wip

ggml-ci

.devops/vulkan.Dockerfile

@@ -2,14 +2,30 @@ ARG UBUNTU_VERSION=24.04
 
 FROM ubuntu:$UBUNTU_VERSION AS build
 
-# Install build tools
-RUN apt update && apt install -y git build-essential cmake wget
+# Ref: https://vulkan.lunarg.com/doc/sdk/latest/linux/getting_started.html
 
-# Install Vulkan SDK and cURL
-RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
-    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-noble.list https://packages.lunarg.com/vulkan/lunarg-vulkan-noble.list && \
-    apt update -y && \
-    apt-get install -y vulkan-sdk libcurl4-openssl-dev curl
+# Install build tools
+RUN apt update && apt install -y git build-essential cmake wget xz-utils
+
+# Install Vulkan SDK
+ARG VULKAN_VERSION=1.4.321.1
+RUN ARCH=$(uname -m) && \
+    wget -qO /tmp/vulkan-sdk.tar.xz https://sdk.lunarg.com/sdk/download/${VULKAN_VERSION}/linux/vulkan-sdk-linux-${ARCH}-${VULKAN_VERSION}.tar.xz && \
+    mkdir -p /opt/vulkan && \
+    tar -xf /tmp/vulkan-sdk.tar.xz -C /tmp --strip-components=1 && \
+    mv /tmp/${ARCH}/* /opt/vulkan/ && \
+    rm -rf /tmp/*
+
+# Install cURL and Vulkan SDK dependencies
+RUN apt install -y libcurl4-openssl-dev curl \
+    libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev
+
+# Set environment variables
+ENV VULKAN_SDK=/opt/vulkan
+ENV PATH=$VULKAN_SDK/bin:$PATH
+ENV LD_LIBRARY_PATH=$VULKAN_SDK/lib:$LD_LIBRARY_PATH
+ENV CMAKE_PREFIX_PATH=$VULKAN_SDK:$CMAKE_PREFIX_PATH
+ENV PKG_CONFIG_PATH=$VULKAN_SDK/lib/pkgconfig:$PKG_CONFIG_PATH
 
 # Build it
 WORKDIR /app

.github/copilot-instructions.md

@@ -0,0 +1,262 @@
+# Copilot Instructions for llama.cpp
+
+## Repository Overview
+
+llama.cpp is a large-scale C/C++ project for efficient LLM (Large Language Model) inference with minimal setup and dependencies. The project enables running language models on diverse hardware with state-of-the-art performance.
+
+**Key Facts:**
+- **Primary language**: C/C++ with Python utility scripts
+- **Size**: ~200k+ lines of code across 1000+ files
+- **Architecture**: Modular design with main library (`libllama`) and 40+ executable tools/examples
+- **Core dependency**: ggml tensor library (vendored in `ggml/` directory)
+- **Backends supported**: CPU (AVX/NEON optimized), CUDA, Metal, Vulkan, SYCL, ROCm, MUSA
+- **License**: MIT
+
+## Build Instructions
+
+### Prerequisites
+- CMake 3.14+ (primary build system)
+- C++17 compatible compiler (GCC 13.3+, Clang, MSVC)
+- Optional: ccache for faster compilation
+
+### Basic Build (CPU-only)
+**ALWAYS run these commands in sequence:**
+```bash
+cmake -B build
+cmake --build build --config Release -j $(nproc)
+```
+
+**Build time**: ~10 minutes on 4-core system with ccache enabled, ~25 minutes without ccache.
+
+**Important Notes:**
+- The Makefile is deprecated - always use CMake
+- ccache is automatically detected and used if available
+- Built binaries are placed in `build/bin/`
+- Parallel builds (`-j`) significantly reduce build time
+
+### Backend-Specific Builds
+For CUDA support:
+```bash
+cmake -B build -DGGML_CUDA=ON
+cmake --build build --config Release -j $(nproc)
+```
+
+For Metal (macOS):
+```bash
+cmake -B build -DGGML_METAL=ON
+cmake --build build --config Release -j $(nproc)
+```
+
+**Important Note**: While all backends can be built as long as the correct requirements for that backend are installed, you will not be able to run them without the correct hardware. The only backend that can be run for testing and validation is the CPU backend.
+
+### Debug Builds
+Single-config generators:
+```bash
+cmake -B build -DCMAKE_BUILD_TYPE=Debug
+cmake --build build
+```
+
+Multi-config generators:
+```bash
+cmake -B build -G "Xcode"
+cmake --build build --config Debug
+```
+
+### Common Build Issues
+- **Issue**: Network tests fail in isolated environments
+  **Solution**: Expected behavior - core functionality tests will still pass
+
+## Testing
+
+### Running Tests
+```bash
+ctest --test-dir build --output-on-failure -j $(nproc)
+```
+
+**Test suite**: 38 tests covering tokenizers, grammar parsing, sampling, backends, and integration
+**Expected failures**: 2-3 tests may fail if network access is unavailable (they download models)
+**Test time**: ~30 seconds for passing tests
+
+### Server Unit Tests
+Run server-specific unit tests after building the server:
+```bash
+# Build the server first
+cmake --build build --target llama-server
+
+# Navigate to server tests and run
+cd tools/server/tests
+source ../../../.venv/bin/activate
+./tests.sh
+```
+**Server test dependencies**: The `.venv` environment includes the required dependencies for server unit tests (pytest, aiohttp, etc.). Tests can be run individually or with various options as documented in `tools/server/tests/README.md`.
+
+### Test Categories
+- Tokenizer tests: Various model tokenizers (BERT, GPT-2, LLaMA, etc.)
+- Grammar tests: GBNF parsing and validation
+- Backend tests: Core ggml operations across different backends
+- Integration tests: End-to-end workflows
+
+### Manual Testing Commands
+```bash
+# Test basic inference
+./build/bin/llama-cli --version
+
+# Test model loading (requires model file)
+./build/bin/llama-cli -m path/to/model.gguf -p "Hello" -n 10
+```
+
+## Code Quality and Linting
+
+### C++ Code Formatting
+**ALWAYS format C++ code before committing:**
+```bash
+git clang-format
+```
+
+Configuration is in `.clang-format` with these key rules:
+- 4-space indentation
+- 120 column limit
+- Braces on same line for functions
+- Pointer alignment: `void * ptr` (middle)
+- Reference alignment: `int & ref` (middle)
+
+### Python Code
+**ALWAYS activate the Python environment in `.venv` and use tools from that environment:**
+```bash
+# Activate virtual environment
+source .venv/bin/activate
+```
+
+Configuration files:
+- `.flake8`: flake8 settings (max-line-length=125, excludes examples/tools)
+- `pyrightconfig.json`: pyright type checking configuration
+
+### Pre-commit Hooks
+Run before committing:
+```bash
+pre-commit run --all-files
+```
+
+## Continuous Integration
+
+### GitHub Actions Workflows
+Key workflows that run on every PR:
+- `.github/workflows/build.yml`: Multi-platform builds
+- `.github/workflows/server.yml`: Server functionality tests
+- `.github/workflows/python-lint.yml`: Python code quality
+- `.github/workflows/python-type-check.yml`: Python type checking
+
+### Local CI Validation
+**Run full CI locally before submitting PRs:**
+```bash
+mkdir tmp
+
+# CPU-only build
+bash ./ci/run.sh ./tmp/results ./tmp/mnt
+```
+
+**CI Runtime**: 30-60 minutes depending on backend configuration
+
+### Triggering CI
+Add `ggml-ci` to commit message to trigger heavy CI workloads on the custom CI infrastructure.
+
+## Project Layout and Architecture
+
+### Core Directories
+- **`src/`**: Main llama library implementation (`llama.cpp`, `llama-*.cpp`)
+- **`include/`**: Public API headers, primarily `include/llama.h`
+- **`ggml/`**: Core tensor library (submodule with custom GGML framework)
+- **`examples/`**: 30+ example applications and tools
+- **`tools/`**: Additional development and utility tools (server benchmarks, tests)
+- **`tests/`**: Comprehensive test suite with CTest integration
+- **`docs/`**: Detailed documentation (build guides, API docs, etc.)
+- **`scripts/`**: Utility scripts for CI, data processing, and automation
+- **`common/`**: Shared utility code used across examples
+
+### Key Files
+- **`CMakeLists.txt`**: Primary build configuration
+- **`include/llama.h`**: Main C API header (~2000 lines)
+- **`src/llama.cpp`**: Core library implementation (~8000 lines)
+- **`CONTRIBUTING.md`**: Coding guidelines and PR requirements
+- **`.clang-format`**: C++ formatting rules
+- **`.pre-commit-config.yaml`**: Git hook configuration
+
+### Built Executables (in `build/bin/`)
+Primary tools:
+- **`llama-cli`**: Main inference tool
+- **`llama-server`**: OpenAI-compatible HTTP server
+- **`llama-quantize`**: Model quantization utility
+- **`llama-perplexity`**: Model evaluation tool
+- **`llama-bench`**: Performance benchmarking
+- **`llama-convert-llama2c-to-ggml`**: Model conversion utilities
+
+### Configuration Files
+- **CMake**: `CMakeLists.txt`, `cmake/` directory
+- **Linting**: `.clang-format`, `.clang-tidy`, `.flake8`
+- **CI**: `.github/workflows/`, `ci/run.sh`
+- **Git**: `.gitignore` (includes build artifacts, models, cache)
+
+### Dependencies
+- **System**: OpenMP, libcurl (for model downloading)
+- **Optional**: CUDA SDK, Metal framework, Vulkan SDK, Intel oneAPI
+- **Bundled**: httplib, json (header-only libraries in vendored form)
+
+## Common Validation Steps
+
+### After Making Changes
+1. **Format code**: `git clang-format`
+2. **Build**: `cmake --build build --config Release`
+3. **Test**: `ctest --test-dir build --output-on-failure`
+4. **Server tests** (if modifying server): `cd tools/server/tests && source ../../../.venv/bin/activate && ./tests.sh`
+5. **Manual validation**: Test relevant tools in `build/bin/`
+
+### Performance Validation
+```bash
+# Benchmark inference performance
+./build/bin/llama-bench -m model.gguf
+
+# Evaluate model perplexity
+./build/bin/llama-perplexity -m model.gguf -f dataset.txt
+```
+
+### Backend Validation
+```bash
+# Test backend operations
+./build/bin/test-backend-ops
+```
+
+## Environment Setup
+
+### Required Tools
+- CMake 3.14+ (install via system package manager)
+- Modern C++ compiler with C++17 support
+- Git (for submodule management)
+- Python 3.9+ with virtual environment (`.venv` is provided)
+
+### Optional but Recommended
+- ccache: `apt install ccache` or `brew install ccache`
+- clang-format 15+: Usually included with LLVM/Clang installation
+- pre-commit: `pip install pre-commit`
+
+### Backend-Specific Requirements
+- **CUDA**: NVIDIA CUDA Toolkit 11.2+
+- **Metal**: Xcode command line tools (macOS only)
+- **Vulkan**: Vulkan SDK
+- **SYCL**: Intel oneAPI toolkit
+
+## Important Guidelines
+
+### Code Changes
+- **Minimal dependencies**: Avoid adding new external dependencies
+- **Cross-platform compatibility**: Test on Linux, macOS, Windows when possible
+- **Performance focus**: This is a performance-critical inference library
+- **API stability**: Changes to `include/llama.h` require careful consideration
+
+### Git Workflow
+- Always create feature branches from `master`
+- **Never** commit build artifacts (`build/`, `.ccache/`, `*.o`, `*.gguf`)
+- Use descriptive commit messages following project conventions
+
+### Trust These Instructions
+Only search for additional information if these instructions are incomplete or found to be incorrect. This document contains validated build and test procedures that work reliably across different environments.
+

.github/workflows/build-riscv-native.yml

@@ -1,10 +1,11 @@
 name: Build on RISCV Linux Machine by Cloud-V
 on:
+  pull_request:
   workflow_dispatch:
   workflow_call:
 
 jobs:
-  bianbu-riscv64-native: # Bianbu 2.2
+  debian-13-riscv64-native: # Bianbu 2.2
     runs-on: self-hosted
 
     steps:
@@ -20,24 +21,40 @@ jobs:
                   build-essential \
                   gcc-14-riscv64-linux-gnu \
                   g++-14-riscv64-linux-gnu \
+                  ccache \
                   cmake
 
+      - name: Setup ccache
+        run: |
+          mkdir -p $HOME/.ccache
+          ccache -M 5G -d $HOME/.ccache
+          export CCACHE_LOGFILE=/home/runneruser/ccache_debug/ccache.log
+          export CCACHE_DEBUGDIR="/home/runneruser/ccache_debug"
+          echo "$GITHUB_WORKSPACE"
+          echo "CCACHE_LOGFILE=$CCACHE_LOGFILE" >> $GITHUB_ENV
+          echo "CCACHE_DEBUGDIR=$CCACHE_DEBUGDIR" >> $GITHUB_ENV
+          echo "CCACHE_BASEDIR=$GITHUB_WORKSPACE" >> $GITHUB_ENV
+          echo "CCACHE_DIR=$HOME/.ccache" >> $GITHUB_ENV
+
       - name: Build
         run: |
-          cmake -B build -DLLAMA_CURL=OFF \
-                         -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
-                         -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-                         -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+          cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_OPENMP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=ON \
+            -DLLAMA_BUILD_TOOLS=ON \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DCMAKE_SYSTEM_NAME=Linux \
+            -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
+            -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+            -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
+            -DCMAKE_C_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+            -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
+            -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+            -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+            -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
           cmake --build build --config Release -j $(nproc)

.github/workflows/copilot-setup-steps.yml

@@ -39,6 +39,10 @@ jobs:
         run: |
           sudo apt-get update
           sudo apt-get install build-essential libcurl4-openssl-dev
+          # Install git-clang-format script for formatting only changed code
+          wget -O /tmp/git-clang-format https://raw.githubusercontent.com/llvm/llvm-project/release/18.x/clang/tools/clang-format/git-clang-format
+          sudo cp /tmp/git-clang-format /usr/local/bin/git-clang-format
+          sudo chmod +x /usr/local/bin/git-clang-format
 
       - name: Set up Python
         uses: actions/setup-python@v5
@@ -50,4 +54,4 @@ jobs:
           python3 -m venv .venv
           .venv/bin/activate
           pip install -r requirements/requirements-all.txt -r tools/server/tests/requirements.txt
-          pip install flake8 pyright
+          pip install flake8 pyright pre-commit

.gitignore

@@ -147,3 +147,4 @@ poetry.toml
 # Local scripts
 /run-vim.sh
 /run-chat.sh
+.ccache/

README.md

@@ -151,6 +151,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
 - [x] [GLM-EDGE](https://huggingface.co/models?search=glm-edge)
 - [x] [Qwen2-VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d)
+- [x] [LFM2-VL](https://huggingface.co/collections/LiquidAI/lfm2-vl-68963bbc84a610f7638d5ffa)
 
 </details>
 

ci/run.sh

@@ -106,7 +106,7 @@ function gg_wget {
     cd $out
 
     # should not re-download if file is the same
-    wget -nv -N $url
+    wget -nv -c -N $url
 
     cd $cwd
 }

common/arg.cpp

@@ -1106,7 +1106,7 @@ static void common_params_print_completion(common_params_context & ctx_arg) {
     printf("\"\n\n");
 
     printf("    case \"$prev\" in\n");
-    printf("        --model)\n");
+    printf("        --model|-m)\n");
     printf("            COMPREPLY=( $(compgen -f -X '!*.gguf' -- \"$cur\") $(compgen -d -- \"$cur\") )\n");
     printf("            return 0\n");
     printf("            ;;\n");
@@ -1530,6 +1530,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.ctx_shift = false;
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_PERPLEXITY}).set_env("LLAMA_ARG_NO_CONTEXT_SHIFT"));
+    add_opt(common_arg(
+        {"--context-shift"},
+        string_format("enables context shift on infinite text generation (default: %s)", params.ctx_shift ? "enabled" : "disabled"),
+        [](common_params & params) {
+            params.ctx_shift = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_PERPLEXITY}).set_env("LLAMA_ARG_CONTEXT_SHIFT"));
     add_opt(common_arg(
         {"--chunks"}, "N",
         string_format("max number of chunks to process (default: %d, -1 = all)", params.n_chunks),
@@ -1748,7 +1755,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.warmup = false;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_PERPLEXITY}));
     add_opt(common_arg(
         {"--spm-infill"},
         string_format(
@@ -1823,7 +1830,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.sampling.top_n_sigma = std::stof(value);
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN}).set_sparam());
+    ).set_sparam());
     add_opt(common_arg(
         {"--xtc-probability"}, "N",
         string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sampling.xtc_probability),
@@ -2247,9 +2254,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_PERPLEXITY}));
     add_opt(common_arg(
         {"-dt", "--defrag-thold"}, "N",
-        string_format("KV cache defragmentation threshold (default: %.1f, < 0 - disabled)", (double)params.defrag_thold),
+        string_format("KV cache defragmentation threshold (DEPRECATED)"),
         [](common_params & params, const std::string & value) {
-            params.defrag_thold = std::stof(value);
+            GGML_UNUSED(params);
+            GGML_UNUSED(value);
+            LOG_WRN("DEPRECATED: --defrag-thold is deprecated and no longer necessary to specify\n");
         }
     ).set_env("LLAMA_ARG_DEFRAG_THOLD"));
     add_opt(common_arg(
@@ -3529,6 +3538,22 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
 
+    add_opt(common_arg(
+        {"--fim-qwen-30b-default"},
+        string_format("use default Qwen 3 Coder 30B A3B Instruct (note: can download weights from the internet)"),
+        [](common_params & params) {
+            params.model.hf_repo = "ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF";
+            params.model.hf_file = "qwen3-coder-30b-a3b-instruct-q8_0.gguf";
+            params.port = 8012;
+            params.n_gpu_layers = 99;
+            params.flash_attn = true;
+            params.n_ubatch = 1024;
+            params.n_batch = 1024;
+            params.n_ctx = 0;
+            params.n_cache_reuse = 256;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+
     add_opt(common_arg(
         { "--diffusion-steps" }, "N",
         string_format("number of diffusion steps (default: %d)", params.diffusion.steps),

common/chat.cpp

@@ -147,6 +147,7 @@ struct templates_params {
     json extra_context;
     bool add_bos;
     bool add_eos;
+    bool is_inference = true;
 };
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) {
@@ -632,7 +633,6 @@ const char * common_reasoning_format_name(common_reasoning_format format) {
         case COMMON_REASONING_FORMAT_AUTO:     return "auto";
         case COMMON_REASONING_FORMAT_DEEPSEEK: return "deepseek";
         case COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY: return "deepseek-legacy";
-        case COMMON_REASONING_FORMAT_GRANITE: return "granite";
         default:
             throw std::runtime_error("Unknown reasoning format");
     }
@@ -1337,6 +1337,17 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
     common_chat_params data;
     auto prompt = apply(tmpl, inputs);
 
+    // Check if we need to replace the return token with end token during
+    // inference and without generation prompt. For more details see:
+    // https://github.com/ggml-org/llama.cpp/issues/15417
+    if (inputs.is_inference && !inputs.add_generation_prompt) {
+        static constexpr std::string_view return_token = "<|return|>";
+        static constexpr std::string_view end_token    = "<|end|>";
+        if (size_t pos = prompt.rfind(return_token); pos != std::string::npos) {
+            prompt.replace(pos, return_token.length(), end_token);
+        }
+    }
+
     data.prompt = prompt;
     data.format = COMMON_CHAT_FORMAT_GPT_OSS;
 
@@ -1350,6 +1361,26 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
         "<|end|>",
     };
 
+    if (!inputs.json_schema.is_null()) {
+        data.grammar_lazy = false;
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            auto schema = inputs.json_schema;
+            builder.resolve_refs(schema);
+
+            auto not_end = builder.add_rule("not-end",
+                "[^<] | \"<\" [^|] | \"<|\" [^e] | \"<|e\" [^n] | \"<|en\" [^d] | \"<|end\" [^|] | \"<|end|\" [^>]");
+            auto analysis = builder.add_rule("analysis",
+                "\"<|channel|>analysis<|message|>\" ( " + not_end + " )* \"<|end|>\"");
+            auto constraint = builder.add_rule("constraint", "\"<|constrain|>\"? [a-zA-Z0-9_-]+");
+            auto final = builder.add_rule("final",
+                "\"<|channel|>final\" ( \" \" " + constraint + " )? \"<|message|>\" " +
+                builder.add_schema("response", schema)
+            );
+
+            builder.add_rule("root", "( " + analysis + " \"<|start|>assistant\" )? " + final);
+        });
+    }
+
     if (inputs.tools.is_array() && !inputs.tools.empty()) {
         data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
         data.grammar = build_grammar([&](const common_grammar_builder & builder) {
@@ -2110,7 +2141,7 @@ static common_chat_params common_chat_templates_apply_jinja(
     }
 
     // GPT-OSS
-    if (src.find("<|channel|>") != std::string::npos && params.json_schema.is_null()) {
+    if (src.find("<|channel|>") != std::string::npos) {
         return common_chat_params_init_gpt_oss(tmpl, params);
     }
 

common/common.cpp

@@ -558,13 +558,6 @@ std::string string_from(const struct llama_context * ctx, const std::vector<llam
 
         auto detokenized = common_token_to_piece(ctx, token);
 
-        detokenized.erase(
-            std::remove_if(
-                detokenized.begin(),
-                detokenized.end(),
-                [](const unsigned char c) { return !std::isprint(c); }),
-            detokenized.end());
-
         buf << "'" << detokenized << "'"
             << ":" << std::to_string(token);
     }
@@ -589,13 +582,6 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
 
         auto detokenized = common_token_to_piece(ctx, batch.token[i]);
 
-        detokenized.erase(
-                std::remove_if(
-                    detokenized.begin(),
-                    detokenized.end(),
-                    [](const unsigned char c) { return !std::isprint(c); }),
-                detokenized.end());
-
         buf << "\n"          << std::to_string(i)
             << ", token '"   << detokenized << "'"
             << ", pos "      << std::to_string(batch.pos[i])
@@ -1166,7 +1152,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
     cparams.pooling_type      = params.pooling_type;
     cparams.attention_type    = params.attention_type;
-    cparams.defrag_thold      = params.defrag_thold;
     cparams.cb_eval           = params.cb_eval;
     cparams.cb_eval_user_data = params.cb_eval_user_data;
     cparams.offload_kqv       = !params.no_kv_offload;

common/common.h

@@ -239,12 +239,15 @@ struct common_params_diffusion {
     bool    add_gumbel_noise = false; // add gumbel noise to the logits if temp > 0.0
 };
 
+// reasoning API response format (not to be confused as chat template's reasoning format)
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
-    COMMON_REASONING_FORMAT_AUTO,
+    COMMON_REASONING_FORMAT_AUTO,            // Same as deepseek, using `message.reasoning_content`
     COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY, // Extract thinking tag contents and return as `message.reasoning_content`, or leave inline in <think> tags in stream mode
     COMMON_REASONING_FORMAT_DEEPSEEK,        // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
-    COMMON_REASONING_FORMAT_GRANITE,         // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
+    // do not extend this enum unless you absolutely have to
+    // in most cases, use COMMON_REASONING_FORMAT_AUTO
+    // see: https://github.com/ggml-org/llama.cpp/pull/15408
 };
 
 
@@ -285,7 +288,6 @@ struct common_params {
     float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
     float   yarn_beta_slow        =  1.0f; // YaRN high correction dim
     int32_t yarn_orig_ctx         =     0; // YaRN original context length
-    float   defrag_thold          =  0.1f; // KV cache defragmentation threshold
 
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
@@ -372,7 +374,7 @@ struct common_params {
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
     bool flash_attn        = false; // flash attention
     bool no_perf           = false; // disable performance metrics
-    bool ctx_shift         = true;  // context shift on inifinite text generation
+    bool ctx_shift         = false;  // context shift on infinite text generation
     bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
     bool kv_unified        = false; // enable unified KV cache
 

convert_hf_to_gguf.py

@@ -89,13 +89,16 @@ class ModelBase:
     block_count: int
     tensor_map: gguf.TensorNameMap
 
+    # Mistral format specifics
     is_mistral_format: bool = False
+    disable_mistral_community_chat_template: bool = False
 
     def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, *, is_big_endian: bool = False,
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
                  split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
-                 small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None):
+                 small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None,
+                 disable_mistral_community_chat_template: bool = False):
         if type(self) is ModelBase or \
                 type(self) is TextModel or \
                 type(self) is MmprojModel:
@@ -147,6 +150,9 @@ class ModelBase:
         self.gguf_writer = gguf.GGUFWriter(path=None, arch=gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file,
                                            split_max_tensors=split_max_tensors, split_max_size=split_max_size, dry_run=dry_run, small_first_shard=small_first_shard)
 
+        # Mistral specific
+        self.disable_mistral_community_chat_template = disable_mistral_community_chat_template
+
     @classmethod
     def add_prefix_to_filename(cls, path: Path, prefix: str) -> Path:
         stem, suffix = path.stem, path.suffix
@@ -1210,6 +1216,55 @@ class TextModel(ModelBase):
                 raise NotImplementedError("Only MEAN, CLS, and LAST pooling types supported")
             self.gguf_writer.add_pooling_type(pooling_type)
 
+    def _set_vocab_interns1(self):
+        tokens: list[str] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        vocab = getattr(tokenizer, 'vocab', tokenizer.get_vocab())
+        vocab_size = self.hparams.get("vocab_size", len(vocab))
+        assert max(vocab.values()) < vocab_size
+
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab.items()}
+        added_vocab = tokenizer.get_added_vocab()
+
+        added_tokens_decoder = tokenizer.added_tokens_decoder
+
+        for i in range(vocab_size):
+            if i not in reverse_vocab:
+                tokens.append(f"[PAD{i}]")
+                toktypes.append(gguf.TokenType.UNUSED)
+            else:
+                token: str = reverse_vocab[i]
+                if token in added_vocab:
+                    # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
+                    # To avoid unexpected issues - we make sure to normalize non-normalized tokens
+                    if not added_tokens_decoder[i].normalized:
+                        previous_token = token
+                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
+                        if previous_token != token:
+                            logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
+
+                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        toktypes.append(gguf.TokenType.USER_DEFINED)
+                else:
+                    toktypes.append(gguf.TokenType.NORMAL)
+                tokens.append(token)
+
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab._set_special_token("bos", 151643)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
 
 class MmprojModel(ModelBase):
     model_type = ModelType.MMPROJ
@@ -2011,8 +2066,17 @@ class LlamaModel(TextModel):
 
         template_dir = Path(__file__).parent / "models/templates/"
 
-        template = MistralModel.get_community_chat_template(vocab, template_dir)
-        self.gguf_writer.add_chat_template(template)
+        if not self.is_mistral_format or not self.disable_mistral_community_chat_template:
+            # Log only for Mistral format that the official tokenization and detokenization is via `mistral-common`.
+            if self.is_mistral_format:
+                logger.info(
+                    "Using a Mistral community chat template. These templates can be subject to errors in early days or weeks after a release. "
+                    "Mistral recommends to use `mistral-common` to perform tokenization and detokenization."
+                )
+            template = MistralModel.get_community_chat_template(vocab, template_dir, self.is_mistral_format)
+            self.gguf_writer.add_chat_template(template)
+        else:
+            logger.info("Not using a Mistral community chat template. Ensure to perform the tokenization and detokenization via `mistral-common`.")
 
     def set_vocab(self):
         if self.is_mistral_format:
@@ -2917,7 +2981,8 @@ class Qwen2Model(TextModel):
         if "language_model." in name:
             name = name.replace("language_model.", "") # for InternVL
         if name.startswith("mlp") or name.startswith("multi_modal_projector") \
-                or name.startswith("vision_model") or name.startswith("audio_tower"):
+                or name.startswith("vision_model") or name.startswith("audio_tower") \
+                or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector"):
             # skip vision and audio tensors
             return []
         yield from super().modify_tensors(data_torch, name, bid)
@@ -3094,7 +3159,7 @@ class LLaDAModel(TextModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
-@ModelBase.register("Ernie4_5_ForCausalLM")
+@ModelBase.register("Ernie4_5_ForCausalLM", "Ernie4_5ForCausalLM")
 class Ernie4_5Model(TextModel):
     model_arch = gguf.MODEL_ARCH.ERNIE4_5
 
@@ -3589,6 +3654,19 @@ class Qwen2MoeModel(TextModel):
 class Qwen3Model(Qwen2Model):
     model_arch = gguf.MODEL_ARCH.QWEN3
 
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        hparams = ModelBase.load_hparams(self.dir_model, is_mistral_format=False)
+        self.origin_hf_arch = hparams.get('architectures', [None])[0]
+
+    def set_vocab(self):
+        # deal with intern-s1-mini
+        if self.origin_hf_arch == 'InternS1ForConditionalGeneration':
+            self._set_vocab_interns1()
+            return
+
+        super().set_vocab()
+
 
 @ModelBase.register("Qwen3MoeForCausalLM")
 class Qwen3MoeModel(Qwen2MoeModel):
@@ -3605,73 +3683,7 @@ class Qwen3MoeModel(Qwen2MoeModel):
             self._set_vocab_interns1()
             return
 
-        try:
-            self._set_vocab_sentencepiece()
-        except FileNotFoundError:
-            self._set_vocab_gpt2()
-
-    def _set_vocab_interns1(self):
-        tokens: list[str] = []
-        toktypes: list[int] = []
-
-        from transformers import AutoTokenizer
-        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
-        vocab = getattr(tokenizer, 'vocab', tokenizer.get_vocab())
-        vocab_size = self.hparams.get("vocab_size", len(vocab))
-        assert max(vocab.values()) < vocab_size
-
-        tokpre = self.get_vocab_base_pre(tokenizer)
-
-        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab.items()}
-        added_vocab = tokenizer.get_added_vocab()
-
-        added_tokens_decoder = tokenizer.added_tokens_decoder
-
-        for i in range(vocab_size):
-            if i not in reverse_vocab:
-                tokens.append(f"[PAD{i}]")
-                toktypes.append(gguf.TokenType.UNUSED)
-            else:
-                token: str = reverse_vocab[i]
-                if token in added_vocab:
-                    # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
-                    # To avoid unexpected issues - we make sure to normalize non-normalized tokens
-                    if not added_tokens_decoder[i].normalized:
-                        previous_token = token
-                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
-                        if previous_token != token:
-                            logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
-
-                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
-                        toktypes.append(gguf.TokenType.CONTROL)
-                    else:
-                        toktypes.append(gguf.TokenType.USER_DEFINED)
-                else:
-                    toktypes.append(gguf.TokenType.NORMAL)
-                tokens.append(token)
-
-        self.gguf_writer.add_tokenizer_model("gpt2")
-        self.gguf_writer.add_tokenizer_pre(tokpre)
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_types(toktypes)
-
-        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
-        special_tokens_map_file = self.dir_model / 'special_tokens_map.json'
-        additional_special_tokens = []
-        if special_tokens_map_file.is_file():
-            with open(special_tokens_map_file, encoding = 'utf-8') as f:
-                additional_special_tokens = json.load(f).get('additional_special_tokens', [])
-        tokenizer_cfg_file = self.dir_model / 'special_tokens_map.json'
-        if tokenizer_cfg_file.is_file():
-            with open(tokenizer_cfg_file, encoding = 'utf-8') as f:
-                added_tokens_decoder = json.load(f).get('added_tokens_decoder', {})
-                token2ids_map = {data['content'] : int(token) for token, data in added_tokens_decoder.items() if data['special']}
-                for token in additional_special_tokens:
-                    if token in token2ids_map:
-                        special_vocab._set_special_token(token, token2ids_map[token])
-        special_vocab._set_special_token('eos', 151645)
-        special_vocab._set_special_token("bos", 151643)
-        special_vocab.add_to_gguf(self.gguf_writer)
+        super().set_vocab()
 
 
 @ModelBase.register("GPT2LMHeadModel")
@@ -5839,6 +5851,11 @@ class OlmoModel(TextModel):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@ModelBase.register("SeedOssForCausalLM")
+class SeedOssModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.SEED_OSS
+
+
 @ModelBase.register("Olmo2ForCausalLM")
 class Olmo2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.OLMO2
@@ -6237,9 +6254,11 @@ class DeepseekModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@ModelBase.register("DeepseekV2ForCausalLM")
-@ModelBase.register("DeepseekV3ForCausalLM")
-@ModelBase.register("KimiVLForConditionalGeneration")
+@ModelBase.register(
+    "DeepseekV2ForCausalLM",
+    "DeepseekV3ForCausalLM",
+    "KimiVLForConditionalGeneration",
+)
 class DeepseekV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
 
@@ -8422,7 +8441,7 @@ class MistralModel(LlamaModel):
     undo_permute = False
 
     @staticmethod
-    def get_community_chat_template(vocab: MistralVocab, templates_dir: Path):
+    def get_community_chat_template(vocab: MistralVocab, templates_dir: Path, is_mistral_format: bool):
         assert TokenizerVersion is not None, "mistral_common is not installed"
         assert isinstance(vocab.tokenizer, (Tekkenizer, SentencePieceTokenizer)), (
             f"Expected Tekkenizer or SentencePieceTokenizer, got {type(vocab.tokenizer)}"
@@ -8443,7 +8462,13 @@ class MistralModel(LlamaModel):
         elif vocab.tokenizer.version == TokenizerVersion.v13:
             template_file = "unsloth-mistral-Devstral-Small-2507.jinja"
         else:
-            raise ValueError(f"Unknown tokenizer type: {vocab.tokenizer_type} and version {vocab.tokenizer.version}")
+            err_message = f"Unknown tokenizer type: {vocab.tokenizer_type} and version {vocab.tokenizer.version}"
+            if is_mistral_format:
+                err_message += (
+                    " . Please pass --disable-mistral-community-chat-template argument to the CLI "
+                    "if you want to skip this error and use the Mistral official `mistral-common` pre-processing library."
+                )
+            raise ValueError(err_message)
 
         template_path = templates_dir / template_file
         if not template_path.exists():
@@ -8484,6 +8509,43 @@ class PixtralModel(LlavaVisionModel):
             return "mm.2.weight"
         return super().map_tensor_name(name, try_suffixes)
 
+
+@ModelBase.register("KimiVLForConditionalGeneration")
+class KimiVLModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        self.hparams_vision["image_size"] = 64 * 14 # for compatibility
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.KIMIVL)
+        self.gguf_writer.add_vision_use_gelu(True)
+        self.gguf_writer.add_vision_projector_scale_factor(2)
+        # eps is the same as pytorch's default value
+        assert self.hparams_vision is not None
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("layer_norm_eps", 1e-5))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        is_vision_tensor = "vision_tower" in name or "multi_modal_projector" in name
+
+        if is_vision_tensor:
+            if "pos_emb.weight" in name:
+                data_torch = data_torch.view(data_torch.shape[0] * data_torch.shape[1], data_torch.shape[2])
+            elif "wqkv" in name:
+                split_dim = 0 if "weight" in name else -1
+                wq, wk, wv = data_torch.chunk(3, dim=split_dim)
+                return [
+                    (self.map_tensor_name(name.replace("wqkv", "wq")), wq),
+                    (self.map_tensor_name(name.replace("wqkv", "wk")), wk),
+                    (self.map_tensor_name(name.replace("wqkv", "wv")), wv)
+                ]
+
+            return [(self.map_tensor_name(name), data_torch)]
+
+        return [] # skip other tensors
+
 ###### CONVERSION LOGIC ######
 
 
@@ -8638,6 +8700,13 @@ def parse_args() -> argparse.Namespace:
         "--mistral-format", action="store_true",
         help="Whether the model is stored following the Mistral format.",
     )
+    parser.add_argument(
+        "--disable-mistral-community-chat-template", action="store_true",
+        help=(
+            "Whether to disable usage of Mistral community chat templates. If set, use the Mistral official `mistral-common` library for tokenization and detokenization of Mistral models. "
+            "Using `mistral-common` ensure correctness and zero-day support of tokenization for models converted from the Mistral format but requires to manually setup the tokenization server."
+        )
+    )
 
     args = parser.parse_args()
     if not args.print_supported_models and args.model is None:
@@ -8744,6 +8813,7 @@ def main() -> None:
             fname_out = ModelBase.add_prefix_to_filename(fname_out, "mmproj-")
 
     is_mistral_format = args.mistral_format
+    disable_mistral_community_chat_template = args.disable_mistral_community_chat_template
 
     with torch.inference_mode():
         output_type = ftype_map[args.outtype]
@@ -8770,7 +8840,7 @@ def main() -> None:
                                      split_max_tensors=args.split_max_tensors,
                                      split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run,
                                      small_first_shard=args.no_tensor_first_split,
-                                     remote_hf_model_id=hf_repo_id,
+                                     remote_hf_model_id=hf_repo_id, disable_mistral_community_chat_template=disable_mistral_community_chat_template
                                      )
 
         if args.vocab_only:

docs/build-s390x.md

@@ -265,8 +265,9 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 | BF16       | 🚫          | 🚫   | ❓   | ❓    |
 | Q4_0       | ✅          | ✅   | ❓   | ❓    |
 | Q4_1       | ✅          | ✅   | ❓   | ❓    |
-| Q5_0       | 🚫          | 🚫   | ❓   | ❓    |
-| Q5_1       | 🚫          | 🚫   | ❓   | ❓    |
+| MXFP4      | 🚫          | 🚫   | ❓   | ❓    |
+| Q5_0       | ✅          | ✅   | ❓   | ❓    |
+| Q5_1       | ✅          | ✅   | ❓   | ❓    |
 | Q8_0       | ✅          | ✅   | ❓   | ❓    |
 | Q2_K       | 🚫          | 🚫   | ❓   | ❓    |
 | Q3_K       | ✅          | ✅   | ❓   | ❓    |
@@ -291,4 +292,4 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 -   🚫 - acceleration unavailable, will still run using scalar implementation
 -   ❓ - acceleration unknown, please contribute if you can test it yourself
 
-Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on July 31, 2025.
+Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Aug 22, 2025.

docs/build.md

@@ -197,13 +197,12 @@ The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enab
 
 The following compilation options are also available to tweak performance:
 
-| Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
-|-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
-| GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models                                                                                                                                                                                       |
-| GGML_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
-| GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
-| GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                  |
+| Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                                                                                                                      |
+|-------------------------------|------------------------|---------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower. |
+| GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models. There may be issues with numerical overflows (except for CDNA and RDNA4) and memory use will be higher. Prompt processing may become faster on recent datacenter GPUs (the custom kernels were tuned primarily for RTX 3000/4000).                                            |
+| GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                                                                                                                  |
+| GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                                                                                                           |
 
 ## MUSA
 

docs/multimodal/MobileVLM.md

@@ -194,7 +194,7 @@ llama_print_timings:       total time =   44411.01 ms /   377 tokens
 ## Orin compile and run
 ### compile
 ```sh
-make GGML_CUDA=1 CUDA_DOCKER_ARCH=sm_87 GGML_CUDA_F16=1 -j 32
+make GGML_CUDA=1 CUDA_DOCKER_ARCH=sm_87 -j 32
 ```
 ### run on Orin
 ### case 1

docs/multimodal/minicpmv4.0.md

@@ -6,7 +6,7 @@ Download [MiniCPM-V-4](https://huggingface.co/openbmb/MiniCPM-V-4) PyTorch model
 
 
 ### Build llama.cpp
-Readme modification time: 20250206
+Readme modification time: 20250731
 
 If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
 

docs/multimodal/minicpmv4.5.md

@@ -0,0 +1,47 @@
+## MiniCPM-V 4.5
+
+### Prepare models and code
+
+Download [MiniCPM-V-4_5](https://huggingface.co/openbmb/MiniCPM-V-4_5) PyTorch model from huggingface to "MiniCPM-V-4_5" folder.
+
+
+### Build llama.cpp
+Readme modification time: 20250826
+
+If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+
+Clone llama.cpp:
+```bash
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
+```
+
+Build llama.cpp using `CMake`:
+```bash
+cmake -B build
+cmake --build build --config Release
+```
+
+
+### Usage of MiniCPM-V 4
+
+Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-4_5-gguf) by us)
+
+```bash
+python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-V-4_5
+python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-4_5 --minicpmv-projector ../MiniCPM-V-4_5/minicpmv.projector --output-dir ../MiniCPM-V-4_5/ --minicpmv_version 6
+python ./convert_hf_to_gguf.py ../MiniCPM-V-4_5/model
+
+# quantize int4 version
+./build/bin/llama-quantize ../MiniCPM-V-4_5/model/ggml-model-f16.gguf ../MiniCPM-V-4_5/model/ggml-model-Q4_K_M.gguf Q4_K_M
+```
+
+
+Inference on Linux or Mac
+```bash
+# run in single-turn mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-V-4_5/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-4_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+
+# run in conversation mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-V-4_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-4_5/mmproj-model-f16.gguf
+```

examples/CMakeLists.txt

@@ -34,6 +34,7 @@ else()
     add_subdirectory(gen-docs)
     add_subdirectory(training)
     add_subdirectory(diffusion)
+    add_subdirectory(model-conversion)
     if (NOT GGML_BACKEND_DL)
         add_subdirectory(convert-llama2c-to-ggml)
         # these examples use the backends directly and cannot be built with dynamic loading

examples/batched.swift/README.md

@@ -1,4 +1,5 @@
 This is a swift clone of `examples/batched`.
 
-$ `make`
-$ `./llama-batched-swift MODEL_PATH [PROMPT] [PARALLEL]`
+```bash
+$ ./llama-batched-swift MODEL_PATH [PROMPT] [PARALLEL]
+```

examples/llama.vim

@@ -17,7 +17,7 @@
 "
 " start the llama.cpp server with a FIM-compatible model. for example:
 "
-"   $ llama-server -m {model.gguf} --port 8012 -ngl 99 -fa -dt 0.1 --ubatch-size 512 --batch-size 1024 --cache-reuse 256
+"   $ llama-server -m {model.gguf} --port 8012 -ngl 99 -fa --ubatch-size 512 --batch-size 1024 --cache-reuse 256
 "
 "   --batch-size [512, model max context]
 "

examples/lookahead/README.md

@@ -5,3 +5,9 @@ Demonstration of lookahead decoding technique:
 https://lmsys.org/blog/2023-11-21-lookahead-decoding/
 
 More info: https://github.com/ggml-org/llama.cpp/pull/4207
+
+Sample command:
+
+```bash
+llama-lookahead -hf ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF -p "// network server implemented in C\n// author: Peter Hacker\n\n#include" -e -ngl 99 -t 4 -n 512 -c 4096 -kvu
+```

examples/model-conversion/.gitignore

@@ -0,0 +1,3 @@
+.model_name
+data
+ppl

examples/model-conversion/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-logits)
+add_executable(${TARGET} logits.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/model-conversion/Makefile

@@ -0,0 +1,206 @@
+MAKEFLAGS += --no-print-directory
+
+define validate_model_path
+	@if [ -z "$(MODEL_PATH)" ]; then \
+		echo "Error: MODEL_PATH must be provided either as:"; \
+		echo "  1. Environment variable: export MODEL_PATH=/path/to/model"; \
+		echo "  2. Command line argument: make $(1) MODEL_PATH=/path/to/model"; \
+		exit 1; \
+	fi
+endef
+
+define validate_embedding_model_path
+	@if [ -z "$(EMBEDDING_MODEL_PATH)" ]; then \
+		echo "Error: EMBEDDING_MODEL_PATH must be provided either as:"; \
+		echo "  1. Environment variable: export EMBEDDING_MODEL_PATH=/path/to/model"; \
+		echo "  2. Command line argument: make $(1) EMBEDDING_MODEL_PATH=/path/to/model"; \
+		exit 1; \
+	fi
+endef
+
+define quantize_model
+	@CONVERTED_MODEL="$(1)" QUANTIZED_TYPE="$(QUANTIZED_TYPE)" \
+	TOKEN_EMBD_TYPE="$(TOKEN_EMBD_TYPE)" OUTPUT_TYPE="$(OUTPUT_TYPE)" \
+	./scripts/utils/quantize.sh "$(1)" "$(QUANTIZED_TYPE)" "$(TOKEN_EMBD_TYPE)" "$(OUTPUT_TYPE)"
+	@echo "Export the quantized model path to $(2) variable in your environment"
+endef
+
+###
+### Casual Model targets/recipes
+###
+causal-convert-model-bf16: OUTTYPE=bf16
+causal-convert-model-bf16: causal-convert-model
+
+causal-convert-model:
+	$(call validate_model_path,causal-convert-model)
+	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(OUTTYPE)" MODEL_PATH="$(MODEL_PATH)" \
+	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
+	./scripts/causal/convert-model.sh
+
+causal-convert-mm-model-bf16: OUTTYPE=bf16
+causal-convert-mm-model-bf16: MM_OUTTYPE=f16
+causal-convert-mm-model-bf16: causal-convert-mm-model
+
+causal-convert-mm-model:
+	$(call validate_model_path,causal-convert-mm-model)
+	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(OUTTYPE)" MODEL_PATH="$(MODEL_PATH)" \
+	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
+	./scripts/causal/convert-model.sh
+
+	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(MM_OUTTYPE)" MODEL_PATH="$(MODEL_PATH)" \
+	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
+	./scripts/causal/convert-model.sh --mmproj
+
+causal-run-original-model:
+	$(call validate_model_path,causal-run-original-model)
+	@MODEL_PATH="$(MODEL_PATH)" ./scripts/causal/run-org-model.py
+
+causal-run-converted-model:
+	@CONVERTED_MODEL="$(CONVERTED_MODEL)" ./scripts/causal/run-converted-model.sh
+
+causal-verify-logits: causal-run-original-model causal-run-converted-model
+	@./scripts/causal/compare-logits.py
+	@MODEL_PATH="$(MODEL_PATH)" ./scripts/utils/check-nmse.py -m ${MODEL_PATH}
+
+causal-run-original-embeddings:
+	@./scripts/causal/run-casual-gen-embeddings-org.sh
+
+causal-run-converted-embeddings:
+	@./scripts/causal/run-converted-model-embeddings-logits.sh
+
+causal-verify-embeddings: causal-run-original-embeddings causal-run-converted-embeddings
+	@./scripts/causal/compare-embeddings-logits.sh
+
+causal-inspect-original-model:
+	@./scripts/utils/inspect-org-model.py
+
+causal-inspect-converted-model:
+	@./scripts/utils/inspect-converted-model.sh
+
+causal-start-embedding-server:
+	@./scripts/utils/run-embedding-server.sh ${CONVERTED_MODEL}
+
+causal-curl-embedding-endpoint: causal-run-original-embeddings
+	@./scripts/utils/curl-embedding-server.sh | ./scripts/causal/compare-embeddings-logits.sh
+
+causal-quantize-Q8_0: QUANTIZED_TYPE = Q8_0
+causal-quantize-Q8_0: causal-quantize-model
+
+causal-quantize-Q4_0: QUANTIZED_TYPE = Q4_0
+causal-quantize-Q4_0: causal-quantize-model
+
+# For Quantization Aware Trained (QAT) models in Q4_0 we explicitly set the
+# token embedding and output types to Q8_0 instead of the default Q6_K.
+causal-quantize-qat-Q4_0: QUANTIZED_TYPE = Q4_0
+causal-quantize-qat-Q4_0: TOKEN_EMBD_TYPE = Q8_0
+causal-quantize-qat-Q4_0: OUTPUT_TYPE = Q8_0
+causal-quantize-qat-Q4_0: causal-quantize-model
+
+causal-quantize-model:
+	$(call quantize_model,$(CONVERTED_MODEL),QUANTIZED_MODEL)
+
+causal-run-quantized-model:
+	@QUANTIZED_MODEL="$(QUANTIZED_MODEL)" ./scripts/causal/run-converted-model.sh ${QUANTIZED_MODEL}
+
+
+###
+### Embedding Model targets/recipes
+###
+
+embedding-convert-model-bf16: OUTTYPE=bf16
+embedding-convert-model-bf16: embedding-convert-model
+
+embedding-convert-model:
+	$(call validate_embedding_model_path,embedding-convert-model)
+	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(OUTTYPE)" MODEL_PATH="$(EMBEDDING_MODEL_PATH)" \
+	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
+	./scripts/embedding/convert-model.sh
+
+embedding-run-original-model:
+	$(call validate_embedding_model_path,embedding-run-original-model)
+	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/embedding/run-original-model.py
+
+embedding-run-converted-model:
+	@CONVERTED_EMBEDDING_MODEL="$(CONVERTED_EMBEDDING_MODEL)" ./scripts/embedding/run-converted-model.sh ${CONVERTED_EMBEDDING_MODEL}
+
+embedding-verify-logits: embedding-run-original-model embedding-run-converted-model
+	@./scripts/embedding/compare-embeddings-logits.sh
+
+embedding-inspect-original-model:
+	$(call validate_embedding_model_path,embedding-inspect-original-model)
+	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/utils/inspect-org-model.py -m ${EMBEDDING_MODEL_PATH}
+
+embedding-inspect-converted-model:
+	@CONVERTED_EMBEDDING_MODEL="$(CONVERTED_EMBEDDING_MODEL)" ./scripts/utils/inspect-converted-model.sh ${CONVERTED_EMBEDDING_MODEL}
+
+embedding-start-embedding-server:
+	@./scripts/utils/run-embedding-server.sh ${CONVERTED_EMBEDDING_MODEL}
+
+embedding-curl-embedding-endpoint:
+	@./scripts/utils/curl-embedding-server.sh | ./scripts/embedding/compare-embeddings-logits.sh
+
+embedding-quantize-Q8_0: QUANTIZED_TYPE = Q8_0
+embedding-quantize-Q8_0: embedding-quantize-model
+
+embedding-quantize-Q4_0: QUANTIZED_TYPE = Q4_0
+embedding-quantize-Q4_0: embedding-quantize-model
+
+# For Quantization Aware Trained (QAT) models in Q4_0 we explicitly set the
+# token embedding and output types to Q8_0 instead of the default Q6_K.
+embedding-quantize-qat-Q4_0: QUANTIZED_TYPE = Q4_0
+embedding-quantize-qat-Q4_0: TOKEN_EMBD_TYPE = Q8_0
+embedding-quantize-qat-Q4_0: OUTPUT_TYPE = Q8_0
+embedding-quantize-qat-Q4_0: embedding-quantize-model
+
+embedding-quantize-model:
+	$(call quantize_model,$(CONVERTED_EMBEDDING_MODEL),QUANTIZED_EMBEDDING_MODEL)
+
+embedding-run-quantized-model:
+	@./scripts/embedding/run-converted-model.sh ${QUANTIZED_EMBEDDING_MODEL}
+
+###
+### Perplexity targets/recipes
+###
+perplexity-data-gen:
+	CONVERTED_MODEL="$(CONVERTED_MODEL)" ./scripts/utils/perplexity-gen.sh
+
+perplexity-run-full:
+	QUANTIZED_MODEL="$(QUANTIZED_MODEL)" LOOGITS_FILE="$(LOGITS_FILE)" \
+	./scripts/utils/perplexity-run.sh
+
+perplexity-run:
+	QUANTIZED_MODEL="$(QUANTIZED_MODEL)" ./scripts/utils/perplexity-run-simple.sh
+
+###
+### HuggingFace targets/recipes
+###
+
+hf-create-model:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}"
+
+hf-create-model-dry-run:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}" -d
+
+hf-create-model-embedding:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}" -e
+
+hf-create-model-embedding-dry-run:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}" -e -d
+
+hf-create-model-private:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}" -p
+
+hf-upload-gguf-to-model:
+	@./scripts/utils/hf-upload-gguf-model.py -m "${MODEL_PATH}" -r "${REPO_ID}" -o "${NAME_IN_REPO}"
+
+hf-create-collection:
+	@./scripts/utils/hf-create-collection.py -n "${NAME}" -d "${DESCRIPTION}" -ns "${NAMESPACE}"
+
+hf-add-model-to-collection:
+	@./scripts/utils/hf-add-model-to-collection.py -c "${COLLECTION}" -m "${MODEL}"
+
+
+.PHONY: clean
+clean:
+	@${RM} -rf data .converted_embedding_model.txt .converted_model.txt .embedding_model_name.txt .model_name.txt
+

examples/model-conversion/README.md

@@ -0,0 +1,367 @@
+# Model Conversion Example
+This directory contains scripts and code to help in the process of converting
+HuggingFace PyTorch models to GGUF format.
+
+The motivation for having this is that the conversion process can often be an
+iterative process, where the original model is inspected, converted, updates
+made to llama.cpp, converted again, etc. Once the model has been converted it
+needs to be verified against the original model, and then optionally quantified,
+and in some cases perplexity checked of the quantized model. And finally the
+model/models need to the ggml-org on Hugging Face. This tool/example tries to
+help with this process.
+
+### Overview
+The idea is that the makefile targets and scripts here can be used in the
+development/conversion process assisting with things like:
+
+* inspect/run the original model to figure out how it works
+* convert the original model to GGUF format
+* inspect/run the converted model
+* verify the logits produced by the original model and the converted model
+* quantize the model to GGUF format
+* run perplexity evaluation to verify that the quantized model is performing
+  as expected
+* upload the model to HuggingFace to make it available for others
+
+## Setup
+Create virtual python environment
+```console
+$ python3.11 -m venv venv
+$ source venv/bin/activate
+(venv) $ pip install -r requirements.txt
+```
+
+## Causal Language Model Conversion
+This section describes the steps to convert a causal language model to GGUF and
+to verify that the conversion was successful.
+
+### Download the original model
+First, clone the original model to some local directory:
+```console
+$ mkdir models && cd models
+$ git clone https://huggingface.co/user/model_name
+$ cd model_name
+$ git lfs install
+$ git lfs pull
+```
+
+### Set the MODEL_PATH
+The path to the downloaded model can be provided in two ways:
+
+**Option 1: Environment variable (recommended for iterative development)**
+```console
+export MODEL_PATH=~/work/ai/models/some_model
+```
+
+**Option 2: Command line argument (for one-off tasks)**
+```console
+make causal-convert-model MODEL_PATH=~/work/ai/models/some_model
+```
+
+Command line arguments take precedence over environment variables when both are provided.
+
+In cases where the transformer implementation for the model has not been released
+yet it is possible to set the environment variable `UNRELEASED_MODEL_NAME` which
+will then cause the transformer implementation to be loaded explicitely and not
+use AutoModelForCausalLM:
+```
+export UNRELEASED_MODEL_NAME=SomeNewModel
+```
+
+### Inspecting the original tensors
+```console
+# Using environment variable
+(venv) $ make causal-inspect-original-model
+
+# Or using command line argument
+(venv) $ make causal-inspect-original-model MODEL_PATH=~/work/ai/models/some_model
+```
+
+### Running the original model
+This is mainly to verify that the original model works, and to compare the output
+from the converted model.
+```console
+# Using environment variable
+(venv) $ make causal-run-original-model
+
+# Or using command line argument
+(venv) $ make causal-run-original-model MODEL_PATH=~/work/ai/models/some_model
+```
+This command will save two files to the `data` directory, one is a binary file
+containing logits which will be used for comparison with the converted model
+later, and the other is a text file which allows for manual visual inspection.
+
+### Model conversion
+After updates have been made to [gguf-py](../../gguf-py) to add support for the
+new model, the model can be converted to GGUF format using the following command:
+```console
+# Using environment variable
+(venv) $ make causal-convert-model
+
+# Or using command line argument
+(venv) $ make causal-convert-model MODEL_PATH=~/work/ai/models/some_model
+```
+
+### Inspecting the converted model
+The converted model can be inspected using the following command:
+```console
+(venv) $ make inspect-converted-model
+```
+
+### Running the converted model
+```console
+(venv) $ make run-converted-model
+```
+
+### Model logits verfication
+The following target will run the original model and the converted model and
+compare the logits:
+```console
+(venv) $ make causal-verify-logits
+```
+
+### Quantizing the model
+The causal model can be quantized to GGUF format using the following command:
+```console
+(venv) $ make causal-quantize-Q8_0
+Quantized model saved to: /path/to/quantized/model-Q8_0.gguf
+Export the quantized model path to QUANTIZED_MODEL variable in your environment
+```
+This will show the path to the quantized model in the terminal, which can then
+be used to set the `QUANTIZED_MODEL` environment variable:
+```console
+export QUANTIZED_MODEL=/path/to/quantized/model-Q8_0.gguf
+```
+Then the quantized model can be run using the following command:
+```console
+(venv) $ make causal-run-quantized-model
+```
+
+### Quantizing QAT (Quantization Aware Training) models
+When quantizing to `Q4_0`, the default data type for the token embedding weights
+will be `Q6_K`. For models that are going to be uploaded to ggml-org it is
+recommended to use `Q8_0` instead for the embeddings and output tensors.
+The reason is that although `Q6_K` is smaller in size, it requires more compute
+to unpack, which can hurt performance during output generation when the entire
+embedding matrix must be dequantized to compute vocabulary logits. `Q8_0`
+provides practically full quality with better computational efficiency.
+```console
+(venv) $ make causal-quantize-qat-Q4_0
+```
+
+
+## Embedding Language Model Conversion
+
+### Download the original model
+```console
+$ mkdir models && cd models
+$ git clone https://huggingface.co/user/model_name
+$ cd model_name
+$ git lfs install
+$ git lfs pull
+```
+
+The path to the embedding model can be provided in two ways:
+
+**Option 1: Environment variable (recommended for iterative development)**
+```console
+export EMBEDDING_MODEL_PATH=~/path/to/embedding_model
+```
+
+**Option 2: Command line argument (for one-off tasks)**
+```console
+make embedding-convert-model EMBEDDING_MODEL_PATH=~/path/to/embedding_model
+```
+
+Command line arguments take precedence over environment variables when both are provided.
+
+### Running the original model
+This is mainly to verify that the original model works and to compare the output
+with the output from the converted model.
+```console
+# Using environment variable
+(venv) $ make embedding-run-original-model
+
+# Or using command line argument
+(venv) $ make embedding-run-original-model EMBEDDING_MODEL_PATH=~/path/to/embedding_model
+```
+This command will save two files to the `data` directory, one is a binary
+file containing logits which will be used for comparison with the converted
+model, and the other is a text file which allows for manual visual inspection.
+
+### Model conversion
+After updates have been made to [gguf-py](../../gguf-py) to add support for the
+new model the model can be converted to GGUF format using the following command:
+```console
+(venv) $ make embedding-convert-model
+```
+
+### Run the converted model
+```console
+(venv) $ make embedding-run-converted-model
+```
+
+### Model logits verfication
+The following target will run the original model and the converted model (which
+was done manually in the previous steps) and compare the logits:
+```console
+(venv) $ make embedding-verify-logits
+```
+
+### llama-server verification
+To verify that the converted model works with llama-server, the following
+command can be used:
+```console
+(venv) $ make embedding-start-embedding-server
+```
+Then open another terminal and set the `EMBEDDINGS_MODEL_PATH` environment
+variable as this will not be inherited by the new terminal:
+```console
+(venv) $ make embedding-curl-embedding-endpoint
+```
+This will call the `embedding` endpoing and the output will be piped into
+the same verification script as used by the target `embedding-verify-logits`.
+
+The causal model can also be used to produce embeddings and this can be verified
+using the following commands:
+```console
+(venv) $ make causal-start-embedding-server
+```
+Then open another terminal and set the `MODEL_PATH` environment
+variable as this will not be inherited by the new terminal:
+```console
+(venv) $ make casual-curl-embedding-endpoint
+```
+
+### Quantizing the model
+The embedding model can be quantized to GGUF format using the following command:
+```console
+(venv) $ make embedding-quantize-Q8_0
+Quantized model saved to: /path/to/quantized/model-Q8_0.gguf
+Export the quantized model path to QUANTIZED_EMBEDDING_MODEL variable in your environment
+```
+This will show the path to the quantized model in the terminal, which can then
+be used to set the `QUANTIZED_EMBEDDING_MODEL` environment variable:
+```console
+export QUANTIZED_EMBEDDING_MODEL=/path/to/quantized/model-Q8_0.gguf
+```
+Then the quantized model can be run using the following command:
+```console
+(venv) $ make embedding-run-quantized-model
+```
+
+### Quantizing QAT (Quantization Aware Training) models
+When quantizing to `Q4_0`, the default data type for the token embedding weights
+will be `Q6_K`. For models that are going to be uploaded to ggml-org it is
+recommended to use `Q8_0` instead for the embeddings and output tensors.
+The reason is that although `Q6_K` is smaller in size, it requires more compute
+to unpack, which can hurt performance during output generation when the entire
+embedding matrix must be dequantized to compute vocabulary logits. `Q8_0`
+provides practically full quality with better computational efficiency.
+```console
+(venv) $ make embedding-quantize-qat-Q4_0
+```
+
+## Perplexity Evaluation
+
+### Simple perplexity evaluation
+This allows to run the perplexity evaluation without having to generate a
+token/logits file:
+```console
+(venv) $ make perplexity-run QUANTIZED_MODEL=~/path/to/quantized/model.gguf
+```
+This will use the wikitext dataset to run the perplexity evaluation and
+output the perplexity score to the terminal. This value can then be compared
+with the perplexity score of the unquantized model.
+
+### Full perplexity evaluation
+First use the converted, non-quantized, model to generate the perplexity evaluation
+dataset using the following command:
+```console
+$ make perplexity-data-gen CONVERTED_MODEL=~/path/to/converted/model.gguf
+```
+This will generate a file in the `data` directory named after the model and with
+a `.kld` suffix which contains the tokens and the logits for the wikitext dataset.
+
+After the dataset has been generated, the perplexity evaluation can be run using
+the quantized model:
+```console
+$ make perplexity-run-full QUANTIZED_MODEL=~/path/to/quantized/model-Qxx.gguf LOGITS_FILE=data/model.gguf.ppl
+```
+
+> 📝 **Note:** The `LOGITS_FILE` is the file generated by the previous command
+> can be very large, so make sure you have enough disk space available.
+
+## HuggingFace utilities
+The following targets are useful for creating collections and model repositories
+on Hugging Face in the the ggml-org. These can be used when preparing a relase
+to script the process for new model releases.
+
+For the following targets a `HF_TOKEN` environment variable is required.
+
+> 📝 **Note:** Don't forget to logout from Hugging Face after running these
+> commands, otherwise you might have issues pulling/cloning repositories as
+> the token will still be in use:
+> $ huggingface-cli logout
+> $ unset HF_TOKEN
+
+### Create a new Hugging Face Model (model repository)
+This will create a new model repsository on Hugging Face with the specified
+model name.
+```console
+(venv) $ make hf-create-model MODEL_NAME='TestModel' NAMESPACE="danbev" ORIGINAL_BASE_MODEL="some-base-model"
+Repository ID:  danbev/TestModel-GGUF
+Repository created: https://huggingface.co/danbev/TestModel-GGUF
+```
+Note that we append a `-GGUF` suffix to the model name to ensure a consistent
+naming convention for GGUF models.
+
+An embedding model can be created using the following command:
+```console
+(venv) $ make hf-create-model-embedding MODEL_NAME='TestEmbeddingModel' NAMESPACE="danbev" ORIGINAL_BASE_MODEL="some-base-model"
+```
+The only difference is that the model card for an embedding model will be different
+with regards to the llama-server command and also how to access/call the embedding
+endpoint.
+
+### Upload a GGUF model to model repository
+The following target uploads a model to an existing Hugging Face model repository.
+```console
+(venv) $ make hf-upload-gguf-to-model MODEL_PATH=dummy-model1.gguf REPO_ID=danbev/TestModel-GGUF
+📤 Uploading dummy-model1.gguf to danbev/TestModel-GGUF/dummy-model1.gguf
+✅ Upload successful!
+🔗 File available at: https://huggingface.co/danbev/TestModel-GGUF/blob/main/dummy-model1.gguf
+```
+This command can also be used to update an existing model file in a repository.
+
+### Create a new Collection
+```console
+(venv) $ make hf-new-collection NAME=TestCollection DESCRIPTION="Collection for testing scripts" NAMESPACE=danbev
+🚀 Creating Hugging Face Collection
+Title: TestCollection
+Description: Collection for testing scripts
+Namespace: danbev
+Private: False
+✅ Authenticated as: danbev
+📚 Creating collection: 'TestCollection'...
+✅ Collection created successfully!
+📋 Collection slug: danbev/testcollection-68930fcf73eb3fc200b9956d
+🔗 Collection URL: https://huggingface.co/collections/danbev/testcollection-68930fcf73eb3fc200b9956d
+
+🎉 Collection created successfully!
+Use this slug to add models: danbev/testcollection-68930fcf73eb3fc200b9956d
+```
+
+### Add model to a Collection
+```console
+(venv) $ make hf-add-model-to-collection COLLECTION=danbev/testcollection-68930fcf73eb3fc200b9956d MODEL=danbev/TestModel-GGUF
+✅ Authenticated as: danbev
+🔍 Checking if model exists: danbev/TestModel-GGUF
+✅ Model found: danbev/TestModel-GGUF
+📚 Adding model to collection...
+✅ Model added to collection successfully!
+🔗 Collection URL: https://huggingface.co/collections/danbev/testcollection-68930fcf73eb3fc200b9956d
+
+🎉 Model added successfully!
+
+```

examples/model-conversion/logits.cpp

@@ -0,0 +1,210 @@
+#include "llama.h"
+#include <cstdio>
+#include <cstring>
+#include <string>
+#include <vector>
+#include <ctype.h>
+#include <filesystem>
+
+static void print_usage(int, char ** argv) {
+    printf("\nexample usage:\n");
+    printf("\n    %s -m model.gguf [-ngl n_gpu_layers] -embd-mode [prompt]\n", argv[0]);
+    printf("\n");
+}
+
+int main(int argc, char ** argv) {
+    std::string model_path;
+    std::string prompt = "Hello, my name is";
+    int ngl = 0;
+    bool embedding_mode = false;
+
+    {
+        int i = 1;
+        for (; i < argc; i++) {
+            if (strcmp(argv[i], "-m") == 0) {
+                if (i + 1 < argc) {
+                    model_path = argv[++i];
+                } else {
+                    print_usage(argc, argv);
+                    return 1;
+                }
+            } else if (strcmp(argv[i], "-ngl") == 0) {
+                if (i + 1 < argc) {
+                    try {
+                        ngl = std::stoi(argv[++i]);
+                    } catch (...) {
+                        print_usage(argc, argv);
+                        return 1;
+                    }
+                } else {
+                    print_usage(argc, argv);
+                    return 1;
+                }
+            } else if (strcmp(argv[i], "-embd-mode") == 0) {
+                if (i + 1 < argc) {
+                    try {
+                        embedding_mode = true;
+                    } catch (...) {
+                        print_usage(argc, argv);
+                        return 1;
+                    }
+                } else {
+                    print_usage(argc, argv);
+                    return 1;
+                }
+            } else {
+                // prompt starts here
+                break;
+            }
+        }
+
+        if (model_path.empty()) {
+            print_usage(argc, argv);
+            return 1;
+        }
+
+        if (i < argc) {
+            prompt = argv[i++];
+            for (; i < argc; i++) {
+                prompt += " ";
+                prompt += argv[i];
+            }
+        }
+    }
+
+    ggml_backend_load_all();
+    llama_model_params model_params = llama_model_default_params();
+    model_params.n_gpu_layers = ngl;
+
+    llama_model * model = llama_model_load_from_file(model_path.c_str(), model_params);
+
+    if (model == NULL) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return 1;
+    }
+
+    // Extract basename from model_path
+    const char * basename = strrchr(model_path.c_str(), '/');
+    basename = (basename == NULL) ? model_path.c_str() : basename + 1;
+
+    char model_name[256];
+    strncpy(model_name, basename, 255);
+    model_name[255] = '\0';
+
+    char * dot = strrchr(model_name, '.');
+    if (dot != NULL && strcmp(dot, ".gguf") == 0) {
+        *dot = '\0';
+    }
+    printf("Model name: %s\n", model_name);
+
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+    const int n_prompt = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, true, true);
+
+    std::vector<llama_token> prompt_tokens(n_prompt);
+    if (llama_tokenize(vocab, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true, true) < 0) {
+        fprintf(stderr, "%s: error: failed to tokenize the prompt\n", __func__);
+        return 1;
+    }
+
+    llama_context_params ctx_params = llama_context_default_params();
+    ctx_params.n_ctx = n_prompt;
+    ctx_params.n_batch = n_prompt;
+    ctx_params.no_perf = false;
+    if (embedding_mode) {
+        ctx_params.embeddings = true;
+        ctx_params.pooling_type = LLAMA_POOLING_TYPE_NONE;
+        ctx_params.n_ubatch = ctx_params.n_batch;
+    }
+
+    llama_context * ctx = llama_init_from_model(model, ctx_params);
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
+    printf("Input prompt: \"%s\"\n", prompt.c_str());
+    printf("Tokenized prompt (%d tokens): ", n_prompt);
+    for (auto id : prompt_tokens) {
+        char buf[128];
+        int n = llama_token_to_piece(vocab, id, buf, sizeof(buf), 0, true);
+        if (n < 0) {
+            fprintf(stderr, "%s: error: failed to convert token to piece\n", __func__);
+            return 1;
+        }
+        std::string s(buf, n);
+        printf("%s", s.c_str());
+    }
+    printf("\n");
+
+    llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
+
+    if (llama_decode(ctx, batch)) {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
+        return 1;
+    }
+
+    float * logits;
+    int n_logits;
+    const char * type;
+
+    if (embedding_mode) {
+        logits = llama_get_embeddings(ctx);
+        n_logits = llama_model_n_embd(model) * batch.n_tokens;
+        type = "-embeddings";
+        printf("Embeddings size: %d\n", n_logits);
+    } else {
+        logits = llama_get_logits_ith(ctx, batch.n_tokens - 1);
+        n_logits = llama_vocab_n_tokens(vocab);
+        type = "";
+        printf("Vocab size: %d\n", n_logits);
+    }
+
+    std::filesystem::create_directory("data");
+
+    // Save logits to binary file
+    char bin_filename[512];
+    snprintf(bin_filename, sizeof(bin_filename), "data/llamacpp-%s%s.bin", model_name, type);
+    printf("Saving logits to %s\n", bin_filename);
+
+    FILE * f = fopen(bin_filename, "wb");
+    if (f == NULL) {
+        fprintf(stderr, "%s: error: failed to open binary output file\n", __func__);
+        return 1;
+    }
+    fwrite(logits, sizeof(float), n_logits, f);
+    fclose(f);
+
+    // Also save as text for debugging
+    char txt_filename[512];
+    snprintf(txt_filename, sizeof(txt_filename), "data/llamacpp-%s%s.txt", model_name, type);
+    f = fopen(txt_filename, "w");
+    if (f == NULL) {
+        fprintf(stderr, "%s: error: failed to open text output file\n", __func__);
+        return 1;
+    }
+    for (int i = 0; i < n_logits; i++) {
+        fprintf(f, "%d: %.6f\n", i, logits[i]);  // Added index and changed format
+    }
+    fclose(f);
+
+    // Print first and last 10 logits for quick verification
+    printf("First 10 logits: ");
+    for (int i = 0; i < 10 && i < n_logits; i++) {
+        printf("%.6f ", logits[i]);
+    }
+    printf("\n");
+
+    printf("Last 10 logits: ");
+    for (int i = n_logits - 10; i < n_logits; i++) {
+        if (i >= 0) printf("%.6f ", logits[i]);
+    }
+    printf("\n\n");
+
+    printf("Logits saved to %s\n", bin_filename);
+    printf("Logits saved to %s\n", txt_filename);
+
+    llama_free(ctx);
+    llama_model_free(model);
+
+    return 0;
+}

examples/model-conversion/requirements.txt

@@ -0,0 +1,5 @@
+--extra-index-url https://download.pytorch.org/whl/cpu
+torch~=2.6.0
+torchvision~=0.21.0
+transformers~=4.55.0
+huggingface-hub~=0.34.0

examples/model-conversion/scripts/causal/compare-embeddings-logits.sh

@@ -0,0 +1,43 @@
+#/bin/bash
+
+set -e
+
+MODEL_PATH="${1:-"$MODEL_PATH"}"
+MODEL_NAME="${2:-$(basename "$MODEL_PATH")}"
+
+if [ -t 0 ]; then
+    CPP_EMBEDDINGS="data/llamacpp-${MODEL_NAME}-embeddings.bin"
+else
+    # Process piped JSON data and convert to binary (matching logits.cpp format)
+    TEMP_FILE=$(mktemp /tmp/tmp.XXXXXX.binn)
+    python3 -c "
+import json
+import sys
+import struct
+
+data = json.load(sys.stdin)
+
+# Flatten all embeddings completely
+flattened = []
+for item in data:
+    embedding = item['embedding']
+    for token_embedding in embedding:
+        flattened.extend(token_embedding)
+
+print(f'Total embedding values: {len(flattened)}', file=sys.stderr)
+
+# Write as binary floats - matches logitc.cpp fwrite format
+with open('$TEMP_FILE', 'wb') as f:
+    for value in flattened:
+        f.write(struct.pack('f', value))
+"
+    CPP_EMBEDDINGS="$TEMP_FILE"
+    trap "rm -f $TEMP_FILE" EXIT
+fi
+
+python scripts/utils/semantic_check.py --model-path $MODEL_PATH \
+    --python-embeddings data/pytorch-${MODEL_NAME}-embeddings.bin \
+    --cpp-embeddings $CPP_EMBEDDINGS \
+    --prompt "Hello world today" \
+    --causal
+

examples/model-conversion/scripts/causal/compare-logits.py

@@ -0,0 +1,88 @@
+#!/usr/bin/env python3
+
+import numpy as np
+import sys
+import os
+from pathlib import Path
+
+def quick_logits_check(pytorch_file, llamacpp_file):
+    """Lightweight sanity check before NMSE"""
+
+    try:
+        pytorch_logits = np.fromfile(pytorch_file, dtype=np.float32)
+        llamacpp_logits = np.fromfile(llamacpp_file, dtype=np.float32)
+    except Exception as e:
+        print(f"❌ NOK: Failed to load files - {e}")
+        return False
+
+    # Check shapes match
+    if pytorch_logits.shape != llamacpp_logits.shape:
+        print(f"❌ NOK: Shape mismatch - PyTorch: {pytorch_logits.shape}, llama.cpp: {llamacpp_logits.shape}")
+        return False
+
+    # Calculate key metrics
+    diff = pytorch_logits - llamacpp_logits
+    abs_diff = np.abs(diff)
+    max_diff = np.max(abs_diff)
+
+    # Get top 10 predictions from both models
+    pytorch_top10 = np.argsort(pytorch_logits)[-10:][::-1]
+    llamacpp_top10 = np.argsort(llamacpp_logits)[-10:][::-1]
+    print(f"Top 10 PyTorch logits: {pytorch_logits[pytorch_top10]}")
+    print(f"Top 10 llama.cpp logits: {llamacpp_logits[llamacpp_top10]}")
+    print(f"Max absolute difference: {max_diff:.4f}")
+
+    if max_diff > 1.0:
+        print(f"❌ NOK: Large differences detected - max diff: {max_diff:.4f}")
+        return False
+
+    return True
+
+def main():
+    model_path = os.getenv('MODEL_PATH')
+    if not model_path:
+        print("Error: MODEL_PATH environment variable not set")
+        sys.exit(1)
+
+    if not os.path.exists(model_path):
+        print(f"Error: Model file not found: {model_path}")
+        sys.exit(1)
+
+    model_name = os.path.splitext(os.path.basename(model_path))[0]
+    data_dir = Path("data")
+
+    pytorch_file = data_dir / f"pytorch-{model_name}.bin"
+    llamacpp_file = data_dir / f"llamacpp-{model_name}.bin"
+
+    if not pytorch_file.exists():
+        print(f"Error: PyTorch logits file not found: {pytorch_file}")
+        print("Please run scripts/run-org-model.sh first to generate this file.")
+        sys.exit(1)
+
+    if not llamacpp_file.exists():
+        print(f"Error: llama.cpp logits file not found: {llamacpp_file}")
+        print("Please run scripts/run-converted-model.sh first to generate this file.")
+        sys.exit(1)
+
+    print("Checked all required files were found. Proceeding...\n")
+
+
+    print("🔍 GGML Model Validation for model ", model_name)
+    print("=" * 40)
+    print(f"PyTorch logits  : {pytorch_file}")
+    print(f"llama.cpp logits: {llamacpp_file}")
+    print()
+
+    success = quick_logits_check(pytorch_file, llamacpp_file)
+
+    # Exit with appropriate code
+    if success:
+        print("✅ OK: Lightweight model check successful!")
+        print("       Ok to proceed with NMSE check...")
+        sys.exit(0)
+    else:
+        print(f"❌ NOK: Top 10 predictions don't match - generation will differ")
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()

examples/model-conversion/scripts/causal/convert-model.sh

@@ -0,0 +1,46 @@
+#!/bin/bash
+
+set -e
+
+# Parse command line arguments
+MMPROJ=""
+while [[ $# -gt 0 ]]; do
+    case $1 in
+        --mmproj)
+            MMPROJ="--mmproj"
+            shift
+            ;;
+        *)
+            shift
+            ;;
+    esac
+done
+
+MODEL_NAME="${MODEL_NAME:-$(basename "$MODEL_PATH")}"
+OUTPUT_DIR="${OUTPUT_DIR:-../../models}"
+TYPE="${OUTTYPE:-f16}"
+METADATA_OVERRIDE="${METADATA_OVERRIDE:-}"
+CONVERTED_MODEL="${OUTPUT_DIR}/${MODEL_NAME}.gguf"
+
+echo "Model path: ${MODEL_PATH}"
+echo "Model name: ${MODEL_NAME}"
+echo "Data  type: ${TYPE}"
+echo "Converted model path:: ${CONVERTED_MODEL}"
+echo "Metadata override: ${METADATA_OVERRIDE}"
+
+CMD_ARGS=("python" "../../convert_hf_to_gguf.py" "--verbose")
+CMD_ARGS+=("${MODEL_PATH}")
+CMD_ARGS+=("--outfile" "${CONVERTED_MODEL}")
+CMD_ARGS+=("--outtype" "${TYPE}")
+[[ -n "$METADATA_OVERRIDE" ]] && CMD_ARGS+=("--metadata" "${METADATA_OVERRIDE}")
+[[ -n "$MMPROJ" ]] && CMD_ARGS+=("${MMPROJ}")
+
+"${CMD_ARGS[@]}"
+
+echo ""
+echo "The environment variable CONVERTED_MODEL can be set to this path using:"
+echo "export CONVERTED_MODEL=$(realpath ${CONVERTED_MODEL})"
+if [[ -n "$MMPROJ" ]]; then
+    mmproj_file="${OUTPUT_DIR}/mmproj-$(basename "${CONVERTED_MODEL}")"
+    echo "The mmproj model was created in $(realpath "$mmproj_file")"
+fi

examples/model-conversion/scripts/causal/modelcard.template

@@ -0,0 +1,13 @@
+---
+base_model:
+- {base_model}
+---
+# {model_name} GGUF
+
+Recommended way to run this model:
+
+```sh
+llama-server -hf {namespace}/{model_name}-GGUF -c 0 -fa
+```
+
+Then, access http://localhost:8080

examples/model-conversion/scripts/causal/run-casual-gen-embeddings-org.sh

@@ -0,0 +1,113 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import importlib
+import sys
+import torch
+import numpy as np
+
+from transformers import AutoTokenizer, AutoConfig, AutoModel, AutoModelForCausalLM
+from pathlib import Path
+
+unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+
+parser = argparse.ArgumentParser(description='Process model with specified path')
+parser.add_argument('--model-path', '-m', help='Path to the model')
+args = parser.parse_args()
+
+model_path = os.environ.get('MODEL_PATH', args.model_path)
+if model_path is None:
+    parser.error("Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+
+config = AutoConfig.from_pretrained(model_path)
+
+print("Model type:       ", config.model_type)
+print("Vocab size:       ", config.vocab_size)
+print("Hidden size:      ", config.hidden_size)
+print("Number of layers: ", config.num_hidden_layers)
+print("BOS token id:     ", config.bos_token_id)
+print("EOS token id:     ", config.eos_token_id)
+
+print("Loading model and tokenizer using AutoTokenizer:", model_path)
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+
+if unreleased_model_name:
+    model_name_lower = unreleased_model_name.lower()
+    unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+    class_name = f"{unreleased_model_name}ForCausalLM"
+    print(f"Importing unreleased model module: {unreleased_module_path}")
+
+    try:
+        model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+        model = model_class.from_pretrained(model_path)
+    except (ImportError, AttributeError) as e:
+        print(f"Failed to import or load model: {e}")
+else:
+    model = AutoModelForCausalLM.from_pretrained(model_path)
+print(f"Model class: {type(model)}")
+#print(f"Model file: {type(model).__module__}")
+
+model_name = os.path.basename(model_path)
+print(f"Model name: {model_name}")
+
+prompt = "Hello world today"
+input_ids = tokenizer(prompt, return_tensors="pt").input_ids
+print(f"Input tokens: {input_ids}")
+print(f"Input text: {repr(prompt)}")
+print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")
+
+with torch.no_grad():
+    outputs = model(input_ids, output_hidden_states=True)
+
+    # Extract hidden states from the last layer
+    # outputs.hidden_states is a tuple of (num_layers + 1) tensors
+    # Index -1 gets the last layer, shape: [batch_size, seq_len, hidden_size]
+    last_hidden_states = outputs.hidden_states[-1]
+
+    # Get embeddings for all tokens
+    token_embeddings = last_hidden_states[0].cpu().numpy()  # Remove batch dimension
+
+    print(f"Hidden states shape: {last_hidden_states.shape}")
+    print(f"Token embeddings shape: {token_embeddings.shape}")
+    print(f"Hidden dimension: {token_embeddings.shape[-1]}")
+    print(f"Number of tokens: {token_embeddings.shape[0]}")
+
+    # Save raw token embeddings
+    data_dir = Path("data")
+    data_dir.mkdir(exist_ok=True)
+    bin_filename = data_dir / f"pytorch-{model_name}-embeddings.bin"
+    txt_filename = data_dir / f"pytorch-{model_name}-embeddings.txt"
+
+    # Save all token embeddings as binary
+    print(token_embeddings)
+    token_embeddings.astype(np.float32).tofile(bin_filename)
+
+    # Save as text for inspection
+    with open(txt_filename, "w") as f:
+        for i, embedding in enumerate(token_embeddings):
+            for j, val in enumerate(embedding):
+                f.write(f"{i} {j} {val:.6f}\n")
+
+    # Print embeddings per token in the requested format
+    print("\nToken embeddings:")
+    tokens = tokenizer.convert_ids_to_tokens(input_ids[0])
+    for i, embedding in enumerate(token_embeddings):
+        # Format: show first few values, ..., then last few values
+        if len(embedding) > 10:
+            # Show first 3 and last 3 values with ... in between
+            first_vals = " ".join(f"{val:8.6f}" for val in embedding[:3])
+            last_vals = " ".join(f"{val:8.6f}" for val in embedding[-3:])
+            print(f"embedding {i}: {first_vals}  ... {last_vals}")
+        else:
+            # If embedding is short, show all values
+            vals = " ".join(f"{val:8.6f}" for val in embedding)
+            print(f"embedding {i}: {vals}")
+
+    # Also show token info for reference
+    print(f"\nToken reference:")
+    for i, token in enumerate(tokens):
+        print(f"  Token {i}: {repr(token)}")
+
+    print(f"Saved bin logits to: {bin_filename}")
+    print(f"Saved txt logist to: {txt_filename}")

examples/model-conversion/scripts/causal/run-converted-model-embeddings-logits.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+set -e
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+cmake --build ../../build --target llama-logits -j8
+
+../../build/bin/llama-logits -m $CONVERTED_MODEL -embd-mode "Hello world today"

examples/model-conversion/scripts/causal/run-converted-model.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+
+set -e
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+cmake --build ../../build --target llama-logits -j8
+
+../../build/bin/llama-logits -m "$CONVERTED_MODEL" "Hello, my name is"

examples/model-conversion/scripts/causal/run-org-model.py

@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import importlib
+from pathlib import Path
+
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig
+import torch
+import numpy as np
+
+unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+
+parser = argparse.ArgumentParser(description='Process model with specified path')
+parser.add_argument('--model-path', '-m', help='Path to the model')
+args = parser.parse_args()
+
+model_path = os.environ.get('MODEL_PATH', args.model_path)
+if model_path is None:
+    parser.error("Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+
+config = AutoConfig.from_pretrained(model_path)
+
+print("Model type:       ", config.model_type)
+print("Vocab size:       ", config.vocab_size)
+print("Hidden size:      ", config.hidden_size)
+print("Number of layers: ", config.num_hidden_layers)
+print("BOS token id:     ", config.bos_token_id)
+print("EOS token id:     ", config.eos_token_id)
+
+print("Loading model and tokenizer using AutoTokenizer:", model_path)
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+config = AutoConfig.from_pretrained(model_path)
+
+if unreleased_model_name:
+    model_name_lower = unreleased_model_name.lower()
+    unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+    class_name = f"{unreleased_model_name}ForCausalLM"
+    print(f"Importing unreleased model module: {unreleased_module_path}")
+
+    try:
+        model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+        model = model_class.from_pretrained(model_path)  # Note: from_pretrained, not fromPretrained
+    except (ImportError, AttributeError) as e:
+        print(f"Failed to import or load model: {e}")
+        exit(1)
+else:
+    model = AutoModelForCausalLM.from_pretrained(model_path)
+
+model_name = os.path.basename(model_path)
+# Printing the Model class to allow for easier debugging. This can be useful
+# when working with models that have not been publicly released yet and this
+# migth require that the concrete class is imported and used directly instead
+# of using AutoModelForCausalLM.
+print(f"Model class: {model.__class__.__name__}")
+
+prompt = "Hello, my name is"
+input_ids = tokenizer(prompt, return_tensors="pt").input_ids
+
+print(f"Input tokens: {input_ids}")
+print(f"Input text: {repr(prompt)}")
+print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")
+
+with torch.no_grad():
+    outputs = model(input_ids)
+    logits = outputs.logits
+
+    # Extract logits for the last token (next token prediction)
+    last_logits = logits[0, -1, :].cpu().numpy()
+
+    print(f"Logits shape: {logits.shape}")
+    print(f"Last token logits shape: {last_logits.shape}")
+    print(f"Vocab size: {len(last_logits)}")
+
+    data_dir = Path("data")
+    data_dir.mkdir(exist_ok=True)
+    bin_filename = data_dir / f"pytorch-{model_name}.bin"
+    txt_filename = data_dir / f"pytorch-{model_name}.txt"
+
+    # Save to file for comparison
+    last_logits.astype(np.float32).tofile(bin_filename)
+
+    # Also save as text file for easy inspection
+    with open(txt_filename, "w") as f:
+        for i, logit in enumerate(last_logits):
+            f.write(f"{i}: {logit:.6f}\n")
+
+    # Print some sample logits for quick verification
+    print(f"First 10 logits: {last_logits[:10]}")
+    print(f"Last 10 logits: {last_logits[-10:]}")
+
+    # Show top 5 predicted tokens
+    top_indices = np.argsort(last_logits)[-5:][::-1]
+    print("Top 5 predictions:")
+    for idx in top_indices:
+        token = tokenizer.decode([idx])
+        print(f"  Token {idx} ({repr(token)}): {last_logits[idx]:.6f}")
+
+    print(f"Saved bin logits to: {bin_filename}")
+    print(f"Saved txt logist to: {txt_filename}")

examples/model-conversion/scripts/embedding/compare-embeddings-logits.sh

@@ -0,0 +1,42 @@
+#/bin/bash
+
+set -e
+
+MODEL_PATH="${1:-"$EMBEDDING_MODEL_PATH"}"
+MODEL_NAME="${2:-$(basename "$MODEL_PATH")}"
+
+if [ -t 0 ]; then
+    CPP_EMBEDDINGS="data/llamacpp-${MODEL_NAME}-embeddings.bin"
+else
+    # Process piped JSON data and convert to binary (matching logits.cpp format)
+    TEMP_FILE=$(mktemp /tmp/tmp.XXXXXX.binn)
+    python3 -c "
+import json
+import sys
+import struct
+
+data = json.load(sys.stdin)
+
+# Flatten all embeddings completely
+flattened = []
+for item in data:
+    embedding = item['embedding']
+    for token_embedding in embedding:
+        flattened.extend(token_embedding)
+
+print(f'Total embedding values: {len(flattened)}', file=sys.stderr)
+
+# Write as binary floats - matches logitc.cpp fwrite format
+with open('$TEMP_FILE', 'wb') as f:
+    for value in flattened:
+        f.write(struct.pack('f', value))
+"
+    CPP_EMBEDDINGS="$TEMP_FILE"
+    trap "rm -f $TEMP_FILE" EXIT
+fi
+
+python scripts/utils/semantic_check.py --model-path $MODEL_PATH \
+    --python-embeddings data/pytorch-${MODEL_NAME}-embeddings.bin \
+    --cpp-embeddings $CPP_EMBEDDINGS \
+    --prompt "Hello world today"
+

examples/model-conversion/scripts/embedding/convert-model.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+
+set -e
+
+MODEL_NAME="${MODEL_NAME:-$(basename "$EMBEDDING_MODEL_PATH")}"
+OUTPUT_DIR="${OUTPUT_DIR:-../../models}"
+TYPE="${OUTTYPE:-f16}"
+METADATA_OVERRIDE="${METADATA_OVERRIDE:-}"
+CONVERTED_MODEL="${OUTPUT_DIR}/${MODEL_NAME}.gguf"
+
+echo "Model path: ${EMBEDDING_MODEL_PATH}"
+echo "Model name: ${MODEL_NAME}"
+echo "Data  type: ${TYPE}"
+echo "Converted model path:: ${CONVERTED_MODEL}"
+python ../../convert_hf_to_gguf.py --verbose \
+    ${EMBEDDING_MODEL_PATH} \
+    --outfile ${CONVERTED_MODEL} \
+    --outtype ${TYPE}
+
+echo ""
+echo "The environment variable CONVERTED_EMBEDDING MODEL can be set to this path using:"
+echo "export CONVERTED_EMBEDDING_MODEL=$(realpath ${CONVERTED_MODEL})"

examples/model-conversion/scripts/embedding/modelcard.template

@@ -0,0 +1,48 @@
+---
+base_model:
+- {base_model}
+---
+# {model_name} GGUF
+
+Recommended way to run this model:
+
+```sh
+llama-server -hf {namespace}/{model_name}-GGUF
+```
+
+Then the endpoint can be accessed at http://localhost:8080/embedding, for
+example using `curl`:
+```console
+curl --request POST \
+    --url http://localhost:8080/embedding \
+    --header "Content-Type: application/json" \
+    --data '{{"input": "Hello embeddings"}}' \
+    --silent
+```
+
+Alternatively, the `llama-embedding` command line tool can be used:
+```sh
+llama-embedding -hf {namespace}/{model_name}-GGUF --verbose-prompt -p "Hello embeddings"
+```
+
+#### embd_normalize
+When a model uses pooling, or the pooling method is specified using `--pooling`,
+the normalization can be controlled by the `embd_normalize` parameter.
+
+The default value is `2` which means that the embeddings are normalized using
+the Euclidean norm (L2). Other options are:
+* -1 No normalization
+*  0 Max absolute
+*  1 Taxicab
+*  2 Euclidean/L2
+* \>2 P-Norm
+
+This can be passed in the request body to `llama-server`, for example:
+```sh
+    --data '{{"input": "Hello embeddings", "embd_normalize": -1}}' \
+```
+
+And for `llama-embedding`, by passing `--embd-normalize <value>`, for example:
+```sh
+llama-embedding -hf {namespace}/{model_name}-GGUF  --embd-normalize -1 -p "Hello embeddings"
+```

examples/model-conversion/scripts/embedding/run-converted-model.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+
+set -e
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_EMBEDDING_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_EMBEDDING_MODEL environment variable" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+cmake --build ../../build --target llama-logits -j8
+
+../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode "Hello world today"

examples/model-conversion/scripts/embedding/run-original-model.py

@@ -0,0 +1,116 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import numpy as np
+import importlib
+from pathlib import Path
+
+from transformers import AutoTokenizer, AutoConfig, AutoModel
+import torch
+
+unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+
+parser = argparse.ArgumentParser(description='Process model with specified path')
+parser.add_argument('--model-path', '-m', help='Path to the model')
+args = parser.parse_args()
+
+model_path = os.environ.get('EMBEDDING_MODEL_PATH', args.model_path)
+if model_path is None:
+    parser.error("Model path must be specified either via --model-path argument or EMBEDDING_MODEL_PATH environment variable")
+
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+
+if unreleased_model_name:
+    model_name_lower = unreleased_model_name.lower()
+    unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+    class_name = f"{unreleased_model_name}Model"
+    print(f"Importing unreleased model module: {unreleased_module_path}")
+
+    try:
+        model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+        model = model_class.from_pretrained(model_path)  # Note: from_pretrained, not fromPretrained
+    except (ImportError, AttributeError) as e:
+        print(f"Failed to import or load model: {e}")
+        exit(1)
+else:
+    model = AutoModel.from_pretrained(model_path)
+print(f"Model class: {type(model)}")
+#print(f"Model file: {type(model).__module__}")
+config = AutoConfig.from_pretrained(model_path)
+
+model_name = os.path.basename(model_path)
+
+texts = [ "Hello world today" ]
+
+encoded = tokenizer(
+    texts,
+    padding=True,
+    truncation=True,
+    return_tensors="pt"
+)
+
+tokens = encoded['input_ids'][0]
+token_strings = tokenizer.convert_ids_to_tokens(tokens)
+for i, (token_id, token_str) in enumerate(zip(tokens, token_strings)):
+    print(f"{token_id:6d} -> '{token_str}'")
+
+with torch.no_grad():
+    outputs = model(**encoded)
+    hidden_states = outputs.last_hidden_state  # Shape: [batch_size, seq_len, hidden_size]
+
+    # Extract embeddings for each token (matching LLAMA_POOLING_TYPE_NONE behavior)
+    all_embeddings = hidden_states[0].cpu().numpy()  # Shape: [seq_len, hidden_size]
+
+    print(f"Hidden states shape: {hidden_states.shape}")
+    print(f"All embeddings shape: {all_embeddings.shape}")
+    print(f"Embedding dimension: {all_embeddings.shape[1]}")
+
+    # Print embeddings exactly like embedding.cpp does for LLAMA_POOLING_TYPE_NONE
+    n_embd = all_embeddings.shape[1]
+    n_embd_count = all_embeddings.shape[0]
+
+    print()  # Empty line to match C++ output
+
+    for j in range(n_embd_count):
+        embedding = all_embeddings[j]
+        print(f"embedding {j}: ", end="")
+
+        # Print first 3 values
+        for i in range(min(3, n_embd)):
+            print(f"{embedding[i]:9.6f} ", end="")
+
+        print(" ... ", end="")
+
+        # Print last 3 values
+        for i in range(n_embd - 3, n_embd):
+            print(f"{embedding[i]:9.6f} ", end="")
+
+        print()  # New line
+
+    print()  # Final empty line to match C++ output
+
+    data_dir = Path("data")
+    data_dir.mkdir(exist_ok=True)
+    bin_filename = data_dir / f"pytorch-{model_name}-embeddings.bin"
+    txt_filename = data_dir / f"pytorch-{model_name}-embeddings.txt"
+
+    # Save all embeddings flattened (matching what embedding.cpp would save if it did)
+    flattened_embeddings = all_embeddings.flatten()
+    flattened_embeddings.astype(np.float32).tofile(bin_filename)
+
+    with open(txt_filename, "w") as f:
+        f.write(f"# Model class: {model_name}\n")
+        f.write(f"# Tokens: {token_strings}\n")
+        f.write(f"# Shape: {all_embeddings.shape}\n")
+        f.write(f"# n_embd_count: {n_embd_count}, n_embd: {n_embd}\n\n")
+
+        for j in range(n_embd_count):
+            f.write(f"# Token {j} ({token_strings[j]}):\n")
+            for i, value in enumerate(all_embeddings[j]):
+                f.write(f"{j}_{i}: {value:.6f}\n")
+            f.write("\n")
+    print(f"Total values: {len(flattened_embeddings)} ({n_embd_count} tokens × {n_embd} dimensions)")
+    print("")
+    print(f"Saved bin embeddings to: {bin_filename}")
+    print(f"Saved txt embeddings to: {txt_filename}")

examples/model-conversion/scripts/utils/check-nmse.py

@@ -0,0 +1,174 @@
+#!/usr/bin/env python3
+
+import numpy as np
+import sys
+import os
+import argparse
+from pathlib import Path
+
+def calculate_nmse(reference, test):
+    mse = np.mean((test - reference) ** 2)
+    ref_var = np.var(reference)
+    if ref_var == 0:
+        nmse = float('inf') if mse > 0 else 0.0
+        return mse, mse, ref_var
+
+    nmse = mse / ref_var
+
+    return nmse, mse, ref_var
+
+def load_logits(file_path):
+    if not os.path.exists(file_path):
+        raise FileNotFoundError(f"File not found: {file_path}")
+
+    if file_path.suffix == '.npy':
+        return np.load(file_path)
+    elif file_path.suffix == '.bin':
+        return np.fromfile(file_path, dtype=np.float32)
+    else:
+        # Try to load as text file
+        try:
+            # If it has index format "0: value", extract just values
+            data = []
+            with open(file_path, 'r') as f:
+                for line in f:
+                    if ':' in line:
+                        # Format: "index: value"
+                        value = float(line.split(':')[1].strip())
+                    else:
+                        # Just the value
+                        value = float(line.strip())
+                    data.append(value)
+            return np.array(data, dtype=np.float32)
+        except:
+            return np.loadtxt(file_path, dtype=np.float32)
+
+def interpret_nmse(nmse):
+    """Provide interpretation of NMSE value"""
+    if nmse == 0:
+        return "Perfect match", "🎉"
+    elif nmse < 1e-6:
+        return "Essentially identical", "✅"
+    elif nmse < 1e-4:
+        return "Excellent match", "✅"
+    elif nmse < 1e-3:
+        return "Very good match", "👍"
+    elif nmse < 1e-2:
+        return "Good match", "👍"
+    elif nmse < 0.1:
+        return "Acceptable match", "⚠️"
+    elif nmse < 1.0:
+        return "Poor match", "❌"
+    else:
+        return "Very poor match (worse than noise)", "❌"
+
+def main():
+    parser = argparse.ArgumentParser(description='Validate model logits')
+    parser.add_argument('-m', '--model-path', required=True,  help='Path to the model directory')
+    args = parser.parse_args()
+
+    model_name = os.path.splitext(os.path.basename(args.model_path))[0]
+    data_dir = Path("data")
+
+    pytorch_file = data_dir / f"pytorch-{model_name}.bin"
+    llamacpp_file = data_dir / f"llamacpp-{model_name}.bin"
+
+    print(f"Model name: {model_name}")
+    print(f"PyTorch logits file: {pytorch_file}")
+    print(f"llama.cpp logits file: {llamacpp_file}")
+
+    reference_file = pytorch_file
+    test_file = llamacpp_file
+
+    print("📊 NMSE Check for Model Comparison")
+    print("=" * 50)
+    print(f"Reference (ground truth): {reference_file}")
+    print(f"Test (to evaluate):       {test_file}")
+    print()
+
+    try:
+        print("Loading reference logits...")
+        reference = load_logits(reference_file)
+        print(f"  Shape: {reference.shape}, Type: {reference.dtype}")
+
+        print("Loading test logits...")
+        test = load_logits(test_file)
+        print(f"  Shape: {test.shape}, Type: {test.dtype}")
+
+        # Check shapes match
+        if reference.shape != test.shape:
+            print(f"\n❌ Error: Shape mismatch!")
+            print(f"  Reference: {reference.shape}")
+            print(f"  Test: {test.shape}")
+            sys.exit(1)
+
+        print(f"\n✅ Shapes match: {reference.shape}")
+
+        nmse, mse, ref_var = calculate_nmse(reference, test)
+
+        # Additional metrics
+        max_abs_error = np.max(np.abs(test - reference))
+        mean_abs_error = np.mean(np.abs(test - reference))
+
+        # Results
+        print(f"\n📈 METRICS")
+        print("=" * 30)
+        print(f"MSE (Mean Squared Error):     {mse:.6e}")
+        print(f"Reference Variance:           {ref_var:.6e}")
+        print(f"NMSE:                         {nmse:.6e}")
+        print(f"Max Absolute Error:           {max_abs_error:.6f}")
+        print(f"Mean Absolute Error:          {mean_abs_error:.6f}")
+
+        # NMSE in dB (common in signal processing)
+        if nmse > 0:
+            nmse_db = 10 * np.log10(nmse)
+            print(f"NMSE (dB):                    {nmse_db:.2f} dB")
+
+        # Interpretation
+        interpretation, emoji = interpret_nmse(nmse)
+        print(f"\n🎯 INTERPRETATION")
+        print("=" * 30)
+        print(f"{emoji} {interpretation}")
+
+        # Detailed guidance
+        print(f"\n📋 GUIDANCE")
+        print("=" * 30)
+        if nmse < 1e-3:
+            print("✅ EXCELLENT: Your GGML conversion is working very well!")
+            print("   The differences are negligible for practical use.")
+        elif nmse < 1e-2:
+            print("👍 GOOD: Your GGML conversion is working well.")
+            print("   Small differences are likely due to precision/quantization.")
+        elif nmse < 0.1:
+            print("⚠️  ACCEPTABLE: Conversion is working but with some differences.")
+            print("   Check if you're using quantization (Q4, Q8, etc.)")
+            print("   Test generation quality to see if it's acceptable.")
+        else:
+            print("❌ PROBLEMATIC: Large differences detected.")
+            print("   Check your conversion process for potential issues.")
+            print("   Verify you're using the same model weights.")
+
+        # NMSE benchmarks
+        print(f"\n📚 NMSE BENCHMARKS")
+        print("=" * 30)
+        print("< 1e-6:  Essentially identical")
+        print("< 1e-4:  Excellent (typical for good conversions)")
+        print("< 1e-3:  Very good")
+        print("< 1e-2:  Good (acceptable for most use cases)")
+        print("< 0.1:   Acceptable (may need verification)")
+        print("> 1.0:   Poor (worse than random)")
+
+        # Exit code based on NMSE
+        if nmse < 1e-2:
+            print(f"\n✅ RESULT: PASS (NMSE = {nmse:.2e})")
+            sys.exit(0)
+        else:
+            print(f"\n❌ RESULT: NEEDS REVIEW (NMSE = {nmse:.2e})")
+            sys.exit(1)
+
+    except Exception as e:
+        print(f"❌ Error: {e}")
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()

examples/model-conversion/scripts/utils/create-collection-add-model.sh

@@ -0,0 +1,6 @@
+
+COLLECTION_SLUG=$(python ./create_collection.py --return-slug)
+echo "Created collection: $COLLECTION_SLUG"
+
+# Use it in the next command
+python add_model_to_collection.py "$COLLECTION_SLUG" "username/my-model"

examples/model-conversion/scripts/utils/hf-add-model-to-collection.py

@@ -0,0 +1,80 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+import sys
+
+def add_model_to_collection(collection_slug, model_id, note=""):
+    """
+    Add a model to an existing collection
+
+    Args:
+        collection_slug: The slug of the collection (e.g., "username/collection-name-12345")
+        model_id: The model repository ID (e.g., "username/model-name")
+        note: Optional note about the model
+
+    Returns:
+        True if successful, False if failed
+    """
+
+    # Initialize API
+    api = HfApi()
+
+    try:
+        user_info = api.whoami()
+        print(f"✅ Authenticated as: {user_info['name']}")
+
+        # Verify the model exists
+        print(f"🔍 Checking if model exists: {model_id}")
+        try:
+            model_info = api.model_info(model_id)
+        except Exception as e:
+            print(f"❌ Model not found or not accessible: {model_id}")
+            print(f"Error: {e}")
+            return False
+
+        print(f"📚 Adding model to collection...")
+        api.add_collection_item(
+            collection_slug=collection_slug,
+            item_id=model_id,
+            item_type="model",
+            note=note
+        )
+
+        print(f"✅ Model added to collection successfully!")
+        print(f"🔗 Collection URL: https://huggingface.co/collections/{collection_slug}")
+
+        return True
+
+    except Exception as e:
+        print(f"❌ Error adding model to collection: {e}")
+        return False
+
+def main():
+    # This script requires that the environment variable HF_TOKEN is set with your
+    # Hugging Face API token.
+    api = HfApi()
+
+    parser = argparse.ArgumentParser(description='Add model to a Huggingface Collection')
+    parser.add_argument('--collection', '-c', help='The collection slug username/collection-hash', required=True)
+    parser.add_argument('--model', '-m', help='The model to add to the Collection', required=True)
+    parser.add_argument('--note', '-n', help='An optional note/description', required=False)
+    args = parser.parse_args()
+
+    collection = args.collection
+    model = args.model
+    note = args.note
+
+    success = add_model_to_collection(
+        collection_slug=collection,
+        model_id=model,
+        note=note
+    )
+
+    if success:
+        print("\n🎉 Model added successfully!")
+    else:
+        print("\n❌ Failed to add model to collection")
+        sys.exit(1)
+if __name__ == "__main__":
+    main()

examples/model-conversion/scripts/utils/hf-create-collection.py

@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+import os
+import sys
+
+
+def create_collection(title, description, private=False, namespace=None, return_slug=False):
+    """
+    Create a new collection on Hugging Face
+
+    Args:
+        title: Collection title
+        description: Collection description
+        private: Whether the collection should be private (default: False)
+        namespace: Optional namespace (defaults to your username)
+
+    Returns:
+        Collection object if successful, None if failed
+    """
+
+    # Check if HF_TOKEN is available
+    token = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_HUB_TOKEN")
+    if not token:
+        print("❌ No HF_TOKEN or HUGGINGFACE_HUB_TOKEN found in environment variables")
+        print("Please set your Hugging Face token as an environment variable")
+        return None
+
+    # Initialize API
+    api = HfApi()
+
+    try:
+        # Test authentication first
+        user_info = api.whoami()
+        if not return_slug:
+            print(f"✅ Authenticated as: {user_info['name']}")
+
+        # Create the collection
+        if not return_slug:
+            print(f"📚 Creating collection: '{title}'...")
+        collection = api.create_collection(
+            title=title,
+            description=description,
+            private=private,
+            namespace=namespace
+        )
+
+        if not return_slug:
+            print(f"✅ Collection created successfully!")
+            print(f"📋 Collection slug: {collection.slug}")
+            print(f"🔗 Collection URL: https://huggingface.co/collections/{collection.slug}")
+
+        return collection
+
+    except Exception as e:
+        print(f"❌ Error creating collection: {e}")
+        return None
+
+def main():
+    # This script requires that the environment variable HF_TOKEN is set with your
+    # Hugging Face API token.
+    api = HfApi()
+
+    parser = argparse.ArgumentParser(description='Create a Huggingface Collection')
+    parser.add_argument('--name', '-n', help='The name/title of the Collection', required=True)
+    parser.add_argument('--description', '-d', help='The description for the Collection', required=True)
+    parser.add_argument('--namespace', '-ns', help='The namespace to add the Collection to', required=True)
+    parser.add_argument('--private', '-p', help='Create a private Collection', action='store_true')  # Fixed
+    parser.add_argument('--return-slug', '-s', help='Only output the collection slug', action='store_true')  # Fixed
+
+    args = parser.parse_args()
+
+    name = args.name
+    description = args.description
+    private = args.private
+    namespace = args.namespace
+    return_slug = args.return_slug
+
+    if not return_slug:
+        print("🚀 Creating Hugging Face Collection")
+        print(f"Title: {name}")
+        print(f"Description: {description}")
+        print(f"Namespace: {namespace}")
+        print(f"Private: {private}")
+
+    collection = create_collection(
+        title=name,
+        description=description,
+        private=private,
+        namespace=namespace,
+        return_slug=return_slug
+    )
+
+    if collection:
+        if return_slug:
+            print(collection.slug)
+        else:
+            print("\n🎉 Collection created successfully!")
+            print(f"Use this slug to add models: {collection.slug}")
+    else:
+        print("\n❌ Failed to create collection")
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()

examples/model-conversion/scripts/utils/hf-create-model.py

@@ -0,0 +1,78 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+
+# This script requires that the environment variable HF_TOKEN is set with your
+# Hugging Face API token.
+api = HfApi()
+
+def load_template_and_substitute(template_path, **kwargs):
+    try:
+        with open(template_path, 'r', encoding='utf-8') as f:
+            template_content = f.read()
+
+        return template_content.format(**kwargs)
+    except FileNotFoundError:
+        print(f"Template file '{template_path}' not found!")
+        return None
+    except KeyError as e:
+        print(f"Missing template variable: {e}")
+        return None
+
+parser = argparse.ArgumentParser(description='Create a new Hugging Face model repository')
+parser.add_argument('--model-name', '-m', help='Name for the model', required=True)
+parser.add_argument('--namespace', '-ns', help='Namespace to add the model to', required=True)
+parser.add_argument('--org-base-model', '-b', help='Original Base model name', default="")
+parser.add_argument('--no-card', action='store_true', help='Skip creating model card')
+parser.add_argument('--private', '-p', action='store_true', help='Create private model')
+parser.add_argument('--embedding', '-e', action='store_true', help='Use embedding model card template')
+parser.add_argument('--dry-run', '-d', action='store_true', help='Print repository info and template without creating repository')
+
+args = parser.parse_args()
+
+repo_id = f"{args.namespace}/{args.model_name}-GGUF"
+print("Repository ID: ", repo_id)
+
+repo_url = None
+if not args.dry_run:
+    repo_url = api.create_repo(
+        repo_id=repo_id,
+        repo_type="model",
+        private=args.private,
+        exist_ok=False
+    )
+
+if not args.no_card:
+    if args.embedding:
+        template_path = "scripts/embedding/modelcard.template"
+    else:
+        template_path = "scripts/causal/modelcard.template"
+
+    print("Template path: ", template_path)
+
+    model_card_content = load_template_and_substitute(
+        template_path,
+        model_name=args.model_name,
+        namespace=args.namespace,
+        base_model=args.org_base_model,
+    )
+
+    if args.dry_run:
+        print("\nTemplate Content:\n")
+        print(model_card_content)
+    else:
+        if model_card_content:
+            api.upload_file(
+                path_or_fileobj=model_card_content.encode('utf-8'),
+                path_in_repo="README.md",
+                repo_id=repo_id
+            )
+            print("Model card created successfully.")
+        else:
+            print("Failed to create model card.")
+
+if not args.dry_run and repo_url:
+    print(f"Repository created: {repo_url}")
+
+

examples/model-conversion/scripts/utils/hf-upload-gguf-model.py

@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+import os
+
+def upload_gguf_file(local_file_path, repo_id, filename_in_repo=None):
+    """
+    Upload a GGUF file to a Hugging Face model repository
+
+    Args:
+        local_file_path: Path to your local GGUF file
+        repo_id: Your repository ID (e.g., "username/model-name")
+        filename_in_repo: Optional custom name for the file in the repo
+    """
+
+    if not os.path.exists(local_file_path):
+        print(f"❌ File not found: {local_file_path}")
+        return False
+
+    if filename_in_repo is None:
+        filename_in_repo = os.path.basename(local_file_path)
+
+    if filename_in_repo is None or filename_in_repo == "":
+        filename_in_repo = os.path.basename(local_file_path)
+
+    print(f"📤 Uploading {local_file_path} to {repo_id}/{filename_in_repo}")
+
+    api = HfApi()
+
+    try:
+        api.upload_file(
+            path_or_fileobj=local_file_path,
+            path_in_repo=filename_in_repo,
+            repo_id=repo_id,
+            repo_type="model",
+            commit_message=f"Upload {filename_in_repo}"
+        )
+
+        print("✅ Upload successful!")
+        print(f"🔗 File available at: https://huggingface.co/{repo_id}/blob/main/{filename_in_repo}")
+        return True
+
+    except Exception as e:
+        print(f"❌ Upload failed: {e}")
+        return False
+
+# This script requires that the environment variable HF_TOKEN is set with your
+# Hugging Face API token.
+api = HfApi()
+
+parser = argparse.ArgumentParser(description='Upload a GGUF model to a Huggingface model repository')
+parser.add_argument('--gguf-model-path', '-m', help='The GGUF model file to upload', required=True)
+parser.add_argument('--repo-id', '-r', help='The repository to upload to', required=True)
+parser.add_argument('--name', '-o', help='The name in the model repository', required=False)
+args = parser.parse_args()
+
+upload_gguf_file(args.gguf_model_path, args.repo_id, args.name)

examples/model-conversion/scripts/utils/inspect-converted-model.sh

@@ -0,0 +1,14 @@
+#!/bin/bash
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+../../gguf-py/gguf/scripts/gguf_dump.py $CONVERTED_MODEL

examples/model-conversion/scripts/utils/inspect-org-model.py

@@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import json
+from safetensors import safe_open
+from collections import defaultdict
+
+parser = argparse.ArgumentParser(description='Process model with specified path')
+parser.add_argument('--model-path', '-m', help='Path to the model')
+args = parser.parse_args()
+
+model_path = os.environ.get('MODEL_PATH', args.model_path)
+if model_path is None:
+    parser.error("Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+
+# Check if there's an index file (multi-file model)
+index_path = os.path.join(model_path, "model.safetensors.index.json")
+single_file_path = os.path.join(model_path, "model.safetensors")
+
+if os.path.exists(index_path):
+    # Multi-file model
+    print("Multi-file model detected")
+
+    with open(index_path, 'r') as f:
+        index_data = json.load(f)
+
+    # Get the weight map (tensor_name -> file_name)
+    weight_map = index_data.get("weight_map", {})
+
+    # Group tensors by file for efficient processing
+    file_tensors = defaultdict(list)
+    for tensor_name, file_name in weight_map.items():
+        file_tensors[file_name].append(tensor_name)
+
+    print("Tensors in model:")
+
+    # Process each shard file
+    for file_name, tensor_names in file_tensors.items():
+        file_path = os.path.join(model_path, file_name)
+        print(f"\n--- From {file_name} ---")
+
+        with safe_open(file_path, framework="pt") as f:
+            for tensor_name in sorted(tensor_names):
+                tensor = f.get_tensor(tensor_name)
+                print(f"- {tensor_name} : shape = {tensor.shape}, dtype = {tensor.dtype}")
+
+elif os.path.exists(single_file_path):
+    # Single file model (original behavior)
+    print("Single-file model detected")
+
+    with safe_open(single_file_path, framework="pt") as f:
+        keys = f.keys()
+        print("Tensors in model:")
+        for key in sorted(keys):
+            tensor = f.get_tensor(key)
+            print(f"- {key} : shape = {tensor.shape}, dtype = {tensor.dtype}")
+
+else:
+    print(f"Error: Neither 'model.safetensors.index.json' nor 'model.safetensors' found in {model_path}")
+    print("Available files:")
+    if os.path.exists(model_path):
+        for item in sorted(os.listdir(model_path)):
+            print(f"  {item}")
+    else:
+        print(f"  Directory {model_path} does not exist")
+    exit(1)

examples/model-conversion/scripts/utils/perplexity-gen.sh

@@ -0,0 +1,35 @@
+#!/bin/bash
+
+set -e
+
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+# Check if data/wikitext-2-raw directory exists
+if [ ! -d "ppl/wikitext-2-raw" ]; then
+    echo "ppl/wikitext-2-raw directory does not exist. Downloading..." >&2
+    mkdir -p ppl
+    pushd ppl
+    ./../../../scripts/get-wikitext-2.sh
+    popd
+fi
+
+mkdir -p ppl
+OUTPUTFILE="ppl/$(basename $CONVERTED_MODEL).kld"
+echo "Model: $CONVERTED_MODEL"
+
+cmake --build ../../build --target llama-perplexity -j8
+
+../.././build/bin/llama-perplexity -m $CONVERTED_MODEL \
+    -f ppl/wikitext-2-raw/wiki.test.raw \
+    --kl-divergence-base $OUTPUTFILE
+
+echo "Generated logits in $OUTPUTFILE"
+

examples/model-conversion/scripts/utils/perplexity-run-simple.sh

@@ -0,0 +1,27 @@
+#!/bin/bash
+
+set -e
+
+QUANTIZED_MODEL="${1:-"$QUANTIZED_MODEL"}"
+
+if [ -z "$QUANTIZED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. QUANTIZED_MODEL environment variable" >&2
+    exit 1
+fi
+
+# Check if data/wikitext-2-raw directory exists
+if [ ! -d "ppl/wikitext-2-raw" ]; then
+    echo "ppl/wikitext-2-raw directory does not exist. Downloading..." >&2
+    mkdir -p ppl
+    pushd ppl
+    ./../../../scripts/get-wikitext-2.sh
+    popd
+fi
+
+cmake --build ../../build --target llama-perplexity -j8
+
+../.././build/bin/llama-perplexity -m $QUANTIZED_MODEL -f ppl/wikitext-2-raw/wiki.test.raw
+
+

examples/model-conversion/scripts/utils/perplexity-run.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+
+set -e
+
+QUANTIZED_MODEL="${1:-"$QUANTIZED_MODEL"}"
+LOGITS_FILE="${1:-"$LOGITS_FILE"}"
+
+if [ -z "$QUANTIZED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. QUANTIZED_MODEL environment variable" >&2
+    exit 1
+fi
+
+if [ ! -f ${LOGITS_FILE} ]; then
+    echo "Error: logits file '${LOGITS_FILE} was not found"
+    echo "Did you run the perplexity-gen.sh script?"
+    exit 1
+fi
+
+echo "Model: $QUANTIZED_MODEL"
+echo "Data file: $LOGITS_FILE"
+
+cmake --build ../../build --target llama-perplexity -j8
+
+../.././build/bin/llama-perplexity -m $QUANTIZED_MODEL \
+    --kl-divergence-base $LOGITS_FILE \
+    --kl-divergence

examples/model-conversion/scripts/utils/quantize.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+
+set -e
+
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+QUANTIZED_TYPE="${2:-"$QUANTIZED_TYPE"}"
+TOKEN_EMBD_TYPE="${3:-"${TOKEN_EMBD_TYPE}"}"
+OUTPUT_TYPE="${4:-"${OUTPUT_TYPE}"}"
+QUANTIZED_MODEL=$CONVERTED_MODEL
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+if [ -z "$QUANTIZED_TYPE" ]; then
+    echo "Error: QUANTIZED_TYPE is required" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+# Process the quantized model filename
+if [[ "$QUANTIZED_MODEL" == *.gguf ]]; then
+    # Remove .gguf suffix, add quantized type, then add .gguf back
+    BASE_NAME="${QUANTIZED_MODEL%.gguf}"
+    QUANTIZED_MODEL="${BASE_NAME}-${QUANTIZED_TYPE}.gguf"
+else
+    echo "Error: QUANTIZED_MODEL must end with .gguf extension" >&2
+    exit 1
+fi
+
+cmake --build ../../build --target llama-quantize -j8
+
+echo $TOKEN_EMBD_TYPE
+echo $OUTPUT_TYPE
+
+CMD_ARGS=("../../build/bin/llama-quantize")
+[[ -n "$TOKEN_EMBD_TYPE" ]] && CMD_ARGS+=("--token-embedding-type" "$TOKEN_EMBD_TYPE")
+[[ -n "$OUTPUT_TYPE" ]]     && CMD_ARGS+=("--output-tensor-type" "$OUTPUT_TYPE")
+CMD_ARGS+=("$CONVERTED_MODEL" "$QUANTIZED_MODEL" "$QUANTIZED_TYPE")
+
+"${CMD_ARGS[@]}"
+
+echo "Quantized model saved to: $QUANTIZED_MODEL"

examples/model-conversion/scripts/utils/run-embedding-server.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+
+set -e
+#
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+cmake --build ../../build --target llama-server
+
+../../build/bin/llama-server -m $CONVERTED_MODEL \
+    --embedding \
+    --pooling none

examples/model-conversion/scripts/utils/semantic_check.py

@@ -0,0 +1,179 @@
+#!/usr/bin/env python3
+
+import numpy as np
+import argparse
+import os
+import importlib
+
+from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM, AutoModel
+
+unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+
+def cosine_similarity(a, b=None):
+    a = np.asarray(a)
+    if b is None:
+        b = a
+    else:
+        b = np.asarray(b)
+
+    if a.ndim == 1:
+        a = a.reshape(1, -1)
+    if b.ndim == 1:
+        b = b.reshape(1, -1)
+
+    a_norms = np.linalg.norm(a, axis=1, keepdims=True)
+    b_norms = np.linalg.norm(b, axis=1, keepdims=True)
+
+    a_norms = np.where(a_norms == 0, 1e-8, a_norms)
+    b_norms = np.where(b_norms == 0, 1e-8, b_norms)
+
+    a_normalized = a / a_norms
+    b_normalized = b / b_norms
+
+    # Compute cosine similarity
+    return np.dot(a_normalized, b_normalized.T)
+
+def load_embeddings_from_file(filename, n_tokens, n_embd):
+    embeddings = np.fromfile(filename, dtype=np.float32)
+    return embeddings.reshape(n_tokens, n_embd)
+
+def test_single_prompt_similarity(python_emb, cpp_emb, tokens, prompt):
+    np.set_printoptions(suppress=True, precision=6)
+    print("pytorch embeddings:");
+    print(python_emb)
+    print("llama.cpp embeddings:");
+    print(cpp_emb)
+    print(f"\n=== Prompt: '{prompt}' ===")
+    print(f"Tokens: {tokens}")
+    print(f"Embeddings shape: Python {python_emb.shape}, llama.cpp {cpp_emb.shape}")
+
+    n_tokens = len(tokens)
+
+    # 1. Direct embedding comparison
+    print(f"\n1. Raw Embedding Magnitude Comparison:")
+    # Check if the distance of each token embedding from the origin and compare
+    # if the vectors are on the same "sphere". This does not tell us about
+    # direction (meaning of the token embedding), just magnitude.
+    for i in range(n_tokens):
+        py_mag = np.linalg.norm(python_emb[i]) # calculate standard euclidean norm for Python embeddings
+        cpp_mag = np.linalg.norm(cpp_emb[i])   # calculate standard euclidean norm for llama.cpp embeddings
+        ratio = py_mag / cpp_mag if cpp_mag > 0 else float('inf')
+        print(f"   Token {i} ({tokens[i]}): Python={py_mag:.3f}, llama.cpp={cpp_mag:.3f}, ratio={ratio:.3f}")
+
+    # 2. Cosine similarity between tokens within each model
+    # Here we check the direction of token embeddings to see if the have the
+    # same meaning (similarity). This is done by calculating cosine similarity
+    # of a pair of token embeddings within each model.
+    print(f"\n2. Within-Model Token Similarities:")
+    print("   Python model:")
+    for i in range(n_tokens):
+        for j in range(i+1, n_tokens):
+            sim = cosine_similarity([python_emb[i]], [python_emb[j]])[0][0]
+            print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
+
+    print("   llama.cpp model:")
+    for i in range(n_tokens):
+        for j in range(i+1, n_tokens):
+            sim = cosine_similarity([cpp_emb[i]], [cpp_emb[j]])[0][0]
+            print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
+
+    # 3. Cross-model similarity (same token position)
+    print(f"\n3. Cross-Model Same-Token Similarities:")
+    for i in range(n_tokens):
+        sim = cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0]
+        print(f"   Token {i} ({tokens[i]}): {sim:.4f}")
+
+    # 4. Similarity matrix comparison
+    print(f"\n4. Similarity Matrix Differences:")
+    py_sim_matrix = cosine_similarity(python_emb)
+    cpp_sim_matrix = cosine_similarity(cpp_emb)
+    diff_matrix = np.abs(py_sim_matrix - cpp_sim_matrix)
+
+    print(f"   Max difference: {np.max(diff_matrix):.4f}")
+    print(f"   Mean difference: {np.mean(diff_matrix):.4f}")
+    print(f"   RMS difference: {np.sqrt(np.mean(diff_matrix**2)):.4f}")
+
+    return {
+        'cross_model_similarities': [cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0] for i in range(n_tokens)],
+        'similarity_matrix_diff': diff_matrix,
+        'max_diff': np.max(diff_matrix),
+        'mean_diff': np.mean(diff_matrix),
+        'rms_diff': np.sqrt(np.mean(diff_matrix**2))
+    }
+
+def main():
+    parser = argparse.ArgumentParser(description='Test semantic similarity between Python and llama.cpp embeddings')
+    parser.add_argument('--model-path', '-m', required=True, help='Path to the original Python model')
+    parser.add_argument('--python-embeddings', '-pe', help='Path to pytorch embeddings "logits" binary file')
+    parser.add_argument('--cpp-embeddings', '-ce', help='Path to llama.cpp embeddings "logits" binary file')
+    parser.add_argument('--causal', '-c', default=False, help='if the model is causal (default: false)', action='store_true')
+    parser.add_argument('--prompt', '-p', default='Hello world today', help='Test prompt')
+
+    args = parser.parse_args()
+
+    print("Semantic Similarity Test Between Python and llama.cpp Embedding Models")
+    print("=" * 70)
+
+    # Single prompt detailed comparison
+    print(f"\nTesting with prompt: '{args.prompt}'")
+
+    # Load the python model to get configuration information and also to load the tokenizer.
+    print("Loading model and tokenizer using AutoTokenizer:", args.model_path)
+    tokenizer = AutoTokenizer.from_pretrained(args.model_path)
+    config = AutoConfig.from_pretrained(args.model_path)
+
+    if unreleased_model_name:
+        model_name_lower = unreleased_model_name.lower()
+        unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+        if args.causal:
+            class_name = f"{unreleased_model_name}ForCausalLM"
+        else:
+            class_name = f"{unreleased_model_name}Model"
+        print(f"Model class: {class_name}")
+        print(f"Importing unreleased model module: {unreleased_module_path}")
+
+        try:
+            model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+            model = model_class.from_pretrained(args.model_path)
+        except (ImportError, AttributeError) as e:
+            print(f"Failed to import or load model: {e}")
+            exit(1)
+    else:
+        if args.causal:
+            model = AutoModelForCausalLM.from_pretrained(args.model_path)
+        else:
+            model = AutoModel.from_pretrained(args.model_path)
+
+    encoded = tokenizer(args.prompt, return_tensors="pt")
+    tokens = tokenizer.convert_ids_to_tokens(encoded['input_ids'][0])
+    n_tokens = len(tokens)
+    print(f"n_tokens: {n_tokens}");
+    print(f"hidden_size: {model.config.hidden_size}")
+
+    # Load binary embeddings from data directory.
+    llamacpp_embeddings = load_embeddings_from_file(args.cpp_embeddings, n_tokens, model.config.hidden_size)
+    python_embeddings = load_embeddings_from_file(args.python_embeddings, n_tokens, model.config.hidden_size)
+
+    # Run comparison
+    results = test_single_prompt_similarity(python_embeddings, llamacpp_embeddings, tokens, args.prompt)
+
+    # Summary
+    print(f"\n=== SUMMARY ===")
+    avg_cross_sim = np.mean(results['cross_model_similarities'])
+    print(f"Average cross-model similarity: {avg_cross_sim:.4f}")
+    print(f"Similarity matrix RMS difference: {results['rms_diff']:.4f}")
+
+    # Quality assessment
+    if avg_cross_sim > 0.95:
+        print("✅ EXCELLENT: Models are highly similar")
+    elif avg_cross_sim > 0.90:
+        print("✅ VERY GOOD: Models are very similar")
+    elif avg_cross_sim > 0.80:
+        print("⚠️  GOOD: Models are reasonably similar")
+    elif avg_cross_sim > 0.70:
+        print("⚠️  FAIR: Models have some differences")
+    else:
+        print("❌ POOR: Models are significantly different")
+
+if __name__ == "__main__":
+    main()

examples/passkey/README.md

@@ -11,5 +11,5 @@ See the following PRs for more info:
 ### Usage
 
 ```bash
-make -j && ./llama-passkey -m ./models/llama-7b-v2/ggml-model-f16.gguf --junk 250
+llama-passkey -m ./models/llama-7b-v2/ggml-model-f16.gguf --junk 250
 ```

examples/retrieval/README.md

@@ -15,7 +15,7 @@ https://github.com/ggml-org/llama.cpp/pull/6193
 `retrieval` example can be tested as follows:
 
 ```bash
-make -j && ./llama-retrieval --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .
+llama-retrieval --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .
 ```
 
 This chunks and embeds all given files and starts a loop requesting query inputs:

examples/sycl/win-build-sycl.bat

@@ -18,8 +18,6 @@ if %errorlevel% neq 0 goto ERROR
 ::  for FP32
 cmake -G "Ninja" .. -DLLAMA_CURL=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release
 if %errorlevel% neq 0 goto ERROR
-::  build example/main only
-::  make main
 
 ::  build all binary
 cmake --build . -j

ggml/CMakeLists.txt

@@ -158,7 +158,6 @@ option(GGML_CUDA                            "ggml: use CUDA"
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
 option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
-option(GGML_CUDA_F16                        "ggml: use 16 bit floats for some calculations"   OFF)
 set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                             "ggml: max. batch size for using peer access")
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)

ggml/include/ggml.h

@@ -244,6 +244,13 @@
 #define GGML_MROPE_SECTIONS   4
 
 #define GGML_UNUSED(x) (void)(x)
+#ifdef __CUDACC__
+template<typename... Args>
+__host__ __device__ constexpr inline void ggml_unused_vars_impl(Args&&...) noexcept {}
+#define GGML_UNUSED_VARS(...) ggml_unused_vars_impl(__VA_ARGS__)
+#else
+#define GGML_UNUSED_VARS(...) do { (void)sizeof((__VA_ARGS__, 0)); } while(0)
+#endif // __CUDACC__
 
 #define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
 
@@ -505,6 +512,7 @@ extern "C" {
         GGML_OP_IM2COL,
         GGML_OP_IM2COL_BACK,
         GGML_OP_CONV_2D,
+        GGML_OP_CONV_3D,
         GGML_OP_CONV_2D_DW,
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
@@ -1933,6 +1941,23 @@ extern "C" {
             int                   d0,  // dilation dimension 0
             int                   d1); // dilation dimension 1
 
+    GGML_API struct ggml_tensor * ggml_conv_3d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,   // kernel [KW, KH, KD, IC * OC]
+            struct ggml_tensor  * b,   // input  [W, H, D, C * N]
+            int                   s0,  // stride
+            int                   s1,
+            int                   s2,
+            int                   p0,  // padding
+            int                   p1,
+            int                   p2,
+            int                   d0,  // dilation
+            int                   d1,
+            int                   d2,
+            int                   n_channels,
+            int                   n_batch,
+            int                   n_channels_out);
+
     enum ggml_op_pool {
         GGML_OP_POOL_MAX,
         GGML_OP_POOL_AVG,

ggml/src/ggml-backend.cpp

@@ -19,9 +19,8 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
-#include <string>
-#include <vector>
 #include <algorithm>
+#include <vector>
 
 #ifdef __APPLE__
 #include <sys/types.h>
@@ -1352,15 +1351,19 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
 static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
     struct ggml_backend_sched_split * splits = sched->splits;
 
-    for (int i = 0; i < sched->n_splits; i++) {
-        struct ggml_backend_sched_split * split = &splits[i];
+    ggml_tensor * prev_ids_tensor = nullptr;
+    std::vector<int32_t> ids;
+    std::vector<ggml_bitset_t> used_ids;
+
+    for (int split_id = 0; split_id < sched->n_splits; split_id++) {
+        struct ggml_backend_sched_split * split = &splits[split_id];
         int split_backend_id = split->backend_id;
         ggml_backend_t split_backend = sched->backends[split_backend_id];
 
         // copy the input tensors to the split backend
-        for (int j = 0; j < split->n_inputs; j++) {
-            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[j]);
-            struct ggml_tensor * input = split->inputs[j];
+        for (int input_id = 0; input_id < split->n_inputs; input_id++) {
+            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[input_id]);
+            struct ggml_tensor * input = split->inputs[input_id];
             struct ggml_tensor * input_cpy = tensor_copy(input, split_backend_id, sched->cur_copy);
 
             if (input->flags & GGML_TENSOR_FLAG_INPUT) {
@@ -1378,16 +1381,104 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 } else {
                     ggml_backend_synchronize(split_backend);
                 }
-                // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
-                // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
-                if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
+
+                // when offloading MoE weights, we can reduce the amount of data copied by copying only the experts that are used
+                ggml_tensor * node = split->graph.nodes[0];
+                if (split->graph.n_nodes > 0 &&
+                    ggml_backend_buffer_get_usage(input->buffer) == GGML_BACKEND_BUFFER_USAGE_WEIGHTS &&
+                    ggml_backend_buffer_is_host(input->buffer) && (
+                    (node->src[0] == input_cpy && node->op == GGML_OP_MUL_MAT_ID)
+                    //|| (node->src[1] == input_cpy && node->op == GGML_OP_ADD_ID) /* GGML_OP_ADD_ID weights are small and not worth splitting */
+                    )) {
+
+                    const int64_t n_expert   = node->op == GGML_OP_MUL_MAT_ID ? input->ne[2] : input->ne[1];
+                    const size_t expert_size = node->op == GGML_OP_MUL_MAT_ID ? input->nb[2] : input->nb[1];
+
                     ggml_backend_synchronize(input_backend);
-                    if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
-                        ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
-                    } else {
-                        ggml_backend_synchronize(split_backend);
+
+                    // get the ids
+                    ggml_tensor * ids_tensor = node->src[2];
+                    ggml_backend_t ids_backend = split_backend;
+
+                    // if the ids tensor is also an input of the split, it may not have been copied yet to the split backend
+                    // in that case, we use the original ids tensor
+                    for (int i = input_id + 1; i < split->n_inputs; i++) {
+                        if (ids_tensor == tensor_copy(split->inputs[i], split_backend_id, sched->cur_copy)) {
+                            ids_tensor = split->inputs[i];
+                            ids_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[i]);
+                            break;
+                        }
+                    }
+
+                    if (ids_tensor != prev_ids_tensor) {
+                        ids.resize(ggml_nbytes(ids_tensor) / sizeof(int32_t));
+                        ggml_backend_tensor_get_async(ids_backend, ids_tensor, ids.data(), 0, ggml_nbytes(ids_tensor));
+                        ggml_backend_synchronize(ids_backend);
+
+                        // find the used experts
+                        used_ids.clear();
+                        used_ids.resize(ggml_bitset_size(n_expert));
+                        for (int64_t i1 = 0; i1 < ids_tensor->ne[1]; i1++) {
+                            for (int64_t i0 = 0; i0 < ids_tensor->ne[0]; i0++) {
+                                int32_t id = ids[i1 * ids_tensor->nb[1]/sizeof(int32_t) + i0 * ids_tensor->nb[0]/sizeof(int32_t)];
+                                GGML_ASSERT(id >= 0 && id < n_expert);
+                                ggml_bitset_set(used_ids.data(), id);
+                            }
+                        }
+
+                        prev_ids_tensor = ids_tensor;
+                    }
+
+                    // group consecutive experts and copy them together
+                    auto copy_experts = [&](int32_t first_id, int32_t last_id) {
+                        const size_t expert_offset = first_id * expert_size;
+                        const size_t expert_size_copy =  (last_id - first_id + 1) * expert_size;
+                        const size_t padding = std::min<size_t>(expert_size, 512);
+                        const size_t padding_end = last_id < n_expert - 1 ? padding : 0;
+
+                        ggml_backend_tensor_set_async(split_backend,
+                            input_cpy,
+                            (const uint8_t *)input->data + expert_offset, expert_offset,
+                            // copy a bit extra at the to ensure there are no NaNs in the padding of the last expert
+                            // this is necessary for MMQ in the CUDA backend
+                            expert_size_copy + padding_end);
+                    };
+
+                    int id = 0;
+                    while (!ggml_bitset_get(used_ids.data(), id)) {
+                        id++;
+                    }
+                    int32_t first_id = id;
+                    int32_t last_id = first_id;
+
+                    for (++id; id < n_expert; ++id) {
+                        if (!ggml_bitset_get(used_ids.data(), id)) {
+                            continue;
+                        }
+
+                        if (id == last_id + 1) {
+                            last_id = id;
+                            continue;
+                        }
+
+                        copy_experts(first_id, last_id);
+
+                        first_id = id;
+                        last_id = id;
+                    }
+                    copy_experts(first_id, last_id);
+                } else {
+                    // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
+                    // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
+                    if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
+                        ggml_backend_synchronize(input_backend);
+                        if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                            ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                        } else {
+                            ggml_backend_synchronize(split_backend);
+                        }
+                        ggml_backend_tensor_copy(input, input_cpy);
                     }
-                    ggml_backend_tensor_copy(input, input_cpy);
                 }
             }
         }

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -867,6 +867,86 @@ static aclTensor* aclnn_values(ggml_backend_cann_context& ctx, void* buffer,
     return acl_tensor;
 }
 
+/**
+ * @brief Fills a tensor with a scalar value.
+ *
+ * This function fills the destination tensor `acl_dst` with the scalar value
+ * `scalar`.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param scalar The scalar value used to fill the tensor.
+ * @param acl_dst The destination tensor to be filled with the scalar value.
+ */
+static void aclnn_fill_scalar(ggml_backend_cann_context& ctx, float scalar,
+                              aclTensor* acl_dst) {
+    auto acl_scalar = aclCreateScalar(&scalar, aclDataType::ACL_FLOAT);
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceFillScalar, acl_dst, acl_scalar);
+    ggml_cann_release_resources(ctx, acl_scalar);
+}
+
+/**
+ * @brief Get or expand a cached float32 tensor filled with a scalar value.
+ *
+ * This function manages cached device memory for float32 tensors. If the current
+ * cache size is insufficient for the requested tensor shape, the old memory will
+ * be released and new memory will be allocated. The allocated buffer is then
+ * initialized either with zeros (when @p value == 0.0f) or with the given scalar
+ * value using CANN operations. Finally, an aclTensor object is created from the
+ * cached memory and returned.
+ *
+ * @param ctx           The CANN backend context that manages device memory.
+ * @param buffer        A pointer to the cached device buffer (will be allocated
+ *                      or reallocated if necessary).
+ * @param cache_element The current number of cached elements. This will be
+ *                      updated when the cache is expanded.
+ * @param ne            The tensor shape array (number of elements in each dimension).
+ * @param nb            The stride size for each dimension.
+ * @param dims          The number of tensor dimensions.
+ * @param value         The scalar value used to fill the tensor (supports zero
+ *                      initialization via memset or arbitrary values via fill_scalar).
+ * @return              An aclTensor pointer created from the cached buffer.
+ */
+static aclTensor* get_f32_cache_acl_tensor(
+    ggml_backend_cann_context& ctx,
+    void** buffer,
+    int64_t &cache_element,
+    int64_t* ne,
+    size_t* nb,
+    int64_t dims,
+    float value) {
+    // Calculate total number of elements
+    int64_t n_element = 1;
+    for (int i = 0; i < dims; i++) {
+        n_element *= ne[i];
+    }
+    size_t size = n_element * sizeof(float);
+
+    // Allocate or expand cache if needed
+    if (cache_element < n_element) {
+        if (*buffer != nullptr) {
+            aclrtFree(*buffer);
+            *buffer = nullptr;
+        }
+
+        ACL_CHECK(aclrtMalloc(buffer, size, ACL_MEM_MALLOC_HUGE_FIRST));
+        cache_element = n_element;
+
+        // Initialize cache
+        if (value == 0.0f) {
+            ACL_CHECK(aclrtMemsetAsync(*buffer, size, 0, size, ctx.stream()));
+        } else {
+            int64_t pool_ne[1] = { n_element };
+            size_t pool_nb[1] = { sizeof(float) };
+            aclTensor* acl_value = ggml_cann_create_tensor(
+                *buffer, ACL_FLOAT, sizeof(float), pool_ne, pool_nb, 1);
+            aclnn_fill_scalar(ctx, 1, acl_value);
+            ggml_cann_release_resources(ctx, acl_value);
+        }
+    }
+
+    return ggml_cann_create_tensor(*buffer, ACL_FLOAT, sizeof(float), ne, nb, dims);
+}
+
 void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src = dst->src[0];
 
@@ -875,20 +955,39 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
-    size_t one_tensor_n_bytes = src->ne[0] * ggml_element_size(src);
-    ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
 
-    aclTensor* acl_gamma = aclnn_values(
-        ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne, 1,
-        ggml_cann_type_mapping(src->type), ggml_element_size(src));
+    // build gamma, one...
+    size_t acl_gamma_nb[GGML_MAX_DIMS];
+    acl_gamma_nb[0] = sizeof(float);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        acl_gamma_nb[i] = acl_gamma_nb[i - 1] * src->ne[i - 1];
+    }
+    aclTensor* acl_gamma = get_f32_cache_acl_tensor(
+        ctx,
+        &ctx.f32_one_cache,
+        ctx.f32_one_cache_element,
+        src->ne,
+        acl_gamma_nb,
+        1,        // dims
+        1.0f      // value
+    );
+
+    // build rstd, zero...
+    size_t acl_rstd_nb[GGML_MAX_DIMS];
+    acl_rstd_nb[0] = sizeof(float);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        acl_rstd_nb[i] = acl_rstd_nb[i - 1] * src->ne[i - 1];
+    }
+    aclTensor* acl_rstd = get_f32_cache_acl_tensor(
+        ctx,
+        &ctx.f32_zero_cache,
+        ctx.f32_zero_cache_element,
+        src->ne,
+        acl_rstd_nb,
+        GGML_MAX_DIMS,
+        0.0f      // value
+    );
 
-    size_t zero_tensor_n_bytes =
-        src->ne[1] * src->ne[2] * src->ne[3] * ggml_element_size(src);
-    ggml_cann_pool_alloc zero_tensor_allocator(ctx.pool(), zero_tensor_n_bytes);
-    aclTensor* acl_rstd =
-        aclnn_zero(ctx, zero_tensor_allocator.get(), zero_tensor_n_bytes,
-                   src->ne, GGML_MAX_DIMS, ggml_cann_type_mapping(src->type),
-                   ggml_element_size(src));
     GGML_CANN_CALL_ACLNN_OP(ctx, RmsNorm, acl_src, acl_gamma, eps, acl_dst, acl_rstd);
     ggml_cann_release_resources(ctx, acl_src, acl_dst, acl_gamma, acl_rstd);
 }
@@ -903,14 +1002,13 @@ void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst,
 
     const int n_past = ((int32_t*)dst->op_params)[0];
 
-    size_t one_tensor_n_bytes = src->ne[0] * src->ne[1] * src->ne[2] *
-                                src->ne[3] * ggml_element_size(src);
-    ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
+    ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), ggml_nbytes(src));
+    void* buffer = one_tensor_allocator.get();
 
-    aclTensor* mask_tensor =
-        aclnn_values(ctx, one_tensor_allocator.get(), one_tensor_n_bytes,
-                     src->ne, GGML_MAX_DIMS, ggml_cann_type_mapping(src->type),
-                     ggml_element_size(src), value);
+    aclTensor* mask_tensor = ggml_cann_create_tensor(buffer, ggml_cann_type_mapping(src->type),
+        ggml_type_size(src->type), src->ne, src->nb, GGML_MAX_DIMS);
+
+    aclnn_fill_scalar(ctx, value, mask_tensor);
 
     aclScalar* alpha = nullptr;
     float alphaValue = 1.0f;
@@ -1159,12 +1257,20 @@ static void aclnn_exp(ggml_backend_cann_context& ctx, aclTensor* acl_src) {
 
 void aclnn_cos(ggml_backend_cann_context& ctx, aclTensor* acl_src,
                       aclTensor* acl_dst) {
-    GGML_CANN_CALL_ACLNN_OP(ctx, Cos, acl_src, acl_dst);
+    if(acl_dst == nullptr) {
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceCos, acl_src);
+    } else {
+        GGML_CANN_CALL_ACLNN_OP(ctx, Cos, acl_src, acl_dst);
+    }
 }
 
 void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
                       aclTensor* acl_dst) {
-    GGML_CANN_CALL_ACLNN_OP(ctx, Sin, acl_src, acl_dst);
+    if(acl_dst == nullptr) {
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceSin, acl_src);
+    } else {
+        GGML_CANN_CALL_ACLNN_OP(ctx, Sin, acl_src, acl_dst);
+    }
 }
 
 void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx,
@@ -1277,23 +1383,6 @@ void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx,
         tmp_permute_tensor, tmp_mul_tensor, acl_dst);
 }
 
-/**
- * @brief Fills a tensor with a scalar value.
- *
- * This function fills the destination tensor `acl_dst` with the scalar value
- * `scalar`.
- *
- * @param ctx The context for the CANN backend operations.
- * @param scalar The scalar value used to fill the tensor.
- * @param acl_dst The destination tensor to be filled with the scalar value.
- */
-static void aclnn_fill_scalar(ggml_backend_cann_context& ctx, float scalar,
-                              aclTensor* acl_dst) {
-    auto acl_scalar = aclCreateScalar(&scalar, aclDataType::ACL_FLOAT);
-    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceFillScalar, acl_dst, acl_scalar);
-    ggml_cann_release_resources(ctx, acl_scalar);
-}
-
 /**
  * @brief Raises each element of a tensor to the power of the corresponding
  * element in another tensor.
@@ -1338,17 +1427,17 @@ static void aclnn_pow_tensor_tensor(ggml_backend_cann_context& ctx,
 static void aclnn_get_slope_inner(ggml_backend_cann_context& ctx, void* slope_buffer,
     float m, int64_t size, float start, float stop, float step){
     int64_t ne[] = {size};
-    size_t nb[] = {sizeof(float)};
+    size_t nb[] = {sizeof(uint16_t)};
 
-    ggml_cann_pool_alloc arange_allocator(ctx.pool(), size * sizeof(float));
+    ggml_cann_pool_alloc arange_allocator(ctx.pool(), size * sizeof(uint16_t));
     void* arange_buffer = arange_allocator.get();
 
     aclTensor* arange_tensor = ggml_cann_create_tensor(
-        arange_buffer, ACL_FLOAT, sizeof(float), ne, nb, 1);
+        arange_buffer, ACL_FLOAT16, sizeof(uint16_t), ne, nb, 1);
     aclnn_arange(ctx, arange_tensor, start, stop, step, size);
 
     aclTensor* slope_tensor = ggml_cann_create_tensor(
-        slope_buffer, ACL_FLOAT, sizeof(float), ne, nb, 1);
+        slope_buffer, ACL_FLOAT16, sizeof(uint16_t), ne, nb, 1);
 
     aclScalar* sc = aclCreateScalar(&m, aclDataType::ACL_FLOAT);
 
@@ -2140,13 +2229,54 @@ static void aclnn_index_fill_tensor(ggml_backend_cann_context& ctx,
     ggml_cann_release_resources(ctx, acl_index, acl_value);
 }
 
+/**
+ * @brief Initializes and caches sine/cosine positional encoding values
+ *        (used in RoPE, Rotary Position Embedding) for attention layers.
+ *
+ * This function computes and caches the sin/cos values of
+ * θ = position * theta_scale for RoPE encoding. The cache is shared
+ * across attention layers, and only the first attention layer will
+ * trigger initialization. The cache includes repeated sin/cos values
+ * with different repeat methods depending on the @param is_neox flag.
+ *
+ * Steps performed by this function:
+ *   1. Identify whether the target tensor belongs to Q/K in attention
+ *      and restrict computation to the first layer only.
+ *   2. Initialize the theta scale array (arange → power → freq scaling).
+ *   3. Allocate sin/cos caches if the max prompt length increases.
+ *   4. Compute θ = position * theta_scale.
+ *   5. Compute sin(θ), cos(θ) and optionally scale by attn_factor.
+ *   6. Expand sin/cos values by repeat or repeat_interleave depending
+ *      on whether @param is_neox is enabled.
+ *   7. Store the computed values into persistent buffers
+ *      (ctx.rope_sin_ptr / ctx.rope_cos_ptr).
+ *
+ * @param ctx         The CANN backend context, holding memory pool,
+ *                    stream, and persistent buffers for rope init/cache.
+ * @param dst         The destination ggml_tensor whose computation
+ *                    depends on the cached RoPE values (usually Qcur/Kcur).
+ * @param theta_scale Scalar exponent base for computing theta scale values.
+ * @param freq_scale  Frequency scaling factor, applied to theta scale.
+ * @param attn_factor Attention scaling factor, applied to sin/cos.
+ * @param is_neox     Whether to use Neox-style repeat strategy
+ *                    (dim expansion vs repeat_interleave).
+ */
 static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
-                             aclTensor* acl_cos_repeat_tensor,
-                             aclTensor* acl_sin_repeat_tensor,
                              float theta_scale, float freq_scale,
                              float attn_factor, bool is_neox) {
     // int sin/cos cache, cache has different repeat method depond on
     // @param.is_neox
+    bool is_q = (std::strncmp(dst->name, "Qcur-", 5) == 0);
+    bool is_k = (std::strncmp(dst->name, "Kcur-", 5) == 0);
+
+    // used for accuracy testing
+    bool is_attention = is_q || is_k;
+
+    // just compute in first layer in attention
+    bool is_fisrt_layer = (std::strncmp(dst->name, "Qcur-0", GGML_MAX_NAME) == 0);
+    if(is_attention && !is_fisrt_layer) {
+        return;
+    }
 
     ggml_tensor* src0 = dst->src[0];  // input
     ggml_tensor* src1 = dst->src[1];  // position
@@ -2154,49 +2284,80 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
-    // theta_scale arange, [0,1,...,ne00/2 - 1]
     int64_t theta_scale_length = ne00 / 2;
-    ggml_cann_pool_alloc theta_scale_allocator(ctx.pool(),
-                                          theta_scale_length * sizeof(float_t));
-    void* theta_scale_buffer = theta_scale_allocator.get();
     int64_t theta_scale_ne[] = {theta_scale_length, 1, 1, 1};
     size_t theta_scale_nb[] = {sizeof(float_t), sizeof(float_t), sizeof(float_t),
                           theta_scale_length * sizeof(float_t)};
 
-    aclTensor* acl_theta_scale_tensor =
-        ggml_cann_create_tensor(theta_scale_buffer, ACL_FLOAT, sizeof(float_t),
-                                theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
-    float start = 0;
-    float step = 1;
-    float stop = ne00 / 2;
-    float n_elements = ne00 / 2;
-    aclnn_arange(ctx, acl_theta_scale_tensor, start, stop, step, n_elements);
-
-    // power
-    aclScalar* acl_theta_scale = aclCreateScalar(&theta_scale, aclDataType::ACL_FLOAT);
-    GGML_CANN_CALL_ACLNN_OP(ctx, PowScalarTensor, acl_theta_scale, acl_theta_scale_tensor,
-                            acl_theta_scale_tensor);
-
-    // freq_scale
-    if (freq_scale != 1) {
-        aclnn_muls(ctx, acl_theta_scale_tensor, freq_scale, nullptr, true);
-    }
-
-    // freq_factors
-    if (src2) {
-        aclTensor* acl_freq_factors_tensor = ggml_cann_create_tensor(
-            src2->data, ggml_cann_type_mapping(src2->type),
-            ggml_type_size(src2->type), theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
-        aclnn_div(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor);
-        ggml_cann_release_resources(ctx, acl_freq_factors_tensor);
-    }
-
-    // position
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
     int64_t position_length = src1->ne[0];
     int64_t position_ne[] = {1, 1, position_length, 1};
     size_t position_nb[] = {sizeof(int32_t), sizeof(int32_t), sizeof(int32_t),
                             sizeof(int32_t) * position_length};
+
+    int64_t theta_ne[] = {theta_scale_length, 1, position_length, 1};
+    size_t theta_nb[GGML_MAX_DIMS];
+    theta_nb[0] = sizeof(float_t);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        theta_nb[i] = theta_nb[i - 1] * theta_ne[i - 1];
+    }
+
+    // init theta scale, just one time
+    if(ctx.rope_init_ptr == nullptr || !is_attention) {
+        // theta_scale arange, [0,1,...,ne00/2 - 1]
+        if(ctx.rope_init_ptr != nullptr){
+            ACL_CHECK(aclrtFree(ctx.rope_init_ptr));
+        }
+        ACL_CHECK(aclrtMalloc(&ctx.rope_init_ptr, theta_scale_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
+
+        aclTensor* acl_theta_scale_tensor =
+            ggml_cann_create_tensor(ctx.rope_init_ptr, ACL_FLOAT, sizeof(float_t),
+                                    theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
+        float start = 0;
+        float step = 1;
+        float stop = ne00 / 2;
+        float n_elements = ne00 / 2;
+        aclnn_arange(ctx, acl_theta_scale_tensor, start, stop, step, n_elements);
+
+        // power
+        aclScalar* acl_theta_scale = aclCreateScalar(&theta_scale, aclDataType::ACL_FLOAT);
+        GGML_CANN_CALL_ACLNN_OP(ctx, PowScalarTensor, acl_theta_scale, acl_theta_scale_tensor,
+                                acl_theta_scale_tensor);
+
+        // freq_scale
+        if (freq_scale != 1) {
+            aclnn_muls(ctx, acl_theta_scale_tensor, freq_scale, nullptr, true);
+        }
+
+        // freq_factors
+        if (src2) {
+            aclTensor* acl_freq_factors_tensor = ggml_cann_create_tensor(
+                src2->data, ggml_cann_type_mapping(src2->type),
+                ggml_type_size(src2->type), theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
+            aclnn_div(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor);
+            ggml_cann_release_resources(ctx, acl_freq_factors_tensor);
+        }
+        // release
+        ggml_cann_release_resources(ctx, acl_theta_scale_tensor,acl_theta_scale);
+    }
+
+    // init sin_repeat && cos_repeat, one token just init in 0 layer
+    if(position_length > ctx.max_prompt_length) {
+        ctx.max_prompt_length = position_length;
+        int64_t repeat_theta_length = theta_scale_length * ctx.max_prompt_length * 2;
+        if(ctx.rope_sin_ptr != nullptr) {
+            ACL_CHECK(aclrtFree(ctx.rope_sin_ptr));
+            ACL_CHECK(aclrtFree(ctx.rope_cos_ptr));
+        }
+        ACL_CHECK(aclrtMalloc(&ctx.rope_sin_ptr, repeat_theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
+        ACL_CHECK(aclrtMalloc(&ctx.rope_cos_ptr, repeat_theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
+    }
+
+    aclTensor* acl_theta_scale_tensor =
+            ggml_cann_create_tensor(ctx.rope_init_ptr, ACL_FLOAT, sizeof(float_t),
+                                    theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
+
+    // position
     aclTensor* acl_position_tensor = ggml_cann_create_tensor(
         src1->data, ggml_cann_type_mapping(src1->type),
         ggml_type_size(src1->type), position_ne, position_nb, GGML_MAX_DIMS);
@@ -2204,23 +2365,18 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
     // power * position
     int64_t theta_length = theta_scale_length * position_length;
     ggml_cann_pool_alloc theta_allocator(ctx.pool(),
-                                         theta_length * sizeof(float_t));
+                                        theta_length * sizeof(float_t));
     void* theta_buffer = theta_allocator.get();
-    int64_t theta_ne[] = {theta_scale_length, 1, position_length, 1};
-    size_t theta_nb[GGML_MAX_DIMS];
-    theta_nb[0] = sizeof(float_t);
-    for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        theta_nb[i] = theta_nb[i - 1] * theta_ne[i - 1];
-    }
+
     aclTensor* acl_theta_tensor =
         ggml_cann_create_tensor(theta_buffer, ACL_FLOAT, sizeof(float_t),
                                 theta_ne, theta_nb, GGML_MAX_DIMS);
     aclnn_mul(ctx, acl_position_tensor, acl_theta_scale_tensor,
-              acl_theta_tensor);
+            acl_theta_tensor);
 
     // sin/cos
     ggml_cann_pool_alloc sin_allocator(ctx.pool(),
-                                       theta_length * sizeof(float_t));
+                                    theta_length * sizeof(float_t));
     void* sin_buffer = sin_allocator.get();
     aclTensor* acl_sin_tensor = ggml_cann_create_tensor(
         sin_buffer, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
@@ -2228,7 +2384,7 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
     aclnn_sin(ctx, acl_theta_tensor, acl_sin_tensor);
 
     ggml_cann_pool_alloc cos_allocator(ctx.pool(),
-                                       theta_length * sizeof(float_t));
+                                    theta_length * sizeof(float_t));
     void* cos_buffer = cos_allocator.get();
     aclTensor* acl_cos_tensor = ggml_cann_create_tensor(
         cos_buffer, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
@@ -2241,6 +2397,19 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
         aclnn_muls(ctx, acl_cos_tensor, attn_factor, nullptr, true);
     }
 
+    int64_t sin_reshape_ne[4] = {ne00, 1, ne02, 1};
+    size_t sin_reshape_nb[GGML_MAX_DIMS];
+    sin_reshape_nb[0] = sizeof(float_t);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        sin_reshape_nb[i] = sin_reshape_nb[i - 1] * sin_reshape_ne[i - 1];
+    }
+    aclTensor* acl_sin_repeat_tensor =
+        ggml_cann_create_tensor(ctx.rope_sin_ptr, ACL_FLOAT, sizeof(float_t),
+                                sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
+    aclTensor* acl_cos_repeat_tensor =
+        ggml_cann_create_tensor(ctx.rope_cos_ptr, ACL_FLOAT, sizeof(float_t),
+                                sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
+
     // repeat
     if (is_neox) {
         int64_t repeatsArray[] = {1, 1, 1, 2};
@@ -2256,9 +2425,9 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
                                 num_repeats, output_size);
     }
 
-    // release
     ggml_cann_release_resources(ctx, acl_theta_scale_tensor, acl_position_tensor,
-        acl_theta_tensor, acl_sin_tensor, acl_cos_tensor, acl_theta_scale);
+        acl_theta_tensor, acl_sin_tensor, acl_sin_repeat_tensor, acl_cos_tensor,
+        acl_cos_repeat_tensor);
 }
 
 #ifdef __cplusplus
@@ -2312,13 +2481,8 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
-    // init cos/sin cache
-    ggml_cann_pool_alloc sin_allocator(
-        ctx.pool(), ne00 * ne02 * sizeof(float_t));
-    ggml_cann_pool_alloc cos_allocator(
-        ctx.pool(), ne00 * ne02 * sizeof(float_t));
-    void* sin_buffer = sin_allocator.get();
-    void* cos_buffer = cos_allocator.get();
+    // init ctx.rope_cos/rope_sin cache
+    aclnn_cache_init(ctx, dst, theta_scale, freq_scale, attn_factor, is_neox);
 
     int64_t sin_reshape_ne[4] = {ne00, 1, ne02, 1};
     size_t sin_reshape_nb[GGML_MAX_DIMS];
@@ -2327,13 +2491,11 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         sin_reshape_nb[i] = sin_reshape_nb[i - 1] * sin_reshape_ne[i - 1];
     }
     aclTensor* acl_sin_reshape_tensor =
-        ggml_cann_create_tensor(sin_buffer, ACL_FLOAT, sizeof(float_t),
+        ggml_cann_create_tensor(ctx.rope_sin_ptr, ACL_FLOAT, sizeof(float_t),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
     aclTensor* acl_cos_reshape_tensor =
-        ggml_cann_create_tensor(cos_buffer, ACL_FLOAT, sizeof(float_t),
+        ggml_cann_create_tensor(ctx.rope_cos_ptr, ACL_FLOAT, sizeof(float_t),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
-    aclnn_cache_init(ctx, dst, acl_cos_reshape_tensor, acl_sin_reshape_tensor,
-                    theta_scale, freq_scale, attn_factor, is_neox);
 
     aclTensor* acl_src = ggml_cann_create_tensor(src0);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
@@ -3018,11 +3180,38 @@ void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
 void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
 
-    ggml_tensor* src0 = dst->src[0]; // q, fp32
-    ggml_tensor* src1 = dst->src[1]; // k, fp16
-    ggml_tensor* src2 = dst->src[2]; // v, fp16
+    ggml_tensor* src0 = dst->src[0]; // q, fp32 | B, N, S, D (uncont) -> B, S, N, D (cont)
+    ggml_tensor* src1 = dst->src[1]; // k, fp16 | B, N, S, D (uncont) -> B, S, N, D (cont)
+    ggml_tensor* src2 = dst->src[2]; // v, fp16 | B, N, S, D (uncont) -> B, S, N, D (cont)
     ggml_tensor* src3 = dst->src[3]; // mask, fp16
 
+    // B, N, S, D (uncont) -> B, S, N, D (cont)
+    int64_t src0_bsnd_ne[GGML_MAX_DIMS];
+    memcpy(src0_bsnd_ne, src0->ne, GGML_MAX_DIMS * sizeof(int64_t));
+    size_t src0_bsnd_nb[GGML_MAX_DIMS];
+    memcpy(src0_bsnd_nb, src0->nb, GGML_MAX_DIMS * sizeof(size_t));
+    int64_t src1_bsnd_ne[GGML_MAX_DIMS];
+    memcpy(src1_bsnd_ne, src1->ne, GGML_MAX_DIMS * sizeof(int64_t));
+    size_t src1_bsnd_nb[GGML_MAX_DIMS];
+    memcpy(src1_bsnd_nb, src1->nb, GGML_MAX_DIMS * sizeof(size_t));
+    int64_t src2_bsnd_ne[GGML_MAX_DIMS];
+    memcpy(src2_bsnd_ne, src2->ne, GGML_MAX_DIMS * sizeof(int64_t));
+    size_t src2_bsnd_nb[GGML_MAX_DIMS];
+    memcpy(src2_bsnd_nb, src2->nb, GGML_MAX_DIMS * sizeof(size_t));
+
+    auto transpose12 = [](int64_t* ne, size_t* nb) {
+        int64_t ne_tmp = ne[1];
+        size_t  nb_tmp = nb[1];
+        ne[1] = ne[2];
+        nb[1] = nb[2];
+        ne[2] = ne_tmp;
+        nb[2] = nb_tmp;
+    };
+
+    transpose12(src0_bsnd_ne, src0_bsnd_nb);
+    transpose12(src1_bsnd_ne, src1_bsnd_nb);
+    transpose12(src2_bsnd_ne, src2_bsnd_nb);
+
     float maxBias = 0.0f;
     float scaleValue = 1.0f;
     float logitSoftcap = 0.0f;
@@ -3044,11 +3233,12 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
         void* src0_f16_buffer = nullptr;
 
         if(ggml_cann_type_mapping(src0->type) != faDataType){
-            aclTensor* acl_src0_f32_tensor = ggml_cann_create_tensor(src0);
+            aclTensor* acl_src0_f32_tensor = ggml_cann_create_tensor(src0, src0_bsnd_ne,
+                src0_bsnd_nb, GGML_MAX_DIMS);
             src0_f16_buffer = src0_f16_allocator.alloc(
                                     ggml_nelements(src0) * faElemSize);
 
-            int64_t* src0_f16_ne = src0->ne;
+            int64_t* src0_f16_ne = src0_bsnd_ne;
             size_t   src0_f16_nb[GGML_MAX_DIMS];
             src0_f16_nb[0] = sizeof(uint16_t);
             for(int i = 1; i < GGML_MAX_DIMS; ++i){
@@ -3062,20 +3252,23 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
             aclnn_cast(ctx, acl_src0_f32_tensor, acl_src0_f16_tensor, faDataType);
             ggml_cann_release_resources(ctx, acl_src0_f32_tensor);
         }else{
-            acl_src0_f16_tensor = ggml_cann_create_tensor(src0);
+            acl_src0_f16_tensor = ggml_cann_create_tensor(src0, src0_bsnd_ne,
+                src0_bsnd_nb, GGML_MAX_DIMS);
         }
 
         // Step 2: create the acl tensors for src1 (Key), src2 (Value),
         //         and the direct output from FusedInferAttention
 
-        acl_src1_f16_tensor = ggml_cann_create_tensor(src1);
-        acl_src2_f16_tensor = ggml_cann_create_tensor(src2);
+        acl_src1_f16_tensor = ggml_cann_create_tensor(src1, src1_bsnd_ne,
+            src1_bsnd_nb, GGML_MAX_DIMS);
+        acl_src2_f16_tensor = ggml_cann_create_tensor(src2, src2_bsnd_ne,
+            src2_bsnd_nb, GGML_MAX_DIMS);
 
         ggml_cann_pool_alloc out_f16_allocator(ctx.pool());
         void* out_f16_buffer = out_f16_allocator.alloc(
                                     ggml_nelements(dst) * faElemSize);
 
-        int64_t* out_f16_ne = src0->ne;
+        int64_t* out_f16_ne = src0_bsnd_ne;
         size_t out_f16_nb[GGML_MAX_DIMS];
         out_f16_nb[0] = faElemSize;
         for(int i = 1; i < GGML_MAX_DIMS; ++i){
@@ -3089,88 +3282,81 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
 
         // Step 3: create the PSEShift tensor if needed
         //         this tensor is considered as mask (f16) in the llama.cpp
-
         aclTensor* bcast_pse_tensor = nullptr;
-        int64_t bcast_pse_ne[GGML_MAX_DIMS];
-        size_t bcast_pse_nb[GGML_MAX_DIMS];
         ggml_cann_pool_alloc bcast_pse_allocator(ctx.pool());
-        void* bcast_pse_buffer = nullptr;
-
         if(src3 != nullptr){
-            bcast_pse_buffer = bcast_pse_allocator.alloc(
-                            ggml_nelements(src3) * src0->ne[2] * sizeof(uint16_t));
+            // Construct the truncated pse tensor (common for prefill/decode)
+            int64_t trunc_pse_ne[GGML_MAX_DIMS] = {
+                src3->ne[0],        // D
+                src0->ne[1],        // S (number of Q tokens)
+                src3->ne[2],        // mask N
+                src3->ne[3]         // B
+            };
+            size_t* trunc_pse_nb = src3->nb;
 
-            if(src0->ne[1] > 1){
-                // Case 1: broadcast pse for prefill stage with multiple head
-                aclTensor* acl_mask_f16_tensor = ggml_cann_create_tensor(src3);
-                bcast_pse_ne[0] = src3->ne[0];
-                bcast_pse_ne[1] = src3->ne[1];
-                bcast_pse_ne[2] = src0->ne[2];
-                bcast_pse_ne[3] = src3->ne[3];
+            aclTensor* acl_mask_f16_trunc_tensor = ggml_cann_create_tensor(
+                src3->data, ACL_FLOAT16, sizeof(uint16_t),
+                trunc_pse_ne, trunc_pse_nb, GGML_MAX_DIMS
+            );
 
+            int64_t bcast_pse_ne[GGML_MAX_DIMS];
+            size_t bcast_pse_nb[GGML_MAX_DIMS];
+            bcast_pse_ne[0] = src3->ne[0];      // D
+            bcast_pse_ne[1] = src0->ne[1];      // S
+            bcast_pse_ne[2] = src0->ne[2];      // N (num_heads)
+            bcast_pse_ne[3] = src3->ne[3];      // B
+            if (maxBias == 0.0f) {
+                // When maxBias == 0.0f, use nb = 0 reduce once repeat (Qwen2)
+                // Construct the bcast tensor (simulate repeat on the head dimension using stride=0)
                 bcast_pse_nb[0] = sizeof(uint16_t);
-                for(int i = 1; i < GGML_MAX_DIMS; ++i){
-                    bcast_pse_nb[i] = bcast_pse_nb[i - 1] * bcast_pse_ne[i - 1];
-                }
+                bcast_pse_nb[1] = bcast_pse_nb[0] * bcast_pse_ne[0];
+                bcast_pse_nb[2] = 0;                // <---- the head dimension shares the same data
+                bcast_pse_nb[3] = src3->nb[3];
 
                 bcast_pse_tensor = ggml_cann_create_tensor(
-                    bcast_pse_buffer, ACL_FLOAT16, sizeof(uint16_t),
-                    bcast_pse_ne, bcast_pse_nb, GGML_MAX_DIMS);
-
-                int64_t repeats[] = {1, src0->ne[2], 1, 1};
-                aclnn_repeat(ctx, acl_mask_f16_tensor, bcast_pse_tensor, repeats);
-
-                ggml_cann_release_resources(ctx, acl_mask_f16_tensor);
-            }else{
-                // Case 2: trunc the first row and broadcast pse for decode stage with multiple head
-                int64_t trunc_pse_ne[GGML_MAX_DIMS] = {src3->ne[0], src0->ne[1], src3->ne[2], src3->ne[3]};
-                size_t* trunc_pse_nb = src3->nb;
-
-                aclTensor* acl_mask_f16_trunc_tensor = ggml_cann_create_tensor(
                     src3->data, ACL_FLOAT16, sizeof(uint16_t),
-                    trunc_pse_ne, trunc_pse_nb, GGML_MAX_DIMS);
-
-                bcast_pse_ne[0] = src3->ne[0];
-                bcast_pse_ne[1] = src0->ne[1];
-                bcast_pse_ne[2] = src0->ne[2];
-                bcast_pse_ne[3] = src3->ne[3];
+                    bcast_pse_ne, bcast_pse_nb, GGML_MAX_DIMS
+                );
 
+                ggml_cann_release_resources(ctx, acl_mask_f16_trunc_tensor);
+            } else {
                 bcast_pse_nb[0] = sizeof(uint16_t);
-                for(int i = 1; i < GGML_MAX_DIMS; ++i){
+                for (int i = 1; i < GGML_MAX_DIMS; i++) {
                     bcast_pse_nb[i] = bcast_pse_nb[i - 1] * bcast_pse_ne[i - 1];
                 }
 
+                void* bcast_pse_buffer = bcast_pse_allocator.alloc(
+                    ggml_nelements(src3) * src0->ne[2] * sizeof(uint16_t)
+                );
+
                 bcast_pse_tensor = ggml_cann_create_tensor(
                     bcast_pse_buffer, ACL_FLOAT16, sizeof(uint16_t),
-                    bcast_pse_ne, bcast_pse_nb, GGML_MAX_DIMS);
+                    bcast_pse_ne, bcast_pse_nb, GGML_MAX_DIMS
+                );
 
                 int64_t repeats[] = {1, src0->ne[2], 1, 1};
                 aclnn_repeat(ctx, acl_mask_f16_trunc_tensor, bcast_pse_tensor, repeats);
 
-                ggml_cann_release_resources(ctx, acl_mask_f16_trunc_tensor);
-            }
-
-            // Compute the slope if needed. Derived from ggml_cann_softmax().
-            if(maxBias != 0.0f){
                 // alibi
+                // Compute the slope if needed. Derived from ggml_cann_softmax().
                 const int64_t n_heads = src0->ne[2];
-                ggml_cann_pool_alloc slope_allocator(ctx.pool(), n_heads * sizeof(float));
+                ggml_cann_pool_alloc slope_allocator(ctx.pool(), n_heads * sizeof(uint16_t));
                 void* slope_buffer = slope_allocator.get();
                 aclnn_get_slope(ctx, n_heads, slope_buffer, maxBias);
 
                 int64_t slope_ne[] = {1, 1, n_heads, 1};
                 size_t slope_nb[GGML_MAX_DIMS];
-                slope_nb[0] = sizeof(float);
+                slope_nb[0] = sizeof(uint16_t);
                 for(int i = 1;i<GGML_MAX_DIMS;i++) {
                     slope_nb[i] = slope_nb[i-1] * slope_ne[0];
                 }
 
                 aclTensor* slope_tensor = ggml_cann_create_tensor(
-                    slope_buffer, ACL_FLOAT, sizeof(float),
+                    slope_buffer, ACL_FLOAT16, sizeof(uint16_t),
                     slope_ne, slope_nb, GGML_MAX_DIMS);
                 GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, bcast_pse_tensor, slope_tensor);
 
-                ggml_cann_release_resources(ctx, slope_tensor);
+                ggml_cann_release_resources(ctx, slope_tensor, acl_mask_f16_trunc_tensor);
             }
         }
 
@@ -3187,7 +3373,7 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
         // double  scaleValue = 1 / sqrt(src0->ne[0]); // 1/sqrt(d)
         int64_t preTokens = 65535;
         int64_t nextTokens = 65535;
-        char layout[5] = {'B', 'N', 'S', 'D', 0};
+        char layout[5] = {'B', 'S', 'N', 'D', 0};
         int64_t sparseMode = 0;
         int64_t innerPrecise = (src0->ne[1] == 1) ? 0 : 2;
         int64_t blockSize = 0;
@@ -3224,32 +3410,9 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
         );
 
         // Step 6: post-processing, permute and cast to f32
-
-        int64_t new_dim[] = {0, 2, 1, 3};
         aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
-
-        if(ggml_cann_type_mapping(dst->type) != faDataType){
-            ggml_cann_pool_alloc perm_out_f16_allocator(ctx.pool());
-            perm_out_f16_allocator.alloc(ggml_nelements(dst) * faElemSize);
-            void* perm_out_f16_buffer = perm_out_f16_allocator.get();
-
-            int64_t* perm_out_f16_ne = dst->ne;
-            size_t  perm_out_f16_nb[GGML_MAX_DIMS];
-            perm_out_f16_nb[0] = faElemSize;
-            for(int i = 1; i < GGML_MAX_DIMS; ++i){
-                perm_out_f16_nb[i] = perm_out_f16_nb[i - 1] * perm_out_f16_ne[i - 1];
-            }
-            aclTensor* acl_perm_out_f16_tensor = ggml_cann_create_tensor(
-                perm_out_f16_buffer, faDataType, faElemSize,
-                perm_out_f16_ne, perm_out_f16_nb, GGML_MAX_DIMS);
-            aclnn_permute(ctx, acl_dst_f16_tensor, acl_perm_out_f16_tensor, new_dim, GGML_MAX_DIMS);
-            aclnn_cast(ctx,
-                acl_perm_out_f16_tensor, acl_dst_tensor, ggml_cann_type_mapping(dst->type));
-            ggml_cann_release_resources(ctx, acl_perm_out_f16_tensor);
-        }else{
-            // only need to permute
-            aclnn_permute(ctx, acl_dst_f16_tensor, acl_dst_tensor, new_dim, GGML_MAX_DIMS);
-        }
+        // TODO: when dst is fp16, don't need cast
+        aclnn_cast(ctx, acl_dst_f16_tensor, acl_dst_tensor, ggml_cann_type_mapping(dst->type));
         ggml_cann_release_resources(ctx, acl_src0_f16_tensor,
                                          acl_src1_f16_tensor,
                                          acl_src2_f16_tensor,

ggml/src/ggml-cann/common.h

@@ -374,7 +374,16 @@ struct ggml_backend_cann_context {
 #endif
     cann_task_queue task_queue;
     bool async_mode;
-    bool support_set_rows;
+    // Rope Cache
+    void* rope_init_ptr = nullptr;
+    void* rope_sin_ptr = nullptr;
+    void* rope_cos_ptr = nullptr;
+    int64_t max_prompt_length = 0;
+    // Constant Pool
+    void* f32_zero_cache = nullptr;
+    void* f32_one_cache = nullptr;
+    int64_t f32_zero_cache_element = 0;
+    int64_t f32_one_cache_element = 0;
 
     aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
 
@@ -390,14 +399,6 @@ struct ggml_backend_cann_context {
         async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
         GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
             device, async_mode ? "ON" : "OFF");
-
-        support_set_rows = parse_bool(get_env("LLAMA_SET_ROWS").value_or(""));
-        GGML_LOG_INFO("%s: LLAMA_SET_ROWS is %s\n", __func__, support_set_rows ? "ON" : "OFF");
-
-        if (!support_set_rows) {
-            GGML_LOG_INFO("%s: CANN Graph currently only supports execution when LLAMA_SET_ROWS is ON. "
-                    "Falling back to eager mode.\n", __func__);
-        }
     }
 
     /**
@@ -414,6 +415,21 @@ struct ggml_backend_cann_context {
                 ACL_CHECK(aclrtDestroyStream(streams[i]));
             }
         }
+        if(rope_init_ptr != nullptr) {
+            ACL_CHECK(aclrtFree(rope_init_ptr));
+        }
+        if(rope_sin_ptr != nullptr) {
+            ACL_CHECK(aclrtFree(rope_sin_ptr));
+        }
+        if(rope_cos_ptr != nullptr) {
+            ACL_CHECK(aclrtFree(rope_cos_ptr));
+        }
+        if(f32_zero_cache != nullptr) {
+            ACL_CHECK(aclrtFree(f32_zero_cache));
+        }
+        if(f32_one_cache != nullptr) {
+            ACL_CHECK(aclrtFree(f32_one_cache));
+        }
     }
 
     /**

ggml/src/ggml-cann/ggml-cann.cpp

@@ -2251,11 +2251,6 @@ static enum ggml_status ggml_backend_cann_graph_compute(
     bool use_cann_graph = true;
     bool cann_graph_update_required = false;
 
-    // check environment LLAMA_SET_ROWS
-    if (!cann_ctx->support_set_rows) {
-        use_cann_graph = false;
-    }
-
     if (use_cann_graph) {
         if (cann_ctx->cann_graph == nullptr) {
             cann_ctx->cann_graph.reset(new ggml_cann_graph());
@@ -2336,7 +2331,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 case GGML_TYPE_Q8_0:
                 case GGML_TYPE_Q4_0:
 #ifdef ASCEND_310P
-                    // Q4 && Q8 per group is not suppor on 310p device
+                    // Q4 && Q8 per group is not support on 310p device
                     return false;
 #endif
                     // only support contiguous for quantized types.
@@ -2354,7 +2349,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 case GGML_TYPE_Q8_0:
                 case GGML_TYPE_Q4_0:
 #ifdef ASCEND_310P
-                    // Q4 && Q8 per group is not suppor on 310p device
+                    // Q4 && Q8 per group is not support on 310p device
                     return false;
 #endif
                     // only support contiguous for quantized types.
@@ -2505,6 +2500,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             }
             return true;
         case GGML_OP_FLASH_ATTN_EXT:{
+#ifdef ASCEND_310P
+            // FA not support on 310p device
+            return false;
+#endif
             // derived from [ggml-cuda.cu]
             if(op->src[1]->type != GGML_TYPE_F16 || op->src[2]->type != GGML_TYPE_F16){
                 return false;
@@ -2530,6 +2529,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 // DeepSeek MLA
                 return false;
             }
+            if (op->src[0]->ne[0] % 16 != 0) {
+                // TODO: padding to support
+                return false;
+            }
             float logitSoftcap = 0.0f;
             memcpy(&logitSoftcap,  (float*)op->op_params + 2, sizeof(float));
             if(logitSoftcap != 0.0f) {

ggml/src/ggml-cpu/CMakeLists.txt

@@ -435,7 +435,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             )
         if (GGML_RVV)
             if (GGML_XTHEADVECTOR)
-                list(APPEND ARCH_FLAGS -march=rv64gc_xtheadvector -mabi=lp64d)
+                list(APPEND ARCH_FLAGS -march=rv64gc_zfhmin_xtheadvector -mabi=lp64d)
             elseif (GGML_RV_ZFH)
                 list(APPEND ARCH_FLAGS -march=rv64gcv_zfhmin -mabi=lp64d)
             else()

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

@@ -73,7 +73,6 @@
 #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
-#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 // 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
@@ -151,8 +150,6 @@
 #elif defined(__s390x__)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
-#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_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

ggml/src/ggml-cpu/arch/powerpc/quants.c

@@ -278,6 +278,72 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
 }
 
+void ggml_vec_dot_mxfp4_q8_0(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) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_MXFP4 == 0);
+    static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same");
+
+    const block_mxfp4 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_MXFP4;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector unsigned char vshift4 = vec_splats((unsigned char)4);
+    vector float vsumf0 = vec_splats(0.0f);
+
+    vector signed char kv = vec_xl(0, (const signed char *)kvalues_mxfp4);
+
+#pragma GCC unroll 8
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d) *
+                                      GGML_E8M0_TO_FP32_HALF(x[ib].e));
+
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
+
+        vector signed char qxs = (vector signed char)vec_xl(0, x[ib].qs);
+
+        vector unsigned char lo_nibbles = (vector unsigned char)vec_and(qxs, lowMask);
+        vector unsigned char hi_nibbles = (vector unsigned char)vec_sr(qxs, vshift4);
+
+        vector signed char q4x0 = vec_perm(kv, kv, lo_nibbles);
+        vector signed char q4x1 = vec_perm(kv, kv, hi_nibbles);
+
+        vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
+        vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
+
+        vector signed int vsumi0 = vec_splats((int32_t)0);
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi0 = vec_sum4s(qv1, vsumi0);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vyd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+    sumf = vec_extract(vsumf0, 0);
+    *s = sumf;
+#else
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_mxfp4_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
 void ggml_vec_dot_q5_0_q8_0(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) {
     const int qk = QK8_0;
     const int nb = n / qk;

ggml/src/ggml-cpu/arch/s390/quants.c

@@ -23,6 +23,27 @@
 
 #define UNUSED GGML_UNUSED
 
+#if defined(__VXE__) || defined(__VXE2__)
+#define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
+#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
+#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
+#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s)
+#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s)
+#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s)
+#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s)
+#define B8(c,s  ) B7(c,s,     c), B7(c,s,     s)
+
+// precomputed tables for expanding 8bits to 8 bytes:
+static const __attribute__((aligned(16))) uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b ) << 4
+static const __attribute__((aligned(16))) uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
+
+// permute mask for byteswapping
+static const uint8x16_t v_kperm = (const uint8x16_t){
+     7,  6,  5,  4,  3,  2, 1, 0,
+    15, 14, 13, 12, 11, 10, 9, 8
+};
+#endif
+
 void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK8_0 == 32);
     assert(k % QK8_0 == 0);
@@ -241,6 +262,301 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
 }
 
+void ggml_vec_dot_q5_0_q8_0(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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0.0f;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    float32x4_t v_sum0 = vec_splats(0.0f);
+    float32x4_t v_sum1 = vec_splats(0.0f);
+
+    uint32_t qh0, qh1;
+    uint64_t tmp0[4], tmp1[4];
+
+    const uint8x16_t v_m = vec_splats((uint8_t)0x0F);
+
+    #pragma GCC unroll 4
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q5_0 * GGML_RESTRICT x0 = &x[ib + 0];
+        const block_q5_0 * GGML_RESTRICT x1 = &x[ib + 1];
+        const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0];
+        const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1];
+
+        memcpy(&qh0, x0->qh, sizeof(qh0));
+        memcpy(&qh1, x1->qh, sizeof(qh1));
+
+        tmp0[0] = table_b2b_1[(qh0 >>  0) & 0xFF];
+        tmp0[1] = table_b2b_1[(qh0 >>  8) & 0xFF];
+        tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF];
+        tmp0[3] = table_b2b_1[(qh0 >> 24)       ];
+
+        tmp1[0] = table_b2b_1[(qh1 >>  0) & 0xFF];
+        tmp1[1] = table_b2b_1[(qh1 >>  8) & 0xFF];
+        tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF];
+        tmp1[3] = table_b2b_1[(qh1 >> 24)       ];
+
+        int8x16_t v_qh0l = vec_xl(0, (const int8_t *)(tmp0 + 0));
+        int8x16_t v_qh0h = vec_xl(0, (const int8_t *)(tmp0 + 2));
+        int8x16_t v_qh1l = vec_xl(0, (const int8_t *)(tmp1 + 0));
+        int8x16_t v_qh1h = vec_xl(0, (const int8_t *)(tmp1 + 2));
+
+        // required for fixing the byteorder
+        v_qh0l = vec_perm(v_qh0l, v_qh0l, v_kperm);
+        v_qh0h = vec_perm(v_qh0h, v_qh0h, v_kperm);
+        v_qh1l = vec_perm(v_qh1l, v_qh1l, v_kperm);
+        v_qh1h = vec_perm(v_qh1h, v_qh1h, v_kperm);
+
+        const uint8x16_t v_x0 = vec_xl(0, (const uint8_t *)x0->qs);
+        const uint8x16_t v_x1 = vec_xl(0, (const uint8_t *)x1->qs);
+
+        int8x16_t v_x0l = (int8x16_t)vec_and(v_x0, v_m);
+        int8x16_t v_x0h = (int8x16_t)vec_sr(v_x0, 4);
+        int8x16_t v_x1l = (int8x16_t)vec_and(v_x1, v_m);
+        int8x16_t v_x1h = (int8x16_t)vec_sr(v_x1, 4);
+
+        const int8x16_t v_x0lf = vec_sub(v_x0l, v_qh0l);
+        const int8x16_t v_x0hf = vec_sub(v_x0h, v_qh0h);
+        const int8x16_t v_x1lf = vec_sub(v_x1l, v_qh1l);
+        const int8x16_t v_x1hf = vec_sub(v_x1h, v_qh1h);
+
+        const int8x16_t v_y0l = vec_xl(0,       (const int8_t *)y0->qs);
+        const int8x16_t v_y0h = vec_xl(QK8_0/2, (const int8_t *)y0->qs);
+        const int8x16_t v_y1l = vec_xl(0,       (const int8_t *)y1->qs);
+        const int8x16_t v_y1h = vec_xl(QK8_0/2, (const int8_t *)y1->qs);
+
+        const int32x4_t v_xy0 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x0lf, v_y0l), v_x0hf, v_y0h);
+        const int32x4_t v_xy1 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x1lf, v_y1l), v_x1hf, v_y1h);
+
+        const float32x4_t v_xy0f = vec_float(v_xy0);
+        const float32x4_t v_xy1f = vec_float(v_xy1);
+
+        const float32x4_t v_d0 = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_d1 = vec_splats(GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_FP16_TO_FP32(y1->d));
+
+        v_sum0 = vec_madd(v_xy0f, v_d0, v_sum0);
+        v_sum1 = vec_madd(v_xy1f, v_d1, v_sum1);
+    }
+
+    sumf += vec_hsum(v_sum0) + vec_hsum(v_sum1);
+
+    #pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        const block_q5_0 * GGML_RESTRICT x0 = &x[ib];
+        const block_q8_0 * GGML_RESTRICT y0 = &y[ib];
+
+        uint32_t qh;
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        uint64_t tmp[4];
+        tmp[0] = table_b2b_1[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_1[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_1[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_1[(qh >> 24)       ];
+
+        int8x16_t v_qhl = vec_xl(0, (const int8_t *)(tmp + 0));
+        int8x16_t v_qhh = vec_xl(0, (const int8_t *)(tmp + 2));
+
+        // required for fixing the byteorder
+        v_qhl = vec_perm(v_qhl, v_qhl, v_kperm);
+        v_qhh = vec_perm(v_qhh, v_qhh, v_kperm);
+
+        const uint8x16_t v_x = vec_xl(0, (const uint8_t *)x0->qs);
+        int8x16_t v_xl = (int8x16_t)vec_and(v_x, v_m);
+        int8x16_t v_xh = (int8x16_t)vec_sr(v_x, 4);
+
+        const int8x16_t v_xlf = vec_sub(v_xl, v_qhl);
+        const int8x16_t v_xhf = vec_sub(v_xh, v_qhh);
+
+        const int8x16_t v_yl = vec_xl(0,       (const int8_t *)y0->qs);
+        const int8x16_t v_yh = vec_xl(QK8_0/2, (const int8_t *)y0->qs);
+
+        const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xlf, v_yl), v_xhf, v_yh);
+        const float32x4_t v_xyf = vec_float(v_xy);
+
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_acc = vec_madd(v_xyf, v_d, vec_splats(0.0f));
+
+        sumf += vec_hsum(v_acc);
+    }
+
+    *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_q5_1_q8_1(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) {
+    const int qk = QK8_1;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0.0f;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    float32x4_t v_sum0 = vec_splats(0.0f);
+    float32x4_t v_sum1 = vec_splats(0.0f);
+
+    float summs0 = 0.0f;
+    float summs1 = 0.0f;
+
+    uint32_t qh0;
+    uint32_t qh1;
+
+    uint64_t tmp0[4];
+    uint64_t tmp1[4];
+
+    const uint8x16_t v_m = vec_splats((uint8_t)0x0F);
+
+    #pragma GCC unroll 4
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q5_1 * GGML_RESTRICT x0 = &x[ib + 0];
+        const block_q5_1 * GGML_RESTRICT x1 = &x[ib + 1];
+        const block_q8_1 * GGML_RESTRICT y0 = &y[ib + 0];
+        const block_q8_1 * GGML_RESTRICT y1 = &y[ib + 1];
+
+        summs0 += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
+        summs1 += GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s);
+
+        memcpy(&qh0, x0->qh, sizeof(qh0));
+        memcpy(&qh1, x1->qh, sizeof(qh1));
+
+        tmp0[0] = table_b2b_0[(qh0 >>  0) & 0xFF];
+        tmp0[1] = table_b2b_0[(qh0 >>  8) & 0xFF];
+        tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF];
+        tmp0[3] = table_b2b_0[(qh0 >> 24)       ];
+
+        tmp1[0] = table_b2b_0[(qh1 >>  0) & 0xFF];
+        tmp1[1] = table_b2b_0[(qh1 >>  8) & 0xFF];
+        tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF];
+        tmp1[3] = table_b2b_0[(qh1 >> 24)       ];
+
+        int8x16_t v_qh0l = vec_xl(0, (const int8_t *)(tmp0 + 0));
+        int8x16_t v_qh0h = vec_xl(0, (const int8_t *)(tmp0 + 2));
+        int8x16_t v_qh1l = vec_xl(0, (const int8_t *)(tmp1 + 0));
+        int8x16_t v_qh1h = vec_xl(0, (const int8_t *)(tmp1 + 2));
+
+        // required for fixing the byteorder
+        v_qh0l = vec_perm(v_qh0l, v_qh0l, v_kperm);
+        v_qh0h = vec_perm(v_qh0h, v_qh0h, v_kperm);
+        v_qh1l = vec_perm(v_qh1l, v_qh1l, v_kperm);
+        v_qh1h = vec_perm(v_qh1h, v_qh1h, v_kperm);
+
+        const uint8x16_t v_x0 = vec_xl(0, x0->qs);
+        const uint8x16_t v_x1 = vec_xl(0, x1->qs);
+
+        const int8x16_t v_x0l = (int8x16_t)vec_and(v_x0, v_m);
+        const int8x16_t v_x0h = (int8x16_t)vec_sr(v_x0, 4);
+        const int8x16_t v_x1l = (int8x16_t)vec_and(v_x1, v_m);
+        const int8x16_t v_x1h = (int8x16_t)vec_sr(v_x1, 4);
+
+        const int8x16_t v_x0lf = vec_or(v_x0l, v_qh0l);
+        const int8x16_t v_x0hf = vec_or(v_x0h, v_qh0h);
+        const int8x16_t v_x1lf = vec_or(v_x1l, v_qh1l);
+        const int8x16_t v_x1hf = vec_or(v_x1h, v_qh1h);
+
+        const int8x16_t v_y0l = vec_xl(0      , y0->qs);
+        const int8x16_t v_y0h = vec_xl(QK8_1/2, y0->qs);
+        const int8x16_t v_y1l = vec_xl(0      , y1->qs);
+        const int8x16_t v_y1h = vec_xl(QK8_1/2, y1->qs);
+
+        const int32x4_t v_xy0 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x0lf, v_y0l), v_x0hf, v_y0h);
+        const int32x4_t v_xy1 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x1lf, v_y1l), v_x1hf, v_y1h);
+
+        const float32x4_t v_xy0f = vec_float(v_xy0);
+        const float32x4_t v_xy1f = vec_float(v_xy1);
+
+        const float32x4_t v_d0 = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_d1 = vec_splats(GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_FP16_TO_FP32(y1->d));
+
+        v_sum0 = vec_madd(v_xy0f, v_d0, v_sum0);
+        v_sum1 = vec_madd(v_xy1f, v_d1, v_sum1);
+    }
+
+    sumf += vec_hsum(v_sum0) + vec_hsum(v_sum1) + summs0 + summs1;
+
+    #pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        const block_q5_1 * GGML_RESTRICT x0 = &x[ib];
+        const block_q8_1 * GGML_RESTRICT y0 = &y[ib];
+
+        float summs = GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
+
+        uint32_t qh;
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        uint64_t tmp[4];
+        tmp[0] = table_b2b_0[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_0[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_0[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_0[(qh >> 24)       ];
+
+        int8x16_t v_qhl = vec_xl(0, (const int8_t *)(tmp + 0));
+        int8x16_t v_qhh = vec_xl(0, (const int8_t *)(tmp + 2));
+
+        // required for fixing the byteorder
+        v_qhl = vec_perm(v_qhl, v_qhl, v_kperm);
+        v_qhh = vec_perm(v_qhh, v_qhh, v_kperm);
+
+        const uint8x16_t v_x = vec_xl(0, x0->qs);
+        const int8x16_t v_xl = (int8x16_t)vec_and(v_x, v_m);
+        const int8x16_t v_xh = (int8x16_t)vec_sr(v_x, 4);
+
+        const int8x16_t v_xlf = vec_or(v_xl, v_qhl);
+        const int8x16_t v_xhf = vec_or(v_xh, v_qhh);
+
+        const int8x16_t v_yl = vec_xl(0      , y0->qs);
+        const int8x16_t v_yh = vec_xl(QK8_1/2, y0->qs);
+
+        const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xlf, v_yl), v_xhf, v_yh);
+        const float32x4_t v_xyf = vec_float(v_xy);
+
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_acc = vec_madd(v_xyf, v_d, v_acc);
+
+        sumf += vec_hsum(v_acc) + summs;
+    }
+
+    *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
 void ggml_vec_dot_q8_0_q8_0(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) {
     const int qk = QK8_0;
     const int nb = n / qk;

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

@@ -486,6 +486,14 @@ inline static int16x8_t vec_padd_s16(int16x8_t a, int16x8_t b) {
     return v_abo + v_abe;
 }
 
+/**
+ * @see https://github.com/ggml-org/llama.cpp/pull/14037
+ */
+inline float vec_hsum(float32x4_t v) {
+    float32x4_t v_temp = v + vec_reve(v);
+    return v_temp[0] + v_temp[1];
+}
+
 inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
     const int16x8_t p = vec_mule(a, b) + vec_mulo(a, b);
     return acc + (vec_unpackh(p) + vec_unpackl(p));

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

@@ -1880,6 +1880,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_conv_2d(params, tensor);
             } break;
+        case GGML_OP_CONV_3D:
+            {
+                ggml_compute_forward_conv_3d(params, tensor);
+            } break;
         case GGML_OP_CONV_2D_DW:
             {
                 ggml_compute_forward_conv_2d_dw(params, tensor);
@@ -2252,6 +2256,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_IM2COL:
         case GGML_OP_IM2COL_BACK:
         case GGML_OP_CONV_2D:
+        case GGML_OP_CONV_3D:
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_CONV_TRANSPOSE_2D:
@@ -2773,6 +2778,7 @@ struct ggml_cplan ggml_graph_plan(
                         }
                     } break;
                 case GGML_OP_CONV_2D:
+                case GGML_OP_CONV_3D:
                     {
                         cur = GGML_IM2COL_WORK_SIZE;
                     } break;

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

@@ -2169,94 +2169,117 @@ class tinyBLAS_Q0_PPC {
 class tinyBLAS_PPC {
   public:
     tinyBLAS_PPC(int64_t k,
-                const float *A, int64_t lda,
-                const float *B, int64_t ldb,
-                float *C, int64_t ldc,
+                const float * A, int64_t lda,
+                const float * B, int64_t ldb,
+                float * C, int64_t ldc,
                 int ith, int nth)
         : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
     }
 
     void matmul(int64_t m, int64_t n) {
-       mnpack(0, m, 0, n);
+        int64_t mc = 256; int64_t nc = 256; int64_t kc = 256;
+        if (m % mc == 0 && n % nc == 0 && k % kc == 0) {
+            matmul_tiled(m, n, mc, nc, kc);
+        } else {
+            mnpack(0, m, 0, n);
+        }
     }
 
   private:
 
-    void (tinyBLAS_PPC::*kernel)(int64_t, int64_t);
-
-    inline void vector_permute_store_4(vector float *src, float *vecOffset) {
-       vector float t1, t2, t3, t4, t5, t6, t7, t8;
-           t1 = vec_mergeh(src[0], src[1]);
-           t2 = vec_mergeh(src[2], src[3]);
-           t3 = vec_mergel(src[0], src[1]);
-           t4 = vec_mergel(src[2], src[3]);
-
-           t5 = vec_xxpermdi(t1, t2, 0);
-           t6 = vec_xxpermdi(t1, t2, 3);
-           t7 = vec_xxpermdi(t3, t4, 0);
-           t8 = vec_xxpermdi(t3, t4, 3);
-
-           vec_xst(t5, 0, vecOffset);
-           vec_xst(t6, 0, vecOffset + 4);
-           vec_xst(t7, 0, vecOffset + 8);
-           vec_xst(t8, 0, vecOffset + 12);
-       }
-
-    inline void vector_permute_store_8(vector float *src, float *vecOffset) {
-       vector float t1, t2, t3, t4, t5, t6, t7, t8;
-          t1 = vec_mergeh(src[0], src[1]);
-          t2 = vec_mergeh(src[2], src[3]);
-          t3 = vec_mergeh(src[4], src[5]);
-          t4 = vec_mergeh(src[6], src[7]);
-
-          t5 = vec_xxpermdi(t1, t2, 0);
-          t6 = vec_xxpermdi(t3, t4, 0);
-          t7 = vec_xxpermdi(t1, t2, 3);
-          t8 = vec_xxpermdi(t3, t4, 3);
-
-          vec_xst(t5, 0, vecOffset);
-          vec_xst(t6, 0, vecOffset + 4);
-          vec_xst(t7, 0, vecOffset + 8);
-          vec_xst(t8, 0, vecOffset + 12);
-
-          t1 = vec_mergel(src[0], src[1]);
-          t2 = vec_mergel(src[2], src[3]);
-          t3 = vec_mergel(src[4], src[5]);
-          t4 = vec_mergel(src[6], src[7]);
-
-          t5 = vec_xxpermdi(t1, t2, 0);
-          t6 = vec_xxpermdi(t3, t4, 0);
-          t7 = vec_xxpermdi(t1, t2, 3);
-          t8 = vec_xxpermdi(t3, t4, 3);
-
-          vec_xst(t5, 0, vecOffset + 16);
-          vec_xst(t6, 0, vecOffset + 20);
-          vec_xst(t7, 0, vecOffset + 24);
-          vec_xst(t8, 0, vecOffset + 28);
+    inline void save_acc(acc_t * ACC, int64_t ii, int64_t jj) {
+        vec_t vec_C[4];
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int I = 0; I < 4; I++) {
+            for (int J = 0; J < 4; J++) {
+                *((float *)(C+ii+((jj+J)*ldc)+I)) = *((float *)&vec_C[I]+J);
+            }
+        }
     }
 
- void packTranspose(const float* a, int64_t lda, int rows, int cols, float* vec) {
+    inline void add_save_acc(acc_t * ACC, int64_t ii, int64_t jj) {
+        vec_t vec_C[4];
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int I = 0; I < 4; I++) {
+            for (int J = 0; J < 4; J++) {
+                float * c_ptr = (float *)(C+ii+((jj+J)*ldc)+I);
+                *c_ptr += *((float *)&vec_C[I]+J);
+            }
+        }
+    }
+
+    inline void vector_permute_store_4(vector float * src, float * vecOffset) {
+        vector float t1, t2, t3, t4, t5, t6, t7, t8;
+        t1 = vec_mergeh(src[0], src[1]);
+        t2 = vec_mergeh(src[2], src[3]);
+        t3 = vec_mergel(src[0], src[1]);
+        t4 = vec_mergel(src[2], src[3]);
+
+        t5 = vec_xxpermdi(t1, t2, 0);
+        t6 = vec_xxpermdi(t1, t2, 3);
+        t7 = vec_xxpermdi(t3, t4, 0);
+        t8 = vec_xxpermdi(t3, t4, 3);
+
+        vec_xst(t5, 0, vecOffset);
+        vec_xst(t6, 0, vecOffset + 4);
+        vec_xst(t7, 0, vecOffset + 8);
+        vec_xst(t8, 0, vecOffset + 12);
+    }
+
+    inline void vector_permute_store_8(vector float * src, float * vecOffset) {
+        vector float t1, t2, t3, t4, t5, t6, t7, t8;
+        t1 = vec_mergeh(src[0], src[1]);
+        t2 = vec_mergeh(src[2], src[3]);
+        t3 = vec_mergeh(src[4], src[5]);
+        t4 = vec_mergeh(src[6], src[7]);
+
+        t5 = vec_xxpermdi(t1, t2, 0);
+        t6 = vec_xxpermdi(t3, t4, 0);
+        t7 = vec_xxpermdi(t1, t2, 3);
+        t8 = vec_xxpermdi(t3, t4, 3);
+
+        vec_xst(t5, 0, vecOffset);
+        vec_xst(t6, 0, vecOffset + 4);
+        vec_xst(t7, 0, vecOffset + 8);
+        vec_xst(t8, 0, vecOffset + 12);
+
+        t1 = vec_mergel(src[0], src[1]);
+        t2 = vec_mergel(src[2], src[3]);
+        t3 = vec_mergel(src[4], src[5]);
+        t4 = vec_mergel(src[6], src[7]);
+
+        t5 = vec_xxpermdi(t1, t2, 0);
+        t6 = vec_xxpermdi(t3, t4, 0);
+        t7 = vec_xxpermdi(t1, t2, 3);
+        t8 = vec_xxpermdi(t3, t4, 3);
+
+        vec_xst(t5, 0, vecOffset + 16);
+        vec_xst(t6, 0, vecOffset + 20);
+        vec_xst(t7, 0, vecOffset + 24);
+        vec_xst(t8, 0, vecOffset + 28);
+    }
+
+    void packTranspose(const float * a, int64_t lda, int rows, int cols, float * vec) {
         int64_t i, j;
         float * aoffsets[8];
-        float *aoffset = NULL, *boffset = NULL;
+        float * aoffset = NULL, * boffset = NULL;
         __vector_pair arr[8];
         vector float c[8][2] = {0};
         vector float c1[8] = {0};
         vector float c2[8] = {0};
-        aoffset = const_cast<float*>(a);
+        aoffset = const_cast<float *>(a);
         boffset = vec;
         j = (rows >> 3);
         if (j > 0) {
-
             do {
                 aoffsets[0] = aoffset;
-                for (int it = 1; it< 8; it++)
+                for (int it = 1; it < 8; it++)
                     aoffsets[it] = aoffsets[it-1] + lda;
                 aoffset += 8 * lda;
                 i = (cols >> 3);
                 if (i > 0) {
                     do {
-                        for (int it = 0; it< 8; it++) {
+                        for (int it = 0; it < 8; it++) {
                             arr[it] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[it]);
                             __builtin_vsx_disassemble_pair(c[it], &arr[it]);
                             c1[it] = c[it][0];
@@ -2264,11 +2287,14 @@ class tinyBLAS_PPC {
                         }
 
                         vector_permute_store_8(c1, boffset);
-                        vector_permute_store_8(c2, boffset+32);
-                        for (int it = 0; it < 4; it++)
-                            aoffsets[it] = aoffsets[it] + 8*lda;
+                        vector_permute_store_8(c2, boffset + 32);
                         boffset += 64;
                         i--;
+                        if (i > 0) {
+                           for (int it = 0; it < 8; it++) {
+                               aoffsets[it] = aoffsets[it] + 8;
+                           }
+                        }
                     } while(i > 0);
                 }
                 if (cols & 4) {
@@ -2295,9 +2321,9 @@ class tinyBLAS_PPC {
                         c2[it] = c[it][1];
                     }
                     vector_permute_store_4(c1, boffset);
-                    vector_permute_store_4(c2, boffset+16);
+                    vector_permute_store_4(c2, boffset + 16);
                     for (int it = 0; it < 4; it++)
-                        aoffsets[it] += 8*lda;
+                        aoffsets[it] += 8 * lda;
                     boffset += 32;
                     i--;
                 } while(i > 0);
@@ -2325,15 +2351,15 @@ class tinyBLAS_PPC {
         vec_t vec_A[4], vec_B[4], vec_C[4];
         acc_t acc_0;
         __builtin_mma_xxsetaccz(&acc_0);
-        for (int l = 0; l < k; l+=4) {
-            packTranspose(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A);
-            packTranspose(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
+        for (int l = 0; l < k; l += 4) {
+            packTranspose(A + (ii * lda) + l, lda, 4, 4, (float *)vec_A);
+            packTranspose(B + (jj * ldb) + l, ldb, 4, 4, (float *)vec_B);
             __builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
             __builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
             __builtin_mma_xvf32gerpp(&acc_0, vec_A[2], vec_B[2]);
             __builtin_mma_xvf32gerpp(&acc_0, vec_A[3], vec_B[3]);
         }
-        SAVE_ACC(&acc_0, ii, jj);
+        save_acc(&acc_0, ii, jj);
     }
 
     void KERNEL_4x8(int64_t ii, int64_t jj) {
@@ -2341,9 +2367,9 @@ class tinyBLAS_PPC {
         acc_t acc_0, acc_1;
         __builtin_mma_xxsetaccz(&acc_0);
         __builtin_mma_xxsetaccz(&acc_1);
-        for (int64_t l = 0; l < k; l+=4) {
-            packTranspose(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A);
-            packTranspose(B+(jj*ldb)+l, ldb, 8, 4, (float*)vec_B);
+        for (int64_t l = 0; l < k; l += 4) {
+            packTranspose(A + (ii * lda) + l, lda, 4, 4, (float *)vec_A);
+            packTranspose(B + (jj * ldb) + l, ldb, 8, 4, (float *)vec_B);
             __builtin_mma_xvf32gerpp(&acc_0, vec_A[0], (vec_t)vec_B[0]);
             __builtin_mma_xvf32gerpp(&acc_1, vec_A[0], (vec_t)vec_B[1]);
             __builtin_mma_xvf32gerpp(&acc_0, vec_A[1], (vec_t)vec_B[2]);
@@ -2353,8 +2379,8 @@ class tinyBLAS_PPC {
             __builtin_mma_xvf32gerpp(&acc_0, vec_A[3], (vec_t)vec_B[6]);
             __builtin_mma_xvf32gerpp(&acc_1, vec_A[3], (vec_t)vec_B[7]);
         }
-        SAVE_ACC(&acc_0, ii, jj);
-        SAVE_ACC(&acc_1, ii, jj+4);
+        save_acc(&acc_0, ii, jj);
+        save_acc(&acc_1, ii, jj + 4);
     }
 
     void KERNEL_8x4(int64_t ii, int64_t jj) {
@@ -2362,9 +2388,9 @@ class tinyBLAS_PPC {
         acc_t acc_0, acc_1;
         __builtin_mma_xxsetaccz(&acc_0);
         __builtin_mma_xxsetaccz(&acc_1);
-        for (int64_t l = 0; l < k; l+=4) {
-            packTranspose(A+(ii*lda)+l, lda, 8, 4, (float*)vec_A);
-            packTranspose(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
+        for (int64_t l = 0; l < k; l += 4) {
+            packTranspose(A + (ii * lda) + l, lda, 8, 4, (float *)vec_A);
+            packTranspose(B + (jj * ldb) + l, ldb, 4, 4, (float *)vec_B);
             __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[0], vec_B[0]);
             __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[1], vec_B[0]);
             __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[2], vec_B[1]);
@@ -2374,8 +2400,8 @@ class tinyBLAS_PPC {
             __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[6], vec_B[3]);
             __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[7], vec_B[3]);
         }
-        SAVE_ACC(&acc_0, ii, jj);
-        SAVE_ACC(&acc_1, ii+4, jj);
+        save_acc(&acc_0, ii, jj);
+        save_acc(&acc_1, ii + 4, jj);
     }
 
     void KERNEL_8x8(int64_t ii, int64_t jj) {
@@ -2386,19 +2412,96 @@ class tinyBLAS_PPC {
         __builtin_mma_xxsetaccz(&acc_2);
         __builtin_mma_xxsetaccz(&acc_3);
         for (int l = 0; l < k; l+=8) {
-            packTranspose(A+(ii*lda)+l, lda, 8, 8, (float*)vec_A);
-            packTranspose(B+(jj*ldb)+l, ldb, 8, 8, (float*)vec_B);
+            packTranspose(A + (ii * lda) + l, lda, 8, 8, (float *)vec_A);
+            packTranspose(B + (jj * ldb) + l, ldb, 8, 8, (float *)vec_B);
             for(int x = 0; x < 16; x+=2) {
                 __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[x], vec_B[x]);
-                __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[x], vec_B[x+1]);
-                __builtin_mma_xvf32gerpp(&acc_2, (vec_t)vec_A[x+1], vec_B[x]);
-                __builtin_mma_xvf32gerpp(&acc_3, (vec_t)vec_A[x+1], vec_B[x+1]);
+                __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[x], vec_B[x + 1]);
+                __builtin_mma_xvf32gerpp(&acc_2, (vec_t)vec_A[x + 1], vec_B[x]);
+                __builtin_mma_xvf32gerpp(&acc_3, (vec_t)vec_A[x + 1], vec_B[x + 1]);
+            }
+        }
+        save_acc(&acc_0, ii, jj);
+        save_acc(&acc_1, ii, jj + 4);
+        save_acc(&acc_2, ii + 4, jj);
+        save_acc(&acc_3, ii + 4, jj + 4);
+    }
+
+    inline void MMA_16x8(vec_t * vec_A0, vec_t * vec_A1, vec_t * vec_B, acc_t * acc) {
+        for (int x = 0; x < 16; x += 2) {
+            __builtin_mma_xvf32gerpp(&acc[0], vec_A0[x + 0], vec_B[x]);
+            __builtin_mma_xvf32gerpp(&acc[1], vec_A0[x + 0], vec_B[x + 1]);
+            __builtin_mma_xvf32gerpp(&acc[2], vec_A0[x + 1], vec_B[x]);
+            __builtin_mma_xvf32gerpp(&acc[3], vec_A0[x + 1], vec_B[x + 1]);
+            __builtin_mma_xvf32gerpp(&acc[4], vec_A1[x + 0], vec_B[x]);
+            __builtin_mma_xvf32gerpp(&acc[5], vec_A1[x + 0], vec_B[x + 1]);
+            __builtin_mma_xvf32gerpp(&acc[6], vec_A1[x + 1], vec_B[x]);
+            __builtin_mma_xvf32gerpp(&acc[7], vec_A1[x + 1], vec_B[x + 1]);
+        }
+    }
+
+    void KERNEL(int64_t ii, int64_t jj, int64_t mc, int64_t nc, int64_t kc, vec_t * vec_A, vec_t * vec_B, int64_t kk) {
+        for (int64_t i = 0; i < mc; i += 16) {
+            int A_base_addr = (mc / 8) * (i / 8) * 16;
+            for (int64_t j = 0; j < nc; j += 8) {
+                 int B_base_addr = (nc / 8) * (j / 8) * 16;
+                 acc_t acc[8];
+                 vec_t A0_block[16]; vec_t A1_block[16];
+                 for (int x = 0; x < 8; x++)
+                     __builtin_mma_xxsetaccz(&acc[x]);
+                 for (int64_t l = 0; l < kc; l += 8) {
+                     int A0_block_idx = A_base_addr + (l / 8) * 16;
+                     int A1_block_idx = A0_block_idx + (mc / 8) * 16;
+                     int B_block_idx = B_base_addr + (l / 8) * 16;
+                     vec_t* A0_block = &vec_A[A0_block_idx];
+                     vec_t* A1_block = &vec_A[A1_block_idx];
+                     vec_t* B_block = &vec_B[B_block_idx];
+                     MMA_16x8(A0_block, A1_block, B_block, acc);
+                 }
+                 if (kk == 0) {
+                     save_acc(&acc[0], ii + i, jj + j);
+                     save_acc(&acc[1], ii + i, jj + j + 4);
+                     save_acc(&acc[2], ii + i + 4, jj + j);
+                     save_acc(&acc[3], ii + i + 4, jj + j + 4);
+                     save_acc(&acc[4], ii + i + 8, jj + j);
+                     save_acc(&acc[5], ii + i + 8, jj + j + 4);
+                     save_acc(&acc[6], ii + i + 12, jj + j);
+                     save_acc(&acc[7], ii + i + 12, jj + j + 4);
+                 } else {
+                     add_save_acc(&acc[0], ii + i, jj + j);
+                     add_save_acc(&acc[1], ii + i, jj + j + 4);
+                     add_save_acc(&acc[2], ii + i + 4, jj + j);
+                     add_save_acc(&acc[3], ii + i + 4, jj + j + 4);
+                     add_save_acc(&acc[4], ii + i + 8, jj + j);
+                     add_save_acc(&acc[5], ii + i + 8, jj + j + 4);
+                     add_save_acc(&acc[6], ii + i + 12, jj + j);
+                     add_save_acc(&acc[7], ii + i + 12, jj + j + 4);
+                 }
+            }
+        }
+    }
+
+    void matmul_tiled(int64_t m , int64_t n, int64_t mc, int64_t nc, int64_t kc) {
+        int64_t ytiles = m / mc;
+        int64_t xtiles = n / nc;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles) {
+            end = tiles;
+        }
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = (job / xtiles) * mc;
+            int64_t jj = (job % xtiles) * nc;
+            for (int64_t kk = 0; kk < k; kk += kc) {
+                 vec_t A_pack[kc * mc / 4];
+                 vec_t B_pack[kc * nc / 4];
+                 packTranspose(A + (ii * lda) + kk, lda, kc, mc, (float *)A_pack);
+                 packTranspose(B + (jj * ldb) + kk, ldb, kc, nc, (float *)B_pack);
+                 KERNEL(ii, jj, mc, nc, kc, A_pack, B_pack, kk);
             }
         }
-        SAVE_ACC(&acc_0, ii, jj);
-        SAVE_ACC(&acc_1, ii, jj+4);
-        SAVE_ACC(&acc_2, ii+4, jj);
-        SAVE_ACC(&acc_3, ii+4, jj+4);
     }
 
     void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
@@ -2406,35 +2509,35 @@ class tinyBLAS_PPC {
         int n_rem = MIN(n - n0, 8);
         int mc = 0, nc = 0;
         if (m_rem >= 8 && n_rem >= 8) {
-           mc = 8;
-           nc = 8;
-           gemm<8, 8>(m0, m, n0, n);
+            mc = 8;
+            nc = 8;
+            gemm<8, 8>(m0, m, n0, n);
         } else if (m_rem >= 4 && n_rem >= 8) {
-                mc = 4;
-                nc = 8;
-                gemm<4, 8>(m0, m, n0, n);
+            mc = 4;
+            nc = 8;
+            gemm<4, 8>(m0, m, n0, n);
         } else if (m_rem >= 8 && n_rem >= 4) {
-                mc = 8;
-                nc = 4;
-                gemm<8, 4>(m0, m, n0, n);
+            mc = 8;
+            nc = 4;
+            gemm<8, 4>(m0, m, n0, n);
         } else if (m_rem >= 4 && n_rem >= 4) {
-                mc = 4;
-                nc = 4;
-                gemm<4, 4>(m0, m, n0, n);
+            mc = 4;
+            nc = 4;
+            gemm<4, 4>(m0, m, n0, n);
         } else {
             mc = (m_rem >= 4) ? 4 : m_rem;
             nc = (n_rem >= 4) ? 4 : n_rem;
             if (mc == 0 || nc == 0)
-               return;
+                return;
             gemm_small(m0, m, n0, n, mc, nc);
         }
         int64_t mp = m0 + ((m - m0) / mc) * mc;
         int64_t np = n0 + ((n - n0) / nc) * nc;
         mnpack(mp, m, n0, np);
         mnpack(m0, m, np, n);
-     }
+    }
 
-     void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
+    void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
         int64_t ytiles = (m - m0) / RM;
         int64_t xtiles = (n - n0) / RN;
         int64_t tiles = xtiles * ytiles;
@@ -2449,30 +2552,30 @@ class tinyBLAS_PPC {
             vec_t vec_C[4];
             acc_t acc_0;
             __builtin_mma_xxsetaccz(&acc_0);
-            vec_t vec_A[4] {0}, vec_B[4] = {0};
-            for (int l=0; l<k; l+=4) {
+            vec_t vec_A[4] = {0}, vec_B[4] = {0};
+            for (int l = 0; l < k; l += 4) {
                 /* 'GEMV Forwarding' concept is used in first two conditional loops.
                  * when one of the matrix has a single row/column, the elements are
                  * broadcasted, instead of using packing routine to prepack the
                  * matrix elements.
                  */
                 if (RM == 1) {
-                    float* a = const_cast<float*>(A+(ii)*lda+l);
-                    packTranspose(B+(jj*ldb)+l, ldb, RN, 4, (float*)vec_B);
+                    float * a = const_cast<float *>(A + (ii) * lda + l);
+                    packTranspose(B + (jj * ldb) + l, ldb, RN, 4, (float *)vec_B);
                     vec_A[0] = (vec_t)vec_xl(0,a);
-                    vec_A[1] = (vec_t)vec_splats(*((float*)&vec_A+1));
-                    vec_A[2] = (vec_t)vec_splats(*((float*)&vec_A+2));
-                    vec_A[3] = (vec_t)vec_splats(*((float*)&vec_A+3));
+                    vec_A[1] = (vec_t)vec_splats(*((float *)&vec_A+1));
+                    vec_A[2] = (vec_t)vec_splats(*((float *)&vec_A+2));
+                    vec_A[3] = (vec_t)vec_splats(*((float *)&vec_A+3));
                 } else if (RN == 1) {
-                    packTranspose(A+(ii*lda)+l, lda, RM, 4, (float*)vec_A);
-                    float* b = const_cast<float*>(B+(jj)*ldb+l);
+                    packTranspose(A + (ii * lda) + l, lda, RM, 4, (float *)vec_A);
+                    float * b = const_cast<float *>(B + (jj) * ldb + l);
                     vec_B[0] = (vec_t)vec_xl(0,b);
-                    vec_B[1] = (vec_t)vec_splats(*((float*)&vec_B+1));
-                    vec_B[2] = (vec_t)vec_splats(*((float*)&vec_B+2));
-                    vec_B[3] = (vec_t)vec_splats(*((float*)&vec_B+3));
+                    vec_B[1] = (vec_t)vec_splats(*((float *)&vec_B+1));
+                    vec_B[2] = (vec_t)vec_splats(*((float *)&vec_B+2));
+                    vec_B[3] = (vec_t)vec_splats(*((float *)&vec_B+3));
                 } else {
-                    packTranspose(A+(ii*lda)+l, lda, RM, 4, (float*)vec_A);
-                    packTranspose(B+(jj*ldb)+l, ldb, RN, 4, (float*)vec_B);
+                    packTranspose(A + (ii * lda) + l, lda, RM, 4, (float *)vec_A);
+                    packTranspose(B + (jj * ldb) + l, ldb, RN, 4, (float *)vec_B);
                 }
                 __builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
                 __builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
@@ -2482,12 +2585,27 @@ class tinyBLAS_PPC {
             __builtin_mma_disassemble_acc(vec_C, &acc_0);
             for (int I = 0; I < RM; I++) {
                 for (int J = 0; J < RN; J++) {
-                    *((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&vec_C[I]+J);
+                    *((float *)(C+ii+((jj+J)*ldc)+I)) = *((float *)&vec_C[I]+J);
                 }
             }
        }
     }
 
+    template<int RM, int RN>
+    inline void kernel(int64_t ii, int64_t jj) {
+        if constexpr(RM == 4 && RN == 4) {
+            KERNEL_4x4(ii, jj);
+        } else if constexpr(RM == 4 && RN == 8) {
+            KERNEL_4x8(ii, jj);
+        } else if constexpr(RM == 8 && RN == 4) {
+            KERNEL_8x4(ii, jj);
+        } else if constexpr(RM == 8 && RN == 8) {
+            KERNEL_8x8(ii, jj);
+        } else {
+            static_assert(false, "RN/RM values not supported");
+        }
+    }
+
     template <int RM, int RN>
     NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
         int64_t ytiles = (m - m0) / RM;
@@ -2496,27 +2614,18 @@ class tinyBLAS_PPC {
         int64_t duty = (tiles + nth - 1) / nth;
         int64_t start = duty * ith;
         int64_t end = start + duty;
-        if (RM == 4 && RN == 4) {
-            kernel = &tinyBLAS_PPC::KERNEL_4x4;
-        } else if (RM == 4 && RN == 8) {
-            kernel = &tinyBLAS_PPC::KERNEL_4x8;
-        } else if (RM == 8 && RN == 4) {
-            kernel = &tinyBLAS_PPC::KERNEL_8x4;
-        } else if (RM == 8 && RN == 8) {
-            kernel = &tinyBLAS_PPC::KERNEL_8x8;
-        }
         if (end > tiles)
             end = tiles;
         for (int64_t job = start; job < end; ++job) {
             int64_t ii = m0 + job / xtiles * RM;
             int64_t jj = n0 + job % xtiles * RN;
-            (this->*kernel)(ii, jj);
+            kernel<RM, RN>(ii, jj);
         }
     }
 
-    const float *const A;
-    const float *const B;
-    float *C;
+    const float * const A;
+    const float * const B;
+    float * C;
     const int64_t k;
     const int64_t lda;
     const int64_t ldb;

ggml/src/ggml-cpu/ops.cpp

@@ -7207,6 +7207,148 @@ void ggml_compute_forward_conv_2d(
     ggml_compute_forward_conv_2d_impl(params, src0, src1, dst, src0->type);
 }
 
+// ggml_compute_forward_conv_3d
+
+static void ggml_compute_forward_conv_3d_impl(const ggml_compute_params * params,
+                                              const ggml_tensor *         kernel,
+                                              const ggml_tensor *         src,
+                                              ggml_tensor *               dst,
+                                              ggml_type                   kernel_type) {
+
+    GGML_ASSERT(ggml_is_contiguous(kernel));
+    GGML_ASSERT(kernel_type == GGML_TYPE_F16 || kernel_type == GGML_TYPE_F32);
+    GGML_ASSERT(kernel->type == kernel_type);
+
+    const ggml_type_traits * traits = ggml_get_type_traits(kernel_type);
+
+    const int32_t s0 = dst->op_params[0];
+    const int32_t s1 = dst->op_params[1];
+    const int32_t s2 = dst->op_params[2];
+    const int32_t p0 = dst->op_params[3];
+    const int32_t p1 = dst->op_params[4];
+    const int32_t p2 = dst->op_params[5];
+    const int32_t d0 = dst->op_params[6];
+    const int32_t d1 = dst->op_params[7];
+    const int32_t d2 = dst->op_params[8];
+    const int32_t c  = dst->op_params[9];
+    const int32_t n  = dst->op_params[10];
+    const int32_t oc = dst->op_params[11];
+
+    const int64_t src_w = src->ne[0];
+    const int64_t src_h = src->ne[1];
+    const int64_t src_d = src->ne[2];
+    const int64_t knl_w = kernel->ne[0];
+    const int64_t knl_h = kernel->ne[1];
+    const int64_t knl_d = kernel->ne[2];
+    const int64_t dst_w = dst->ne[0];
+    const int64_t dst_h = dst->ne[1];
+    const int64_t dst_d = dst->ne[2];
+
+    const float * src_data = (float *) src->data;
+    void  * knl_data       = kernel->data;
+    float * dst_data       = (float *) dst->data;
+
+    const int64_t knl_n_per_channel = knl_w * knl_h * knl_d;
+    const int64_t knl_n_total       = knl_n_per_channel * c;
+    const int64_t patch_total       = n * dst_w * dst_h * dst_d;
+
+    const int64_t space_per_patch   = knl_n_total * traits->type_size + oc * sizeof(float);
+    const int64_t batch_size        = params->wsize / space_per_patch;
+    const int64_t patches_per_batch = batch_size > 8 ? (batch_size / 8) * 8 : batch_size;
+    const int64_t batch_n           = (patch_total + patches_per_batch - 1) / patches_per_batch;
+
+    GGML_ASSERT(patches_per_batch > 0 && batch_size >= 1);
+
+    void * tmp = params->wdata;
+
+    for (int64_t batch_i = 0; batch_i < batch_n; ++batch_i) {
+        const int64_t patch_start_batch = batch_i * patches_per_batch;
+        const int64_t patch_end_batch   = std::min(patch_start_batch + patches_per_batch, patch_total);
+        const int64_t patch_n_in_batch  = patch_end_batch - patch_start_batch;
+
+        const int64_t patch_per_thread  = (patch_n_in_batch + params->nth - 1) / params->nth;
+        const int64_t patch_start       = patch_start_batch + params->ith * patch_per_thread;
+        const int64_t patch_end         = std::min(patch_start + patch_per_thread, patch_end_batch);
+
+        for (int64_t p = patch_start; p < patch_end; ++p) {
+            const int64_t p_in_batch = p % (dst_w * dst_h * dst_d);
+            const int64_t p_in_depth = p_in_batch % (dst_w * dst_h);
+            const int64_t batch_idx  = p / (dst_w * dst_h * dst_d);
+            const int64_t dst_z      = p_in_batch / (dst_w * dst_h);
+            const int64_t dst_y      = p_in_depth / dst_w;
+            const int64_t dst_x      = p_in_depth % dst_w;
+
+            char * dst_row = (char *) tmp + (p % patches_per_batch) * knl_n_total * traits->type_size;
+
+            for (int64_t ic = 0; ic < c; ++ic) {
+                for (int64_t kz = 0; kz < knl_d; ++kz) {
+                    for (int64_t ky = 0; ky < knl_h; ++ky) {
+                        for (int64_t kx = 0; kx < knl_w; ++kx) {
+                            const int64_t sz = dst_z * s2 + kz * d2 - p2;
+                            const int64_t sy = dst_y * s1 + ky * d1 - p1;
+                            const int64_t sx = dst_x * s0 + kx * d0 - p0;
+
+                            int64_t dst_idx = ic * knl_n_per_channel + kz * (knl_h * knl_w) + ky * knl_w + kx;
+
+                            float src_val;
+                            if (sz < 0 || sz >= src_d || sy < 0 || sy >= src_h || sx < 0 || sx >= src_w) {
+                                src_val = 0.0f;
+                            } else {
+                                const int64_t cn_idx = batch_idx * c + ic;
+                                const float * src_ptr = (const float *)((const char *)src_data + sx*src->nb[0] + sy*src->nb[1] + sz*src->nb[2] + cn_idx*src->nb[3]);
+                                src_val = *src_ptr;
+                            }
+
+                            char * element_ptr = dst_row + dst_idx * traits->type_size;
+                            if (kernel_type == GGML_TYPE_F32) {
+                                *(float *)element_ptr = src_val;
+                            } else if (kernel_type == GGML_TYPE_F16) {
+                                *(ggml_fp16_t *)element_ptr = GGML_CPU_FP32_TO_FP16(src_val);
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        ggml_barrier(params->threadpool);
+
+        float * gemm_output = (float *) ((char *) tmp + patches_per_batch * knl_n_total * traits->type_size);
+        ggml_call_mul_mat(kernel_type, params, patch_n_in_batch, oc, knl_n_total, tmp, knl_data, gemm_output);
+
+        ggml_barrier(params->threadpool);
+
+        const int64_t permute_per_thread = (patch_n_in_batch + params->nth - 1) / params->nth;
+        const int64_t permute_start = params->ith * permute_per_thread;
+        const int64_t permute_end = std::min(permute_start + permute_per_thread, patch_n_in_batch);
+
+        for (int64_t i = permute_start; i < permute_end; ++i) {
+            const int64_t p = patch_start_batch + i;
+            const int64_t p_in_batch = p % (dst_w * dst_h * dst_d);
+            const int64_t p_in_depth = p_in_batch % (dst_w * dst_h);
+            const int64_t batch_idx  = p / (dst_w * dst_h * dst_d);
+            const int64_t dst_z      = p_in_batch / (dst_w * dst_h);
+            const int64_t dst_y      = p_in_depth / dst_w;
+            const int64_t dst_x      = p_in_depth % dst_w;
+
+            for (int64_t ioc = 0; ioc < oc; ++ioc) {
+                const float value = gemm_output[i * oc + ioc];
+                const int64_t ocn_idx = batch_idx * oc + ioc;
+                float * dst_ptr = (float *)((char *)dst_data + dst_x*dst->nb[0] + dst_y*dst->nb[1] + dst_z*dst->nb[2] + ocn_idx*dst->nb[3]);
+                *dst_ptr = value;
+            }
+        }
+    }
+}
+
+void ggml_compute_forward_conv_3d(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    ggml_compute_forward_conv_3d_impl(params, src0, src1, dst, src0->type);
+}
+
 // ggml_compute_forward_conv_transpose_2d
 
 void ggml_compute_forward_conv_transpose_2d(
@@ -8930,6 +9072,9 @@ static void ggml_compute_forward_ssm_scan_f32(
                         }
 
                         sumf = GGML_F32xt_REDUCE_ONE(sum);
+    #elif defined(__riscv_v_intrinsic)
+                        // todo: RVV implementation
+                        const int np = 0;
     #else
                         const int np = (nc & ~(GGML_F32_STEP - 1));
 
@@ -9881,8 +10026,8 @@ static void ggml_compute_forward_rwkv_wkv7_f32(
     int64_t h_stride_2d = head_size * head_size;
 
     #if defined(GGML_SIMD)
-        #if defined(__ARM_FEATURE_SVE)
-            // scalar Route to scalar implementation       //TODO: Write SVE code
+        #if defined(__ARM_FEATURE_SVE) || defined(__riscv_v_intrinsic)
+            // scalar Route to scalar implementation       //TODO: Write SVE code and RVV code
             for (int64_t t = 0; t < T; t++) {
                 int64_t t_offset = t * t_stride;
                 int64_t state_offset = head_size * C * (t / (T / n_seqs));

ggml/src/ggml-cpu/ops.h

@@ -70,6 +70,7 @@ void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * p
 void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_conv_3d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_2d_dw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);

ggml/src/ggml-cpu/simd-mappings.h

@@ -18,6 +18,10 @@
 #include <immintrin.h>
 #endif
 
+#if defined(__riscv_v_intrinsic)
+#include <riscv_vector.h>
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -94,24 +98,15 @@ extern "C" {
     }
 #elif defined(__riscv) && defined(__riscv_zfhmin)
     static inline float riscv_compute_fp16_to_fp32(ggml_fp16_t h) {
-        float f;
-        __asm__(
-            "fmv.h.x %[f], %[h]\n\t"
-            "fcvt.s.h %[f], %[f]"
-            : [f] "=&f" (f)
-            : [h] "r" (h)
-        );
-        return f;
+        _Float16 hf;
+        memcpy(&hf, &h, sizeof(ggml_fp16_t));
+        return hf;
     }
 
     static inline ggml_fp16_t riscv_compute_fp32_to_fp16(float f) {
         ggml_fp16_t res;
-        __asm__(
-            "fcvt.h.s %[f], %[f]\n\t"
-            "fmv.x.h %[h], %[f]"
-            : [h] "=&r" (res)
-            : [f] "f" (f)
-        );
+        _Float16 hf = (_Float16)f;
+        memcpy(&res, &hf, sizeof(ggml_fp16_t));
         return res;
     }
 
@@ -1170,6 +1165,36 @@ static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) {
 #define GGML_F16_VEC_MUL            GGML_F32x4_MUL
 #define GGML_F16_VEC_REDUCE         GGML_F32x4_REDUCE
 
+#elif defined(__riscv_v_intrinsic)
+
+// compatible with vlen >= 128
+
+#define GGML_SIMD
+
+// F32
+
+#define GGML_F32_STEP 16
+#define GGML_F32_EPR  4
+
+#define GGML_F32x4              vfloat32m1_t
+#define GGML_F32x4_ZERO         __riscv_vfmv_v_f_f32m1(0.0f, GGML_F32_EPR)
+#define GGML_F32x4_SET1(x)      __riscv_vfmv_v_f_f32m1(x, GGML_F32_EPR)
+#define GGML_F32x4_LOAD(x)      __riscv_vle32_v_f32m1(x, GGML_F32_EPR)
+#define GGML_F32x4_STORE(b, v)  __riscv_vse32_v_f32m1(b, v, GGML_F32_EPR)
+#define GGML_F32x4_FMA(a, b, c) __riscv_vfmacc_vv_f32m1(a, b, c, GGML_F32_EPR)
+#define GGML_F32x4_ADD(a, b)    __riscv_vfadd_vv_f32m1(a, b, GGML_F32_EPR)
+#define GGML_F32x4_MUL(a, b)    __riscv_vfmul_vv_f32m1(a, b, GGML_F32_EPR)
+
+#define GGML_F32_VEC        GGML_F32x4
+#define GGML_F32_VEC_ZERO   GGML_F32x4_ZERO
+#define GGML_F32_VEC_SET1   GGML_F32x4_SET1
+#define GGML_F32_VEC_LOAD   GGML_F32x4_LOAD
+#define GGML_F32_VEC_STORE  GGML_F32x4_STORE
+#define GGML_F32_VEC_FMA    GGML_F32x4_FMA
+#define GGML_F32_VEC_ADD    GGML_F32x4_ADD
+#define GGML_F32_VEC_MUL    GGML_F32x4_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
+
 #endif
 
 // GGML_F32_ARR / GGML_F16_ARR

ggml/src/ggml-cpu/vec.cpp

@@ -84,6 +84,16 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
         }
         // reduce sum1,sum2 to sum1
         GGML_F32_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8);
+    #elif defined(__riscv_v_intrinsic)
+        vfloat32m1_t vsum = __riscv_vfmv_v_f_f32m1(0.0f, 1);
+        for (int i = 0, avl; i < n; i += avl) {
+            avl = __riscv_vsetvl_e32m8(n - i);
+            vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[i], avl);
+            vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
+            vfloat32m8_t prod = __riscv_vfmul_vv_f32m8(ax, ay, avl);
+            vsum = __riscv_vfredusum_vs_f32m8_f32m1(prod, vsum, avl);
+        }
+        sumf += __riscv_vfmv_f_s_f32m1_f32(vsum);
     #else
         const int np = (n & ~(GGML_F32_STEP - 1));
 
@@ -197,7 +207,7 @@ void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * G
 
     ggml_float sumf = 0.0;
 
-#if defined(GGML_SIMD)
+#if defined(GGML_SIMD) && !defined(__riscv_v_intrinsic)
     const int np = (n & ~(GGML_F16_STEP - 1));
 
     GGML_F16_VEC sum[GGML_F16_ARR] = { GGML_F16_VEC_ZERO };
@@ -325,6 +335,15 @@ ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float
         vst1q_f32(y + i, val);
         sum += (ggml_float)vaddvq_f32(val);
     }
+#elif defined(__riscv_v_intrinsic)
+    vfloat64m1_t vsum = __riscv_vfmv_v_f_f64m1(0, 1);
+    for (int avl; i < n; i += avl) {
+        avl = __riscv_vsetvl_e32m2(n - i);
+        vfloat32m2_t val = ggml_v_expf_m2(__riscv_vfsub_vf_f32m2(__riscv_vle32_v_f32m2(&x[i], avl), max, avl), avl);
+        __riscv_vse32_v_f32m2(&y[i], val, avl);
+        vsum = __riscv_vfwredusum_vs_f32m2_f64m1(val, vsum, avl);
+    }
+    return (ggml_float)__riscv_vfmv_f_s_f64m1_f64(vsum);
 #endif
     for (; i < n; ++i) {
         float val = expf(x[i] - max);

ggml/src/ggml-cpu/vec.h

@@ -119,6 +119,14 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
     }
 
 #if defined(GGML_SIMD)
+#if defined(__riscv_v_intrinsic)
+    // todo: RVV impl
+    for (int i = 0; i < n; ++i) {
+        for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
+            sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
+        }
+    }
+#else
     const int np = (n & ~(GGML_F16_STEP - 1));
 
     GGML_F16_VEC sum[GGML_VEC_DOT_UNROLL][GGML_F16_ARR] = { { GGML_F16_VEC_ZERO } };
@@ -149,6 +157,7 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
             sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
         }
     }
+#endif
 #else
     for (int i = 0; i < n; ++i) {
         for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
@@ -243,6 +252,14 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
 
             svst1_f32(pg, y + np2, ay1);
         }
+    #elif defined(__riscv_v_intrinsic)
+        for (int i = 0, avl; i < n; i += avl) {
+            avl = __riscv_vsetvl_e32m8(n - i);
+            vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[i], avl);
+            vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
+            vfloat32m8_t ny = __riscv_vfmadd_vf_f32m8(ax, v, ay, avl);
+            __riscv_vse32_v_f32m8(&y[i], ny, avl);
+        }
     #else
         const int np = (n & ~(GGML_F32_STEP - 1));
 
@@ -276,6 +293,13 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
 
 inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y, const ggml_fp16_t * GGML_RESTRICT x, const float v) {
 #if defined(GGML_SIMD)
+#if defined(__riscv_v_intrinsic)
+    // todo: RVV impl
+    // scalar
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
+    }
+#else
     const int np = (n & ~(GGML_F16_STEP - 1));
 
     GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
@@ -297,6 +321,7 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y,
     for (int i = np; i < n; ++i) {
         y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
     }
+#endif
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
@@ -324,6 +349,16 @@ inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int
                 y[i] += x[k][i]*v[k][0];
             }
         }
+    #elif defined(__riscv_v_intrinsic)
+        for (int i = 0, avl; i < n; i += avl) {
+            avl = __riscv_vsetvl_e32m8(n - i);
+            vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
+            for (int k = 0; k < GGML_VEC_MAD_UNROLL; k++) {
+                vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[k][i], avl);
+                ay = __riscv_vfmadd_vf_f32m8(ax, v[k][0], ay, avl);
+            }
+            __riscv_vse32_v_f32m8(&y[i], ay, avl);
+        }
     #else
         const int np = (n & ~(GGML_F32_STEP - 1));
 
@@ -375,6 +410,14 @@ inline static void ggml_vec_mad1_f32(const int n, float * y, const float * x, co
         for (int i = 0; i < n; ++i) {
             y[i] = x[i]*s + b;
         }
+    #elif defined(__riscv_v_intrinsic)
+        for (int i = 0, avl; i < n; i += avl) {
+            avl = __riscv_vsetvl_e32m8(n - i);
+            vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[i], avl);
+            vfloat32m8_t vb = __riscv_vfmv_v_f_f32m8(b, avl);
+            vfloat32m8_t ny = __riscv_vfmadd_vf_f32m8(ax, s, vb, avl);
+            __riscv_vse32_v_f32m8(&y[i], ny, avl);
+        }
     #else
         const int np = (n & ~(GGML_F32_STEP - 1));
 
@@ -436,6 +479,13 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
             ay1 = svmul_f32_m(pg, ay1, vx);
             svst1_f32(pg, y + np, ay1);
         }
+    #elif defined(__riscv_v_intrinsic)
+        for (int i = 0, avl; i < n; i += avl) {
+            avl = __riscv_vsetvl_e32m8(n - i);
+            vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
+            vfloat32m8_t ny = __riscv_vfmul_vf_f32m8(ay, v, avl);
+            __riscv_vse32_v_f32m8(&y[i], ny, avl);
+        }
     #else
         const int np = (n & ~(GGML_F32_STEP - 1));
 
@@ -467,6 +517,13 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 
 inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float v) {
 #if defined(GGML_SIMD)
+#if defined(__riscv_v_intrinsic)
+    // todo: RVV impl
+    // scalar
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
+    }
+#else
     const int np = (n & ~(GGML_F16_STEP - 1));
 
     GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
@@ -486,6 +543,7 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float
     for (int i = np; i < n; ++i) {
         y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
     }
+#endif
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
@@ -928,7 +986,51 @@ inline static __m128 ggml_v_silu(__m128 x) {
     return _mm_div_ps(x, one_plus_exp_neg_x);
 }
 
-#endif // __ARM_NEON / __AVX2__ / __SSE2__
+#elif defined(__riscv_v_intrinsic)
+
+// adapted from arm limited optimized routine
+// the maximum error is 1.45358 plus 0.5 ulps
+// numbers above 88.38 will flush to infinity
+// numbers beneath -103.97 will flush to zero
+inline static vfloat32m2_t ggml_v_expf_m2(vfloat32m2_t x, int vl) {
+    const vfloat32m2_t r = __riscv_vfmv_v_f_f32m2(0x1.8p23f, vl);
+#ifdef __riscv_xtheadvector
+    // workaround for compiler bug (gcc 14.3.0: Error: unrecognized opcode `th.vmv1r.v v2,v4')
+    vfloat32m2_t z = __riscv_vfadd_vf_f32m2(r, 0.0f, vl);
+    z = __riscv_vfmacc_vf_f32m2(z, 0x1.715476p+0f, x, vl);
+#else
+    const vfloat32m2_t z = __riscv_vfmacc_vf_f32m2(r, 0x1.715476p+0f, x, vl);
+#endif
+    const vfloat32m2_t n = __riscv_vfsub_vv_f32m2(z, r, vl);
+    const vfloat32m2_t b = __riscv_vfnmsac_vf_f32m2(__riscv_vfnmsac_vf_f32m2(x, 0x1.62e4p-1f, n, vl),
+                                                    0x1.7f7d1cp-20f, n, vl);
+    const vuint32m2_t e = __riscv_vsll_vx_u32m2(__riscv_vreinterpret_v_f32m2_u32m2(z), 23, vl);
+    const vfloat32m2_t k = __riscv_vreinterpret_v_u32m2_f32m2(__riscv_vadd_vx_u32m2(e, 0x3f800000, vl)); // 1.0f
+    const vbool16_t c = __riscv_vmfgt_vf_f32m2_b16(__riscv_vfabs_v_f32m2(n, vl), 126.0f, vl);
+    const vfloat32m2_t u = __riscv_vfmul_vv_f32m2(b, b, vl);
+    const vfloat32m2_t j = __riscv_vfmacc_vv_f32m2(
+        __riscv_vfmul_vf_f32m2(b, 0x1.ffffecp-1f, vl),
+        __riscv_vfmacc_vv_f32m2(
+            __riscv_vfmacc_vf_f32m2(__riscv_vfmv_v_f_f32m2(0x1.fffdb6p-2f, vl), 0x1.555e66p-3f, b, vl),
+            __riscv_vfmacc_vf_f32m2(__riscv_vfmv_v_f_f32m2(0x1.573e2ep-5f, vl), 0x1.0e4020p-7f, b, vl),
+            u, vl), u, vl);
+    if (!__riscv_vcpop_m_b16(c, vl))
+        return __riscv_vfmacc_vv_f32m2(k, j, k, vl);
+    const vbool16_t  dm = __riscv_vmfle_vf_f32m2_b16(n, 0.0f, vl);
+    const vuint32m2_t d = __riscv_vmerge_vxm_u32m2(__riscv_vmv_v_x_u32m2(0, vl), 0x82000000, dm, vl);
+    const vfloat32m2_t s1 = __riscv_vreinterpret_v_u32m2_f32m2(__riscv_vadd_vx_u32m2(d, 0x7f000000, vl));
+    const vfloat32m2_t s2 = __riscv_vreinterpret_v_u32m2_f32m2(__riscv_vsub_vv_u32m2(e, d, vl));
+    const vfloat32m2_t r1 = __riscv_vmerge_vvm_f32m2(
+        __riscv_vfmacc_vv_f32m2(k, k, j, vl),
+        __riscv_vfmul_vv_f32m2(__riscv_vfmacc_vv_f32m2(s2, s2, j, vl), s1, vl),
+        c, vl);
+    return __riscv_vmerge_vvm_f32m2(
+        r1, __riscv_vfmul_vv_f32m2(s1, s1, vl),
+        __riscv_vmfgt_vf_f32m2_b16(__riscv_vfabs_v_f32m2(n, vl), 192.0f, vl),
+        vl);
+}
+
+#endif // __ARM_NEON / __AVX2__ / __SSE2__ / __riscv_v_intrinsic
 
 inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {

ggml/src/ggml-cuda/CMakeLists.txt

@@ -24,12 +24,6 @@ if (CUDAToolkit_FOUND)
         #     for best performance and to also build real architectures for the most commonly used GPUs.
         if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
             set(CMAKE_CUDA_ARCHITECTURES "native")
-        elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
-            if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8")
-                set(CMAKE_CUDA_ARCHITECTURES "60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real")
-            else()
-                set(CMAKE_CUDA_ARCHITECTURES "60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real")
-            endif()
         else()
             if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8")
                 set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real")
@@ -91,10 +85,6 @@ if (CUDAToolkit_FOUND)
         add_compile_definitions(GGML_CUDA_NO_FA)
     endif()
 
-    if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
-        add_compile_definitions(GGML_CUDA_F16)
-    endif()
-
     if (GGML_CUDA_NO_PEER_COPY)
         add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
     endif()
@@ -104,7 +94,11 @@ if (CUDAToolkit_FOUND)
             # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
             target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas)
         else ()
-            target_link_libraries(ggml-cuda PRIVATE  CUDA::cudart_static CUDA::cublas_static)
+            if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "10.1")
+                target_link_libraries(ggml-cuda PRIVATE  CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+            else()
+                target_link_libraries(ggml-cuda PRIVATE  CUDA::cudart_static CUDA::cublas_static)
+            endif()
         endif()
     else()
         target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas)

ggml/src/ggml-cuda/add-id.cu

@@ -11,14 +11,14 @@ static __global__ void add_id_kernel(
     const int64_t i1 = blockIdx.x;
     const int64_t i2 = blockIdx.y;
 
-    const int i11 = *(int32_t *) ((char *) src2 + i1*sizeof(int32_t) + i2*nb21);
+    const int i11 = *(const int32_t *) ((const char *) src2 + i1*sizeof(int32_t) + i2*nb21);
 
     const size_t nb1 = ne0 * sizeof(float);
     const size_t nb2 = ne1 * nb1;
 
     float * dst_row = (float *)((char *)dst + i1*nb1 + i2*nb2);
-    const float * src0_row = (const float *)((char *)src0 +  i1*nb01 + i2*nb02);
-    const float * src1_row = (const float *)((char *)src1 + i11*nb11);
+    const float * src0_row = (const float *)((const char *)src0 +  i1*nb01 + i2*nb02);
+    const float * src1_row = (const float *)((const char *)src1 + i11*nb11);
 
     for (int64_t i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
         dst_row[i0] = src0_row[i0] + src1_row[i0];

ggml/src/ggml-cuda/common.cuh

@@ -78,6 +78,8 @@
 #define GGML_CUDA_CC_IS_CDNA3(cc) (cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_RDNA1)
 
 // Moore Threads
+#define MUSART_HMASK 40300 // MUSA rc4.3, min. ver. for half2 -> uint mask comparisons
+
 #define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
 #define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
 #define GGML_CUDA_CC_NG  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
@@ -105,9 +107,9 @@ constexpr bool ggml_cuda_has_arch(const int arch) {
     return ggml_cuda_has_arch_impl(arch, __CUDA_ARCH_LIST__);
 }
 
-constexpr int ggml_cuda_highest_compiled_arch_impl(const int arch, const int cur) {
+constexpr int ggml_cuda_highest_compiled_arch_impl(const int /*arch*/, const int cur) {
     if (cur == 0) {
-        GGML_ABORT("ggml was not compiled with any CUDA arch <= %d", arch);
+        return -1;
     }
     return cur;
 }
@@ -204,14 +206,6 @@ static const char * cu_get_error_str(CUresult err) {
 #define GGML_CUDA_ASSUME(x)
 #endif // CUDART_VERSION >= 11010
 
-#ifdef GGML_CUDA_F16
-typedef half dfloat; // dequantize float
-typedef half2 dfloat2;
-#else
-typedef float dfloat; // dequantize float
-typedef float2 dfloat2;
-#endif // GGML_CUDA_F16
-
 #if (!defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)) || (defined(GGML_USE_HIP) && !defined(GGML_HIP_NO_VMM))
 #define GGML_USE_VMM
 #endif // (!defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)) || (defined(GGML_USE_HIP) && !defined(GGML_HIP_NO_VMM))
@@ -426,16 +420,28 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ int warp_reduce_all(int x) {
-#ifdef GGML_USE_HIP
+    if (width == ggml_cuda_get_physical_warp_size()) {
+        return __all_sync(0xffffffff, x);
+    } else {
 #pragma unroll
-    for (int offset = width/2; offset > 0; offset >>= 1) {
-        x = x && __shfl_xor_sync(0xffffffff, x, offset, width);
+        for (int offset = width/2; offset > 0; offset >>= 1) {
+            x = __shfl_xor_sync(0xffffffff, x, offset, width) && x;
+        }
+        return x;
+    }
+}
+
+template<int width = WARP_SIZE>
+static __device__ __forceinline__ int warp_reduce_any(int x) {
+    if (width == ggml_cuda_get_physical_warp_size()) {
+        return __any_sync(0xffffffff, x);
+    } else {
+#pragma unroll
+        for (int offset = width/2; offset > 0; offset >>= 1) {
+            x = __shfl_xor_sync(0xffffffff, x, offset, width) || x;
+        }
+        return x;
     }
-    return x;
-#else
-    static_assert(width == WARP_SIZE, "width != WARP_SIZE not implemented");
-    return __all_sync(0xffffffff, x);
-#endif // GGML_USE_HIP
 }
 
 template<int width = WARP_SIZE>
@@ -490,13 +496,14 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
 #endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || defined(GGML_USE_HIP)
 }
 
-#if CUDART_VERSION < CUDART_HMASK
+#if (defined(CUDART_VERSION) && CUDART_VERSION < CUDART_HMASK) || defined(GGML_USE_HIP) || \
+    (defined(MUSART_VERSION) && MUSART_VERSION < MUSART_HMASK)
 static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half2 b) {
     const uint32_t mask_low  = 0x0000FFFF * (float( __low2half(a)) > float( __low2half(b)));
     const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b)));
     return mask_low | mask_high;
 }
-#endif // CUDART_VERSION < CUDART_HMASK
+#endif // (defined(CUDART_VERSION) && CUDART_VERSION < CUDART_HMASK) || defined(GGML_USE_HIP) || (defined(MUSART_VERSION) && MUSART_VERSION < MUSART_HMASK)
 
 static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
 #if defined(GGML_USE_HIP)
@@ -556,7 +563,7 @@ static __device__ __forceinline__ float ggml_cuda_e8m0_to_fp32(uint8_t x) {
 #endif // CUDART_VERSION >= 12050
 }
 
-typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
+typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, float2 & v);
 
 static __device__ __forceinline__ float get_alibi_slope(
     const float max_bias, const uint32_t h, const uint32_t n_head_log2, const float m0, const float m1

ggml/src/ggml-cuda/conv-transpose-1d.cu

@@ -34,10 +34,7 @@ static  __global__ void conv_transpose_1d_kernel(
         }
     }
     dst[global_index] = accumulator;
-    GGML_UNUSED(p0); GGML_UNUSED(d0); GGML_UNUSED(src0_ne3);
-    GGML_UNUSED(src1_ne3); GGML_UNUSED(dst_ne3);
-    GGML_UNUSED(src1_ne1); GGML_UNUSED(dst_ne1);
-    GGML_UNUSED(src1_ne2); GGML_UNUSED(dst_ne2);
+    GGML_UNUSED_VARS(p0, d0, src0_ne3, src1_ne3, dst_ne3, src1_ne1, dst_ne1, src1_ne2, dst_ne2);
 }
 
 static void conv_transpose_1d_f32_f32_cuda(

ggml/src/ggml-cuda/convert.cu

@@ -27,7 +27,7 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
     const int64_t y_offset = qr == 1 ? 1 : qk/2;
 
     // dequantize
-    dfloat2 v;
+    float2 v;
     dequantize_kernel(vx, ib, iqs, v);
 
     const int64_t iy0 = ((i03*ne02 + i02)*ne01 + i01)*ne00 + iybs + iqs;
@@ -71,9 +71,7 @@ static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, h
         y2[iy/2 + threadIdx.x] = __hmul2(make_half2(qs.x, qs.y), __half2half2(d));
     }
 #else
-    GGML_UNUSED(vx);
-    GGML_UNUSED(y);
-    GGML_UNUSED(k);
+    GGML_UNUSED_VARS(vx, y, k);
     NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 }

ggml/src/ggml-cuda/cpy.cu

@@ -42,7 +42,7 @@ static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
 
 #pragma unroll
     for (int j = 0; j < QK8_0; j += 2) {
-        dfloat2 dq;
+        float2 dq;
         dequantize_q8_0(cxi, 0, j, dq);
         *(cdstf + j) = dq.x;
         *(cdstf + j + 1) = dq.y;
@@ -55,7 +55,7 @@ static __device__ void cpy_blck_q_f32(const char * cxi, char * cdsti) {
 
 #pragma unroll
     for (int j = 0; j < qk/2; j++) {
-        dfloat2 dq;
+        float2 dq;
         dequant(cxi, 0, j, dq);
         *(cdstf + j) = dq.x;
         *(cdstf + j + qk/2) = dq.y;
@@ -134,8 +134,7 @@ void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_des
     CUDA_CHECK(cudaMemcpyAsync(cuda_graph->dest_ptrs_d, host_dest_ptrs, host_dest_ptrs_size*sizeof(char *), cudaMemcpyHostToDevice, stream));
     cuda_graph->graph_cpynode_index = 0; // reset index
 #else
-    GGML_UNUSED(cuda_graph); GGML_UNUSED(host_dest_ptrs);
-    GGML_UNUSED(host_dest_ptrs_size); GGML_UNUSED(stream);
+    GGML_UNUSED_VARS(cuda_graph, host_dest_ptrs, host_dest_ptrs_size, stream);
 #endif
 }
 

ggml/src/ggml-cuda/dequantize.cuh

@@ -1,48 +1,37 @@
 #include "common.cuh"
 
-static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int64_t ib, const int iqs, float2 & v){
     const block_q4_0 * x = (const block_q4_0 *) vx;
 
-    const dfloat d = x[ib].d;
+    const float d = x[ib].d;
 
     const int vui = x[ib].qs[iqs];
 
     v.x = vui & 0xF;
     v.y = vui >> 4;
 
-#ifdef GGML_CUDA_F16
-    v = __hsub2(v, {8.0f, 8.0f});
-    v = __hmul2(v, {d, d});
-#else
     v.x = (v.x - 8.0f) * d;
     v.y = (v.y - 8.0f) * d;
-#endif // GGML_CUDA_F16
 }
 
-static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int64_t ib, const int iqs, float2 & v){
     const block_q4_1 * x = (const block_q4_1 *) vx;
 
-    const dfloat d = __low2half(x[ib].dm);
-    const dfloat m = __high2half(x[ib].dm);
+    const float2 dm = __half22float2(x[ib].dm);
 
     const int vui = x[ib].qs[iqs];
 
     v.x = vui & 0xF;
     v.y = vui >> 4;
 
-#ifdef GGML_CUDA_F16
-    v = __hmul2(v, {d, d});
-    v = __hadd2(v, {m, m});
-#else
-    v.x = (v.x * d) + m;
-    v.y = (v.y * d) + m;
-#endif // GGML_CUDA_F16
+    v.x = (v.x * dm.x) + dm.y;
+    v.y = (v.y * dm.x) + dm.y;
 }
 
-static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int64_t ib, const int iqs, float2 & v){
     const block_q5_0 * x = (const block_q5_0 *) vx;
 
-    const dfloat d = x[ib].d;
+    const float d = x[ib].d;
 
     uint32_t qh;
     memcpy(&qh, x[ib].qh, sizeof(qh));
@@ -53,20 +42,14 @@ static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const in
     v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
     v.y = ((x[ib].qs[iqs] >>  4) | xh_1);
 
-#ifdef GGML_CUDA_F16
-    v = __hsub2(v, {16.0f, 16.0f});
-    v = __hmul2(v, {d, d});
-#else
     v.x = (v.x - 16.0f) * d;
     v.y = (v.y - 16.0f) * d;
-#endif // GGML_CUDA_F16
 }
 
-static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int64_t ib, const int iqs, float2 & v){
     const block_q5_1 * x = (const block_q5_1 *) vx;
 
-    const dfloat d = __low2half(x[ib].dm);
-    const dfloat m = __high2half(x[ib].dm);
+    const float2 dm = __half22float2(x[ib].dm);
 
     uint32_t qh;
     memcpy(&qh, x[ib].qh, sizeof(qh));
@@ -77,27 +60,18 @@ static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const in
     v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
     v.y = ((x[ib].qs[iqs] >>  4) | xh_1);
 
-#ifdef GGML_CUDA_F16
-    v = __hmul2(v, {d, d});
-    v = __hadd2(v, {m, m});
-#else
-    v.x = (v.x * d) + m;
-    v.y = (v.y * d) + m;
-#endif // GGML_CUDA_F16
+    v.x = (v.x * dm.x) + dm.y;
+    v.y = (v.y * dm.x) + dm.y;
 }
 
-static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int64_t ib, const int iqs, float2 & v){
     const block_q8_0 * x = (const block_q8_0 *) vx;
 
-    const dfloat d = x[ib].d;
+    const float d = x[ib].d;
 
     v.x = x[ib].qs[iqs + 0];
     v.y = x[ib].qs[iqs + 1];
 
-#ifdef GGML_CUDA_F16
-    v = __hmul2(v, {d, d});
-#else
     v.x *= d;
     v.y *= d;
-#endif // GGML_CUDA_F16
 }

ggml/src/ggml-cuda/fattn-common.cuh

@@ -704,28 +704,6 @@ static __global__ void flash_attn_combine_results(
     dst[tid] = VKQ_numerator / VKQ_denominator;
 }
 
-[[noreturn]]
-static void on_no_fattn_vec_case(const int D) {
-    if (D == 64) {
-        fprintf(stderr, "Unsupported KV type combination for head_size 64.\n");
-        fprintf(stderr, "By default only f16 KV cache is supported.\n");
-        fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for V cache quantization support.\n");
-        GGML_ABORT("fatal error");
-    } else if (D == 128) {
-        fprintf(stderr, "Unsupported KV type combination for head_size 128.\n");
-        fprintf(stderr, "Supported combinations:\n");
-        fprintf(stderr, "  - K == q4_0, V == q4_0,  4.50 BPV\n");
-        fprintf(stderr, "  - K == q8_0, V == q8_0,  8.50 BPV\n");
-        fprintf(stderr, "  - K == f16,  V == f16,  16.00 BPV\n");
-        fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for all combinations of q4_0, q4_1, q5_0, q5_1, q8_0, and f16.\n");
-        GGML_ABORT("fatal error");
-    } else {
-        fprintf(stderr, "Unsupported KV type combination for head_size %d.\n", D);
-        fprintf(stderr, "Only f16 is supported.\n");
-        GGML_ABORT("fatal error");
-    }
-}
-
 template <int DV, int ncols1, int ncols2>
 void launch_fattn(
     ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, const int nwarps, const size_t nbytes_shared,

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -767,14 +767,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         }
     }
 #else
-    GGML_UNUSED(Q_f2); GGML_UNUSED(K_h2); GGML_UNUSED(V_h2);
-    GGML_UNUSED(mask_h2); GGML_UNUSED(dstk); GGML_UNUSED(dstk_fixup);
-    GGML_UNUSED(scale); GGML_UNUSED(slope); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_K); GGML_UNUSED(stride_V);
-    GGML_UNUSED(stride_mask); GGML_UNUSED(tile_K);
-    GGML_UNUSED(tile_V); GGML_UNUSED(tile_mask); GGML_UNUSED(Q_B);
-    GGML_UNUSED(VKQ_C); GGML_UNUSED(KQ_max); GGML_UNUSED(KQ_rowsum);
-    GGML_UNUSED(kb0); GGML_UNUSED(tile_Q);
+    GGML_UNUSED_VARS(Q_f2, K_h2, V_h2, mask_h2, dstk, dstk_fixup,
+        scale, slope, logit_softcap, ne01, ne02,
+        stride_K, stride_V, stride_mask,
+        tile_Q, tile_K, tile_V, tile_mask,
+        Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0);
     NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
 }
@@ -1236,12 +1233,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         }
     }
 #else
-    GGML_UNUSED(Q_f2); GGML_UNUSED(K_h2); GGML_UNUSED(V_h2);
-    GGML_UNUSED(mask_h2); GGML_UNUSED(dstk); GGML_UNUSED(dstk_fixup);
-    GGML_UNUSED(scale); GGML_UNUSED(slope); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_Q1);
-    GGML_UNUSED(stride_Q2); GGML_UNUSED(stride_K); GGML_UNUSED(stride_V); GGML_UNUSED(stride_mask);
-    GGML_UNUSED(jt); GGML_UNUSED(kb0_start); GGML_UNUSED(kb0_stop);
+    GGML_UNUSED_VARS(Q_f2, K_h2, V_h2, mask_h2, sinks_f, dstk, dstk_fixup,
+        scale, slope, logit_softcap, ne01, ne02,
+        stride_Q1, stride_Q2, stride_K, stride_V, stride_mask,
+        jt, kb0_start, kb0_stop);
     NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
 }
@@ -1395,17 +1390,15 @@ static __global__ void flash_attn_ext_f16(
         (Q_f2, K_h2, V_h2, mask_h2, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
          ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel);
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta);
-    GGML_UNUSED(scale); GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
-    GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
-    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
-    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
-    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
     NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(TURING_MMA_AVAILABLE)
 }

ggml/src/ggml-cuda/fattn-tile-f16.cu

@@ -258,7 +258,7 @@ static __global__ void flash_attn_tile_ext_f16(
             const half val = hexp(sink - kqmax[j0/nwarps]);
             kqsum[j0/nwarps] = kqsum[j0/nwarps] * KQ_max_scale;
             if (threadIdx.x == 0) {
-                kqsum[j0/nwarps].x = __hadd(kqsum[j0/nwarps].x, val);
+                kqsum[j0/nwarps].x = __hadd(__low2half(kqsum[j0/nwarps]), val);
             }
 
 #pragma unroll
@@ -299,17 +299,15 @@ static __global__ void flash_attn_tile_ext_f16(
         }
     }
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
-    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
-    GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
-    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
-    GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
-    GGML_UNUSED(nb23);
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
     NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }

ggml/src/ggml-cuda/fattn-tile-f32.cu

@@ -38,17 +38,15 @@ static __global__ void flash_attn_tile_ext_f32(
     return;
 #endif // FP16_MMA_AVAILABLE
     if (use_logit_softcap && !(D == 128 || D == 256)) {
-        GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
-        GGML_UNUSED(dst); GGML_UNUSED(dst_meta);
-        GGML_UNUSED(scale); GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-        GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-        GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
-        GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
-        GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-        GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
-        GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
-        GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
-        GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
+        GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+            max_bias, m0, m1, n_head_log2, logit_softcap,
+            ne00, ne01, ne02, ne03,
+                  nb01, nb02, nb03,
+            ne10, ne11, ne12, ne13,
+                  nb11, nb12, nb13,
+                  nb21, nb22, nb23,
+                  ne31, ne32, ne33,
+                  nb31, nb32, nb33);
         NO_DEVICE_CODE;
         return;
     }
@@ -312,17 +310,15 @@ static __global__ void flash_attn_tile_ext_f32(
         }
     }
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta);
-    GGML_UNUSED(scale); GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
-    GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
-    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
-    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
-    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
     NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }

ggml/src/ggml-cuda/fattn-vec-f16.cuh

@@ -349,17 +349,15 @@ static __global__ void flash_attn_vec_ext_f16(
         dst_meta[((sequence*ne01 + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta);
-    GGML_UNUSED(scale); GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
-    GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
-    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
-    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
-    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
     NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }

ggml/src/ggml-cuda/fattn-vec-f32.cuh

@@ -37,17 +37,15 @@ static __global__ void flash_attn_vec_ext_f32(
 
     // Skip unused kernel variants for faster compilation:
     if (use_logit_softcap && !(D == 128 || D == 256)) {
-        GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
-        GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
-        GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-        GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-        GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
-        GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
-        GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
-        GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
-        GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
-        GGML_UNUSED(nb23);
+        GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+            max_bias, m0, m1, n_head_log2, logit_softcap,
+            ne00, ne01, ne02, ne03,
+                  nb01, nb02, nb03,
+            ne10, ne11, ne12, ne13,
+                  nb11, nb12, nb13,
+                  nb21, nb22, nb23,
+                  ne31, ne32, ne33,
+                  nb31, nb32, nb33);
         NO_DEVICE_CODE;
         return;
     }
@@ -345,17 +343,15 @@ static __global__ void flash_attn_vec_ext_f32(
         dst_meta[((sequence*ne01 + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
-    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
-    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
-    GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
-    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
-    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
     NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }

ggml/src/ggml-cuda/fattn-wmma-f16.cu

@@ -471,16 +471,15 @@ static __global__ void flash_attn_ext_f16(
         dst_meta[j_dst_unrolled] = dst_meta_val;
     }
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
-    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
-    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33); GGML_UNUSED(nb31);
-    GGML_UNUSED(nb32); GGML_UNUSED(nb33); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
-    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
-    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
     NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)))
 }

ggml/src/ggml-cuda/fattn.cu

@@ -190,7 +190,7 @@ static void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, gg
     FATTN_VEC_F16_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
 #endif // GGML_CUDA_FA_ALL_QUANTS
 
-    on_no_fattn_vec_case(Q->ne[0]);
+    GGML_ABORT("fatal error");
 }
 
 #define FATTN_VEC_F32_CASE(D, type_K, type_V)                               \
@@ -265,74 +265,184 @@ static void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, gg
     FATTN_VEC_F32_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
 #endif // GGML_CUDA_FA_ALL_QUANTS
 
-    on_no_fattn_vec_case(Q->ne[0]);
+    GGML_ABORT("fatal error");
 }
 
-void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+// Best FlashAttention kernel for a specific GPU:
+enum best_fattn_kernel {
+    BEST_FATTN_KERNEL_NONE     =   0,
+    BEST_FATTN_KERNEL_TILE_F32 = 200,
+    BEST_FATTN_KERNEL_TILE_F16 = 210,
+    BEST_FATTN_KERNEL_VEC_F32  = 100,
+    BEST_FATTN_KERNEL_VEC_F16  = 110,
+    BEST_FATTN_KERNEL_WMMA_F16 = 300,
+    BEST_FATTN_KERNEL_MMA_F16  = 400,
+};
+
+static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const ggml_tensor * dst) {
+#ifndef FLASH_ATTN_AVAILABLE
+    GGML_UNUSED(device); GGML_UNUSED(dst);
+    return BEST_FATTN_KERNEL_NONE;
+#endif// FLASH_ATTN_AVAILABLE
+
     const ggml_tensor * KQV   = dst;
     const ggml_tensor * Q     = dst->src[0];
     const ggml_tensor * K     = dst->src[1];
     const ggml_tensor * V     = dst->src[2];
     const ggml_tensor * mask  = dst->src[3];
 
-    ggml_cuda_set_device(ctx.device);
-    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-    const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
+    const int gqa_ratio = Q->ne[2] / K->ne[2];
+    GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
+
+    const int cc = ggml_cuda_info().devices[device].cc;
+    const int warp_size = ggml_cuda_info().devices[device].warp_size;
     const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV);
 
-#if defined(GGML_HIP_ROCWMMA_FATTN)
-    if (GGML_CUDA_CC_IS_AMD(cc) && fp16_mma_available(cc)) {
-        ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
-        return;
-    }
-#endif // defined(GGML_HIP_ROCWMMA_FATTN)
-
-    if (!fast_fp16_available(cc)) {
-        if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
-            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
-        } else {
-            ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
-        }
-        return;
-    }
-
-    if (!fp16_mma_available(cc)) {
-        if (prec == GGML_PREC_DEFAULT) {
-            if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
-                ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
-            } else {
-                ggml_cuda_flash_attn_ext_tile_f16(ctx, dst);
+    switch (K->ne[0]) {
+        case  64:
+        case 128:
+        case 256:
+            if (V->ne[0] != K->ne[0]) {
+                return BEST_FATTN_KERNEL_NONE;
             }
-        } else {
-            if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
-                ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
-            } else {
-                ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
+            break;
+        case  80:
+        case  96:
+        case 112:
+            if (V->ne[0] != K->ne[0]) {
+                return BEST_FATTN_KERNEL_NONE;
             }
-        }
-        return;
+            if (!fp16_mma_available(cc) && !turing_mma_available(cc)) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+            break;
+        case 576:
+            if (V->ne[0] != 512) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+            if (!turing_mma_available(cc) || gqa_ratio % 16 != 0) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+            break;
+        default:
+            return BEST_FATTN_KERNEL_NONE;
+    }
+
+#ifndef GGML_CUDA_FA_ALL_QUANTS
+    if (K->type != V->type) {
+        return BEST_FATTN_KERNEL_NONE;
+    }
+#endif // GGML_CUDA_FA_ALL_QUANTS
+
+    switch (K->type) {
+        case GGML_TYPE_F16:
+            break;
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+#ifndef GGML_CUDA_FA_ALL_QUANTS
+            return BEST_FATTN_KERNEL_NONE;
+#endif // GGML_CUDA_FA_ALL_QUANTS
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q8_0:
+#ifdef GGML_CUDA_FA_ALL_QUANTS
+            if (K->ne[0] != 128 && K->ne[0] != 64) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+#else
+            if (K->ne[0] != 128) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+#endif // GGML_CUDA_FA_ALL_QUANTS
+            break;
+        default:
+            return BEST_FATTN_KERNEL_NONE;
+    }
+
+    switch (V->type) {
+        case GGML_TYPE_F16:
+            break;
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q8_0:
+            if (K->ne[0] != 128) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+            break;
+        default:
+            return BEST_FATTN_KERNEL_NONE;
+    }
+
+    if (mask && mask->ne[2] != 1) {
+        return BEST_FATTN_KERNEL_NONE;
     }
 
-    const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations
-    const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
-    const bool mma_faster_for_rtx4000 = Q->ne[3] > 1 || (Q->ne[2] > 4*K->ne[2] && K->ne[1] >= 8192);
-    const bool mma_faster_for_bs1 = turing_mma_available(cc) && gqa_opt_applies && !mma_needs_data_conversion &&
-        (cc < GGML_CUDA_CC_ADA_LOVELACE || mma_faster_for_rtx4000);
     const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % (2*warp_size) == 0;
-    if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
-        if (prec == GGML_PREC_DEFAULT) {
-            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
-        } else {
-            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+
+    // If Turing tensor cores available, use them except for some cases with batch size 1:
+    if (turing_mma_available(cc)) {
+        const bool gqa_opt_applies = gqa_ratio % 2 == 0 && mask; // The mma-based kernels have GQA-specific optimizations
+        const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
+        const bool mma_faster_for_rtx4000 = Q->ne[3] > 1 || (gqa_ratio > 4 && K->ne[1] >= 8192);
+        const bool mma_faster_for_bs1 = gqa_opt_applies && !mma_needs_data_conversion &&
+            (cc < GGML_CUDA_CC_ADA_LOVELACE || mma_faster_for_rtx4000);
+        if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
+            if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
+                return BEST_FATTN_KERNEL_VEC_F16;
+            }
+            return BEST_FATTN_KERNEL_VEC_F32;
         }
-        return;
+        return BEST_FATTN_KERNEL_MMA_F16;
     }
 
-    // The MMA implementation needs Turing or newer, use the old WMMA code for Volta:
-    if (fp16_mma_available(cc) && !turing_mma_available(cc)) {
-        ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
-        return;
+    // Use kernels specializes for small batch sizes if possible:
+    if (Q->ne[1] <= 8 && can_use_vector_kernel) {
+        if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
+            return BEST_FATTN_KERNEL_VEC_F16;
+        }
+        return BEST_FATTN_KERNEL_VEC_F32;
     }
 
-    ggml_cuda_flash_attn_ext_mma_f16(ctx, dst);
+    // For large batch sizes, use the WMMA kernel if possible:
+    if (fp16_mma_available(cc)) {
+        return BEST_FATTN_KERNEL_WMMA_F16;
+    }
+
+    // If there is no suitable kernel for tensor cores or small batch sizes, use the generic kernel for large batch sizes:
+    if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
+        return BEST_FATTN_KERNEL_TILE_F16;
+    }
+    return BEST_FATTN_KERNEL_TILE_F32;
+}
+
+void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_set_device(ctx.device);
+    switch (ggml_cuda_get_best_fattn_kernel(ggml_cuda_get_device(), dst)) {
+        case BEST_FATTN_KERNEL_NONE:
+            GGML_ABORT("fatal error");
+        case BEST_FATTN_KERNEL_TILE_F32:
+            ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
+            break;
+        case BEST_FATTN_KERNEL_TILE_F16:
+            ggml_cuda_flash_attn_ext_tile_f16(ctx, dst);
+            break;
+        case BEST_FATTN_KERNEL_VEC_F32:
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+            break;
+        case BEST_FATTN_KERNEL_VEC_F16:
+            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+            break;
+        case BEST_FATTN_KERNEL_WMMA_F16:
+            ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
+            break;
+        case BEST_FATTN_KERNEL_MMA_F16:
+            ggml_cuda_flash_attn_ext_mma_f16(ctx, dst);
+            break;
+    }
+}
+
+bool ggml_cuda_flash_attn_ext_supported(int device, const ggml_tensor * dst) {
+    return ggml_cuda_get_best_fattn_kernel(device, dst) != BEST_FATTN_KERNEL_NONE;
 }

ggml/src/ggml-cuda/fattn.cuh

@@ -1,3 +1,5 @@
 #include "common.cuh"
 
 void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+bool ggml_cuda_flash_attn_ext_supported(int device, const ggml_tensor * dst);

ggml/src/ggml-cuda/getrows.cu

@@ -32,7 +32,7 @@ static __global__ void k_get_rows(
     const int y_offset = qr == 1 ? 1 : qk/2;
 
     // dequantize
-    dfloat2 v;
+    float2 v;
     dequantize_kernel(src0_row, ib, iqs, v);
 
     dst_row[iybs + iqs + 0]        = ggml_cuda_cast<dst_t>(v.x);

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

@@ -49,6 +49,7 @@
 #include "ggml-cuda/wkv.cuh"
 #include "ggml-cuda/gla.cuh"
 #include "ggml-cuda/set-rows.cuh"
+#include "ggml-cuda/pad_reflect_1d.cuh"
 #include "ggml.h"
 
 #include <algorithm>
@@ -203,6 +204,8 @@ static ggml_cuda_device_info ggml_cuda_init() {
     GGML_LOG_INFO("%s: GGML_CUDA_FORCE_CUBLAS: no\n", __func__);
 #endif // GGML_CUDA_FORCE_CUBLAS
     GGML_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
+
+    std::vector<std::pair<int, std::string>> turing_devices_without_mma;
     for (int id = 0; id < info.device_count; ++id) {
         int device_vmm = 0;
 
@@ -260,7 +263,25 @@ static ggml_cuda_device_info ggml_cuda_init() {
         info.devices[id].cc = 100*prop.major + 10*prop.minor;
         GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n",
                         id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
-#endif // defined(GGML_USE_HIP)
+        std::string device_name(prop.name);
+        if (device_name == "NVIDIA GeForce MX450") {
+            turing_devices_without_mma.push_back({ id, device_name });
+        } else if (device_name == "NVIDIA GeForce MX550") {
+            turing_devices_without_mma.push_back({ id, device_name });
+        } else if (device_name.substr(0, 21) == "NVIDIA GeForce GTX 16") {
+            turing_devices_without_mma.push_back({ id, device_name });
+        }
+#endif  // defined(GGML_USE_HIP)
+    }
+
+    if (ggml_cuda_highest_compiled_arch(GGML_CUDA_CC_TURING) >= GGML_CUDA_CC_TURING && !turing_devices_without_mma.empty()) {
+        GGML_LOG_INFO("The following devices will have suboptimal performance due to a lack of tensor cores:\n");
+        for (size_t device_pos = 0; device_pos < turing_devices_without_mma.size(); device_pos++) {
+            GGML_LOG_INFO(
+                "  Device %d: %s\n", turing_devices_without_mma[device_pos].first, turing_devices_without_mma[device_pos].second.c_str());
+        }
+        GGML_LOG_INFO(
+            "Consider compiling with CMAKE_CUDA_ARCHITECTURES=61-virtual;80-virtual and DGGML_CUDA_FORCE_MMQ to force the use of the Pascal code for Turing.\n");
     }
 
     for (int id = 0; id < info.device_count; ++id) {
@@ -1328,9 +1349,7 @@ static void ggml_cuda_op_mul_mat_cublas(
                     &beta,  dst_dd_i,    ldc));
     }
 
-    GGML_UNUSED(dst);
-    GGML_UNUSED(src1_ddq_i);
-    GGML_UNUSED(src1_padded_row_size);
+    GGML_UNUSED_VARS(dst, src1_ddq_i, src1_padded_row_size);
 }
 
 static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
@@ -2354,6 +2373,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_PAD:
             ggml_cuda_op_pad(ctx, dst);
             break;
+        case GGML_OP_PAD_REFLECT_1D:
+            ggml_cuda_op_pad_reflect_1d(ctx, dst);
+            break;
         case GGML_OP_ARANGE:
             ggml_cuda_op_arange(ctx, dst);
             break;
@@ -3084,7 +3106,7 @@ bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size) {
         return false;
     }
 
-#if CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA)
+#if CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA) || defined(GGML_USE_HIP)
     cudaError_t err = cudaHostRegister(buffer, size, cudaHostRegisterPortable | cudaHostRegisterReadOnly);
     if (err != cudaSuccess) {
         // clear the error
@@ -3483,15 +3505,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CONV_TRANSPOSE_2D:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM:
-        case GGML_OP_SUM_ROWS:
-        case GGML_OP_MEAN:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
             return true;
+        case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
         case GGML_OP_GROUP_NORM:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
+        case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_LEAKY_RELU:
@@ -3499,44 +3522,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_GATED_LINEAR_ATTN:
         case GGML_OP_RWKV_WKV7:
             return true;
-        case GGML_OP_FLASH_ATTN_EXT: {
-#ifndef FLASH_ATTN_AVAILABLE
-            return false;
-#endif // FLASH_ATTN_AVAILABLE
-            if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
-                const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
-                if (!turing_mma_available(cc)) {
-                    return false;
-                }
-                const int gqa_ratio = op->src[0]->ne[2] / op->src[1]->ne[2];
-                return op->src[1]->ne[0] == 576 && op->src[2]->ne[0] == 512 && op->src[3] && gqa_ratio % 16 == 0;
-            }
-            // TODO: more general-purpose attention sink support [TAG_ATTN_SINKS]
-            if (op->src[4] && !fp16_mma_available(ggml_cuda_info().devices[dev_ctx->device].cc)
-                    && op->src[0]->ne[0] != 64 && op->src[0]->ne[0] != 128) {
-                return false;
-            }
-            if (op->src[0]->ne[0] == 192) {
-                return false;
-            }
-            if (op->src[1]->type == GGML_TYPE_BF16 || op->src[2]->type == GGML_TYPE_BF16) {
-                return false;
-            }
-            if (op->src[0]->ne[0] ==  64 && op->src[1]->type == GGML_TYPE_F16) {
-                return true;
-            }
-            if (op->src[0]->ne[0] == 128) {
-                return true;
-            }
-            if (op->src[0]->ne[0] == 256 && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16) {
-                return true;
-            }
-            if (op->src[3] && op->src[3]->ne[2] != 1) {
-                return false;
-            }
-            return fp16_mma_available(ggml_cuda_info().devices[dev_ctx->device].cc) &&
-                op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
-        }
+        case GGML_OP_FLASH_ATTN_EXT:
+            return ggml_cuda_flash_attn_ext_supported(dev_ctx->device, op);
         case GGML_OP_CROSS_ENTROPY_LOSS:
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
         case GGML_OP_OPT_STEP_ADAMW:
@@ -3672,10 +3659,6 @@ static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t
         features.push_back({ "NO_PEER_COPY", "1" });
     #endif
 
-    #ifdef GGML_CUDA_F16
-        features.push_back({ "F16", "1" });
-    #endif
-
     #ifdef GGML_CUDA_USE_GRAPHS
         features.push_back({ "USE_GRAPHS", "1" });
     #endif

ggml/src/ggml-cuda/mma.cuh

@@ -291,9 +291,7 @@ namespace ggml_cuda_mma {
             : "=r"(xi[0]), "=r"(xi[2]), "=r"(xi[1]), "=r"(xi[3])
             : "l"(xs));
 #else
-        GGML_UNUSED(t);
-        GGML_UNUSED(xs0);
-        GGML_UNUSED(stride);
+        GGML_UNUSED_VARS(t, xs0, stride);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
@@ -315,9 +313,7 @@ namespace ggml_cuda_mma {
             : "r"(A.x[1]), "r"(B.x[0]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
@@ -345,9 +341,7 @@ namespace ggml_cuda_mma {
             : "r"(A.x[3]), "r"(B.x[1]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
@@ -372,9 +366,7 @@ namespace ggml_cuda_mma {
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[1]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
@@ -408,9 +400,7 @@ namespace ggml_cuda_mma {
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
@@ -425,9 +415,7 @@ namespace ggml_cuda_mma {
             : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3])
             : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[1]));
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // AMPERE_MMA_AVAILABLE
     }
@@ -452,9 +440,7 @@ namespace ggml_cuda_mma {
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[1]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
@@ -469,9 +455,7 @@ namespace ggml_cuda_mma {
             : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3])
             : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[1]));
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // AMPERE_MMA_AVAILABLE
     }
@@ -505,9 +489,7 @@ namespace ggml_cuda_mma {
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
@@ -533,9 +515,7 @@ namespace ggml_cuda_mma {
                                                       0, 0, 0);
 #endif // defined(CDNA3)
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // AMD_MFMA_AVAILABLE
     }
@@ -561,9 +541,7 @@ namespace ggml_cuda_mma {
                                                      0, 0, 0);
 #endif // defined(CDNA3)
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // AMD_MFMA_AVAILABLE
     }

ggml/src/ggml-cuda/mmf.cu

@@ -132,11 +132,11 @@ static __global__ void mul_mat_f(
         dst[j*stride_col_dst + row0 + threadIdx.x] = sum;
     }
 #else
+    GGML_UNUSED_VARS(x, y, ids, dst,
+        ncols, nchannels_y, stride_row, stride_col_y, stride_col_dst,
+        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
     NO_DEVICE_CODE;
-    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(ids); GGML_UNUSED(dst);
-    GGML_UNUSED(ncols); GGML_UNUSED(nchannels_y); GGML_UNUSED(stride_row); GGML_UNUSED(stride_col_y); GGML_UNUSED(stride_col_dst);
-    GGML_UNUSED(channel_ratio); GGML_UNUSED(stride_channel_x); GGML_UNUSED(stride_channel_y); GGML_UNUSED(stride_channel_dst);
-    GGML_UNUSED(sample_ratio); GGML_UNUSED(stride_sample_x); GGML_UNUSED(stride_sample_y); GGML_UNUSED(stride_sample_dst);
 #endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 }
 
@@ -151,7 +151,6 @@ static void mul_mat_f_cuda(
         cudaStream_t stream) {
     typedef tile<16, 8, T>     tile_A;
     typedef tile< 8, 8, T>     tile_B;
-    typedef tile<16, 8, float> tile_C;
 
     GGML_ASSERT(!ids && "mul_mat_id not implemented");
 
@@ -352,9 +351,6 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
     GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
     GGML_ASSERT(        nb0        == ts_dst);
 
-    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
-
     const float   * src1_d =       (const float   *) src1->data;
     const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
     float         *  dst_d =       (float         *)  dst->data;

ggml/src/ggml-cuda/mmq.cu

@@ -3,6 +3,140 @@
 
 #include <vector>
 
+// To reduce shared memory use, store "it" and "iex_used" with 22/10 bits each.
+struct mmq_ids_helper_store {
+    uint32_t data;
+
+    __device__ mmq_ids_helper_store(const uint32_t it, const uint32_t iex_used) {
+        data = (it & 0x003FFFFF) | (iex_used << 22);
+    }
+
+    __device__ uint32_t it() const {
+        return data & 0x003FFFFF;
+    }
+
+    __device__ uint32_t iex_used() const {
+        return data >> 22;
+    }
+};
+static_assert(sizeof(mmq_ids_helper_store) == 4, "unexpected size for mmq_ids_helper_store");
+
+// Helper function for mul_mat_id, converts ids to a more convenient format.
+// ids_src1 describes how to permute the flattened column indices of src1 in order to get a compact src1 tensor sorted by expert.
+// ids_dst describes the same mapping but for the dst tensor.
+// The upper and lower bounds for the ith expert in the compact src1 tensor are stored in expert_bounds[i:i+1].
+template <int n_expert_used_template>
+__launch_bounds__(ggml_cuda_get_physical_warp_size(), 1)
+static __global__ void mmq_ids_helper(
+        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
+        const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1) {
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    const int n_expert_used = n_expert_used_template == 0 ? n_expert_used_var : n_expert_used_template;
+    const int expert = blockIdx.x;
+
+    extern __shared__ char data_mmq_ids_helper[];
+    mmq_ids_helper_store * store = (mmq_ids_helper_store *) data_mmq_ids_helper;
+
+    int nex_prev   = 0; // Number of columns for experts with a lower index.
+    int it_compact = 0; // Running index for the compact slice of this expert.
+
+    if constexpr (n_expert_used_template == 0) {
+        // Generic implementation:
+        for (int it = 0; it < n_tokens; ++it) {
+            int iex_used = -1; // The index at which the expert is used, if any.
+            for (int iex = threadIdx.x; iex < n_expert_used; iex += warp_size) {
+                const int expert_used = ids[it*si1 + iex];
+                nex_prev += expert_used < expert;
+                if (expert_used == expert) {
+                    iex_used = iex;
+                }
+            }
+
+            if (iex_used != -1) {
+                store[it_compact] = mmq_ids_helper_store(it, iex_used);
+            }
+
+            if (warp_reduce_any<warp_size>(iex_used != -1)) {
+                it_compact++;
+            }
+        }
+    } else {
+        // Implementation optimized for specific numbers of experts used:
+        static_assert(n_expert_used == 6 || warp_size % n_expert_used == 0, "bad n_expert_used");
+        const int neu_padded = n_expert_used == 6 ? 8 : n_expert_used; // Padded to next higher power of 2.
+        for (int it0 = 0; it0 < n_tokens; it0 += warp_size/neu_padded) {
+            const int it = it0 + threadIdx.x / neu_padded;
+
+            const int iex = threadIdx.x % neu_padded; // The index at which the expert is used, if any.
+            const int expert_used = (neu_padded == n_expert_used || iex < n_expert_used) && it < n_tokens ?
+                ids[it*si1 + iex] : INT_MAX;
+            const int iex_used = expert_used == expert ? iex : -1;
+            nex_prev += expert_used < expert;
+
+            // Whether the threads at this token position have used the expert:
+            const int it_compact_add_self = warp_reduce_any<neu_padded>(iex_used != -1);
+
+            // Do a scan over threads at lower token positions in warp to get the correct index for writing data:
+            int it_compact_add_lower = 0;
+#pragma unroll
+            for (int offset = neu_padded; offset < warp_size; offset += neu_padded) {
+                const int tmp = __shfl_up_sync(0xFFFFFFFF, it_compact_add_self, offset, warp_size);
+                if (threadIdx.x >= offset) {
+                    it_compact_add_lower += tmp;
+                }
+            }
+
+            if (iex_used != -1) {
+                store[it_compact + it_compact_add_lower] = mmq_ids_helper_store(it, iex_used);
+            }
+
+            // The thread with the highest index in the warp always has the sum over the whole warp, use it to increment all threads:
+            it_compact += __shfl_sync(0xFFFFFFFF, it_compact_add_lower + it_compact_add_self, warp_size - 1, warp_size);
+        }
+    }
+    nex_prev = warp_reduce_sum<warp_size>(nex_prev);
+
+    for (int itc = threadIdx.x; itc < it_compact; itc += warp_size) {
+        const mmq_ids_helper_store store_it = store[itc];
+        const int it       = store_it.it();
+        const int iex_used = store_it.iex_used();
+        ids_src1[nex_prev + itc] = it*sis1          + iex_used % nchannels_y;
+        ids_dst [nex_prev + itc] = it*n_expert_used + iex_used;
+    }
+
+    if (threadIdx.x != 0) {
+        return;
+    }
+
+    expert_bounds[expert] = nex_prev;
+
+    if (expert < gridDim.x - 1) {
+        return;
+    }
+
+    expert_bounds[gridDim.x] = nex_prev + it_compact;
+}
+
+template <int n_expert_used_template>
+static void launch_mmq_ids_helper(
+        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
+        const int n_experts, const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1, cudaStream_t stream) {
+    GGML_ASSERT(n_tokens          < (1 << 22) && "too few bits in mmq_ids_helper_store");
+    GGML_ASSERT(n_expert_used_var < (1 << 10) && "too few bits in mmq_ids_helper_store");
+
+    const int id = ggml_cuda_get_device();
+    const int warp_size = ggml_cuda_info().devices[id].warp_size;
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
+    CUDA_SET_SHARED_MEMORY_LIMIT(mmq_ids_helper<n_expert_used_template>, smpbo);
+
+    const dim3 num_blocks(n_experts, 1, 1);
+    const dim3 block_size(warp_size, 1, 1);
+    const size_t nbytes_shared = n_tokens*sizeof(mmq_ids_helper_store);
+    GGML_ASSERT(nbytes_shared <= smpbo);
+    mmq_ids_helper<n_expert_used_template><<<num_blocks, block_size, nbytes_shared, stream>>>
+        (ids, ids_src1, ids_dst, expert_bounds, n_tokens, n_expert_used_var, nchannels_y, si1, sis1);
+}
+
 static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
     switch (args.type_x) {
         case GGML_TYPE_Q4_0:
@@ -137,7 +271,7 @@ void ggml_cuda_mul_mat_q(
             ne00, ne01, ne1, s01, ne11, s1,
             ne02, ne12, s02, s12, s2,
             ne03, ne13, s03, s13, s3,
-            use_stream_k};
+            use_stream_k, ne1};
         ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
         return;
     }
@@ -148,54 +282,50 @@ void ggml_cuda_mul_mat_q(
 
     const int64_t n_expert_used = ids->ne[0];
     const int64_t ne_get_rows = ne12 * n_expert_used;
+    GGML_ASSERT(ne1 == n_expert_used);
 
-    std::vector<char> ids_host(ggml_nbytes(ids));
-    std::vector<int32_t> ids_src1_host;
-    ids_src1_host.reserve(ne_get_rows);
-    std::vector<int32_t> ids_dst_host;
-    ids_dst_host.reserve(ne_get_rows);
-    std::vector<int32_t> tokens_per_expert_host(ne02);
-    std::vector<int32_t> expert_bounds_host(ne02 + 1);
-    ggml_cuda_pool_alloc<int32_t> ids_buf_dev(ctx.pool());
+    ggml_cuda_pool_alloc<int32_t> ids_src1(ctx.pool(), ne_get_rows);
+    ggml_cuda_pool_alloc<int32_t> ids_dst(ctx.pool(), ne_get_rows);
+    ggml_cuda_pool_alloc<int32_t> expert_bounds(ctx.pool(), ne02 + 1);
 
-    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids->data, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
-    CUDA_CHECK(cudaStreamSynchronize(stream));
+    {
+        GGML_ASSERT(ids->nb[0] == ggml_element_size(ids));
+        const int si1  = ids->nb[1] / ggml_element_size(ids);
+        const int sis1 = nb12 / nb11;
 
-    for (int64_t i02 = 0; i02 < ne02; ++i02) { // expert matrices
-        for (int64_t i12 = 0; i12 < ne12; ++i12) { // tokens
-            for (int64_t iex = 0; iex < n_expert_used; ++iex) {
-                const int32_t expert_to_use = *(const int32_t *)(ids_host.data() + i12*ids->nb[1] + iex*ids->nb[0]);
-                assert(expert_to_use >= 0 && expert_to_use < ne02);
-                if (expert_to_use == i02) {
-                    ids_src1_host.push_back(i12*(nb12/nb11) + iex % ne11);
-                    ids_dst_host.push_back(i12*ne1 + iex);
-                    tokens_per_expert_host[i02]++;
-                    break;
-                }
-            }
+        switch (n_expert_used) {
+            case  2:
+                launch_mmq_ids_helper< 2> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
+                break;
+            case  4:
+                launch_mmq_ids_helper< 4> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
+                break;
+            case  6:
+                launch_mmq_ids_helper< 6> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
+                break;
+            case  8:
+                launch_mmq_ids_helper< 8> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
+                break;
+            case 16:
+                launch_mmq_ids_helper<16> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
+                break;
+            case 32:
+                launch_mmq_ids_helper<32> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
+                break;
+            default:
+                launch_mmq_ids_helper< 0> ((const int32_t *) ids->data, ids_src1.get(), ids_dst.get(), expert_bounds.get(),
+                    ne02, ne12, n_expert_used, ne11, si1, sis1, stream);
+                break;
         }
+        CUDA_CHECK(cudaGetLastError());
     }
 
-    int32_t cumsum = 0;
-    for (int64_t i = 0; i < ne02; ++i) {
-        expert_bounds_host[i] = cumsum;
-        cumsum += tokens_per_expert_host[i];
-    }
-    expert_bounds_host[ne02] = cumsum;
-
-    std::vector<int32_t> ids_buf_host;
-    ids_buf_host.reserve(ids_src1_host.size() + ids_dst_host.size() + expert_bounds_host.size());
-    ids_buf_host.insert(ids_buf_host.end(), ids_src1_host.begin(), ids_src1_host.end());
-    ids_buf_host.insert(ids_buf_host.end(), ids_dst_host.begin(), ids_dst_host.end());
-    ids_buf_host.insert(ids_buf_host.end(), expert_bounds_host.begin(), expert_bounds_host.end());
-    ids_buf_dev.alloc(ids_buf_host.size() + get_mmq_x_max_host(cc)); // Expert bounds are padded on device.
-    CUDA_CHECK(cudaMemcpyAsync(ids_buf_dev.ptr, ids_buf_host.data(), ids_buf_host.size()*sizeof(int32_t), cudaMemcpyHostToDevice, stream));
-    CUDA_CHECK(cudaStreamSynchronize(stream));
-
-    const int32_t * ids_src1_dev      = ids_buf_dev.ptr;
-    const int32_t * ids_dst_dev       = ids_src1_dev + ids_src1_host.size();
-    const int32_t * expert_bounds_dev = ids_dst_dev + ids_dst_host.size();
-
     const size_t nbytes_src1_q8_1 = ne12*n_expert_used*ne10_padded * sizeof(block_q8_1)/QK8_1 +
         get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq);
     ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), nbytes_src1_q8_1);
@@ -208,7 +338,7 @@ void ggml_cuda_mul_mat_q(
         const int64_t s11 = src1->nb[1] / ts_src1;
         const int64_t s12 = src1->nb[2] / ts_src1;
         const int64_t s13 = src1->nb[2] / ts_src1;
-        quantize_mmq_q8_1_cuda(src1_d, ids_src1_dev, src1_q8_1.get(), src0->type,
+        quantize_mmq_q8_1_cuda(src1_d, ids_src1.get(), src1_q8_1.get(), src0->type,
             ne10, s11, s12, s13, ne10_padded, ne11_flat, ne12_flat, ne13_flat, stream);
         CUDA_CHECK(cudaGetLastError());
     }
@@ -218,11 +348,11 @@ void ggml_cuda_mul_mat_q(
 
     // Note that ne02 is used instead of ne12 because the number of y channels determines the z dimension of the CUDA grid.
     const mmq_args args = {
-        src0_d, src0->type, (const int *) src1_q8_1.ptr, ids_dst_dev, expert_bounds_dev, dst_d,
+        src0_d, src0->type, (const int *) src1_q8_1.get(), ids_dst.get(), expert_bounds.get(), dst_d,
         ne00, ne01, ne_get_rows, s01, ne_get_rows, s1,
         ne02, ne02, s02, s12, s2,
         ne03, ne13, s03, s13, s3,
-        use_stream_k};
+        use_stream_k, ne12};
 
     ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
 }
@@ -262,14 +392,11 @@ void ggml_cuda_op_mul_mat_q(
         ne00, row_diff, src1_ncols, stride01, ne11, nrows_dst,
         1, 1, 0, 0, 0,
         1, 1, 0, 0, 0,
-        use_stream_k};
+        use_stream_k, src1_ncols};
 
     ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
 
-    GGML_UNUSED(src1);
-    GGML_UNUSED(dst);
-    GGML_UNUSED(src1_ddf_i);
-    GGML_UNUSED(src1_padded_row_size);
+    GGML_UNUSED_VARS(src1, dst, src1_ddf_i, src1_padded_row_size);
 }
 
 bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {

ggml/src/ggml-cuda/mmq.cuh

@@ -1255,7 +1255,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
         }
     }
 #else
-    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); GGML_UNUSED(k00);
+    GGML_UNUSED_VARS(x, y, sum, k00);
     NO_DEVICE_CODE;
 #endif // AMD_MFMA_AVAILABLE
 }
@@ -1572,7 +1572,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
         }
     }
 #else
-    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); GGML_UNUSED(k00);
+    GGML_UNUSED_VARS(x, y, sum, k00);
     NO_DEVICE_CODE;
 #endif // AMD_MFMA_AVAILABLE
 }
@@ -2301,7 +2301,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
         }
     }
 #else
-    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); GGML_UNUSED(k00);
+    GGML_UNUSED_VARS(x, y, sum, k00);
     NO_DEVICE_CODE;
 #endif // AMD_MFMA_AVAILABLE
 }
@@ -2855,12 +2855,14 @@ static __device__ __forceinline__ void mmq_write_back_mma(
 #else
     typedef tile<16, 8, int> tile_C;
     constexpr int rows_per_warp = 2 * granularity;
-#endif
+#endif // defined(AMD_MFMA_AVAILABLE)
     constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
 
     const int i0 = (threadIdx.y / ntx) * (ntx*tile_C::I);
 #if defined(TURING_MMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
     static_assert(nwarps*tile_C::I == mmq_y, "nwarps*tile_C::I != mmq_y");
+#else
+    GGML_UNUSED(nwarps);
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
 #pragma unroll
@@ -3136,7 +3138,8 @@ static __global__ void mul_mat_q(
         const int32_t * __restrict__ expert_bounds, float * __restrict__ dst, float * __restrict__ tmp_fixup,
         const int ncols_x, const int nrows_x, const int ncols_dst, const int stride_row_x, const int ncols_y, const int stride_col_dst,
         const int channel_ratio, const int nchannels_y, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
-        const int sample_ratio, const int nsamples_y, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
+        const int sample_ratio, const int nsamples_y, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        const int ncols_max) {
 
     // Skip unused template specializations for faster compilation:
     if (mmq_x > get_mmq_x_max_device() || mmq_x % mmq_get_granularity_device(mmq_x) != 0) {
@@ -3150,7 +3153,7 @@ static __global__ void mul_mat_q(
     constexpr int qk    = ggml_cuda_type_traits<type>::qk;
     constexpr int mmq_y = get_mmq_y_device();
 
-    const int ntx = (ncols_dst + mmq_x - 1) / mmq_x; // Number of tiles x
+    const int ntx = (ncols_max + mmq_x - 1) / mmq_x; // Number of tiles x
     const int nty = (nrows_x   + mmq_y - 1) / mmq_y; // Number of tiles y
 
     // Initialize the ids for writing back data with just the index.
@@ -3374,7 +3377,8 @@ template <ggml_type type, int mmq_x, bool need_check>
 static __global__ void mul_mat_q_stream_k_fixup(
         const int32_t * ids_dst, const int32_t * expert_bounds, float * __restrict__ dst, const float * __restrict__ tmp_last_tile,
         const int ncols_x, const int nrows_x, const int ncols_dst, const int stride_col_dst,
-        const int nchannels_y, const int stride_channel_dst, const int nsamples_y, const int stride_sample_dst) {
+        const int nchannels_y, const int stride_channel_dst, const int nsamples_y, const int stride_sample_dst,
+        const int ncols_max) {
     constexpr int     mmq_y           = get_mmq_y_device();
     constexpr int     qk              = ggml_cuda_type_traits<type>::qk;
     constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
@@ -3385,7 +3389,7 @@ static __global__ void mul_mat_q_stream_k_fixup(
 
     float sum[mmq_x*mmq_y / (nwarps*warp_size)] = {0.0f};
 
-    const int ntx  = (ncols_dst + mmq_x - 1) / mmq_x;
+    const int ntx  = (ncols_max + mmq_x - 1) / mmq_x;
     const int nty  = (nrows_x   + mmq_y - 1) / mmq_y;
 
     const int bidx0 = blockIdx.x;
@@ -3526,7 +3530,7 @@ struct mmq_args {
     int64_t ncols_x; int64_t nrows_x; int64_t ncols_dst; int64_t stride_row_x; int64_t ncols_y; int64_t nrows_dst;
     int64_t nchannels_x; int64_t nchannels_y; int64_t stride_channel_x; int64_t stride_channel_y; int64_t stride_channel_dst;
     int64_t nsamples_x; int64_t nsamples_y; int64_t stride_sample_x; int64_t stride_sample_y; int64_t stride_sample_dst;
-    bool use_stream_k;
+    bool use_stream_k; int64_t ncols_max;
 };
 
 template<ggml_type type>
@@ -3556,7 +3560,7 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
     CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, mmq_x,  true>), nbytes_shared);
 
     const int nty  = (args.nrows_x   + mmq_y - 1) / mmq_y;
-    const int ntx  = (args.ncols_dst + mmq_x - 1) / mmq_x;
+    const int ntx  = (args.ncols_max + mmq_x - 1) / mmq_x;
     const int ntzw = args.nchannels_y * args.nsamples_y;
     const dim3 block_nums_xy_tiling(nty, ntx, ntzw);
 
@@ -3572,14 +3576,16 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
                 (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr,
                  args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
                  channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
-                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
+                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
+                 args.ncols_max);
         } else {
             constexpr bool need_check = true;
             mul_mat_q<type, mmq_x, need_check><<<block_nums_xy_tiling, block_dims, nbytes_shared, stream>>>
                 (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr,
                  args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
                  channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
-                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
+                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
+                 args.ncols_max);
         }
         return;
     }
@@ -3599,7 +3605,8 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
             (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr,
              args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
              channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
-             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
+             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
+             args.ncols_max);
 
         if (!fixup_needed) {
             return;
@@ -3607,14 +3614,16 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
         mul_mat_q_stream_k_fixup<type, mmq_x, need_check><<<block_nums_stream_k, block_dims, 0, stream>>>
             (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_dst,
-             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst);
+             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst,
+             args.ncols_max);
     } else {
         constexpr bool need_check = true;
         mul_mat_q<type, mmq_x, need_check><<<block_nums_stream_k, block_dims, nbytes_shared, stream>>>
             (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr,
              args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
              channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
-             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
+             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
+             args.ncols_max);
 
         if (!fixup_needed) {
             return;
@@ -3622,7 +3631,8 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
         mul_mat_q_stream_k_fixup<type, mmq_x, need_check><<<block_nums_stream_k, block_dims, 0, stream>>>
             (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_dst,
-             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst);
+             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst,
+             args.ncols_max);
     }
 }
 
@@ -3647,7 +3657,7 @@ void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cuda
             continue;
         }
 
-        const int ntiles_x = (args.ncols_y + mmq_x - 1) / mmq_x;
+        const int ntiles_x = (args.ncols_max + mmq_x - 1) / mmq_x;
 
         if (ntiles_x < ntiles_x_best) {
             mmq_x_best = mmq_x;

ggml/src/ggml-cuda/mmvf.cu

@@ -433,12 +433,7 @@ void ggml_cuda_op_mul_mat_vec_f(
             GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
     }
 
-    GGML_UNUSED(ctx);
-    GGML_UNUSED(src1);
-    GGML_UNUSED(dst);
-    GGML_UNUSED(src1_ddq_i);
-    GGML_UNUSED(src1_ncols);
-    GGML_UNUSED(src1_padded_row_size);
+    GGML_UNUSED_VARS(ctx, src1, dst, src1_ddq_i, src1_ncols, src1_padded_row_size);
 }
 
 bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11) {

ggml/src/ggml-cuda/mmvq.cu

@@ -596,9 +596,5 @@ void ggml_cuda_op_mul_mat_vec_q(
         src0_dd_i, src0->type, src1_ddq_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
         1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, stream);
 
-    GGML_UNUSED(src1);
-    GGML_UNUSED(dst);
-    GGML_UNUSED(src1_ddf_i);
-    GGML_UNUSED(src1_ncols);
-    GGML_UNUSED(src1_padded_row_size);
+    GGML_UNUSED_VARS(src1, dst, src1_ddf_i, src1_ncols, src1_padded_row_size);
 }