA_FATTN_GFX12)
+endif()
+
 if (NOT GGML_CUDA_FA)
     add_compile_definitions(GGML_CUDA_NO_FA)
 endif()

ggml/src/ggml-metal/CMakeLists.txt

@@ -44,21 +44,22 @@ if (GGML_METAL_EMBED_LIBRARY)
     set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
 
     add_custom_command(
-        OUTPUT ${METALLIB_EMBED_ASM}
+        OUTPUT "${METALLIB_EMBED_ASM}"
         COMMAND echo "Embedding Metal library"
-        COMMAND sed -e '/__embed_ggml-common.h__/r         ${METALLIB_COMMON}' -e '/__embed_ggml-common.h__/d'         < ${METALLIB_SOURCE}           > ${METALLIB_SOURCE_EMBED_TMP}
-        COMMAND sed -e '/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}'   -e '/\#include \"ggml-metal-impl.h\"/d' < ${METALLIB_SOURCE_EMBED_TMP} > ${METALLIB_SOURCE_EMBED}
-        COMMAND echo ".section __DATA,__ggml_metallib"          >  ${METALLIB_EMBED_ASM}
-        COMMAND echo ".globl _ggml_metallib_start"              >> ${METALLIB_EMBED_ASM}
-        COMMAND echo "_ggml_metallib_start:"                    >> ${METALLIB_EMBED_ASM}
-        COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM}
-        COMMAND echo ".globl _ggml_metallib_end"                >> ${METALLIB_EMBED_ASM}
-        COMMAND echo "_ggml_metallib_end:"                      >> ${METALLIB_EMBED_ASM}
+        COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}"       -e "/__embed_ggml-common.h__/d"         < "${METALLIB_SOURCE}"           > "${METALLIB_SOURCE_EMBED_TMP}"
+        COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${METALLIB_SOURCE_EMBED_TMP}" > "${METALLIB_SOURCE_EMBED}"
+        COMMAND echo ".section __DATA,__ggml_metallib"          >  "${METALLIB_EMBED_ASM}"
+        COMMAND echo ".globl _ggml_metallib_start"              >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo "_ggml_metallib_start:"                    >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo .incbin "\"${METALLIB_SOURCE_EMBED}\""     >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo ".globl _ggml_metallib_end"                >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo "_ggml_metallib_end:"                      >> "${METALLIB_EMBED_ASM}"
         DEPENDS ../ggml-common.h ggml-metal.metal ggml-metal-impl.h
         COMMENT "Generate assembly for embedded Metal library"
+        VERBATIM
     )
 
-    target_sources(ggml-metal PRIVATE ${METALLIB_EMBED_ASM})
+    target_sources(ggml-metal PRIVATE "${METALLIB_EMBED_ASM}")
 else()
     if (GGML_METAL_SHADER_DEBUG)
         # custom command to do the following:

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

@@ -3333,8 +3333,6 @@ kernel void kernel_flash_attn_ext(
 
     threadgroup q_t  * sq  = (threadgroup q_t  *) (shmem_f16 +                0*DK); // holds the query data
     threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 +                0*DK); // same as above but in q4_t
-    threadgroup o_t  * so  = (threadgroup o_t  *) (shmem_f16 +                0*DK); // reuse query data for accumulation
-    threadgroup o4_t * so4 = (threadgroup o4_t *) (shmem_f16 +                0*DK); // same as above but in o4_t
     threadgroup s_t  * ss  = (threadgroup s_t  *) (shmem_f16 + 2*sgitg*SH + 2*Q*DK); // scratch buffer for attention, mask and diagonal matrix
 
     threadgroup k_t    * sk    = (threadgroup k_t    *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // scratch buffer to load K in shared memory
@@ -3548,20 +3546,20 @@ kernel void kernel_flash_attn_ext(
 
             // O = diag(ms)*O
             {
-                s8x8_t mm;
-                simdgroup_load(mm, ss + 2*C, TS, 0, false);
+                s8x8_t ms;
+                simdgroup_load(ms, ss + 2*C, TS, 0, false);
 
                 #pragma unroll(DV8)
                 for (short i = 0; i < DV8; ++i) {
-                    simdgroup_multiply(lo[i], mm, lo[i]);
+                    simdgroup_multiply(lo[i], ms, lo[i]);
                 }
             }
 
             // O = O + (Q*K^T)*V
             {
                 for (short cc = 0; cc < C/8; ++cc) {
-                    s8x8_t ms;
-                    simdgroup_load(ms, ss + 8*cc, TS, 0, false);
+                    s8x8_t vs;
+                    simdgroup_load(vs, ss + 8*cc, TS, 0, false);
 
                     if (is_same<vd4x4_t, v4x4_t>::value) {
                         // we can read directly from global memory
@@ -3572,7 +3570,7 @@ kernel void kernel_flash_attn_ext(
                             v8x8_t mv;
                             simdgroup_load(mv, pv + i*8, args.nb21/sizeof(v_t), 0, false); // TODO: use ne20
 
-                            simdgroup_multiply_accumulate(lo[i], ms, mv, lo[i]);
+                            simdgroup_multiply_accumulate(lo[i], vs, mv, lo[i]);
                         }
                     } else {
                         for (short ii = 0; ii < DV16; ii += 4) {
@@ -3593,10 +3591,10 @@ kernel void kernel_flash_attn_ext(
                                     v8x8_t mv;
 
                                     simdgroup_load(mv, sv + 16*k + 0*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], ms, mv, lo[2*(ii + k) + 0]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], vs, mv, lo[2*(ii + k) + 0]);
 
                                     simdgroup_load(mv, sv + 16*k + 1*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], ms, mv, lo[2*(ii + k) + 1]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], vs, mv, lo[2*(ii + k) + 1]);
                                 }
                             } else {
                                 if (ii + tx < DV16) {
@@ -3611,10 +3609,10 @@ kernel void kernel_flash_attn_ext(
                                     v8x8_t mv;
 
                                     simdgroup_load(mv, sv + 16*k + 0*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], ms, mv, lo[2*(ii + k) + 0]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], vs, mv, lo[2*(ii + k) + 0]);
 
                                     simdgroup_load(mv, sv + 16*k + 1*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], ms, mv, lo[2*(ii + k) + 1]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], vs, mv, lo[2*(ii + k) + 1]);
                                 }
                             }
                         }
@@ -3624,83 +3622,80 @@ kernel void kernel_flash_attn_ext(
         }
 
         // these are needed for reducing the results from the simdgroups (reuse the ss buffer)
-        for (short j = 0; j < Q; ++j) {
-            if (tiisg == 0) {
-                ss[j*TS + 0] = S[j];
-                ss[j*TS + 1] = M[j];
-            }
+        for (short j = tiisg; j < Q; j += NW) {
+            ss[j*TS + 0] = S[j];
+            ss[j*TS + 1] = M[j];
         }
     }
 
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    threadgroup float  * so  = (threadgroup float  *) (shmem_f16 + 0*DK); // reuse query data for accumulation
+    threadgroup float4 * so4 = (threadgroup float4 *) (shmem_f16 + 0*DK);
+
+    // store result to shared memory in F32
+    if (sgitg == 0) {
+        for (short i = 0; i < DV8; ++i) {
+            //simdgroup_store(lo[i], so + i*8, DV, 0, false);
+            simdgroup_float8x8 t(1.0f);
+            simdgroup_multiply(t, lo[i], t);
+            simdgroup_store(t, so + i*8, DV, 0, false);
+        }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
     // reduce the warps sequentially
     for (ushort sg = 1; sg < nsg; ++sg) {
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        // each simdgroup stores its output to shared memory, reusing sq
         if (sgitg == sg) {
-            for (short i = 0; i < DV8; ++i) {
-                simdgroup_store(lo[i], so + i*8, DV, 0, false);
-            }
-        }
+            for (short j = tiisg; j < Q; j += NW) {
+                const float S0 = ss[j*TS - 1*SH + 0];
+                const float S1 = ss[j*TS        + 0];
 
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        // the first simdgroup accumulates the results from the other simdgroups
-        if (sgitg == 0) {
-            for (short j = 0; j < Q; ++j) {
-                const float S0 = ss[j*TS +         0];
-                const float S1 = ss[j*TS + sg*SH + 0];
-
-                const float M0 = ss[j*TS +         1];
-                const float M1 = ss[j*TS + sg*SH + 1];
+                const float M0 = ss[j*TS - 1*SH + 1];
+                const float M1 = ss[j*TS        + 1];
 
                 const float M = max(M0, M1);
 
-                const float ms0 = exp(M0 - M);
-                const float ms1 = exp(M1 - M);
+                float ms0 = exp(M0 - M);
+                float ms1 = exp(M1 - M);
 
                 const float S = S0*ms0 + S1*ms1;
 
-                if (tiisg == 0) {
-                    ss[j*TS + 0] = S;
-                    ss[j*TS + 1] = M;
+                ss[j*TS + 0] = S;
+                ss[j*TS + 1] = M;
 
-                    ss[j*TS + 2*C + j        ] = ms0;
-                    ss[j*TS + 2*C + j + sg*SH] = ms1;
-                }
+                ss[j*TS + 2*C + j - 1*SH] = ms0;
+                ss[j*TS + 2*C + j       ] = ms1;
             }
 
+            //simdgroup_barrier(mem_flags::mem_threadgroup);
+
             // O_0 = diag(ms0)*O_0 + diag(ms1)*O_1
             {
                 s8x8_t ms0;
                 s8x8_t ms1;
 
-                simdgroup_load(ms0, ss + 2*C,         TS, 0, false);
-                simdgroup_load(ms1, ss + 2*C + sg*SH, TS, 0, false);
+                simdgroup_load(ms0, ss + 2*C - 1*SH, TS, 0, false);
+                simdgroup_load(ms1, ss + 2*C,        TS, 0, false);
 
                 #pragma unroll(DV8)
                 for (short i = 0; i < DV8; ++i) {
-                    o8x8_t t;
+                    simdgroup_float8x8 t;
 
                     simdgroup_load    (t, so + i*8, DV, 0, false);
-                    simdgroup_multiply(t, ms1, t);
+                    simdgroup_multiply(t, ms0, t);
 
-                    simdgroup_multiply_accumulate(lo[i], ms0, lo[i], t);
+                    simdgroup_multiply_accumulate(t, ms1, lo[i], t);
+                    simdgroup_store(t, so + i*8, DV, 0, false);
                 }
             }
         }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
-    // store result to shared memory (reuse sq)
-    if (sgitg == 0) {
-        for (short i = 0; i < DV8; ++i) {
-            simdgroup_store(lo[i], so + i*8, DV, 0, false);
-        }
-    }
-
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    threadgroup s_t * sf = (threadgroup s_t *) (shmem_f16 + 2*Q*DK);
+    threadgroup s_t * sf = (threadgroup s_t *) (shmem_f16 + 2*(nsg-1)*SH + 2*Q*DK);
 
     // final rescale with 1/S and store to global memory
     for (short j = sgitg; j < Q && iq1 + j < args.ne01; j += nsg) {
@@ -3723,8 +3718,8 @@ kernel void kernel_flash_attn_ext(
     half,   half4x4,   simdgroup_half8x8,  \
     float,             simdgroup_float8x8, \
     float,             simdgroup_float8x8, \
-    float,  float4,    simdgroup_float8x8
-    //half,   half4,     simdgroup_half8x8
+    half,   half4,     simdgroup_half8x8
+    //float,  float4,    simdgroup_float8x8
 
 #define FA_TYPES_BF \
     bfloat, bfloat4,   simdgroup_bfloat8x8, \
@@ -3732,8 +3727,8 @@ kernel void kernel_flash_attn_ext(
     bfloat, bfloat4x4, simdgroup_bfloat8x8, \
     float,             simdgroup_float8x8,  \
     float,             simdgroup_float8x8,  \
-    float,  float4,    simdgroup_float8x8
-    //half,   half4,     simdgroup_half8x8
+    half,   half4,     simdgroup_half8x8
+    //float,  float4,    simdgroup_float8x8
 
 typedef decltype(kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 64, 64>) flash_attn_ext_t;
 

ggml/src/ggml-opencl/CMakeLists.txt

@@ -80,6 +80,7 @@ set(GGML_OPENCL_KERNELS
     mul_mv_q4_0_f32_1d_8x_flat
     mul_mv_q4_0_f32_1d_16x_flat
     mul_mv_q6_k
+    mul_mv_id_q4_0_f32_8x_flat
     mul
     norm
     relu

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

@@ -321,6 +321,7 @@ struct ggml_backend_opencl_context {
     cl_program program_upscale;
     cl_program program_concat;
     cl_program program_tsembd;
+    cl_program program_mul_mv_id_q4_0_f32_8x_flat;
 
     cl_kernel kernel_add, kernel_add_row;
     cl_kernel kernel_mul, kernel_mul_row;
@@ -366,6 +367,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_concat_f32_contiguous;
     cl_kernel kernel_concat_f32_non_contiguous;
     cl_kernel kernel_timestep_embedding;
+    cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
 
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     // Transpose kernels
@@ -1112,7 +1114,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
-        // repeat
+    // repeat
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
         const std::string kernel_src {
@@ -1256,6 +1258,22 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         }
     }
 
+    // mul_mv_id_q4_0_f32_8x_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_q4_0_f32_8x_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_q4_0_f32_8x_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_id_q4_0_f32_8x_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_q4_0_f32_8x_flat = clCreateKernel(backend_ctx->program_mul_mv_id_q4_0_f32_8x_flat, "kernel_mul_mv_id_q4_0_f32_8x_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // Adreno kernels
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     // transpose
@@ -2178,6 +2196,13 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
             }
             return false;
+        case GGML_OP_MUL_MAT_ID:
+            if (op->src[0]->type == GGML_TYPE_Q4_0) {
+                if (op->src[1]->type == GGML_TYPE_F32) {
+                    return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
+                }
+            }
+            return false;
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
@@ -5536,6 +5561,136 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     }
 }
 
+static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    const ggml_tensor * src2 = dst->src[2];
+    GGML_ASSERT(src2);
+    GGML_ASSERT(src2->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offset2 = extra2->offset + src2->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+#ifdef GGML_OPENCL_SOA_Q
+    ggml_tensor_extra_cl_q4_0 * extra0_q4_0 = (ggml_tensor_extra_cl_q4_0 *)src0->extra;
+#endif
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb02 = src0->nb[2];
+
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3];
+
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+
+    const int ne20 = src2->ne[0];
+    const int ne21 = src2->ne[1];
+
+    const cl_ulong nb21 = src2->nb[1];
+
+    const int ne0 = dst->ne[0];
+    const int ne1 = dst->ne[1];
+
+    const int r2 = ne12/ne02;
+    const int r3 = ne13/ne03;
+    const int dst_rows = ne20*ne21; // ne20 = n_used_experts, ne21 = n_rows
+
+    GGML_ASSERT(ne00 == ne10);
+
+    int sgs   = 32; // subgroup size
+    int nsg   = 1;  // number of subgroups
+    int nrows = 1;  // number of row in src1
+    int ndst  = 4;  // number of values produced by each subgroup
+
+    cl_kernel kernel;
+
+    // subgroup mat vec
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0: {
+            kernel = backend_ctx->kernel_mul_mv_id_q4_0_f32_8x_flat;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 1;
+                ndst = 8;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 1;
+                ndst = 8;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q4_0->q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q4_0->d));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb00));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &r3));
+
+            break;
+        }
+        default:
+            GGML_ASSERT(false && "not implemented");;
+    }
+
+    int _ne1 = 1;
+    int ne123 = dst_rows;
+
+    size_t global_work_size[] = {(size_t)(ne01+ndst*nsg-1)/(ndst*nsg)*sgs, (size_t)(_ne1+nrows-1)/nrows*nsg, (size_t)ne123};
+    size_t local_work_size[] = {(size_t)sgs, (size_t)nsg, 1};
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+#endif
+}
+
 static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -6444,6 +6599,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_mul_mat;
             break;
+        case GGML_OP_MUL_MAT_ID:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_mul_mat_id;
+            break;
         case GGML_OP_SCALE:
             if (!any_on_device) {
                 return false;

ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl

@@ -0,0 +1,283 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+// This function requires the original shuffled weights.
+// As a reminder, the original weights are shuffled so that (q[0], q[16]) are
+// packed together in a byte, so are (q[1], q[17]) and so on.
+inline float block_q_4_0_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float           d   = *dh;
+    global ushort * qs  = ((global ushort *)x + il/2);
+    float           acc = 0.f;
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+//
+// This variant outputs 8 values.
+//
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 8 // each SIMD group works on 8 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_DST 8
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32_8x_flat(
+        global char  * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = 0;
+
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_0/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float8 sumf = 0.f;
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_0 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0.f;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 0*nb*QK4_0/2, d + ib + 0*nb, sumy, yl, il);
+        sumf.s1 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 1*nb*QK4_0/2, d + ib + 1*nb, sumy, yl, il);
+        sumf.s2 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 2*nb*QK4_0/2, d + ib + 2*nb, sumy, yl, il);
+        sumf.s3 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 3*nb*QK4_0/2, d + ib + 3*nb, sumy, yl, il);
+
+        sumf.s4 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 4*nb*QK4_0/2, d + ib + 4*nb, sumy, yl, il);
+        sumf.s5 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 5*nb*QK4_0/2, d + ib + 5*nb, sumy, yl, il);
+        sumf.s6 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 6*nb*QK4_0/2, d + ib + 6*nb, sumy, yl, il);
+        sumf.s7 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float8 tot = (float8)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7)
+    );
+
+    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;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_q4_0_f32_8x_flat(
+        global char  *  src0_q,
+        global half  *  src0_d,
+        global float *  src1,
+        ulong           offset1,
+        global char  *  src2,
+        ulong           offset2,
+        global float *  dst,
+        ulong           offsetd,
+        int             ne00,
+        int             ne01,
+        int             ne02,
+        ulong           nb00,
+        ulong           nb02,
+        int             ne10,
+        int             ne11,
+        int             ne12,
+        ulong           nb11,
+        ulong           nb12,
+        int             ne20,
+        int             ne21,
+        ulong           nb21,
+        int             ne0,
+        int             ne1,
+        int             r2,
+        int             r3
+) {
+    src1 = (global float *)((global char *)src1 + offset1);
+    src2 = (global char  *)((global char *)src2 + offset2);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    const int iid1 = get_group_id(2)/ne20;
+    const int idx  = get_group_id(2)%ne20;
+
+    const int i02 = ((global int *)(src2 + iid1*nb21))[idx];
+
+    const int i11 = idx%ne11;
+    const int i12 = iid1;
+
+    const int i1 = idx;
+    const int i2 = i12;
+
+    global char  * src0_q_cur = src0_q + (i02*nb02/nb00)*(QK4_0/2);
+    global half  * src0_d_cur = src0_d + (i02*nb02/nb00);
+    global float * src1_cur   = (global float *)((global char *) src1  + i11*nb11 + i12*nb12);
+    global float * dst_cur    = dst + i1*ne0 + i2*ne1*ne0;
+
+    mul_vec_q_n_f32_8x_flat(src0_q_cur, src0_d_cur, src1_cur, dst_cur, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

ggml/src/ggml-quants.c

@@ -2425,8 +2425,6 @@ void dequantize_row_iq1_m(const block_iq1_m * GGML_RESTRICT x, float * GGML_REST
     }
 }
 
-static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
 void dequantize_row_iq4_nl(const block_iq4_nl * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK4_NL == 0);
     const int64_t nb = k / QK4_NL;

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -53,6 +53,9 @@ struct socket_t {
     }
 };
 
+// macro for nicer error messages on server crash
+#define RPC_STATUS_ASSERT(x) if (!(x)) GGML_ABORT("Remote RPC server crashed or returned malformed response")
+
 // all RPC structures must be packed
 #pragma pack(push, 1)
 // ggml_tensor is serialized into rpc_tensor
@@ -425,7 +428,7 @@ static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cm
 static bool check_server_version(const std::shared_ptr<socket_t> & sock) {
     rpc_msg_hello_rsp response;
     bool status = send_rpc_cmd(sock, RPC_CMD_HELLO, nullptr, 0, &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     if (response.major != RPC_PROTO_MAJOR_VERSION || response.minor > RPC_PROTO_MINOR_VERSION) {
         fprintf(stderr, "RPC server version mismatch: %d.%d.%d\n", response.major, response.minor, response.patch);
         return false;
@@ -481,7 +484,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     rpc_msg_free_buffer_req request = {ctx->remote_ptr};
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_FREE_BUFFER, &request, sizeof(request), nullptr, 0);
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     delete ctx;
 }
 
@@ -493,7 +496,7 @@ static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
     rpc_msg_buffer_get_base_req request = {ctx->remote_ptr};
     rpc_msg_buffer_get_base_rsp response;
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_GET_BASE, &request, sizeof(request), &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     ctx->base_ptr = reinterpret_cast<void *>(response.base_ptr);
     return ctx->base_ptr;
 }
@@ -545,7 +548,7 @@ static enum ggml_status ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_
         request.tensor = serialize_tensor(tensor);
 
         bool status = send_rpc_cmd(ctx->sock, RPC_CMD_INIT_TENSOR, &request, sizeof(request), nullptr, 0);
-        GGML_ASSERT(status);
+        RPC_STATUS_ASSERT(status);
     }
     return GGML_STATUS_SUCCESS;
 }
@@ -560,7 +563,7 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggm
         request.hash = fnv_hash((const uint8_t*)data, size);
         rpc_msg_set_tensor_hash_rsp response;
         bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR_HASH, &request, sizeof(request), &response, sizeof(response));
-        GGML_ASSERT(status);
+        RPC_STATUS_ASSERT(status);
         if (response.result) {
             // the server has the same data, no need to send it
             return;
@@ -573,7 +576,7 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggm
     memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
     memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), data, size);
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR, input.data(), input.size());
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
 }
 
 static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
@@ -583,7 +586,7 @@ static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, con
     request.offset = offset;
     request.size = size;
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_GET_TENSOR, &request, sizeof(request), data, size);
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
 }
 
 static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
@@ -601,7 +604,7 @@ static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, con
     request.dst = serialize_tensor(dst);
     rpc_msg_copy_tensor_rsp response;
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_COPY_TENSOR, &request, sizeof(request), &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     return response.result;
 }
 
@@ -609,7 +612,7 @@ static void ggml_backend_rpc_buffer_clear(ggml_backend_buffer_t buffer, uint8_t
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     rpc_msg_buffer_clear_req request = {ctx->remote_ptr, value};
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_CLEAR, &request, sizeof(request), nullptr, 0);
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
 }
 
 static ggml_backend_buffer_i ggml_backend_rpc_buffer_interface = {
@@ -635,7 +638,7 @@ static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_back
     rpc_msg_alloc_buffer_rsp response;
     auto sock = get_socket(buft_ctx->endpoint);
     bool status = send_rpc_cmd(sock, RPC_CMD_ALLOC_BUFFER, &request, sizeof(request), &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     if (response.remote_ptr != 0) {
         ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft,
             ggml_backend_rpc_buffer_interface,
@@ -650,7 +653,7 @@ static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_back
 static size_t get_alignment(const std::shared_ptr<socket_t> & sock) {
     rpc_msg_get_alignment_rsp response;
     bool status = send_rpc_cmd(sock, RPC_CMD_GET_ALIGNMENT, nullptr, 0, &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     return response.alignment;
 }
 
@@ -662,7 +665,7 @@ static size_t ggml_backend_rpc_buffer_type_get_alignment(ggml_backend_buffer_typ
 static size_t get_max_size(const std::shared_ptr<socket_t> & sock) {
     rpc_msg_get_max_size_rsp response;
     bool status = send_rpc_cmd(sock, RPC_CMD_GET_MAX_SIZE, nullptr, 0, &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     return response.max_size;
 }
 
@@ -683,7 +686,7 @@ static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_ty
 
         rpc_msg_get_alloc_size_rsp response;
         bool status = send_rpc_cmd(sock, RPC_CMD_GET_ALLOC_SIZE, &request, sizeof(request), &response, sizeof(response));
-        GGML_ASSERT(status);
+        RPC_STATUS_ASSERT(status);
 
         return response.alloc_size;
     } else {
@@ -761,7 +764,7 @@ static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, g
     rpc_msg_graph_compute_rsp response;
     auto sock = get_socket(rpc_ctx->endpoint);
     bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_COMPUTE, input.data(), input.size(), &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     return (enum ggml_status)response.result;
 }
 
@@ -835,7 +838,7 @@ bool ggml_backend_is_rpc(ggml_backend_t backend) {
 static void get_device_memory(const std::shared_ptr<socket_t> & sock, size_t * free, size_t * total) {
     rpc_msg_get_device_memory_rsp response;
     bool status = send_rpc_cmd(sock, RPC_CMD_GET_DEVICE_MEMORY, nullptr, 0, &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     *free = response.free_mem;
     *total = response.total_mem;
 }

ggml/src/ggml-sycl/CMakeLists.txt

@@ -142,7 +142,7 @@ else()
         FetchContent_Declare(
             ONEMATH
             GIT_REPOSITORY https://github.com/uxlfoundation/oneMath.git
-            GIT_TAG c255b1b4c41e2ee3059455c1f96a965d6a62568a
+            GIT_TAG 8efe85f5aaebb37f1d8c503b7af66315feabf142
         )
         FetchContent_MakeAvailable(ONEMATH)
         # Create alias to match with find_package targets name

ggml/src/ggml-sycl/common.hpp

@@ -149,8 +149,6 @@ typedef sycl::float2 dfloat2;
 
 #define MMVQ_MAX_BATCH_SIZE  8
 
-static const int8_t kvalues_iq4nl[16]={-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
 static int g_all_sycl_device_count = -1;
 static bool g_ggml_backend_sycl_buffer_type_initialized = false;
 
@@ -515,9 +513,9 @@ constexpr size_t ceil_div(const size_t m, const size_t n) {
 
 bool gpu_has_xmx(sycl::device &dev);
 
-template <int N, class T> void debug_print_array(const std::string & prefix, const T array[N]) {
+template <int N, class T> std::string debug_get_array_str(const std::string & prefix, const T array[N]) {
     if (LIKELY(!g_ggml_sycl_debug)) {
-        return;
+        return "";
     }
     std::stringstream ss;
     ss << prefix << "=[";
@@ -528,29 +526,26 @@ template <int N, class T> void debug_print_array(const std::string & prefix, con
         ss << array[N - 1];
     }
     ss << "]";
-    GGML_SYCL_DEBUG("%s", ss.str().c_str());
+    return ss.str();
 }
 
-inline void debug_print_tensor(const std::string & prefix, const ggml_tensor * tensor,
-                               const std::string & suffix = "") {
-    if (LIKELY(!g_ggml_sycl_debug)) {
-        return;
-    }
-    GGML_SYCL_DEBUG("%s=", prefix.c_str());
+inline std::string debug_get_tensor_str(const std::string &prefix,
+        const ggml_tensor *tensor, const std::string &suffix = "") {
+    std::stringstream ss;
+    if (LIKELY(!g_ggml_sycl_debug)) { return ss.str(); }
+    ss << prefix.c_str() << "=";
     if (tensor) {
-        GGML_SYCL_DEBUG("'%s':type=%s", tensor->name, ggml_type_name(tensor->type));
-        debug_print_array<GGML_MAX_DIMS>(";ne", tensor->ne);
-        debug_print_array<GGML_MAX_DIMS>(";nb", tensor->nb);
-        if (!ggml_is_contiguous(tensor)) {
-            GGML_SYCL_DEBUG(";strided");
-        }
-        if (ggml_is_permuted(tensor)) {
-            GGML_SYCL_DEBUG(";permuted");
-        }
+        ss << "'" << tensor->name << "':type=" << ggml_type_name(tensor->type);
+        ss << debug_get_array_str<GGML_MAX_DIMS>(";ne", tensor->ne);
+        ss << debug_get_array_str<GGML_MAX_DIMS>(";nb", tensor->nb);
+
+        if (!ggml_is_contiguous(tensor)) { ss << ";strided"; }
+        if (ggml_is_permuted(tensor)) { ss << ";permuted"; }
     } else {
-        GGML_SYCL_DEBUG("nullptr");
+        ss << "nullptr";
     }
-    GGML_SYCL_DEBUG("%s", suffix.c_str());
+    ss << suffix;
+    return ss.str();
 }
 
 // Use scope_op_debug_print to log operations coming from running a model
@@ -566,10 +561,10 @@ struct scope_op_debug_print {
             return;
         }
         GGML_SYCL_DEBUG("[SYCL][OP] call %s%s:", func.data(), func_suffix.data());
-        debug_print_tensor(" dst", dst);
+        GGML_SYCL_DEBUG("%s", debug_get_tensor_str(" dst", dst).c_str());
         if (dst) {
             for (std::size_t i = 0; i < num_src; ++i) {
-                debug_print_tensor("\tsrc" + std::to_string(i), dst->src[i]);
+                GGML_SYCL_DEBUG("%s", debug_get_tensor_str("\tsrc" + std::to_string(i), dst->src[i]).c_str());
             }
         }
         GGML_SYCL_DEBUG("%s\n", suffix.data());

ggml/src/ggml-sycl/cpy.cpp

@@ -723,8 +723,7 @@ static void ggml_cpy_q4_1_q4_1(const char * cx, char * cdst, const int ne, const
 
 void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1) try {
     // Unlike other operators ggml_sycl_cpy takes 2 distinct tensors instead of a dst ggml_tensor and rely on its src field
-    scope_op_debug_print scope_dbg_print(__func__, src1, /*num_src=*/0,
-                                         std::string(" src0 type=") + ggml_type_name(src0->type));
+    scope_op_debug_print scope_dbg_print(__func__, src1, /*num_src=*/0, debug_get_tensor_str("\tsrc0", src0));
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 

ggml/src/ggml-sycl/gemm.hpp

@@ -65,6 +65,9 @@ public:
 
         dnnl::primitive_attr primitive_attr;
         primitive_attr.set_scratchpad_mode(dnnl::scratchpad_mode::user);
+#ifdef GGML_SYCL_F16
+        primitive_attr.set_fpmath_mode(dnnl::fpmath_mode::f16);
+#endif
 
         auto a_mem = dnnl::memory(a_in_md, eng, const_cast<void*>(a));
         auto b_mem = dnnl::memory(b_in_md, eng, const_cast<void*>(b));

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -347,7 +347,7 @@ static enum ggml_status
 ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
                                      ggml_tensor *tensor) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor, "\n");
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor, "\n").c_str());
     ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context;
 
     if (tensor->view_src != NULL) {
@@ -385,7 +385,7 @@ static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer,
                                                 const void *data, size_t offset,
                                                 size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
     ggml_sycl_set_device(ctx->device);
@@ -413,7 +413,7 @@ static void ggml_backend_sycl_buffer_get_tensor(ggml_backend_buffer_t buffer,
                                                 void *data, size_t offset,
                                                 size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
 
@@ -444,8 +444,8 @@ ggml_backend_sycl_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
                                     ggml_tensor *dst) try {
     bool is_cpy_supported = ggml_backend_buffer_is_sycl(src->buffer);
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": dst=", dst);
-    debug_print_tensor(" src=", src);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": dst", dst).c_str());
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(" src", src).c_str());
     GGML_SYCL_DEBUG(" is_cpy_supported=%d\n", is_cpy_supported);
     if (is_cpy_supported) {
         ggml_backend_sycl_buffer_context * src_ctx = (ggml_backend_sycl_buffer_context *)src->buffer->context;
@@ -525,7 +525,7 @@ catch (sycl::exception const &exc) {
 static void ggml_backend_sycl_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value,
                                                    size_t offset, size_t size) {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu value=%u\n", size, offset, value);
     ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *) buffer->context;
     SYCL_CHECK(ggml_sycl_set_device(ctx->device));
@@ -805,7 +805,7 @@ static enum ggml_status
 ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
                                            ggml_tensor *tensor) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor, "\n");
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor, "\n").c_str());
     GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported
 
     ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
@@ -891,7 +891,7 @@ ggml_backend_sycl_split_buffer_set_tensor(ggml_backend_buffer_t buffer,
                                           ggml_tensor *tensor, const void *data,
                                           size_t offset, size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     // split tensors must always be set in their entirety at once
     GGML_ASSERT(offset == 0);
@@ -947,7 +947,7 @@ ggml_backend_sycl_split_buffer_get_tensor(ggml_backend_buffer_t buffer,
                                           const ggml_tensor *tensor, void *data,
                                           size_t offset, size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     // split tensors must always be set in their entirety at once
     GGML_ASSERT(offset == 0);
@@ -2127,21 +2127,18 @@ inline void ggml_sycl_op_mul_mat_sycl(
         const sycl::half *src1_ptr = src1->type == GGML_TYPE_F16
                 ? (const sycl::half *)src1->data + src1_padded_row_size
                                          : src1_as_f16.get();
-        ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool(), row_diff * src1_ncols);
 
 #if GGML_SYCL_DNNL
         if (!g_ggml_sycl_disable_dnn) {
             DnnlGemmWrapper::row_gemm(ctx, src1_ncols, row_diff, ne10, src1_ptr,
                                       DnnlGemmWrapper::to_dt<sycl::half>(), src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
-                                      dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>(), stream);
-            scope_op_debug_print scope_dbg_print(__func__, "/to_fp32_sycl", dst, /*num_src=*/2,
-                                                 " : converting dst to fp32");
-            const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
-            to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff* src1_ncols, stream);
+                                      dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
         }
         else
 #endif
         {
+            ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool(), row_diff * src1_ncols);
+
             const sycl::half alpha_f16 = 1.0f;
             const sycl::half beta_f16  = 0.0f;
             SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm(
@@ -3866,7 +3863,7 @@ static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend,
                                                const void *data, size_t offset,
                                                size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
@@ -3887,7 +3884,7 @@ static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend,
                                                void *data, size_t offset,
                                                size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
@@ -3910,8 +3907,8 @@ static bool ggml_backend_sycl_cpy_tensor_async(ggml_backend_t backend,
     bool is_cpy_supported                = dst->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) &&
                             ggml_backend_buffer_is_sycl(src->buffer);
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": dst=", dst);
-    debug_print_tensor(" src=", src);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": dst", dst).c_str());
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(" src", src).c_str());
     GGML_SYCL_DEBUG(" is_cpy_supported=%d\n", is_cpy_supported);
     if (is_cpy_supported) {
         /*

ggml/src/ggml-vulkan/CMakeLists.txt

@@ -49,15 +49,7 @@ if (Vulkan_FOUND)
                              ../../include/ggml-vulkan.h
                             )
 
-    set(VULKAN_SHADER_GEN_CMAKE_ARGS
-        -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR}
-        -DCMAKE_RUNTIME_OUTPUT_DIRECTORY=${CMAKE_RUNTIME_OUTPUT_DIRECTORY}
-    )
-
-    set(VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS "")
-    if (CMAKE_BUILD_TYPE AND CMAKE_BUILD_TYPE MATCHES "Debug|Release|MinSizeRel|RelWithDebInfo")
-        list(APPEND VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS --config=${CMAKE_BUILD_TYPE})
-    endif()
+    set(VULKAN_SHADER_GEN_CMAKE_ARGS "")
 
     # Test all shader extensions
     test_shader_extension_support(
@@ -136,42 +128,39 @@ if (Vulkan_FOUND)
         set(HOST_CMAKE_TOOLCHAIN_FILE "")
     endif()
 
-    # Always use ExternalProject_Add approach
     include(ExternalProject)
 
-    # Add toolchain file if cross-compiling
     if (CMAKE_CROSSCOMPILING)
         list(APPEND VULKAN_SHADER_GEN_CMAKE_ARGS -DCMAKE_TOOLCHAIN_FILE=${HOST_CMAKE_TOOLCHAIN_FILE})
         message(STATUS "vulkan-shaders-gen toolchain file: ${HOST_CMAKE_TOOLCHAIN_FILE}")
     endif()
 
-    # Native build through ExternalProject_Add
     ExternalProject_Add(
         vulkan-shaders-gen
         SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders
-        CMAKE_ARGS ${VULKAN_SHADER_GEN_CMAKE_ARGS}
-        BUILD_COMMAND ${CMAKE_COMMAND} --build . ${VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS}
-        INSTALL_COMMAND ${CMAKE_COMMAND} --install .
-        INSTALL_DIR ${CMAKE_BINARY_DIR}
+        CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR}/$<CONFIG>
+                   -DCMAKE_INSTALL_BINDIR=.
+                   -DCMAKE_BUILD_TYPE=$<CONFIG>
+                   ${VULKAN_SHADER_GEN_CMAKE_ARGS}
+
+        BUILD_COMMAND   ${CMAKE_COMMAND} --build   . --config $<CONFIG>
+        INSTALL_COMMAND ${CMAKE_COMMAND} --install . --config $<CONFIG>
     )
     ExternalProject_Add_StepTargets(vulkan-shaders-gen build install)
 
     set (_ggml_vk_host_suffix $<IF:$<STREQUAL:${CMAKE_HOST_SYSTEM_NAME},Windows>,.exe,>)
-    set (_ggml_vk_genshaders_cmd ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/vulkan-shaders-gen${_ggml_vk_host_suffix})
-    set (_ggml_vk_header     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp)
-    set (_ggml_vk_source     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp)
-    set (_ggml_vk_input_dir  ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders)
-    set (_ggml_vk_output_dir ${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv)
+    set (_ggml_vk_genshaders_dir "${CMAKE_BINARY_DIR}/$<CONFIG>")
+    set (_ggml_vk_genshaders_cmd "${_ggml_vk_genshaders_dir}/vulkan-shaders-gen${_ggml_vk_host_suffix}")
+    set (_ggml_vk_header     "${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp")
+    set (_ggml_vk_source     "${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp")
+    set (_ggml_vk_input_dir  "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders")
+    set (_ggml_vk_output_dir "${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv")
 
-    file(GLOB _ggml_vk_shader_deps "${_ggml_vk_input_dir}/*.comp")
-    set (_ggml_vk_shader_deps ${_ggml_vk_shader_deps} vulkan-shaders-gen)
-
-    # Add build and install dependencies for all builds
-    set(_ggml_vk_shader_deps ${_ggml_vk_shader_deps} vulkan-shaders-gen-build vulkan-shaders-gen-install)
+    file(GLOB _ggml_vk_shader_files CONFIGURE_DEPENDS "${_ggml_vk_input_dir}/*.comp")
 
     add_custom_command(
         OUTPUT ${_ggml_vk_header}
-                ${_ggml_vk_source}
+               ${_ggml_vk_source}
 
         COMMAND ${_ggml_vk_genshaders_cmd}
             --glslc      ${Vulkan_GLSLC_EXECUTABLE}
@@ -181,7 +170,11 @@ if (Vulkan_FOUND)
             --target-cpp ${_ggml_vk_source}
             --no-clean
 
-        DEPENDS ${_ggml_vk_shader_deps}
+        DEPENDS ${_ggml_vk_shader_files}
+                vulkan-shaders-gen
+                vulkan-shaders-gen-build
+                vulkan-shaders-gen-install
+
         COMMENT "Generate vulkan shaders"
     )
 

ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt

@@ -25,15 +25,3 @@ add_executable(${TARGET} vulkan-shaders-gen.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 target_link_libraries(vulkan-shaders-gen PUBLIC Threads::Threads)
-
-# Configure output directories for MSVC builds
-if(MSVC)
-    # Get the main project's runtime output directory if possible
-    if(DEFINED CMAKE_RUNTIME_OUTPUT_DIRECTORY)
-        foreach(CONFIG ${CMAKE_CONFIGURATION_TYPES})
-            string(TOUPPER ${CONFIG} CONFIG)
-            set_target_properties(${TARGET} PROPERTIES
-                RUNTIME_OUTPUT_DIRECTORY_${CONFIG} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
-        endforeach()
-    endif()
-endif()

gguf-py/gguf/constants.py

@@ -291,6 +291,7 @@ class MODEL_ARCH(IntEnum):
     BERT             = auto()
     NOMIC_BERT       = auto()
     NOMIC_BERT_MOE   = auto()
+    NEO_BERT         = auto()
     JINA_BERT_V2     = auto()
     BLOOM            = auto()
     STABLELM         = auto()
@@ -343,6 +344,8 @@ class MODEL_ARCH(IntEnum):
     WAVTOKENIZER_DEC = auto()
     PLM              = auto()
     BAILINGMOE       = auto()
+    DOTS1            = auto()
+    ARCEE            = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -571,6 +574,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.BERT:             "bert",
     MODEL_ARCH.NOMIC_BERT:       "nomic-bert",
     MODEL_ARCH.NOMIC_BERT_MOE:   "nomic-bert-moe",
+    MODEL_ARCH.NEO_BERT:         "neo-bert",
     MODEL_ARCH.JINA_BERT_V2:     "jina-bert-v2",
     MODEL_ARCH.BLOOM:            "bloom",
     MODEL_ARCH.STABLELM:         "stablelm",
@@ -623,6 +627,8 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec",
     MODEL_ARCH.PLM:              "plm",
     MODEL_ARCH.BAILINGMOE:       "bailingmoe",
+    MODEL_ARCH.DOTS1:            "dots1",
+    MODEL_ARCH.ARCEE:            "arcee",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -1077,6 +1083,18 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_UP_EXP,
         MODEL_TENSOR.LAYER_OUT_NORM,
     ],
+    MODEL_ARCH.NEO_BERT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.ENC_OUTPUT_NORM,
+        MODEL_TENSOR.CLS,
+        MODEL_TENSOR.CLS_OUT,
+    ],
     MODEL_ARCH.JINA_BERT_V2: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.TOKEN_EMBD_NORM,
@@ -2044,6 +2062,45 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
     ],
+    MODEL_ARCH.DOTS1: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+    ],
+    MODEL_ARCH.ARCEE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     # TODO
 }
 

gguf-py/gguf/gguf_writer.py

@@ -271,7 +271,7 @@ class GGUFWriter:
 
     def add_key_value(self, key: str, val: Any, vtype: GGUFValueType, sub_type: GGUFValueType | None = None) -> None:
         if any(key in kv_data for kv_data in self.kv_data):
-            raise ValueError(f'Duplicated key name {key!r}')
+            logger.warning(f'Duplicated key name {key!r}, overwriting it with new value {val!r} of type {vtype.name}')
 
         self.kv_data[0][key] = GGUFValue(value=val, type=vtype, sub_type=sub_type)
 

gguf-py/gguf/tensor_mapping.py

@@ -31,6 +31,7 @@ class TensorNameMap:
             "model.embeddings",                          # rwkv7
             "model.word_embeddings",                     # bailingmoe
             "language_model.model.embed_tokens",         # llama4
+            "encoder",                                   # neobert
         ),
 
         # Token type embeddings
@@ -134,6 +135,7 @@ class TensorNameMap:
             "rwkv.blocks.{bid}.ln1",                                # rwkv6
             "model.layers.{bid}.ln1",                               # rwkv7
             "model.layers.{bid}.input_layernorm",                   # llama4
+            "transformer_encoder.{bid}.attention_norm",             # neobert
         ),
 
         # Attention norm 2
@@ -161,6 +163,7 @@ class TensorNameMap:
             "model.layers.{bid}.self_attn.qkv_proj",                               # phi3
             "encoder.layers.{bid}.self_attention.query_key_value",                 # chatglm
             "transformer.layers.{bid}.attn.qkv_proj",                              # openelm
+            "transformer_encoder.{bid}.qkv",                                       # neobert
         ),
 
         # Attention query
@@ -236,6 +239,7 @@ class TensorNameMap:
             "transformer.layers.{bid}.attn.out_proj",                       # openelm
             "transformer.h.{bid}.attn.attention.out_proj",                  # exaone
             "model.layers.{bid}.self_attn.o_proj",                          # llama4
+            "transformer_encoder.{bid}.wo",                                 # neobert
         ),
 
         # Attention output norm
@@ -276,6 +280,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.post_attention_layernorm",                 # chatglm
             "transformer.layers.{bid}.ffn_norm",                             # openelm
             "model.layers.{bid}.post_attention_layernorm",                   # llama4
+            "transformer_encoder.{bid}.ffn_norm",                            # neobert
         ),
 
         # Post feed-forward norm
@@ -305,7 +310,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_EXP_PROBS_B: (
-            "model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3
+            "model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3 dots1
         ),
 
         # Feed-forward up
@@ -333,11 +338,14 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
             "encoder.layers.{bid}.mlp.fc1",                           # nomic-bert-moe
             "model.layers.{bid}.mlp.c_fc",                            # starcoder2
-            "encoder.layer.{bid}.mlp.gated_layers_v",                 # jina-bert-v2
+            "encoder.layer.{bid}.mlp.gated_layers_v",                 # jina-bert-v2 (split up/gate, no longer used)
+            "encoder.layer.{bid}.mlp.gated_layers",                   # jina-bert-v2 (GEGLU)
+            "encoder.layer.{bid}.mlp.up_gated_layer",                 # jina-v2-code (GEGLU)
             "model.layers.{bid}.residual_mlp.w3",                     # arctic
             "encoder.layers.{bid}.mlp.dense_h_to_4h",                 # chatglm
             "transformer.h.{bid}.mlp.c_fc_1",                         # exaone
             "model.layers.{bid}.feed_forward.up_proj",                # llama4
+            "transformer_encoder.{bid}.ffn.w12",                      # neobert
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -370,7 +378,7 @@ class TensorNameMap:
             "model.layers.layers.{bid}.mlp.gate_proj",    # plamo
             "model.layers.{bid}.feed_forward.w1",         # internlm2
             "encoder.layers.{bid}.mlp.fc12",              # nomic-bert
-            "encoder.layer.{bid}.mlp.gated_layers_w",     # jina-bert-v2
+            "encoder.layer.{bid}.mlp.gated_layers_w",     # jina-bert-v2 (split up/gate, no longer used)
             "transformer.h.{bid}.mlp.linear_1",           # refact
             "model.layers.{bid}.residual_mlp.w1",         # arctic
             "transformer.h.{bid}.mlp.c_fc_0",             # exaone
@@ -420,6 +428,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.dense_4h_to_h",                 # chatglm
             "model.layers.h.{bid}.mlp.c_proj",                        # exaone
             "model.layers.{bid}.feed_forward.down_proj",              # llama4
+            "transformer_encoder.{bid}.ffn.w3",                       # neobert
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -830,12 +839,14 @@ class TensorNameMap:
         # TODO: these do not belong to block_mappings_cfg - move them to mappings_cfg
         MODEL_TENSOR.ENC_OUTPUT_NORM: (
             "encoder.final_layer_norm", # t5
+            "layer_norm",               # neobert
         ),
 
         MODEL_TENSOR.CLS: (
             "classifier",       # jina
             "classifier.dense", # roberta
             "pre_classifier",   # distillbert
+            "dense",            # neobert
         ),
 
         MODEL_TENSOR.CLS_OUT: (

include/llama.h

@@ -243,18 +243,21 @@ extern "C" {
 
     typedef bool (*llama_progress_callback)(float progress, void * user_data);
 
-    // Input data for llama_decode
+    // Input data for llama_encode/llama_decode
     // A llama_batch object can contain input about one or many sequences
     // The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens
     //
     // - token  : the token ids of the input (used when embd is NULL)
     // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
     // - pos    : the positions of the respective token in the sequence
-    //            (if set to NULL, the token position will be tracked automatically by llama_decode)
+    //            (if set to NULL, the token position will be tracked automatically by llama_encode/llama_decode)
     // - seq_id : the sequence to which the respective token belongs
     //            (if set to NULL, the sequence ID will be assumed to be 0)
     // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
-    //            (if set to NULL, only the logits for last token will be returned)
+    //            (if set to NULL:
+    //               - if embeddings: all tokens are output
+    //               - if not:        only the last token is output
+    //            )
     //
     typedef struct llama_batch {
         int32_t n_tokens;
@@ -262,8 +265,8 @@ extern "C" {
         llama_token  *  token;
         float        *  embd;
         llama_pos    *  pos;
-        int32_t      *  n_seq_id; // TODO: remove, should belong to only 1 sequence
-        llama_seq_id ** seq_id;   // TODO: become llama_seq_id * seq_id;
+        int32_t      *  n_seq_id;
+        llama_seq_id ** seq_id;
         int8_t       *  logits;   // TODO: rename this to "output"
     } llama_batch;
 
@@ -961,8 +964,8 @@ extern "C" {
     // Get the number of threads used for prompt and batch processing (multiple token).
     LLAMA_API int32_t llama_n_threads_batch(struct llama_context * ctx);
 
-    // Set whether the model is in embeddings mode or not
-    // If true, embeddings will be returned but logits will not
+    // Set whether the context outputs embeddings or not
+    // TODO: rename to avoid confusion with llama_get_embeddings()
     LLAMA_API void llama_set_embeddings(struct llama_context * ctx, bool embeddings);
 
     // Set whether to use causal attention or not

requirements/requirements-compare-llama-bench.txt

@@ -1,2 +1,3 @@
 tabulate~=0.9.0
 GitPython~=3.1.43
+matplotlib~=3.10.0

scripts/compare-llama-bench.py

@@ -19,6 +19,7 @@ except ImportError as e:
     print("the following Python libraries are required: GitPython, tabulate.") # noqa: NP100
     raise e
 
+
 logger = logging.getLogger("compare-llama-bench")
 
 # All llama-bench SQL fields
@@ -122,11 +123,15 @@ help_s = (
 parser.add_argument("--check", action="store_true", help="check if all required Python libraries are installed")
 parser.add_argument("-s", "--show", help=help_s)
 parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
+parser.add_argument("--plot", help="generate a performance comparison plot and save to specified file (e.g., plot.png)")
+parser.add_argument("--plot_x", help="parameter to use as x axis for plotting (default: n_depth)", default="n_depth")
+parser.add_argument("--plot_log_scale", action="store_true", help="use log scale for x axis in plots (off by default)")
 
 known_args, unknown_args = parser.parse_known_args()
 
 logging.basicConfig(level=logging.DEBUG if known_args.verbose else logging.INFO)
 
+
 if known_args.check:
     # Check if all required Python libraries are installed. Would have failed earlier if not.
     sys.exit(0)
@@ -499,7 +504,6 @@ else:
 
 name_compare = bench_data.get_commit_name(hexsha8_compare)
 
-
 # If the user provided columns to group the results by, use them:
 if known_args.show is not None:
     show = known_args.show.split(",")
@@ -544,6 +548,14 @@ else:
             show.remove(prop)
         except ValueError:
             pass
+
+    # Add plot_x parameter to parameters to show if it's not already present:
+    if known_args.plot:
+        for k, v in PRETTY_NAMES.items():
+            if v == known_args.plot_x and k not in show:
+                show.append(k)
+                break
+
     rows_show = bench_data.get_rows(show, hexsha8_baseline, hexsha8_compare)
 
 if not rows_show:
@@ -600,6 +612,161 @@ if "gpu_info" in show:
 headers  = [PRETTY_NAMES[p] for p in show]
 headers += ["Test", f"t/s {name_baseline}", f"t/s {name_compare}", "Speedup"]
 
+if known_args.plot:
+    def create_performance_plot(table_data: list[list[str]], headers: list[str], baseline_name: str, compare_name: str, output_file: str, plot_x_param: str, log_scale: bool = False):
+        try:
+            import matplotlib.pyplot as plt
+            import matplotlib
+            matplotlib.use('Agg')
+        except ImportError as e:
+            logger.error("matplotlib is required for --plot.")
+            raise e
+
+        data_headers = headers[:-4] # Exclude the last 4 columns (Test, baseline t/s, compare t/s, Speedup)
+        plot_x_index = None
+        plot_x_label = plot_x_param
+
+        if plot_x_param not in ["n_prompt", "n_gen", "n_depth"]:
+            pretty_name = PRETTY_NAMES.get(plot_x_param, plot_x_param)
+            if pretty_name in data_headers:
+                plot_x_index = data_headers.index(pretty_name)
+                plot_x_label = pretty_name
+            elif plot_x_param in data_headers:
+                plot_x_index = data_headers.index(plot_x_param)
+                plot_x_label = plot_x_param
+            else:
+                logger.error(f"Parameter '{plot_x_param}' not found in current table columns. Available columns: {', '.join(data_headers)}")
+                return
+
+        grouped_data = {}
+
+        for i, row in enumerate(table_data):
+            group_key_parts = []
+            test_name = row[-4]
+
+            base_test = ""
+            x_value = None
+
+            if plot_x_param in ["n_prompt", "n_gen", "n_depth"]:
+                for j, val in enumerate(row[:-4]):
+                    header_name = data_headers[j]
+                    if val is not None and str(val).strip():
+                        group_key_parts.append(f"{header_name}={val}")
+
+                if plot_x_param == "n_prompt" and "pp" in test_name:
+                    base_test = test_name.split("@")[0]
+                    x_value = base_test
+                elif plot_x_param == "n_gen" and "tg" in test_name:
+                    x_value = test_name.split("@")[0]
+                elif plot_x_param == "n_depth" and "@d" in test_name:
+                    base_test = test_name.split("@d")[0]
+                    x_value = int(test_name.split("@d")[1])
+                else:
+                    base_test = test_name
+
+                if base_test.strip():
+                    group_key_parts.append(f"Test={base_test}")
+            else:
+                for j, val in enumerate(row[:-4]):
+                    if j != plot_x_index:
+                        header_name = data_headers[j]
+                        if val is not None and str(val).strip():
+                            group_key_parts.append(f"{header_name}={val}")
+                    else:
+                        x_value = val
+
+                group_key_parts.append(f"Test={test_name}")
+
+            group_key = tuple(group_key_parts)
+
+            if group_key not in grouped_data:
+                grouped_data[group_key] = []
+
+            grouped_data[group_key].append({
+                'x_value': x_value,
+                'baseline': float(row[-3]),
+                'compare': float(row[-2]),
+                'speedup': float(row[-1])
+            })
+
+        if not grouped_data:
+            logger.error("No data available for plotting")
+            return
+
+        def make_axes(num_groups, max_cols=2, base_size=(8, 4)):
+            from math import ceil
+            cols = 1 if num_groups == 1 else min(max_cols, num_groups)
+            rows = ceil(num_groups / cols)
+
+            # Scale figure size by grid dimensions
+            w, h = base_size
+            fig, ax_arr = plt.subplots(rows, cols,
+                                       figsize=(w * cols, h * rows),
+                                       squeeze=False)
+
+            axes = ax_arr.flatten()[:num_groups]
+            return fig, axes
+
+        num_groups = len(grouped_data)
+        fig, axes = make_axes(num_groups)
+
+        plot_idx = 0
+
+        for group_key, points in grouped_data.items():
+            if plot_idx >= len(axes):
+                break
+            ax = axes[plot_idx]
+
+            try:
+                points_sorted = sorted(points, key=lambda p: float(p['x_value']) if p['x_value'] is not None else 0)
+                x_values = [float(p['x_value']) if p['x_value'] is not None else 0 for p in points_sorted]
+            except ValueError:
+                points_sorted = sorted(points, key=lambda p: group_key)
+                x_values = [p['x_value'] for p in points_sorted]
+
+            baseline_vals = [p['baseline'] for p in points_sorted]
+            compare_vals = [p['compare'] for p in points_sorted]
+
+            ax.plot(x_values, baseline_vals, 'o-', color='skyblue',
+                    label=f'{baseline_name}', linewidth=2, markersize=6)
+            ax.plot(x_values, compare_vals, 's--', color='lightcoral', alpha=0.8,
+                    label=f'{compare_name}', linewidth=2, markersize=6)
+
+            if log_scale:
+                ax.set_xscale('log', base=2)
+                unique_x = sorted(set(x_values))
+                ax.set_xticks(unique_x)
+                ax.set_xticklabels([str(int(x)) for x in unique_x])
+
+            title_parts = []
+            for part in group_key:
+                if '=' in part:
+                    key, value = part.split('=', 1)
+                    title_parts.append(f"{key}: {value}")
+
+            title = ', '.join(title_parts) if title_parts else "Performance comparison"
+
+            ax.set_xlabel(plot_x_label, fontsize=12, fontweight='bold')
+            ax.set_ylabel('Tokens per second (t/s)', fontsize=12, fontweight='bold')
+            ax.set_title(title, fontsize=12, fontweight='bold')
+            ax.legend(loc='best', fontsize=10)
+            ax.grid(True, alpha=0.3)
+
+            plot_idx += 1
+
+        for i in range(plot_idx, len(axes)):
+            axes[i].set_visible(False)
+
+        fig.suptitle(f'Performance comparison: {compare_name} vs. {baseline_name}',
+                     fontsize=14, fontweight='bold')
+        fig.subplots_adjust(top=1)
+
+        plt.tight_layout()
+        plt.savefig(output_file, dpi=300, bbox_inches='tight')
+        plt.close()
+
+    create_performance_plot(table, headers, name_baseline, name_compare, known_args.plot, known_args.plot_x, known_args.plot_log_scale)
+
 print(tabulate( # noqa: NP100
     table,
     headers=headers,

scripts/sync-ggml.last

@@ -1 +1 @@
-94a83ba5a725ae2aee79df75dd99b2119d0478cc
+6a7d170c04789f6ebcf320ed03c1b16973f93bd7

src/llama-arch.cpp

@@ -20,6 +20,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_BERT,             "bert"             },
     { LLM_ARCH_NOMIC_BERT,       "nomic-bert"       },
     { LLM_ARCH_NOMIC_BERT_MOE,   "nomic-bert-moe"   },
+    { LLM_ARCH_NEO_BERT,         "neo-bert"         },
     { LLM_ARCH_JINA_BERT_V2,     "jina-bert-v2"     },
     { LLM_ARCH_BLOOM,            "bloom"            },
     { LLM_ARCH_STABLELM,         "stablelm"         },
@@ -72,6 +73,8 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
     { LLM_ARCH_PLM,              "plm"              },
     { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
+    { LLM_ARCH_DOTS1,            "dots1"            },
+    { LLM_ARCH_ARCEE,            "arcee"            },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -243,6 +246,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_ARCEE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_LLAMA4,
         {
@@ -494,6 +515,21 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_NEO_BERT,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_ENC_OUTPUT_NORM, "enc.output_norm" },
+            { LLM_TENSOR_CLS,             "cls" },
+            { LLM_TENSOR_CLS_OUT,         "cls.output" },
+        },
+    },
     {
         LLM_ARCH_JINA_BERT_V2,
         {
@@ -1555,6 +1591,34 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
         },
     },
+    {
+        LLM_ARCH_DOTS1,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
+        }
+    },
     {
         LLM_ARCH_UNKNOWN,
         {

src/llama-arch.h

@@ -24,6 +24,7 @@ enum llm_arch {
     LLM_ARCH_BERT,
     LLM_ARCH_NOMIC_BERT,
     LLM_ARCH_NOMIC_BERT_MOE,
+    LLM_ARCH_NEO_BERT,
     LLM_ARCH_JINA_BERT_V2,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
@@ -76,6 +77,8 @@ enum llm_arch {
     LLM_ARCH_WAVTOKENIZER_DEC,
     LLM_ARCH_PLM,
     LLM_ARCH_BAILINGMOE,
+    LLM_ARCH_DOTS1,
+    LLM_ARCH_ARCEE,
     LLM_ARCH_UNKNOWN,
 };
 

src/llama-batch.cpp

@@ -1,8 +1,14 @@
 #include "llama-batch.h"
 
+#include "llama-impl.h"
+#include "llama-cparams.h"
+#include "llama-vocab.h"
+#include "llama-memory.h"
+
 #include <cassert>
 #include <cstring>
 #include <algorithm>
+#include <sstream>
 
 llama_ubatch llama_sbatch::reserve_ubatch(size_t n_ubatch, bool has_embd) {
     // clear empty sequences
@@ -105,12 +111,7 @@ void llama_sbatch::add_seq_to_ubatch(llama_ubatch & ubatch, llama_sbatch_seq & s
             ubatch.seq_id = batch->seq_id + seq.offset;
         }
     }
-    if (logits_all) {
-        for (size_t i = 0; i < length; ++i) {
-            ubatch.output[ubatch.n_tokens + i] = 1;
-            out_ids.push_back(ids[seq.offset + i]);
-        }
-    } else if (batch->logits) {
+    if (batch->logits) {
         if (ubatch.equal_seqs) {
             for (size_t i = 0; i < length; ++i) {
                 size_t id = ids[seq.offset + i];
@@ -197,11 +198,10 @@ llama_ubatch llama_sbatch::split_seq(size_t n_ubatch) {
     return ubatch;
 }
 
-llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) {
+llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split) {
     GGML_ASSERT(batch.n_tokens >= 0);
     this->batch = &batch;
     this->n_embd = n_embd;
-    this->logits_all = logits_all;
 
     n_tokens = batch.n_tokens;
     ids.resize(n_tokens);
@@ -285,17 +285,56 @@ llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple
             );
 }
 
-llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0) {
-    batch = in_batch;
-    GGML_ASSERT(batch.n_tokens > 0);
-    if (!batch.pos) {
-        assert(p0 >= 0);
-        pos.resize(batch.n_tokens);
-        for (int32_t i = 0; i < batch.n_tokens; i++) {
-            pos[i] = p0 + i;
-        }
-        batch.pos = pos.data();
+llama_batch_allocr::llama_batch_allocr() {
+    const char * LLAMA_BATCH_DEBUG = getenv("LLAMA_BATCH_DEBUG");
+    debug = LLAMA_BATCH_DEBUG ? atoi(LLAMA_BATCH_DEBUG) : 0;
+
+    seq_pos.resize(LLAMA_MAX_SEQ);
+    seq_cpl.resize(LLAMA_MAX_SEQ);
+    for (auto & cur : seq_cpl) {
+        cur.resize(LLAMA_MAX_SEQ);
     }
+}
+
+bool llama_batch_allocr::init(
+        const llama_batch & batch_inp,
+        const llama_vocab & vocab,
+        const llama_memory_i * memory,
+        bool embd_all) {
+    clear();
+
+    batch = batch_inp;
+
+    GGML_ASSERT(batch.n_tokens > 0);
+
+    //
+    // validate input batch
+    //
+
+    if (batch.token) {
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= vocab.n_tokens()) {
+                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
+                return false;
+            }
+        }
+    }
+
+    if (batch.seq_id) {
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
+                if (batch.seq_id && (batch.seq_id[i][s] < 0 || batch.seq_id[i][s] >= LLAMA_MAX_SEQ)) {
+                    LLAMA_LOG_ERROR("%s: invalid seq_id[%d][%d] = %d > %d\n", __func__, i, s, batch.seq_id[i][s], LLAMA_MAX_SEQ);
+                    return false;
+                }
+            }
+        }
+    }
+
+    //
+    // auto-generate missing fields
+    //
+
     if (!batch.n_seq_id) {
         n_seq_id.resize(batch.n_tokens);
         for (int32_t i = 0; i < batch.n_tokens; i++) {
@@ -303,6 +342,7 @@ llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0
         }
         batch.n_seq_id = n_seq_id.data();
     }
+
     if (!batch.seq_id) {
         seq_id.resize(batch.n_tokens + 1);
         seq_id[batch.n_tokens] = NULL;
@@ -311,10 +351,221 @@ llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0
         }
         batch.seq_id = seq_id.data();
     }
+
+    if (!batch.pos) {
+        pos.resize(batch.n_tokens);
+
+        // initialize the starting position for each sequence based on the positions in the memory
+        llama_pos p0[LLAMA_MAX_SEQ];
+        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+            if (!memory) {
+                p0[s] = 0;
+            } else {
+                p0[s] = memory->seq_pos_max(s) + 1;
+            }
+        }
+
+        for (int32_t i = 0; i < batch.n_tokens; i++) {
+            const llama_seq_id seq_id = batch.seq_id[i][0];
+
+            pos[i] = p0[seq_id];
+
+            for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
+                p0[batch.seq_id[i][s]] = pos[i] + 1;
+            }
+        }
+
+        batch.pos = pos.data();
+    }
+
     if (!batch.logits) {
-        logits.resize(batch.n_tokens);
-        logits[logits.size() - 1] = true;
-        batch.logits = logits.data();
+        if (embd_all) {
+            // return the output for all tokens
+            output.resize(batch.n_tokens, true);
+        } else {
+            // return the output only for the last token
+            output.resize(batch.n_tokens, false);
+            output[output.size() - 1] = true;
+        }
+
+        batch.logits = output.data();
+    } else if (embd_all) {
+        bool warn = false;
+
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            if (batch.logits[i] == 0) {
+                warn = true;
+            }
+        }
+
+        if (warn) {
+            LLAMA_LOG_WARN("%s: embeddings required but some input tokens were not marked as outputs -> overriding\n", __func__);
+
+            output.resize(batch.n_tokens, true);
+            batch.logits = output.data();
+        }
+    }
+
+    //
+    // compute stats
+    //
+
+    for (int32_t i = 0; i < batch.n_tokens; ++i) {
+        n_outputs += batch.logits[i] != 0;
+    }
+
+    // determine coupled sequences
+    // these are pairs of sequences that have at least one token in the input batch that is assigned to both of them
+    for (int32_t i = 0; i < batch.n_tokens; ++i) {
+        for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
+            seq_pos[batch.seq_id[i][s]].insert(batch.pos[i]);
+
+            if (s > 0) {
+                const llama_seq_id s0 = batch.seq_id[i][0];
+                const llama_seq_id s1 = batch.seq_id[i][s];
+
+                // mark that sequence s1 is coupled to s0
+                seq_cpl[s1][s0] = true;
+
+                // note: the other way around is not necessary for now
+                //seq_cpl[s0][s1] = true;
+            }
+        }
+    }
+
+    if (debug > 0) {
+        LLAMA_LOG_DEBUG("%s: input batch info:\n", __func__);
+        LLAMA_LOG_DEBUG("%s:   n_tokens  = %d\n", __func__,          batch.n_tokens);
+        LLAMA_LOG_DEBUG("%s:   token     = %p\n", __func__, (void *) batch.token);
+        LLAMA_LOG_DEBUG("%s:   embd      = %p\n", __func__, (void *) batch.embd);
+        LLAMA_LOG_DEBUG("%s:   pos       = %p\n", __func__, (void *) batch.pos);
+        LLAMA_LOG_DEBUG("%s:   n_seq_id  = %p\n", __func__, (void *) batch.n_seq_id);
+        LLAMA_LOG_DEBUG("%s:   seq_id    = %p\n", __func__, (void *) batch.seq_id);
+        LLAMA_LOG_DEBUG("%s:   logits    = %p\n", __func__, (void *) batch.logits);
+        LLAMA_LOG_DEBUG("%s:   n_outputs = %d\n", __func__, n_outputs);
+
+        if (debug > 1) {
+            int seq_id_max = 0;
+            for (int32_t i = 0; i < batch.n_tokens; ++i) {
+                for (int s = 0; s < batch.n_seq_id[i]; ++s) {
+                    for (int s = 0; s < batch.n_seq_id[i]; ++s) {
+                        seq_id_max = std::max(seq_id_max, batch.seq_id[i][s]);
+                    }
+                }
+            }
+            ++seq_id_max;
+
+            LLAMA_LOG_DEBUG("%s:   token     = [\n", __func__);
+            for (int32_t i = 0; i < batch.n_tokens; ++i) {
+                std::vector<int8_t> seq_id(seq_id_max);
+
+                for (int s = 0; s < batch.n_seq_id[i]; ++s) {
+                    seq_id[batch.seq_id[i][s]] = 1;
+                }
+
+                std::stringstream ss;
+                for (int s = 0; s < seq_id_max; ++s) {
+                    if (seq_id[s]) {
+                        ss << s%10;
+                    } else {
+                        ss << ".";
+                    }
+                }
+
+                LLAMA_LOG_DEBUG("%s:  %4d: id = %6d (%16s), pos = %4d, n_seq_id = %2d, seq_id = [%s], output = %d\n",
+                        __func__, i, batch.token[i], vocab.token_to_piece(batch.token[i]).c_str(),
+                        batch.pos[i], batch.n_seq_id[i], ss.str().c_str(), batch.logits[i]);
+            }
+            LLAMA_LOG_DEBUG("%s:   ]\n", __func__);
+
+            LLAMA_LOG_DEBUG("%s:   seq       = [\n", __func__);
+            for (int s0 = 0; s0 < (int) seq_pos.size(); ++s0) {
+                if (seq_pos[s0].empty()) {
+                    continue;
+                }
+
+                std::stringstream ss;
+                for (int s1 = 0; s1 < (int) seq_cpl[s0].size(); ++s1) {
+                    if (seq_cpl[s0][s1]) {
+                        ss << s1 << " ";
+                    }
+                }
+
+                LLAMA_LOG_DEBUG("%s:  %4d: pos = [%4d, %4d], cpl = %s\n",
+                        __func__, s0, seq_pos_min(s0), seq_pos_max(s0), ss.str().empty() ? "-" : ss.str().c_str());
+            }
+            LLAMA_LOG_DEBUG("%s:   ]\n", __func__);
+        }
+    }
+
+    //
+    // consistency checks
+    //
+
+    for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        if (seq_pos[s].empty()) {
+            continue;
+        }
+
+        if (memory && seq_pos_min(s) != memory->seq_pos_max(s) + 1) {
+            LLAMA_LOG_ERROR("%s: sequence %d does not start from the last position stored in the memory\n", __func__, s);
+            return false;
+        }
+
+        if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) {
+            LLAMA_LOG_ERROR("%s: sequence %d positions are not continuous\n", __func__, s);
+            return false;
+        }
+    }
+
+    if (memory) {
+        for (int32_t s0 = 0; s0 < LLAMA_MAX_SEQ; ++s0) {
+            for (int32_t s1 = 0; s1 < LLAMA_MAX_SEQ; ++s1) {
+                if (seq_cpl[s0][s1]) {
+                    if (memory->seq_pos_min(s0) != memory->seq_pos_min(s1) ||
+                        memory->seq_pos_max(s0) != memory->seq_pos_max(s1)) {
+                        LLAMA_LOG_ERROR("%s: sequence %d is coupled to %d in the input batch, but have divereged\n", __func__, s0, s1);
+                        return false;
+                    }
+                }
+            }
+        }
+    }
+
+    return true;
+}
+
+const llama_batch & llama_batch_allocr::get_batch() const {
+    return batch;
+}
+
+uint32_t llama_batch_allocr::get_n_outputs() const {
+    return n_outputs;
+}
+
+llama_pos llama_batch_allocr::seq_pos_min(llama_seq_id seq_id) const {
+    return seq_pos[seq_id].empty() ? -1 : *seq_pos[seq_id].begin();
+}
+
+llama_pos llama_batch_allocr::seq_pos_max(llama_seq_id seq_id) const {
+    return seq_pos[seq_id].empty() ? -1 : *seq_pos[seq_id].rbegin();
+}
+
+void llama_batch_allocr::clear() {
+    n_outputs = 0;
+
+    batch = {};
+    pos.clear();
+    n_seq_id.clear();
+    seq_id.clear();
+    output.clear();
+
+    for (auto & cur : seq_pos) {
+        cur.clear();
+    }
+
+    for (auto & cur : seq_cpl) {
+        std::fill(cur.begin(), cur.end(), false);
     }
 }
 

src/llama-batch.h

@@ -4,6 +4,7 @@
 
 #include <array>
 #include <vector>
+#include <set>
 
 // very similar to llama_batch,
 // but has more metadata about sequences
@@ -18,8 +19,8 @@ struct llama_ubatch {
     llama_token  *  token;    // [n_tokens]
     float        *  embd;     // [n_embd, n_tokens]
     llama_pos    *  pos;      // [n_tokens]
-    int32_t      *  n_seq_id; // [n_seqs] // TODO: remove, should belong to only 1 sequence
-    llama_seq_id ** seq_id;   // [n_seqs] // TODO: become llama_seq_id * seq_id;
+    int32_t      *  n_seq_id; // [n_seqs]
+    llama_seq_id ** seq_id;   // [n_seqs]
     int8_t       *  output;   // [n_tokens]
 };
 
@@ -39,8 +40,6 @@ struct llama_sbatch {
 
     size_t n_embd;
 
-    bool logits_all; // TODO: remove once lctx.logits_all is removed too
-
     // sorted indices into the batch
     std::vector<int64_t> ids;
     // batch indices of the output
@@ -76,19 +75,45 @@ struct llama_sbatch {
     llama_ubatch split_seq(size_t n_ubatch);
 
     llama_sbatch() = default;
-    llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
+    llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false);
 };
 
-// temporary allocate memory for the input batch if needed
-struct llama_batch_allocr {
-    struct llama_batch batch;
+// a helper for sanitizing and fulfilling a batch
+class llama_batch_allocr {
+public:
+    llama_batch_allocr();
+
+    // sanitize and auto-gen missing data in the input batch
+    // memory is optional. if provided will be used to check for sequence continuity and to determine the positions
+    bool init(
+            const llama_batch & batch_inp,
+            const llama_vocab & vocab,
+            const llama_memory_i * memory,
+            bool embd_all);
+
+    const llama_batch & get_batch() const;
+
+    uint32_t get_n_outputs() const;
+
+    llama_pos seq_pos_min(llama_seq_id seq_id) const;
+    llama_pos seq_pos_max(llama_seq_id seq_id) const;
+
+private:
+    void clear();
+
+    llama_batch batch;
+
+    uint32_t n_outputs;
 
     std::array<llama_seq_id, 1> seq_id_0 = { 0 }; // default sequence id
+
     std::vector<llama_pos>      pos;
     std::vector<int32_t>        n_seq_id;
     std::vector<llama_seq_id *> seq_id;
-    std::vector<int8_t>         logits;
+    std::vector<int8_t>         output;
 
-    // optionally fulfill the batch returned by llama_batch_get_one
-    llama_batch_allocr(struct llama_batch in_batch, llama_pos p0);
+    std::vector<std::set<llama_pos>> seq_pos; // seq_pos[s]: the set of positions in sequence s
+    std::vector<std::vector<bool>>   seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1
+
+    int debug;
 };

src/llama-chat.cpp

@@ -183,6 +183,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_BAILING;
     } else if (tmpl_contains("<|header_start|>") && tmpl_contains("<|header_end|>")) {
         return LLM_CHAT_TEMPLATE_LLAMA4;
+    } else if (tmpl_contains("<|endofuserprompt|>")) {
+        return LLM_CHAT_TEMPLATE_DOTS1;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -643,6 +645,21 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "Assistant:";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_DOTS1) {
+        // dots.llm1.inst (DOTS1)
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "<|system|>" << message->content << "<|endofsystem|>";
+            } else if (role == "user") {
+                ss << "<|userprompt|>" << message->content << "<|endofuserprompt|>";
+            } else {
+                ss << "<|response|>" << message->content << "<|endofresponse|>";
+            }
+        }
+        if (add_ass) {
+            ss << "<|response|>";
+        }
     } else {
         // template not supported
         return -1;

src/llama-chat.h

@@ -43,6 +43,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_BAILING,
     LLM_CHAT_TEMPLATE_LLAMA4,
     LLM_CHAT_TEMPLATE_SMOLVLM,
+    LLM_CHAT_TEMPLATE_DOTS1,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

src/llama-context.cpp

@@ -1,6 +1,7 @@
 #include "llama-context.h"
 
 #include "llama-impl.h"
+#include "llama-batch.h"
 #include "llama-io.h"
 #include "llama-memory.h"
 #include "llama-mmap.h"
@@ -18,7 +19,8 @@
 llama_context::llama_context(
         const llama_model & model,
               llama_context_params params) :
-    model(model) {
+    model(model),
+    batch_allocr(std::make_unique<llama_batch_allocr>()) {
     LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__);
 
     t_start_us = model.t_start_us;
@@ -27,8 +29,8 @@ llama_context::llama_context(
     const auto & hparams = model.hparams;
 
     cparams.n_seq_max = std::max(1u, params.n_seq_max);
-    if (cparams.n_seq_max > LLAMA_MAX_PARALLEL_SEQUENCES) {
-        throw std::runtime_error("n_seq_max must be <= " + std::to_string(LLAMA_MAX_PARALLEL_SEQUENCES));
+    if (cparams.n_seq_max > LLAMA_MAX_SEQ) {
+        throw std::runtime_error("n_seq_max must be <= " + std::to_string(LLAMA_MAX_SEQ));
     }
 
     cparams.n_threads        = params.n_threads;
@@ -494,7 +496,7 @@ float * llama_context::get_logits() {
 }
 
 float * llama_context::get_logits_ith(int32_t i) {
-    int32_t j = -1;
+    int64_t j = -1;
 
     try {
         if (logits == nullptr) {
@@ -517,7 +519,7 @@ float * llama_context::get_logits_ith(int32_t i) {
         }
         if (j >= n_outputs) {
             // This should not happen
-            throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs));
+            throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
         }
 
         return logits + j*model.vocab.n_tokens();
@@ -536,7 +538,7 @@ float * llama_context::get_embeddings() {
 }
 
 float * llama_context::get_embeddings_ith(int32_t i) {
-    int32_t j = -1;
+    int64_t j = -1;
 
     try {
         if (embd == nullptr) {
@@ -559,7 +561,7 @@ float * llama_context::get_embeddings_ith(int32_t i) {
         }
         if (j >= n_outputs) {
             // This should not happen
-            throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs));
+            throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
         }
 
         return embd + j*model.hparams.n_embd;
@@ -719,52 +721,41 @@ llm_graph_result_ptr llama_context::process_ubatch(const llama_ubatch & ubatch,
     return res;
 }
 
-int llama_context::encode(llama_batch & inp_batch) {
-    if (inp_batch.n_tokens == 0) {
+int llama_context::encode(const llama_batch & batch_inp) {
+    if (batch_inp.n_tokens == 0) {
         LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
         return -1;
     }
 
-    // temporary allocate memory for the input batch if needed
     // note: during encode, we always pass the full sequence starting from pos = 0
-    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : 0);
+    if (!batch_allocr->init(batch_inp, model.vocab, nullptr, true)) {
+        LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
+        return -1;
+    }
 
-    const llama_batch & batch = batch_allocr.batch;
-    const int32_t n_tokens = batch.n_tokens;
+    const llama_batch & batch = batch_allocr->get_batch();
 
-    const auto & hparams = model.hparams;
+    const uint32_t n_tokens = batch.n_tokens;
 
     GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
 
-    // TODO: move the validation to the llama_batch_allocr
-    if (batch.token) {
-        for (int32_t i = 0; i < n_tokens; ++i) {
-            if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) {
-                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
-                return -1;
-            }
-
-            if (batch.seq_id && (batch.seq_id[i][0] < 0 || batch.seq_id[i][0] >= LLAMA_MAX_PARALLEL_SEQUENCES)) {
-                LLAMA_LOG_ERROR("%s: invalid seq_id[%d] = %d > %d\n", __func__, i, batch.seq_id[i][0], LLAMA_MAX_PARALLEL_SEQUENCES);
-                throw -1;
-            }
-        }
-    }
-
     // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot
-    GGML_ASSERT(cparams.n_ubatch >= (uint32_t) n_tokens && "encoder requires n_ubatch >= n_tokens");
+    GGML_ASSERT(cparams.n_ubatch >= n_tokens && "encoder requires n_ubatch >= n_tokens");
 
     if (t_compute_start_us == 0) {
         t_compute_start_us = ggml_time_us();
     }
 
+    // TODO: this clear of the buffer can easily be forgotten - need something better
     embd_seq.clear();
 
     n_queued_tokens += n_tokens;
 
+    const auto & hparams = model.hparams;
+
     const int64_t n_embd = hparams.n_embd;
 
-    llama_sbatch sbatch = llama_sbatch(batch, n_embd, /* simple_split */ true, /* logits_all */ true);
+    llama_sbatch sbatch = llama_sbatch(batch, n_embd, /* simple_split */ true);
 
     const llama_ubatch ubatch = sbatch.split_simple(n_tokens);
 
@@ -774,7 +765,7 @@ int llama_context::encode(llama_batch & inp_batch) {
         return -2;
     };
 
-    for (int32_t i = 0; i < n_tokens; ++i) {
+    for (uint32_t i = 0; i < n_tokens; ++i) {
         output_ids[i] = i;
     }
 
@@ -830,7 +821,8 @@ int llama_context::encode(llama_batch & inp_batch) {
 
                     GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits
 
-                    for (int32_t i = 0; i < n_tokens; i++) {
+                    // TODO: fix indexing [UBATCH_IDX]
+                    for (uint32_t i = 0; i < n_tokens; i++) {
                         const llama_seq_id seq_id = ubatch.seq_id[i][0];
                         if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
                             continue;
@@ -845,6 +837,7 @@ int llama_context::encode(llama_batch & inp_batch) {
                     auto & embd_seq_out = embd_seq;
                     const uint32_t n_cls_out = hparams.n_cls_out;
 
+                    // TODO: fix indexing [UBATCH_IDX]
                     for (uint32_t s = 0; s < ubatch.n_seqs; ++s) {
                         const llama_seq_id seq_id = ubatch.seq_id[s][0];
                         if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
@@ -878,10 +871,10 @@ int llama_context::encode(llama_batch & inp_batch) {
 
         // remember the sequence ids used during the encoding - needed for cross attention later
         cross.seq_ids_enc.resize(n_tokens);
-        for (int32_t i = 0; i < n_tokens; i++) {
+        for (uint32_t i = 0; i < n_tokens; i++) {
             cross.seq_ids_enc[i].clear();
-            for (int s = 0; s < ubatch.n_seq_id[i]; s++) {
-                llama_seq_id seq_id = ubatch.seq_id[i][s];
+            for (int s = 0; s < batch.n_seq_id[i]; s++) {
+                llama_seq_id seq_id = batch.seq_id[i][s];
                 cross.seq_ids_enc[i].insert(seq_id);
             }
         }
@@ -890,51 +883,45 @@ int llama_context::encode(llama_batch & inp_batch) {
     return 0;
 }
 
-int llama_context::decode(llama_batch & inp_batch) {
+int llama_context::decode(const llama_batch & batch_inp) {
     if (!memory) {
         LLAMA_LOG_DEBUG("%s: cannot decode batches with this context (calling encode() instead)\n", __func__);
-        return encode(inp_batch);
+        return encode(batch_inp);
     }
 
-    if (inp_batch.n_tokens == 0) {
+    if (batch_inp.n_tokens == 0) {
         LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
         return -1;
     }
 
-    if (!inp_batch.pos) {
-        if (inp_batch.seq_id) {
-            LLAMA_LOG_ERROR("%s: pos == NULL, but seq_id != NULL\n", __func__);
-            return -1;
-        }
+    // when computing embeddings, all tokens are output
+    const bool embd_all = cparams.embeddings;
+
+    if (!batch_allocr->init(batch_inp, model.vocab, memory.get(), embd_all)) {
+        LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
+        return -1;
     }
 
-    // temporary allocate memory for the input batch if needed
-    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : memory->seq_pos_max(0) + 1);
-
-    const llama_batch & batch = batch_allocr.batch;
+    const llama_batch & batch = batch_allocr->get_batch();
 
     const auto & vocab   = model.vocab;
     const auto & hparams = model.hparams;
 
     const int32_t n_vocab = vocab.n_tokens();
+    const int64_t n_embd  = hparams.n_embd;
 
-    const int64_t n_tokens_all = batch.n_tokens;
-    const int64_t n_embd       = hparams.n_embd;
+    const uint32_t n_tokens_all = batch.n_tokens;
 
     GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
 
-    // TODO: move the validation to the llama_batch_allocr
-    if (batch.token) {
-        for (int64_t i = 0; i < n_tokens_all; ++i) {
-            if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) {
-                LLAMA_LOG_ERROR("%s: invalid token[%" PRId64 "] = %d\n", __func__, i, batch.token[i]);
-                return -1;
-            }
+    const uint32_t n_outputs_all = batch_allocr->get_n_outputs();
 
-            if (batch.seq_id && (batch.seq_id[i][0] < 0 || batch.seq_id[i][0] >= LLAMA_MAX_PARALLEL_SEQUENCES)) {
-                LLAMA_LOG_ERROR("%s: invalid seq_id[%" PRId64 "] = %d >= %d\n", __func__, i, batch.seq_id[i][0], LLAMA_MAX_PARALLEL_SEQUENCES);
-                return -1;
-            }
+    if (embd_all) {
+        // require that all tokens are output
+        if (n_outputs_all != n_tokens_all) {
+            LLAMA_LOG_ERROR("%s: pooled embedding requires that all tokens are output (n_outputs_all = %d, n_tokens_all = %d)\n",
+                    __func__, n_outputs_all, n_tokens_all);
+            return -1;
         }
     }
 
@@ -947,25 +934,9 @@ int llama_context::decode(llama_batch & inp_batch) {
     }
     n_queued_tokens += n_tokens_all;
 
-    // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens
-    const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
-
+    // TODO: this clear of the buffer can easily be forgotten - need something better
     embd_seq.clear();
 
-    int64_t n_outputs_all = 0;
-
-    // count outputs
-    if (batch.logits && !embd_pooled) {
-        for (uint32_t i = 0; i < n_tokens_all; ++i) {
-            n_outputs_all += batch.logits[i] != 0;
-        }
-    } else if (embd_pooled) {
-        n_outputs_all = n_tokens_all;
-    } else {
-        // keep last output only
-        n_outputs_all = 1;
-    }
-
     bool did_optimize = false;
 
     // handle any pending defrags/shifts
@@ -974,7 +945,7 @@ int llama_context::decode(llama_batch & inp_batch) {
     llama_memory_state_ptr mstate;
 
     while (true) {
-        mstate = memory->init_batch(batch, cparams.n_ubatch, embd_pooled, /* logits_all */ n_outputs_all == n_tokens_all);
+        mstate = memory->init_batch(batch, cparams.n_ubatch, embd_all);
         if (!mstate) {
             return -2;
         }
@@ -1018,7 +989,7 @@ int llama_context::decode(llama_batch & inp_batch) {
 
     // reserve output buffer
     if (output_reserve(n_outputs_all) < n_outputs_all) {
-        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %" PRId64 " outputs\n", __func__, n_outputs_all);
+        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
         return -2;
     };
 
@@ -1027,7 +998,7 @@ int llama_context::decode(llama_batch & inp_batch) {
     do {
         const auto & ubatch = mstate->get_ubatch();
 
-        // count the outputs in this u_batch
+        // count the outputs in this ubatch
         {
             int32_t n_outputs_new = 0;
 
@@ -1052,18 +1023,19 @@ int llama_context::decode(llama_batch & inp_batch) {
 
         if (!res) {
             // the last ubatch failed or was aborted -> remove all positions of that ubatch from the KV cache
-            llama_pos pos_min[LLAMA_MAX_PARALLEL_SEQUENCES];
-            for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+            llama_pos pos_min[LLAMA_MAX_SEQ];
+            for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
                 pos_min[s] = std::numeric_limits<llama_pos>::max();
             }
 
+            // TODO: fix sequence indexing
             for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
                 const auto & seq_id = ubatch.seq_id[i][0];
 
                 pos_min[seq_id] = std::min(pos_min[seq_id], ubatch.pos[i]);
             }
 
-            for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+            for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
                 if (pos_min[s] == std::numeric_limits<llama_pos>::max()) {
                     continue;
                 }
@@ -1086,7 +1058,7 @@ int llama_context::decode(llama_batch & inp_batch) {
         //    ggml_graph_dump_dot(gf, NULL, "llama.dot");
         //}
 
-        auto * t_logits = cparams.embeddings ? nullptr         : res->get_logits();
+        auto * t_logits = res->get_logits();
         auto * t_embd   = cparams.embeddings ? res->get_embd() : nullptr;
 
         if (t_embd && res->get_embd_pooled()) {
@@ -1170,14 +1142,14 @@ int llama_context::decode(llama_batch & inp_batch) {
     n_outputs = n_outputs_all;
 
     // set output mappings
-    {
+    if (n_outputs > 0) {
         bool sorted_output = true;
 
         auto & out_ids = mstate->out_ids();
 
-        GGML_ASSERT(out_ids.size() == (size_t) n_outputs_all);
+        GGML_ASSERT(out_ids.size() == (size_t) n_outputs);
 
-        for (int64_t i = 0; i < n_outputs_all; ++i) {
+        for (int64_t i = 0; i < n_outputs; ++i) {
             int64_t out_id = out_ids[i];
             output_ids[out_id] = i;
             if (out_id != i) {
@@ -1189,20 +1161,22 @@ int llama_context::decode(llama_batch & inp_batch) {
         // note: this is mostly relevant for recurrent models atm
         if (!sorted_output) {
             const uint32_t n_vocab = model.vocab.n_tokens();
-            const uint32_t n_embd  = model.hparams.n_embd;
+            const uint64_t n_embd  = model.hparams.n_embd;
 
             GGML_ASSERT((size_t) n_outputs == out_ids.size());
 
             // TODO: is there something more efficient which also minimizes swaps?
             // selection sort, to minimize swaps (from https://en.wikipedia.org/wiki/Selection_sort)
-            for (int32_t i = 0; i < n_outputs - 1; ++i) {
-                int32_t j_min = i;
-                for (int32_t j = i + 1; j < n_outputs; ++j) {
+            for (uint32_t i = 0; i < n_outputs - 1; ++i) {
+                uint32_t j_min = i;
+                for (uint32_t j = i + 1; j < n_outputs; ++j) {
                     if (out_ids[j] < out_ids[j_min]) {
                         j_min = j;
                     }
                 }
-                if (j_min == i) { continue; }
+                if (j_min == i) {
+                    continue;
+                }
                 std::swap(out_ids[i], out_ids[j_min]);
                 if (logits_size > 0) {
                     for (uint32_t k = 0; k < n_vocab; k++) {
@@ -1215,8 +1189,10 @@ int llama_context::decode(llama_batch & inp_batch) {
                     }
                 }
             }
+
             std::fill(output_ids.begin(), output_ids.end(), -1);
-            for (int32_t i = 0; i < n_outputs; ++i) {
+
+            for (uint32_t i = 0; i < n_outputs; ++i) {
                 output_ids[out_ids[i]] = i;
             }
         }
@@ -1236,7 +1212,7 @@ int llama_context::decode(llama_batch & inp_batch) {
 // output
 //
 
-int32_t llama_context::output_reserve(int32_t n_outputs) {
+uint32_t llama_context::output_reserve(int32_t n_outputs) {
     const auto & hparams = model.hparams;
     const auto & vocab   = model.vocab;
 
@@ -1246,9 +1222,8 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
     const auto n_vocab = vocab.n_tokens();
     const auto n_embd  = hparams.n_embd;
 
-    // TODO: use a per-batch flag for logits presence instead
-    bool has_logits = !cparams.embeddings;
-    bool has_embd   =  cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE);
+    bool has_logits = true;
+    bool has_embd   = cparams.embeddings;
 
     // TODO: hacky enc-dec support
     if (model.arch == LLM_ARCH_T5) {
@@ -1302,8 +1277,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
     // set all ids as invalid (negative)
     std::fill(output_ids.begin(), output_ids.end(), -1);
 
-    this->n_outputs     = 0;
-    this->n_outputs_max = n_outputs_max;
+    this->n_outputs = 0;
 
     return n_outputs_max;
 }
@@ -1332,7 +1306,7 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
     LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs);
 
     if (n_tokens % n_seqs != 0) {
-        n_tokens = (n_tokens / n_seqs) * n_seqs;
+        n_tokens = ((n_tokens + (n_seqs - 1)) / n_seqs) * n_seqs; // round to next multiple of n_seqs
         n_outputs = std::min(n_outputs, n_tokens);
 
         LLAMA_LOG_DEBUG("%s: making n_tokens a multiple of n_seqs - n_tokens = %u, n_seqs = %u, n_outputs = %u\n", __func__, n_tokens, n_seqs, n_outputs);
@@ -1794,14 +1768,12 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
 
         std::vector<int32_t> w_output_pos;
 
-        GGML_ASSERT(n_outputs <= n_outputs_max);
-
         w_output_pos.resize(n_outputs);
 
         // build a more compact representation of the output ids
         for (size_t i = 0; i < n_batch(); ++i) {
             // map an output id to a position in the batch
-            int32_t pos = output_ids[i];
+            int64_t pos = output_ids[i];
             if (pos >= 0) {
                 GGML_ASSERT(pos < n_outputs);
                 w_output_pos[pos] = i;
@@ -2071,14 +2043,11 @@ void llama_context::opt_epoch_iter(
 
         n_queued_tokens += n_tokens_all;
 
-        // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens
-        const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
-
         embd_seq.clear();
 
-        int64_t n_outputs_all = n_tokens_all;
+        uint32_t n_outputs_all = n_tokens_all;
 
-        auto mstate = memory->init_batch(batch, cparams.n_ubatch, embd_pooled, /* logits_all */ true);
+        auto mstate = memory->init_batch(batch, cparams.n_ubatch, true);
         if (!mstate || mstate->get_status() != LLAMA_MEMORY_STATUS_SUCCESS) {
             LLAMA_LOG_ERROR("%s: could not initialize batch\n", __func__);
             break;
@@ -2086,7 +2055,7 @@ void llama_context::opt_epoch_iter(
 
         // reserve output buffer
         if (output_reserve(n_outputs_all) < n_outputs_all) {
-            LLAMA_LOG_ERROR("%s: could not reserve space for batch with %" PRId64 " outputs\n", __func__, n_outputs_all);
+            LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
             GGML_ABORT("TODO: handle this error");
         };
 

src/llama-context.h

@@ -1,7 +1,6 @@
 #pragma once
 
 #include "llama.h"
-#include "llama-batch.h"
 #include "llama-cparams.h"
 #include "llama-graph.h"
 #include "llama-adapter.h"
@@ -13,6 +12,7 @@
 #include <vector>
 
 struct llama_model;
+class llama_batch_allocr;
 
 class llama_io_read_i;
 class llama_io_write_i;
@@ -102,8 +102,8 @@ struct llama_context {
             llama_memory_state_i * mstate,
                      ggml_status & ret);
 
-    int encode(llama_batch & inp_batch);
-    int decode(llama_batch & inp_batch);
+    int encode(const llama_batch & batch_inp);
+    int decode(const llama_batch & batch_inp);
 
     //
     // state save/load
@@ -181,7 +181,7 @@ private:
 
     // Make sure enough space is available for outputs.
     // Returns max number of outputs for which space was reserved.
-    int32_t output_reserve(int32_t n_outputs);
+    uint32_t output_reserve(int32_t n_outputs);
 
     //
     // graph
@@ -246,8 +246,10 @@ private:
     // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE
     std::map<llama_seq_id, std::vector<float>> embd_seq;
 
-    int32_t n_outputs     = 0; // number of actually-used outputs in the current ubatch or last logical batch
-    int32_t n_outputs_max = 0; // capacity (of tokens positions) for the output buffers
+    // reuse the batch_allocr to avoid unnecessary memory allocations
+    std::unique_ptr<llama_batch_allocr> batch_allocr;
+
+    uint32_t n_outputs = 0; // number of actually-used outputs in the current ubatch or last logical batch
 
     std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
 

src/llama-cparams.cpp

@@ -1,5 +1,5 @@
 #include "llama-cparams.h"
 
 size_t llama_max_parallel_sequences(void) {
-    return LLAMA_MAX_PARALLEL_SEQUENCES;
+    return LLAMA_MAX_SEQ;
 }

src/llama-cparams.h

@@ -4,7 +4,7 @@
 
 #include <cstdint>
 
-#define LLAMA_MAX_PARALLEL_SEQUENCES 64
+#define LLAMA_MAX_SEQ 64
 
 struct llama_cparams {
     uint32_t n_ctx;           // context size used during inference

src/llama-graph.cpp

@@ -139,6 +139,7 @@ void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) {
 
         std::vector<uint64_t> sum(n_tokens, 0);
 
+        // TODO: fix indexing [UBATCH_IDX]
         for (int s = 0; s < n_seqs; ++s) {
             const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
@@ -156,6 +157,7 @@ void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) {
             }
         }
 
+        // TODO: fix indexing [UBATCH_IDX]
         for (int s = 0; s < n_seqs; ++s) {
             const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
@@ -180,6 +182,7 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
         uint32_t * data = (uint32_t *) cls->data;
         memset(cls->data, 0, n_tokens * ggml_element_size(cls));
 
+        // TODO: fix indexing [UBATCH_IDX]
         for (int s = 0; s < n_seqs; ++s) {
             const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
@@ -210,6 +213,7 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
         std::vector<int> last_pos(n_tokens, -1);
         std::vector<int> last_row(n_tokens, -1);
 
+        // TODO: fix indexing [UBATCH_IDX]
         for (int s = 0; s < n_seqs; ++s) {
             const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
@@ -250,22 +254,6 @@ void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) {
     }
 }
 
-void llm_graph_input_s_mask::set_input(const llama_ubatch * ubatch) {
-    GGML_UNUSED(ubatch);
-
-    const int64_t n_kv = kv_state->get_n_kv();
-
-    if (s_mask) {
-        GGML_ASSERT(ggml_backend_buffer_is_host(s_mask->buffer));
-        float * data = (float *) s_mask->data;
-
-        // clear unused states
-        for (int i = 0; i < n_kv; ++i) {
-            data[i] = kv_state->s_mask(i);
-        }
-    }
-}
-
 void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
     GGML_UNUSED(ubatch);
 
@@ -299,6 +287,7 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
                                 const int32_t ti = s0*n_seq_tokens + i;
                                 float f = -INFINITY;
 
+                                // TODO: fix indexing [UBATCH_IDX]
                                 for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) {
                                     if (ubatch->seq_id[s0][s] == seq_id && ubatch->pos[ti] <= ubatch->pos[tj]) {
                                         if (hparams.use_alibi) {
@@ -338,6 +327,7 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
                                 const int32_t ti = s0*n_seq_tokens + i;
                                 float f = -INFINITY;
 
+                                // TODO: fix indexing [UBATCH_IDX]
                                 for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) {
                                     if (ubatch->seq_id[s0][s] == seq_id) {
                                         if (hparams.use_alibi) {
@@ -393,6 +383,7 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
             for (int j = 0; j < n_tokens; ++j) {
                 for (int i = 0; i < n_enc; ++i) {
                     float f = -INFINITY;
+                    // TODO: fix indexing [UBATCH_IDX]
                     for (int s = 0; s < ubatch->n_seq_id[j]; ++s) {
                         const llama_seq_id seq_id = ubatch->seq_id[j][s];
                         if (cross->seq_ids_enc[i].find(seq_id) != cross->seq_ids_enc[i].end()) {
@@ -650,6 +641,7 @@ ggml_tensor * llm_graph_context::build_ffn(
             {
                 // Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
                 int64_t split_point = cur->ne[0] / 2;
+                // TODO: these conts should not be needed, see https://github.com/ggml-org/llama.cpp/pull/14090#discussion_r2137437217
                 ggml_tensor * x0 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], 0));
                 ggml_tensor * x1 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur)));
 
@@ -663,7 +655,7 @@ ggml_tensor * llm_graph_context::build_ffn(
             {
                 // Split into two equal parts
                 int64_t split_point = cur->ne[0] / 2;
-                // TODO: these conts should not be needed
+                // TODO: these conts should not be needed, see https://github.com/ggml-org/llama.cpp/pull/14090#discussion_r2137437217
                 ggml_tensor * x0 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], 0));
                 ggml_tensor * x1 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur)));
 
@@ -986,23 +978,6 @@ ggml_tensor * llm_graph_context::build_inp_s_copy() const {
     return cur;
 }
 
-ggml_tensor * llm_graph_context::build_inp_s_mask() const {
-    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
-
-    auto inp = std::make_unique<llm_graph_input_s_mask>(kv_state);
-
-    const auto n_kv = kv_state->get_n_kv();
-
-    auto & cur = inp->s_mask;
-
-    cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_kv);
-    ggml_set_input(cur);
-
-    res->add_input(std::move(inp));
-
-    return cur;
-}
-
 ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
     auto inp = std::make_unique<llm_graph_input_cross_embd>(cross);
 
@@ -1455,43 +1430,53 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
-ggml_tensor * llm_graph_context::build_copy_mask_state(
+ggml_tensor * llm_graph_context::build_recurrent_state(
          ggml_cgraph * gf,
          ggml_tensor * s,
          ggml_tensor * state_copy,
-         ggml_tensor * state_mask,
-             int32_t   n_state,
-             int32_t   n_seqs) const {
+             int32_t   state_size,
+             int32_t   n_seqs,
+                bool   avoid_copies) const {
     const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
 
     const auto n_kv    = kv_state->get_n_kv();
     const auto kv_head = kv_state->get_head();
+    const auto rs_zero = kv_state->get_rs_z();
 
-    ggml_tensor * states = ggml_reshape_2d(ctx0, s, n_state, kv_state->get_size());
+    ggml_tensor * states = ggml_reshape_2d(ctx0, s, state_size, kv_state->get_size());
 
-    // copy states
-    // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv
-    // this shrinks the tensors's ne[1] to n_kv
-    states = ggml_get_rows(ctx0, states, state_copy);
+    // Clear a single state which will then be copied to the other cleared states.
+    // Note that this is a no-op when the view is zero-sized.
+    ggml_tensor * state_zero = ggml_view_1d(ctx0, states, state_size*(rs_zero >= 0), rs_zero*states->nb[1]*(rs_zero >= 0));
+    ggml_build_forward_expand(gf, ggml_scale_inplace(ctx0, state_zero, 0));
 
-    // clear states of sequences which are starting at the beginning of this batch
-    // FIXME: zero-out NANs?
-    states = ggml_mul(ctx0, states, state_mask);
+    ggml_tensor * output_states;
 
-    // copy states which won't be changed further (between n_seqs and n_kv)
+    if (!avoid_copies) {
+        // copy states
+        // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv
+        // {state_size, kv_size} -> {state_size, n_seqs}
+        output_states = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_seqs, 0));
+        ggml_build_forward_expand(gf, output_states);
+    } else {
+        // FIXME: make the gathering operation happen before the copy below
+        //        (maybe with an optional lambda function passed as a parameter instead of `avoid_copies`?)
+        output_states = states;
+    }
+
+    // copy extra states which won't be changed further (between n_seqs and n_kv)
+    ggml_tensor * states_extra = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_kv - n_seqs, n_seqs*state_copy->nb[0]));
     ggml_build_forward_expand(gf,
         ggml_cpy(ctx0,
-            ggml_view_1d(ctx0, states, n_state*(n_kv - n_seqs), (n_seqs          )*n_state*ggml_element_size(states)),
-            ggml_view_1d(ctx0, s,      n_state*(n_kv - n_seqs), (kv_head + n_seqs)*n_state*ggml_element_size(s))));
+            states_extra,
+            ggml_view_1d(ctx0, s, state_size*(n_kv - n_seqs), (kv_head + n_seqs)*state_size*ggml_element_size(s))));
 
-    // the part of the states that will be used and modified
-    return ggml_view_2d(ctx0, states, n_state, n_seqs, states->nb[1], 0);
+    return output_states;
 }
 
 ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
          ggml_cgraph * gf,
          ggml_tensor * state_copy,
-         ggml_tensor * state_mask,
   const llama_ubatch & ubatch,
                  int   il) const {
     const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
@@ -1502,8 +1487,8 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
 
     ggml_tensor * token_shift_all = kv_state->get_k_l(il);
 
-    ggml_tensor * token_shift = build_copy_mask_state(
-            gf, token_shift_all, state_copy, state_mask,
+    ggml_tensor * token_shift = build_recurrent_state(
+            gf, token_shift_all, state_copy,
             hparams.n_embd_k_s(), n_seqs);
 
     token_shift = ggml_reshape_3d(ctx0, token_shift, hparams.n_embd, token_shift_count, n_seqs);
@@ -1578,23 +1563,30 @@ void llm_graph_context::build_pooling(
                 ggml_tensor * inp_cls = build_inp_cls();
                 inp = ggml_get_rows(ctx0, inp, inp_cls);
 
-                if (cls != nullptr && cls_b != nullptr) {
+                if (cls) {
                     // classification head
                     // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
-                    cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls, inp), cls_b);
+                    cur = ggml_mul_mat(ctx0, cls, inp);
+                    if (cls_b) {
+                        cur = ggml_add(ctx0, cur, cls_b);
+                    }
                     cur = ggml_tanh(ctx0, cur);
 
                     // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
                     // https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896
                     if (cls_out) {
-                        GGML_ASSERT(cls_out_b != nullptr);
-                        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls_out, cur), cls_out_b);
+                        cur = ggml_mul_mat(ctx0, cls_out, cur);
+                        if (cls_out_b) {
+                            cur = ggml_add(ctx0, cur, cls_out_b);
+                        }
                     }
                 } else if (cls_out) {
                     // Single layer classification head (direct projection)
                     // https://github.com/huggingface/transformers/blob/f4fc42216cd56ab6b68270bf80d811614d8d59e4/src/transformers/models/bert/modeling_bert.py#L1476
-                    GGML_ASSERT(cls_out_b != nullptr);
-                    cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls_out, inp), cls_out_b);
+                    cur = ggml_mul_mat(ctx0, cls_out, inp);
+                    if (cls_out_b) {
+                        cur = ggml_add(ctx0, cur, cls_out_b);
+                    }
                 } else {
                     GGML_ABORT("RANK pooling requires either cls+cls_b or cls_out+cls_out_b");
                 }

src/llama-graph.h

@@ -200,18 +200,6 @@ public:
     const llama_kv_cache_recurrent_state * kv_state;
 };
 
-class llm_graph_input_s_mask : public llm_graph_input_i {
-public:
-    llm_graph_input_s_mask(const llama_kv_cache_recurrent_state * kv_state) : kv_state(kv_state) {}
-    virtual ~llm_graph_input_s_mask() = default;
-
-    void set_input(const llama_ubatch * ubatch) override;
-
-    ggml_tensor * s_mask; // F32 [1, n_kv]
-
-    const llama_kv_cache_recurrent_state * kv_state;
-};
-
 class llm_graph_input_cross_embd : public llm_graph_input_i {
 public:
     llm_graph_input_cross_embd(
@@ -390,7 +378,7 @@ struct llm_graph_params {
     const llama_memory_state_i * mstate;
     const llama_cross          * cross;
 
-    int32_t n_outputs;
+    uint32_t n_outputs;
 
     const llm_graph_cb & cb;
 };
@@ -424,8 +412,8 @@ struct llm_graph_context {
     const float norm_eps;
     const float norm_rms_eps;
 
-    const int32_t n_tokens;
-    const int32_t n_outputs;
+    const int64_t n_tokens;
+    const int64_t n_outputs;
     const int32_t n_ctx_orig; // yarn
 
     const enum llama_pooling_type pooling_type;
@@ -521,7 +509,6 @@ struct llm_graph_context {
     ggml_tensor * build_inp_mean() const;
     ggml_tensor * build_inp_cls() const;
     ggml_tensor * build_inp_s_copy() const;
-    ggml_tensor * build_inp_s_mask() const;
 
     ggml_tensor * build_inp_cross_embd() const;
     ggml_tensor * build_inp_pos_bucket_enc() const;
@@ -606,18 +593,17 @@ struct llm_graph_context {
     // recurrent
     //
 
-    ggml_tensor * build_copy_mask_state(
+    ggml_tensor * build_recurrent_state(
              ggml_cgraph * gf,
              ggml_tensor * s,
              ggml_tensor * state_copy,
-             ggml_tensor * state_mask,
-                 int32_t   n_state,
-                 int32_t   n_seqs) const;
+                 int32_t   state_size,
+                 int32_t   n_seqs,
+                    bool   avoid_copies = false) const;
 
     ggml_tensor * build_rwkv_token_shift_load(
              ggml_cgraph * gf,
              ggml_tensor * state_copy,
-             ggml_tensor * state_mask,
       const llama_ubatch & ubatch,
                      int   il) const;
 

src/llama-kv-cache-recurrent.cpp

@@ -359,18 +359,16 @@ llama_pos llama_kv_cache_recurrent::seq_pos_max(llama_seq_id seq_id) const {
     return result;
 }
 
-llama_memory_state_ptr llama_kv_cache_recurrent::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_pooled, bool logits_all) {
-    GGML_UNUSED(embd_pooled);
-
-    auto sbatch = llama_sbatch(batch, hparams.n_embd, false, logits_all);
+llama_memory_state_ptr llama_kv_cache_recurrent::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_all) {
+    auto sbatch = llama_sbatch(batch, hparams.n_embd, false);
 
     std::vector<llama_ubatch> ubatches;
 
     while (sbatch.n_tokens > 0) {
         llama_ubatch ubatch;
 
-        if (embd_pooled) {
-            // Pooled embeddings cannot be split across ubatches (yet)
+        if (embd_all) {
+            // if all tokens are output, split by sequence
             ubatch = sbatch.split_seq(n_ubatch);
         } else {
             ubatch = sbatch.split_equal(n_ubatch);
@@ -406,21 +404,12 @@ bool llama_kv_cache_recurrent::prepare(const std::vector<llama_ubatch> & ubatche
 
     bool success = true;
 
-    // TODO: here we have to verify that all ubatches can fit in the cells
-    //       however, the current implementation is broken because it relies on s_copy() and s_mask() to update the cells
-    //         during the compute of each ubatch. to reproduce, uncomment the following loop and run:
-    //
-    //           $ llama-parallel -m ./mamba-130m/ggml-model-f16.gguf -np 5 -ns 8
-    //
-    //       recovery from failures when the batch does not fit in the KV cache will not work correctly until this is fixed
-    //
-    GGML_UNUSED(ubatches);
-    //for (const auto & ubatch : ubatches) {
-    //    if (!find_slot(ubatch)) {
-    //        success = false;
-    //        break;
-    //    }
-    //}
+    for (const auto & ubatch : ubatches) {
+        if (!find_slot(ubatch)) {
+            success = false;
+            break;
+        }
+    }
 
     // restore the original state
     cells = std::move(org_cells);
@@ -431,14 +420,13 @@ bool llama_kv_cache_recurrent::prepare(const std::vector<llama_ubatch> & ubatche
 }
 
 bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
-    const uint32_t n_tokens = ubatch.n_tokens;
-    const uint32_t n_seqs   = ubatch.n_seqs;
+    const uint32_t n_seqs = ubatch.n_seqs;
 
     const uint32_t n_seq_tokens = ubatch.n_seq_tokens;
 
     // if we have enough unused cells before the current head ->
     //   better to start searching from the beginning of the cache, hoping to fill it
-    if (head > used + 2*n_tokens) {
+    if (head > used + 2*n_seqs) {
         head = 0;
     }
 
@@ -534,16 +522,16 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
                 empty_cell.src = orig_cell.src;
                 orig_cell.seq_id.erase(seq_id);
                 empty_cell.seq_id.insert(seq_id); // will be overwritten
+                GGML_ASSERT(!orig_cell.is_empty()); // has at least one remaining seq_id
             }
             seq_meta.tail = next_empty_cell;
             // find next empty cell
             if (s + 1 < n_seqs) {
-                next_empty_cell += 1;
                 for (uint32_t i = 0; i < size; ++i) {
+                    next_empty_cell += 1;
                     if (next_empty_cell >= size) { next_empty_cell -= size; }
                     kv_cell & cell = cells[next_empty_cell];
                     if (cell.is_empty()) { break; }
-                    next_empty_cell += 1;
                 }
             }
         }
@@ -553,8 +541,8 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
 
     // gather and re-order
     for (uint32_t s = 0; s < n_seqs; ++s) {
-        int32_t dst_id = s + min;
-        int32_t src_id = cells[ubatch.seq_id[s][0]].tail;
+        const int32_t dst_id = s + min;
+        const int32_t src_id = cells[ubatch.seq_id[s][0]].tail;
         if (dst_id != src_id) {
             kv_cell & dst_cell = cells[dst_id];
             kv_cell & src_cell = cells[src_id];
@@ -563,12 +551,14 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
             std::swap(dst_cell.src, src_cell.src);
             std::swap(dst_cell.seq_id, src_cell.seq_id);
 
-            // swap tails (assuming they NEVER overlap)
-            for (const llama_seq_id seq_id : src_cell.seq_id) {
-                cells[seq_id].tail = src_id;
-            }
-            for (const llama_seq_id seq_id : dst_cell.seq_id) {
-                cells[seq_id].tail = dst_id;
+            // swap tails
+            for (uint32_t i = 0; i < size; ++i) {
+                int32_t & tail = cells[i].tail;
+                if (tail == src_id) {
+                    tail = dst_id;
+                } else if (tail == dst_id) {
+                    tail = src_id;
+                }
             }
         }
     }
@@ -576,7 +566,7 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
     // update the pos of the used seqs
     for (uint32_t s = 0; s < n_seqs; ++s) {
         const llama_pos last_pos = ubatch.pos[n_seq_tokens * s + n_seq_tokens - 1];
-        int32_t cell_id = s + min;
+        const int32_t cell_id = s + min;
         kv_cell & cell = cells[cell_id];
 
         if (cell.pos >= 0 && last_pos != cell.pos + (llama_pos) n_seq_tokens) {
@@ -594,6 +584,38 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
         }
     }
 
+    // Find first cell without src refs, to use as the zero-ed state
+    {
+        // TODO: bake-in src refcounts in the cell metadata
+        std::vector<int32_t> refcounts(size, 0);
+        for (size_t i = 0; i < size; ++i) {
+            const int32_t src = cells[i].src;
+            if (src >= 0) {
+                refcounts[src] += 1;
+            }
+        }
+
+        rs_z = -1;
+        for (int i = min; i <= max; ++i) {
+            if (refcounts[i] == 0) {
+                rs_z = i;
+                break;
+            }
+        }
+
+        for (int i = min; i <= max; ++i) {
+            if (cells[i].src < 0) {
+                GGML_ASSERT(rs_z >= 0);
+                cells[i].src0 = rs_z;
+            } else {
+                // Stage the source ids for all used cells to allow correct seq_* behavior
+                // and still make these values available when setting the inputs
+                cells[i].src0 = cells[i].src;
+            }
+            cells[i].src = i; // avoid moving or clearing twice
+        }
+    }
+
     // allow getting the range of used cells, from head to head + n
     head = min;
     n    = max - min + 1;
@@ -605,47 +627,8 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
 }
 
 bool llama_kv_cache_recurrent::get_can_shift() const {
-    return false;
-}
-
-int32_t llama_kv_cache_recurrent::s_copy(int i) const {
-    const uint32_t cell_id = i + head;
-
-    //////////////////////////////////////////////
-    // TODO: this should not mutate the KV cache !
-    kv_cell & cell = const_cast<kv_cell &>(cells[cell_id]);
-
-    // prevent out-of-bound sources
-    if (cell.src < 0 || (uint32_t) cell.src >= size) {
-        cell.src = cell_id;
-    }
-
-    int32_t res = cell.src;
-
-    // TODO: do not mutate the KV cache
-    // ensure copy only happens once
-    if (cell.src != (int32_t) cell_id) {
-        cell.src = cell_id;
-    }
-
-    return res;
-}
-
-float llama_kv_cache_recurrent::s_mask(int i) const {
-    const uint32_t cell_id = i + head;
-
-    //////////////////////////////////////////////
-    // TODO: this should not mutate the KV cache !
-    kv_cell & cell = const_cast<kv_cell &>(cells[cell_id]);
-
-    float res = (float) (cell.src >= 0);
-
-    // only clear once
-    if (cell.src < 0) {
-        cell.src = cell_id;
-    }
-
-    return res;
+    // shifting the pos is trivial for recurrent models
+    return true;
 }
 
 size_t llama_kv_cache_recurrent::total_size() const {
@@ -1111,6 +1094,10 @@ uint32_t llama_kv_cache_recurrent_state::get_head() const {
     return is_full ? 0 : kv->head;
 }
 
+int32_t llama_kv_cache_recurrent_state::get_rs_z() const {
+    return is_full ? 0 : kv->rs_z;
+}
+
 uint32_t llama_kv_cache_recurrent_state::get_size() const {
     return kv->size;
 }
@@ -1124,9 +1111,5 @@ ggml_tensor * llama_kv_cache_recurrent_state::get_v_l(int32_t il) const {
 }
 
 int32_t llama_kv_cache_recurrent_state::s_copy(int i) const {
-    return kv->s_copy(i);
-}
-
-float llama_kv_cache_recurrent_state::s_mask(int i) const {
-    return kv->s_mask(i);
+    return  kv->cells[i + kv->head].src0;
 }

src/llama-kv-cache-recurrent.h

@@ -32,8 +32,7 @@ public:
     llama_memory_state_ptr init_batch(
             const llama_batch & batch,
             uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) override;
+            bool embd_all) override;
 
     llama_memory_state_ptr init_full() override;
 
@@ -57,10 +56,6 @@ public:
 
     bool get_can_shift() const override;
 
-    // TODO: temporary methods - they are not really const as they do const_cast<>, fix this
-    int32_t s_copy(int i) const;
-    float   s_mask(int i) const;
-
     // state write/load
 
     void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
@@ -73,10 +68,14 @@ public:
     // computed before each graph build
     uint32_t n = 0;
 
+    // first zero-ed state
+    int32_t rs_z = -1;
+
     // TODO: optimize for recurrent state needs
     struct kv_cell {
         llama_pos pos  = -1;
-        int32_t   src  = -1; // used to copy states
+        int32_t   src  = -1; // used to know where states should be copied from
+        int32_t   src0 = -1; // like src, but only used when setting the inputs (allowing to copy once)
         int32_t   tail = -1;
 
         std::set<llama_seq_id> seq_id;
@@ -157,13 +156,13 @@ public:
 
     uint32_t get_n_kv() const;
     uint32_t get_head() const;
+    int32_t  get_rs_z() const;
     uint32_t get_size() const;
 
     ggml_tensor * get_k_l(int32_t il) const;
     ggml_tensor * get_v_l(int32_t il) const;
 
     int32_t s_copy(int i) const;
-    float   s_mask(int i) const;
 
 private:
     const llama_memory_status status;

src/llama-kv-cache-unified-iswa.cpp

@@ -95,36 +95,69 @@ llama_pos llama_kv_cache_unified_iswa::seq_pos_max(llama_seq_id seq_id) const {
     return kv_swa->seq_pos_max(seq_id);
 }
 
-llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_pooled, bool logits_all) {
-    GGML_UNUSED(embd_pooled);
+llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_all) {
+    GGML_UNUSED(embd_all);
 
-    // TODO: if we fail with split_simple, we should attempt different splitting strategies
+    // first try simple split
+    do {
+        auto sbatch = llama_sbatch(batch, hparams.n_embd, true);
+
+        std::vector<llama_ubatch> ubatches;
+
+        while (sbatch.n_tokens > 0) {
+            auto ubatch = sbatch.split_simple(n_ubatch);
+
+            ubatches.push_back(ubatch);
+        }
+
+        auto heads_base = kv_base->prepare(ubatches);
+        if (heads_base.empty()) {
+            break;
+        }
+
+        auto heads_swa = kv_swa->prepare(ubatches);
+        if (heads_swa.empty()) {
+            break;
+        }
+
+        assert(heads_base.size() == heads_swa.size());
+
+        return std::make_unique<llama_kv_cache_unified_iswa_state>(
+                this, std::move(sbatch), std::move(heads_base), std::move(heads_swa), std::move(ubatches));
+    } while (false);
+
+    // if it fails, try equal split
+    do {
+        auto sbatch = llama_sbatch(batch, hparams.n_embd, false);
+
+        std::vector<llama_ubatch> ubatches;
+
+        while (sbatch.n_tokens > 0) {
+            auto ubatch = sbatch.split_equal(n_ubatch);
+
+            ubatches.push_back(ubatch);
+        }
+
+        auto heads_base = kv_base->prepare(ubatches);
+        if (heads_base.empty()) {
+            break;
+        }
+
+        auto heads_swa = kv_swa->prepare(ubatches);
+        if (heads_swa.empty()) {
+            break;
+        }
+
+        assert(heads_base.size() == heads_swa.size());
+
+        return std::make_unique<llama_kv_cache_unified_iswa_state>(
+                this, std::move(sbatch), std::move(heads_base), std::move(heads_swa), std::move(ubatches));
+    } while (false);
+
+    // TODO: if we fail again, we should attempt different splitting strategies
     //       but to do that properly, we first have to refactor the batches to be more flexible
 
-    auto sbatch = llama_sbatch(batch, hparams.n_embd, true, logits_all);
-
-    std::vector<llama_ubatch> ubatches;
-
-    while (sbatch.n_tokens > 0) {
-        auto ubatch = sbatch.split_simple(n_ubatch);
-
-        ubatches.push_back(ubatch);
-    }
-
-    auto heads_base = kv_base->prepare(ubatches);
-    if (heads_base.empty()) {
-        return std::make_unique<llama_kv_cache_unified_iswa_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
-    }
-
-    auto heads_swa = kv_swa->prepare(ubatches);
-    if (heads_swa.empty()) {
-        return std::make_unique<llama_kv_cache_unified_iswa_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
-    }
-
-    assert(heads_base.size() == heads_swa.size());
-
-    return std::make_unique<llama_kv_cache_unified_iswa_state>(
-            this, std::move(sbatch), std::move(heads_base), std::move(heads_swa), std::move(ubatches));
+    return std::make_unique<llama_kv_cache_unified_iswa_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
 }
 
 llama_memory_state_ptr llama_kv_cache_unified_iswa::init_full() {

src/llama-kv-cache-unified-iswa.h

@@ -34,8 +34,7 @@ public:
     llama_memory_state_ptr init_batch(
             const llama_batch & batch,
             uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) override;
+            bool embd_all) override;
 
     llama_memory_state_ptr init_full() override;
 

src/llama-kv-cache-unified.cpp

@@ -127,6 +127,9 @@ llama_kv_cache_unified::llama_kv_cache_unified(
                 ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f),
                 ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f));
     }
+
+    const char * LLAMA_KV_CACHE_DEBUG = getenv("LLAMA_KV_CACHE_DEBUG");
+    debug = LLAMA_KV_CACHE_DEBUG ? atoi(LLAMA_KV_CACHE_DEBUG) : 0;
 }
 
 void llama_kv_cache_unified::clear(bool data) {
@@ -307,24 +310,27 @@ llama_pos llama_kv_cache_unified::seq_pos_max(llama_seq_id seq_id) const {
 llama_memory_state_ptr llama_kv_cache_unified::init_batch(
             const llama_batch & batch,
             uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) {
-    GGML_UNUSED(embd_pooled);
+            bool embd_all) {
+    GGML_UNUSED(embd_all);
 
-    auto sbatch = llama_sbatch(batch, hparams.n_embd, true, logits_all);
+    do {
+        auto sbatch = llama_sbatch(batch, hparams.n_embd, true);
 
-    std::vector<llama_ubatch> ubatches;
-    while (sbatch.n_tokens > 0) {
-        ubatches.push_back(sbatch.split_simple(n_ubatch));
-    }
+        std::vector<llama_ubatch> ubatches;
+        while (sbatch.n_tokens > 0) {
+            ubatches.push_back(sbatch.split_simple(n_ubatch));
+        }
 
-    auto heads = prepare(ubatches);
-    if (heads.empty()) {
-        return std::make_unique<llama_kv_cache_unified_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
-    }
+        auto heads = prepare(ubatches);
+        if (heads.empty()) {
+            break;
+        }
 
-    return std::make_unique<llama_kv_cache_unified_state>(
-            this, std::move(sbatch), std::move(heads), std::move(ubatches));
+        return std::make_unique<llama_kv_cache_unified_state>(
+                this, std::move(sbatch), std::move(heads), std::move(ubatches));
+    } while (false);
+
+    return std::make_unique<llama_kv_cache_unified_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
 }
 
 llama_memory_state_ptr llama_kv_cache_unified::init_full() {
@@ -462,7 +468,7 @@ bool llama_kv_cache_unified::update(llama_context * lctx, bool do_shift, const d
             for (uint32_t i = 0; i < n_kv; ++i) {
                 assert(dinfo.ids[i] <= n_kv);
 
-                if (dinfo.ids[i] == n_kv) {
+                if (dinfo.ids[i] == n_kv || dinfo.ids[i] == i) {
                     continue;
                 }
 
@@ -512,43 +518,68 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
         head_cur = 0;
     }
 
-    // otherwise, one cell per token.
-
     if (n_tokens > cells.size()) {
         LLAMA_LOG_ERROR("%s: n_tokens = %d > size = %u\n", __func__, n_tokens, cells.size());
         return -1;
     }
 
-//#define FIND_SLOT_DEBUG 1
-#if FIND_SLOT_DEBUG
-    LLAMA_LOG_WARN("begin: n = %5d, used = %5d, head = %5d, n_swa = %5d\n", cells.used_max_p1(), cells.get_used(), head, n_swa);
+    if (debug > 0) {
+        LLAMA_LOG_DEBUG("%s: n = %5d, used = %5d, head = %5d, size = %5d, n_swa = %5d\n", __func__, cells.used_max_p1(), cells.get_used(), head, get_size(), n_swa);
 
-    // for debugging
-    {
-        std::string ss;
-        if (n_swa > 0) {
+        if ((debug == 2 && n_swa > 0) || debug > 2) {
+            std::string ss;
             for (uint32_t i = 0; i < cells.size(); ++i) {
                 if (cells.is_empty(i)) {
                     ss += '.';
                 } else {
-                    ss += std::to_string(cells.seq_get(i));
+                    assert(cells.seq_count(i) >= 1);
+
+                    if (cells.seq_count(i) == 1) {
+                        ss += std::to_string(cells.seq_get(i));
+                    } else {
+                        ss += 'M';
+                    }
                 }
                 if (i%256 == 255) {
+                    ss += " *";
                     ss += '\n';
                 }
             }
-        }
-        LLAMA_LOG_WARN("\n%s\n", ss.c_str());
-    }
-
-    for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
-        if (cells.seq_pos_min(s) < 0) {
-            continue;
+            LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
         }
 
-        LLAMA_LOG_WARN("kv_cells: n_swa = %4d, min[%d] = %5d, max[%d] = %5d\n", n_swa, s, cells.seq_pos_min(s), s, cells.seq_pos_max(s));
+        if ((debug == 2 && n_swa > 0) || debug > 2) {
+            std::string ss;
+            for (uint32_t i = 0; i < cells.size(); ++i) {
+                std::string cur;
+                if (cells.is_empty(i)) {
+                    cur = '.';
+                } else {
+                    cur = std::to_string(cells.pos_get(i));
+                }
+                const int n = cur.size();
+                for (int j = 0; j < 5 - n; ++j) {
+                    cur += ' ';
+                }
+                ss += cur;
+                if (i%256 == 255) {
+                    ss += " *";
+                }
+                if (i%64 == 63) {
+                    ss += '\n';
+                }
+            }
+            LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
+        }
+
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
+            if (cells.seq_pos_min(s) < 0) {
+                continue;
+            }
+
+            LLAMA_LOG_DEBUG("%s: min[%d] = %5d, max[%d] = %5d\n", __func__, s, cells.seq_pos_min(s), s, cells.seq_pos_max(s));
+        }
     }
-#endif
 
     uint32_t n_tested = 0;
 
@@ -559,21 +590,15 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
             continue;
         }
 
-        // keep track of what the minimum sequence positions would be if we accept the ubatch
-        llama_seq_id seq_pos_min[LLAMA_MAX_PARALLEL_SEQUENCES];
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
-            seq_pos_min[s] = cells.seq_pos_min(s);
-        }
-
         bool found = true;
         for (uint32_t i = 0; i < n_tokens; i++) {
-            const llama_pos    pos    = ubatch.pos[i];
-            const llama_seq_id seq_id = ubatch.seq_id[i][0];
+            //const llama_pos    pos    = ubatch.pos[i];
+            //const llama_seq_id seq_id = ubatch.seq_id[i][0];
 
             // can we use this cell? either:
             //  - the cell is empty
             //  - the cell is occupied only by one sequence:
-            //    - mask causally, if the sequence is the same as the one we are inserting
+            //    - (disabled) mask causally, if the sequence is the same as the one we are inserting
             //    - mask SWA, using current max pos for that sequence in the cache
             //                always insert in the cell with minimum pos
             bool can_use = cells.is_empty(head_cur + i);
@@ -581,21 +606,17 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
             if (!can_use && cells.seq_count(head_cur + i) == 1) {
                 const llama_pos pos_cell = cells.pos_get(head_cur + i);
 
-                // causal mask
-                if (cells.seq_has(head_cur + i, seq_id)) {
-                    can_use = pos_cell >= pos;
-                }
+                // (disabled) causal mask
+                // note: it's better to purge any "future" tokens beforehand
+                //if (cells.seq_has(head_cur + i, seq_id)) {
+                //    can_use = pos_cell >= pos;
+                //}
 
                 if (!can_use) {
                     const llama_seq_id seq_id_cell = cells.seq_get(head_cur + i);
 
                     // SWA mask
-                    // note: we insert only in the cell with minimum pos in order to preserve the invariant that
-                    //       all positions between [pos_min, pos_max] for each sequence will be present in the cache
-                    //       ref: https://github.com/ggml-org/llama.cpp/pull/13746#issuecomment-2916057092
-                    if (pos_cell == seq_pos_min[seq_id_cell] &&
-                        is_masked_swa(pos_cell, cells.seq_pos_max(seq_id_cell) + 1)) {
-                        seq_pos_min[seq_id_cell]++;
+                    if (is_masked_swa(pos_cell, cells.seq_pos_max(seq_id_cell) + 1)) {
                         can_use = true;
                     }
                 }
@@ -623,18 +644,58 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
 }
 
 void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch & ubatch) {
-    for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
-        if (!cells.is_empty(head_cur + i)) {
-            cells.rm(head_cur + i);
-        }
+    if (debug > 0) {
+        LLAMA_LOG_DEBUG("%s: ubatch info:\n", __func__);
+        LLAMA_LOG_DEBUG("%s:   n_tokens = %d, equal_seqs = %d\n", __func__, ubatch.n_tokens, ubatch.equal_seqs);
+        LLAMA_LOG_DEBUG("%s:   n_seq_tokens = %d, n_seqs = %d\n", __func__, ubatch.n_seq_tokens, ubatch.n_seqs);
+    }
 
-        cells.pos_set(head_cur + i, ubatch.pos[i]);
+    // 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];
+    for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        seq_pos_max_rm[s] = -1;
+    }
 
-        for (int32_t j = 0; j < ubatch.n_seq_id[i]; j++) {
-            cells.seq_add(head_cur + i, ubatch.seq_id[i][j]);
+    for (uint32_t s = 0; s < ubatch.n_seqs; ++s) {
+        for (uint32_t j = 0; j < ubatch.n_seq_tokens; ++j) {
+            const uint32_t idx = s*ubatch.n_seq_tokens + j;
+
+            if (!cells.is_empty(head_cur + idx)) {
+                assert(cells.seq_count(head_cur + idx) == 1);
+
+                const llama_seq_id seq_id = cells.seq_get(head_cur + idx);
+                const llama_pos    pos    = cells.pos_get(head_cur + idx);
+
+                seq_pos_max_rm[seq_id] = std::max(seq_pos_max_rm[seq_id], pos);
+
+                cells.rm(head_cur + idx);
+            }
+
+            cells.pos_set(head_cur + idx, ubatch.pos[idx]);
+
+            // TODO: fix indexing [UBATCH_IDX]
+            for (int32_t i = 0; i < ubatch.n_seq_id[s]; i++) {
+                cells.seq_add(head_cur + idx, ubatch.seq_id[s][i]);
+            }
         }
     }
 
+    // note: we want to preserve the invariant that all positions between [pos_min, pos_max] for each sequence
+    //       will be present in the cache. so we have to purge any position which is less than those we would overwrite
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/13746#issuecomment-2916057092
+    for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        if (seq_pos_max_rm[s] == -1) {
+            continue;
+        }
+
+        if (cells.seq_pos_min(s) <= seq_pos_max_rm[s]) {
+            LLAMA_LOG_DEBUG("%s: purging positions [%d, %d] of sequence %d from KV cache\n",
+                    __func__, cells.seq_pos_min(s), seq_pos_max_rm[s], s);
+
+            seq_rm(s, cells.seq_pos_min(s), seq_pos_max_rm[s] + 1);
+        }
+    }
     // move the head at the end of the slot
     head = head_cur + ubatch.n_tokens;
 }
@@ -731,14 +792,14 @@ ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_
 }
 
 void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const {
-    const int64_t n_tokens     = ubatch->n_tokens;
-    const int64_t n_seq_tokens = ubatch->n_seq_tokens;
-    const int64_t n_seqs       = ubatch->n_seqs;
+    const uint32_t n_tokens     = ubatch->n_tokens;
+    const uint32_t n_seq_tokens = ubatch->n_seq_tokens;
+    const uint32_t n_seqs       = ubatch->n_seqs;
 
     GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
     float * data = (float *) dst->data;
 
-    const auto n_kv = dst->ne[0];
+    const int64_t n_kv = dst->ne[0];
 
     // Use only the previous KV cells of the correct sequence for each token of the ubatch.
     // It's assumed that if a token in the batch has multiple sequences, they are equivalent.
@@ -752,12 +813,14 @@ void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ub
     //      xxxxx-----
     //      xxxxx-----
     // To visualize the mask, see https://github.com/ggml-org/llama.cpp/pull/12615
-    for (int h = 0; h < 1; ++h) {
-        for (int s = 0; s < n_seqs; ++s) {
+    for (uint32_t h = 0; h < 1; ++h) {
+        for (uint32_t s = 0; s < n_seqs; ++s) {
             const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
-            for (int j = 0; j < n_seq_tokens; ++j) {
-                const llama_pos p1 = ubatch->pos[s*n_seq_tokens + j];
+            for (uint32_t j = 0; j < n_seq_tokens; ++j) {
+                const uint32_t idx = s*n_seq_tokens + j;
+
+                const llama_pos p1 = ubatch->pos[idx];
 
                 for (uint32_t i = 0; i < n_kv; ++i) {
                     float f = 0.0f;
@@ -787,16 +850,16 @@ void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ub
                         f = -INFINITY;
                     }
 
-                    data[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f;
+                    data[h*(n_kv*n_tokens) + idx*n_kv + i] = f;
                 }
             }
         }
 
         // mask padded tokens
         if (data) {
-            for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
-                for (uint32_t j = 0; j < n_kv; ++j) {
-                    data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY;
+            for (uint32_t j = n_tokens; j < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++j) {
+                for (uint32_t i = 0; i < n_kv; ++i) {
+                    data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY;
                 }
             }
         }
@@ -944,11 +1007,9 @@ llm_graph_result_ptr llama_kv_cache_unified::build_graph_shift(
     const auto & n_embd_head_k = hparams.n_embd_head_k;
   //const auto & n_embd_head_v = hparams.n_embd_head_v;
 
-    //GGML_ASSERT(kv_self->size == n_ctx);
-
     auto inp = std::make_unique<llm_graph_input_k_shift>(this);
 
-    inp->k_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, cparams.n_ctx);
+    inp->k_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, cells.size());
     ggml_set_input(inp->k_shift);
 
     for (const auto & layer : layers) {
@@ -1449,9 +1510,11 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell
         seq_rm(dest_seq_id, -1, -1);
 
         llama_sbatch sbatch;
-        llama_ubatch batch = sbatch.reserve_ubatch(cell_count, /* has_embd */ false);
+        llama_ubatch ubatch = sbatch.reserve_ubatch(cell_count, /* has_embd */ false);
 
-        batch.n_tokens = cell_count;
+        ubatch.n_tokens = cell_count;
+        ubatch.n_seq_tokens = cell_count;
+        ubatch.n_seqs = 1;
 
         for (uint32_t i = 0; i < cell_count; ++i) {
             llama_pos pos;
@@ -1471,18 +1534,18 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell
                 io.read_to(&seq_id, sizeof(seq_id));
             }
 
-            batch.pos[i]      = pos;
-            batch.n_seq_id[i] = n_seq_id;
-            batch.seq_id[i]   = &dest_seq_id;
+            ubatch.pos[i]      = pos;
+            ubatch.n_seq_id[i] = n_seq_id;
+            ubatch.seq_id[i]   = &dest_seq_id;
         }
 
-        const auto head_cur = find_slot(batch);
+        const auto head_cur = find_slot(ubatch);
         if (head_cur < 0) {
             LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__);
             return false;
         }
 
-        apply_ubatch(head_cur, batch);
+        apply_ubatch(head_cur, ubatch);
 
         // keep the head at the old position because we will read the KV data into it in state_read_data()
         head = head_cur;
@@ -1490,8 +1553,8 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell
         // DEBUG CHECK: head_cur should be our first cell, head_cur + cell_count - 1 should be our last cell (verify seq_id and pos values)
         // Assume that this is one contiguous block of cells
         GGML_ASSERT(head_cur + cell_count <= cells.size());
-        GGML_ASSERT(cells.pos_get(head_cur)                  == batch.pos[0]);
-        GGML_ASSERT(cells.pos_get(head_cur + cell_count - 1) == batch.pos[cell_count - 1]);
+        GGML_ASSERT(cells.pos_get(head_cur)                  == ubatch.pos[0]);
+        GGML_ASSERT(cells.pos_get(head_cur + cell_count - 1) == ubatch.pos[cell_count - 1]);
         GGML_ASSERT(cells.seq_has(head_cur,                  dest_seq_id));
         GGML_ASSERT(cells.seq_has(head_cur + cell_count - 1, dest_seq_id));
     } else {
@@ -1676,7 +1739,7 @@ llama_kv_cache_unified_state::llama_kv_cache_unified_state(
         llama_context * lctx,
         bool do_shift,
         defrag_info dinfo) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), lctx(lctx), do_shift(do_shift), dinfo(std::move(dinfo)) {
-    if (!do_shift && dinfo.empty()) {
+    if (!do_shift && this->dinfo.empty()) {
         status = LLAMA_MEMORY_STATUS_NO_UPDATE;
     }
 }

src/llama-kv-cache-unified.h

@@ -59,8 +59,7 @@ public:
     llama_memory_state_ptr init_batch(
             const llama_batch & batch,
             uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) override;
+            bool embd_all) override;
 
     llama_memory_state_ptr init_full() override;
 
@@ -158,6 +157,8 @@ private:
     // SWA
     const uint32_t n_swa = 0;
 
+    int debug = 0;
+
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
     std::vector<ggml_context_ptr>        ctxs;

src/llama-kv-cells.h

@@ -23,7 +23,7 @@ public:
 
         used.clear();
 
-        for (uint32_t s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (uint32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
             seq_pos[s].clear();
         }
     }
@@ -80,6 +80,9 @@ public:
         assert(isrc < pos.size());
         assert(idst < pos.size());
 
+        assert(pos[idst] == -1);
+        assert(pos[isrc] != -1);
+
         pos  [idst] = pos  [isrc];
         shift[idst] = shift[isrc];
         seq  [idst] = seq  [isrc];
@@ -144,9 +147,10 @@ public:
         assert(pos[i] != -1);
 
         seq_pos_rm(i);
+        seq[i].reset();
 
         pos[i] = -1;
-        seq[i].reset();
+        shift[i] = 0;
 
         used.erase(i);
     }
@@ -164,6 +168,7 @@ public:
 
         if (seq[i].none()) {
             pos[i] = -1;
+            shift[i] = 0;
 
             used.erase(i);
 
@@ -192,6 +197,7 @@ public:
             seq[i].reset();
 
             pos[i] = -1;
+            shift[i] = 0;
 
             used.erase(i);
 
@@ -234,7 +240,7 @@ public:
     llama_seq_id seq_get(uint32_t i) const {
         assert(seq[i].count() == 1);
 
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
                 return s;
             }
@@ -247,7 +253,7 @@ public:
     // return -1 if the sequence is not present
     llama_pos seq_pos_min(llama_seq_id seq_id) const {
         assert(seq_id >= 0);
-        assert(seq_id < LLAMA_MAX_PARALLEL_SEQUENCES);
+        assert(seq_id < LLAMA_MAX_SEQ);
 
         if (seq_pos[seq_id].empty()) {
             return -1;
@@ -260,7 +266,7 @@ public:
     // return -1 if the sequence is not present
     llama_pos seq_pos_max(llama_seq_id seq_id) const {
         assert(seq_id >= 0);
-        assert(seq_id < LLAMA_MAX_PARALLEL_SEQUENCES);
+        assert(seq_id < LLAMA_MAX_SEQ);
 
         if (seq_pos[seq_id].empty()) {
             return -1;
@@ -317,21 +323,20 @@ public:
         pos[i]   += d;
         shift[i] += d;
 
-        seq_pos_add(i);
-
         has_shift = true;
 
         if (pos[i] < 0) {
-            seq_pos_rm(i);
-
             seq[i].reset();
             pos[i] = -1;
+            shift[i] = 0;
 
             used.erase(i);
 
             return true;
         }
 
+        seq_pos_add(i);
+
         return false;
     }
 
@@ -379,20 +384,20 @@ private:
     //
     std::vector<llama_pos> shift;
 
-    using bits_t = std::bitset<LLAMA_MAX_PARALLEL_SEQUENCES>;
+    using bits_t = std::bitset<LLAMA_MAX_SEQ>;
 
     // the bitset seq[i] tells us which sequences are currently occupying the i-th cell
     std::vector<bits_t> seq;
 
     // the set seq_pos[s] tells us which positions are currently present for sequence s
     // this way seq_pos[s].begin() and seq_pos[s].rbegin() give us the min/max positions currently in the cache
-    std::set<llama_pos> seq_pos[LLAMA_MAX_PARALLEL_SEQUENCES];
+    std::set<llama_pos> seq_pos[LLAMA_MAX_SEQ];
 
     // helper functions for updating `seq_pos`, once cell at a time:
 
     // remove cell i
     void seq_pos_rm(uint32_t i) {
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
                 seq_pos[s].erase(pos[i]);
             }
@@ -401,7 +406,7 @@ private:
 
     // add cell i
     void seq_pos_add(uint32_t i) {
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
                 seq_pos[s].insert(pos[i]);
             }

src/llama-memory.h

@@ -73,8 +73,7 @@ struct llama_memory_i {
     virtual llama_memory_state_ptr init_batch(
             const llama_batch & batch,
             uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) = 0;
+            bool embd_all) = 0;
 
     // simulate full cache, used for allocating worst-case compute buffers
     virtual llama_memory_state_ptr init_full() = 0;