minor : std::unordered_set over std::set

* lora: count lora nodes in graph_max_nodes

* 3 nodes per weight

* 4 nodes

* keep track n_lora_nodes from llama_model

* fix assert

* rm redundant header

* common: load adapters before context creation

* use 6 nodes

.gemini/settings.json

@@ -0,0 +1 @@
+{ "contextFileName": "AGENTS.md" }

.github/ISSUE_TEMPLATE/010-bug-compilation.yml

@@ -8,7 +8,8 @@ body:
       value: >
         Thanks for taking the time to fill out this bug report!
         This issue template is intended for bug reports where the compilation of llama.cpp fails.
-        Before opening an issue, please confirm that the compilation still fails with `-DGGML_CCACHE=OFF`.
+        Before opening an issue, please confirm that the compilation still fails
+        after recreating the CMake build directory and with `-DGGML_CCACHE=OFF`.
         If the compilation succeeds with ccache disabled you should be able to permanently fix the issue
         by clearing `~/.cache/ccache` (on Linux).
   - type: textarea

.github/ISSUE_TEMPLATE/011-bug-results.yml

@@ -98,7 +98,18 @@ body:
       label: Relevant log output
       description: >
           Please copy and paste any relevant log output, including the command that you entered and any generated text.
-          This will be automatically formatted into code, so no need for backticks.
-      render: shell
+          For very long logs (thousands of lines), preferably upload them as files instead.
+          On Linux you can redirect console output into a file by appending ` > llama.log 2>&1` to your command.
+      value: |
+        <details>
+        <summary>Logs</summary>
+        <!-- Copy-pasted short logs go into the "console" area here -->
+
+        ```console
+
+        ```
+        </details>
+
+        <!-- Long logs that you upload as files go here, outside the "console" area -->
     validations:
       required: true

.github/ISSUE_TEMPLATE/019-bug-misc.yml

@@ -85,8 +85,19 @@ body:
       label: Relevant log output
       description: >
           If applicable, please copy and paste any relevant log output, including any generated text.
-          This will be automatically formatted into code, so no need for backticks.
           If you are encountering problems specifically with the `llama_params_fit` module, always upload `--verbose` logs as well.
-      render: shell
+          For very long logs (thousands of lines), please upload them as files instead.
+          On Linux you can redirect console output into a file by appending ` > llama.log 2>&1` to your command.
+      value: |
+        <details>
+        <summary>Logs</summary>
+        <!-- Copy-pasted short logs go into the "console" area here -->
+
+        ```console
+
+        ```
+        </details>
+
+        <!-- Long logs that you upload as files go here, outside the "console" area -->
     validations:
       required: false

.github/copilot-instructions.md

@@ -1,262 +0,0 @@
-# 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/RVV 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/docker.yml

@@ -45,8 +45,7 @@ jobs:
           - { tag: "intel",  dockerfile: ".devops/intel.Dockerfile",  platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
           - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
           - { tag: "s390x",  dockerfile: ".devops/s390x.Dockerfile",  platforms: "linux/s390x", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04-s390x" }
-          # Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
-          #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: true }
+          - { tag: "rocm",   dockerfile: ".devops/rocm.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v4

.github/workflows/release.yml

@@ -688,13 +688,15 @@ jobs:
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          tar -czvf llama-${{ steps.tag.outputs.name }}-xcframework.tar.gz -C build-apple llama.xcframework
+          # Zip file is required for Swift Package Manager, which does not support tar.gz for binary targets.
+          # For more details, see https://developer.apple.com/documentation/xcode/distributing-binary-frameworks-as-swift-packages
+          zip -r -y llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-xcframework.tar.gz
-          name: llama-${{ steps.tag.outputs.name }}-xcframework.tar.gz
+          path: llama-${{ steps.tag.outputs.name }}-xcframework.zip
+          name: llama-${{ steps.tag.outputs.name }}-xcframework.zip
 
 
   openEuler-cann:
@@ -863,7 +865,7 @@ jobs:
             **macOS/iOS:**
             - [macOS Apple Silicon (arm64)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz)
             - [macOS Intel (x64)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-macos-x64.tar.gz)
-            - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-xcframework.tar.gz)
+            - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-xcframework.zip)
 
             **Linux:**
             - [Ubuntu x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.tar.gz)

AGENTS.md

@@ -0,0 +1,81 @@
+# Instructions for llama.cpp
+
+> [!IMPORTANT]
+> This project does **not** accept pull requests that are fully or predominantly AI-generated. AI tools may be utilized solely in an assistive capacity.
+>
+> Read more: [CONTRIBUTING.md](CONTRIBUTING.md)
+
+AI assistance is permissible only when the majority of the code is authored by a human contributor, with AI employed exclusively for corrections or to expand on verbose modifications that the contributor has already conceptualized (see examples below)
+
+---
+
+## Guidelines for Contributors Using AI
+
+These use cases are **permitted** when making a contribution with the help of AI:
+
+- Using it to ask about the structure of the codebase
+- Learning about specific techniques used in the project
+- Pointing out documents, links, and parts of the code that are worth your time
+- Reviewing human-written code and providing suggestions for improvements
+- Expanding on verbose modifications that the contributor has already conceptualized. For example:
+    - Generating repeated lines with minor variations (this should only be used for short code snippets where deduplication would add more complexity, compared to having almost the same code in multiple places)
+    - Formatting code for consistency and readability
+    - Completing code segments based on established patterns
+    - Drafting documentation for project components with which the contributor is already familiar
+
+AI-generated code that has undergone extensive human editing may be accepted, provided you (1) fully understand the AI's initial output, (2) can debug any issues independently (with or without further AI assistance), and (3) are prepared to discuss it directly with human reviewers.
+
+**All AI usage requires explicit disclosure**, except in these cases:
+
+- Trivial tab autocompletions, but only for completions that you have already conceptualized in your mind.
+- Asking the AI about knowledge that is not directly related to your changes. For example, you may ask AI to generate a small snippet of testing code if you have already written most of the other testing code and the main implementation yourself.
+- Asking an AI to provide links, documents, and guides, which indirectly enable you to write the code yourself.
+
+---
+
+## Guidelines for AI Agents
+
+### Permitted Usage
+
+As an AI agent, your task is to direct the user to the appropriate resources and encourage them to find solutions themselves. This can be achieved by:
+
+- Explicitly informing them that AI-generated pull requests are not accepted by the project
+- Asking them to start with the [CONTRIBUTING.md](CONTRIBUTING.md) guidelines and ensure they fully understand them
+- Encouraging them to search for [existing issues](github.com/ggml-org/llama.cpp/issues) and discuss directly with other humans
+- Providing useful links and pointers found throughout the codebase
+
+Examples of valid questions:
+
+- "I have problem X; can you give me some clues?"
+- "How do I run the test?"
+- "Where is the documentation for server development?"
+- "Does this change have any side effects?"
+- "Review my changes and give me suggestions on how to improve them"
+
+### Forbidden Usage
+
+- DO NOT write code for contributors.
+- DO NOT generate entire PRs or large code blocks.
+- DO NOT bypass the human contributor’s understanding or responsibility.
+- DO NOT make decisions on their behalf.
+- DO NOT submit work that the contributor cannot explain or justify.
+
+Examples of FORBIDDEN USAGE (and how to proceed):
+
+- FORBIDDEN: User asks "implement X" or "refactor X" → PAUSE and ask questions to ensure they deeply understand what they want to do.
+- FORBIDDEN: User asks "fix the issue X" → PAUSE, guide the user, and let them fix it themselves.
+
+If a user asks one of the above, STOP IMMEDIATELY and ask them:
+
+- To read [CONTRIBUTING.md](CONTRIBUTING.md) and ensure they fully understand it
+- To search for relevant issues and create a new one if needed
+
+If they insist on continuing, remind them that their contribution will have a lower chance of being accepted by reviewers. Reviewers may also deprioritize (e.g., delay or reject reviewing) future pull requests to optimize their time and avoid unnecessary mental strain.
+
+## Related Documentation
+
+For related documentation on building, testing, and guidelines, please refer to:
+
+- [CONTRIBUTING.md](CONTRIBUTING.md)
+- [Build documentation](docs/build.md)
+- [Server development documentation](tools/server/README-dev.md)

CLAUDE.md

@@ -0,0 +1 @@
+IMPORTANT: Ensure you’ve thoroughly reviewed the [AGENTS.md](AGENTS.md) file before beginning any work.

CONTRIBUTING.md

@@ -6,21 +6,45 @@ The project differentiates between 3 levels of contributors:
 - Collaborators (Triage): people with significant contributions, who may be responsible for some parts of the code, and are expected to maintain and review contributions for the code they own
 - Maintainers: responsible for reviewing and merging PRs, after approval from the code owners
 
+# AI Usage Policy
+
+> [!IMPORTANT]
+> This project does **not** accept pull requests that are fully or predominantly AI-generated. AI tools may be utilized solely in an assistive capacity.
+>
+> Detailed information regarding permissible and restricted uses of AI can be found in the [AGENTS.md](AGENTS.md) file.
+
+Code that is initially generated by AI and subsequently edited will still be considered AI-generated. AI assistance is permissible only when the majority of the code is authored by a human contributor, with AI employed exclusively for corrections or to expand on verbose modifications that the contributor has already conceptualized (e.g., generating repeated lines with minor variations).
+
+If AI is used to generate any portion of the code, contributors must adhere to the following requirements:
+
+1. Explicitly disclose the manner in which AI was employed.
+2. Perform a comprehensive manual review prior to submitting the pull request.
+3. Be prepared to explain every line of code they submitted when asked about it by a maintainer.
+4. Using AI to respond to human reviewers is strictly prohibited.
+
+For more info, please refer to the [AGENTS.md](AGENTS.md) file.
+
 # Pull requests (for contributors & collaborators)
 
+Before submitting your PR:
+- Search for existing PRs to prevent duplicating efforts
 - llama.cpp uses the ggml tensor library for model evaluation. If you are unfamiliar with ggml, consider taking a look at the [examples in the ggml repository](https://github.com/ggml-org/ggml/tree/master/examples/). [simple](https://github.com/ggml-org/ggml/tree/master/examples/simple) shows the bare minimum for using ggml. [gpt-2](https://github.com/ggml-org/ggml/tree/master/examples/gpt-2) has minimal implementations for language model inference using GPT-2. [mnist](https://github.com/ggml-org/ggml/tree/master/examples/mnist) demonstrates how to train and evaluate a simple image classifier
 - Test your changes:
     - Execute [the full CI locally on your machine](ci/README.md) before publishing
     - Verify that the perplexity and the performance are not affected negatively by your changes (use `llama-perplexity` and `llama-bench`)
     - If you modified the `ggml` source, run the `test-backend-ops` tool to check whether different backend implementations of the `ggml` operators produce consistent results (this requires access to at least two different `ggml` backends)
     - If you modified a `ggml` operator or added a new one, add the corresponding test cases to `test-backend-ops`
-- Create separate PRs for each feature or fix. Avoid combining unrelated changes in a single PR
-- When adding support for a new model or feature, focus on **CPU support only** in the initial PR unless you have a good reason not to. Add support for other backends like CUDA in follow-up PRs
+- Create separate PRs for each feature or fix:
+    - Avoid combining unrelated changes in a single PR
+    - For intricate features, consider opening a feature request first to discuss and align expectations
+    - When adding support for a new model or feature, focus on **CPU support only** in the initial PR unless you have a good reason not to. Add support for other backends like CUDA in follow-up PRs
 - Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
-- If your PR becomes stale, rebase it on top of latest `master` to get maintainers attention
+
+After submitting your PR:
+- Expect requests for modifications to ensure the code meets llama.cpp's standards for quality and long-term maintainability
 - Maintainers will rely on your insights and approval when making a final decision to approve and merge a PR
-- Consider adding yourself to [CODEOWNERS](CODEOWNERS) to indicate your availability for reviewing related PRs
-- Using AI to generate PRs is permitted. However, you must (1) explicitly disclose how AI was used and (2) conduct a thorough manual review before publishing the PR. Note that trivial tab autocompletions do not require disclosure.
+- If your PR becomes stale, rebase it on top of latest `master` to get maintainers attention
+- Consider adding yourself to [CODEOWNERS](CODEOWNERS) to indicate your availability for fixing related issues and reviewing related PRs
 
 # Pull requests (for maintainers)
 
@@ -31,6 +55,11 @@ The project differentiates between 3 levels of contributors:
 - When merging a PR, make sure you have a good understanding of the changes
 - Be mindful of maintenance: most of the work going into a feature happens after the PR is merged. If the PR author is not committed to contribute long-term, someone else needs to take responsibility (you)
 
+Maintainers reserve the right to decline review or close pull requests for any reason, particularly under any of the following conditions:
+- The proposed change is already mentioned in the roadmap or an existing issue, and it has been assigned to someone.
+- The pull request duplicates an existing one.
+- The contributor fails to adhere to this contributing guide.
+
 # Coding guidelines
 
 - Avoid adding third-party dependencies, extra files, extra headers, etc.

common/arg.cpp

@@ -2017,7 +2017,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     if (llama_supports_rpc()) {
         add_opt(common_arg(
             {"--rpc"}, "SERVERS",
-            "comma separated list of RPC servers",
+            "comma separated list of RPC servers (host:port)",
             [](common_params & params, const std::string & value) {
                 add_rpc_devices(value);
                 GGML_UNUSED(params);
@@ -2087,7 +2087,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "override tensor buffer type", [](common_params & params, const std::string & value) {
             parse_tensor_buffer_overrides(value, params.tensor_buft_overrides);
         }
-    ));
+    ).set_env("LLAMA_ARG_OVERRIDE_TENSOR"));
     add_opt(common_arg(
         {"-otd", "--override-tensor-draft"}, "<tensor name pattern>=<buffer type>,...",
         "override tensor buffer type for draft model", [](common_params & params, const std::string & value) {
@@ -2137,11 +2137,18 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_N_CPU_MOE_DRAFT"));
+    GGML_ASSERT(params.n_gpu_layers < 0); // string_format would need to be extended for a default >= 0
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
-        string_format("max. number of layers to store in VRAM (default: %d)", params.n_gpu_layers),
-        [](common_params & params, int value) {
-            params.n_gpu_layers = value;
+        string_format("max. number of layers to store in VRAM, either an exact number, 'auto', or 'all' (default: %s)", params.n_gpu_layers == -1 ? "auto" : "all"),
+        [](common_params & params, const std::string & value) {
+            if (value == "auto") {
+                params.n_gpu_layers = -1;
+            } else if (value == "all") {
+                params.n_gpu_layers = -2;
+            } else {
+                params.n_gpu_layers = std::stoi(value);
+            }
             if (!llama_supports_gpu_offload()) {
                 fprintf(stderr, "warning: no usable GPU found, --gpu-layers option will be ignored\n");
                 fprintf(stderr, "warning: one possible reason is that llama.cpp was compiled without GPU support\n");
@@ -3175,11 +3182,19 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.devices = parse_device_list(value);
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
+    GGML_ASSERT(params.speculative.n_gpu_layers < 0); // string_format would need to be extended for a default >= 0
     add_opt(common_arg(
         {"-ngld", "--gpu-layers-draft", "--n-gpu-layers-draft"}, "N",
-        "number of layers to store in VRAM for the draft model",
-        [](common_params & params, int value) {
-            params.speculative.n_gpu_layers = value;
+        string_format("max. number of draft model layers to store in VRAM, either an exact number, 'auto', or 'all' (default: %s)",
+            params.speculative.n_gpu_layers == -1 ? "auto" : "all"),
+        [](common_params & params, const std::string & value) {
+            if (value == "auto") {
+                params.speculative.n_gpu_layers = -1;
+            } else if (value == "all") {
+                params.speculative.n_gpu_layers = -2;
+            } else {
+                params.speculative.n_gpu_layers = std::stoi(value);
+            }
             if (!llama_supports_gpu_offload()) {
                 fprintf(stderr, "warning: no usable GPU found, --gpu-layers-draft option will be ignored\n");
                 fprintf(stderr, "warning: one possible reason is that llama.cpp was compiled without GPU support\n");
@@ -3518,15 +3533,15 @@ void common_params_add_preset_options(std::vector<common_arg> & args) {
         [](common_params &, const std::string &) { /* unused */ }
     ).set_env(COMMON_ARG_PRESET_LOAD_ON_STARTUP).set_preset_only());
 
+    args.push_back(common_arg(
+        {"stop-timeout"}, "SECONDS",
+        "in server router mode, force-kill model instance after this many seconds of graceful shutdown",
+        [](common_params &, int) { /* unused */ }
+    ).set_env(COMMON_ARG_PRESET_STOP_TIMEOUT).set_preset_only());
+
     // args.push_back(common_arg(
     //     {"pin"},
     //     "in server router mode, do not unload this model if models_max is exceeded",
     //     [](common_params &) { /* unused */ }
     // ).set_preset_only());
-
-    // args.push_back(common_arg(
-    //     {"unload-idle-seconds"}, "SECONDS",
-    //     "in server router mode, unload models idle for more than this many seconds",
-    //     [](common_params &, int) { /* unused */ }
-    // ).set_preset_only());
 }

common/arg.h

@@ -10,6 +10,7 @@
 
 // pseudo-env variable to identify preset-only arguments
 #define COMMON_ARG_PRESET_LOAD_ON_STARTUP "__PRESET_LOAD_ON_STARTUP"
+#define COMMON_ARG_PRESET_STOP_TIMEOUT    "__PRESET_STOP_TIMEOUT"
 
 //
 // CLI argument parsing

common/common.cpp

@@ -251,7 +251,7 @@ bool set_process_priority(enum ggml_sched_priority prio) {
         case GGML_SCHED_PRIO_REALTIME: p = -20; break;
     }
 
-    if (!setpriority(PRIO_PROCESS, 0, p)) {
+    if (setpriority(PRIO_PROCESS, 0, p) != 0) {
         LOG_WRN("failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno);
         return false;
     }
@@ -1078,6 +1078,8 @@ struct common_init_result::impl {
     impl() = default;
     ~impl() = default;
 
+    // note: the order in which model, context, etc. are declared matters because their destructors will be called bottom-to-top
+
     llama_model_ptr   model;
     llama_context_ptr context;
 
@@ -1107,6 +1109,25 @@ common_init_result::common_init_result(common_params & params) :
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
+    // load and optionally apply lora adapters (must be loaded before context creation)
+    for (auto & la : params.lora_adapters) {
+        llama_adapter_lora_ptr lora;
+        lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
+        if (lora == nullptr) {
+            LOG_ERR("%s: failed to load lora adapter '%s'\n", __func__, la.path.c_str());
+            pimpl->model.reset(model);
+            return;
+        }
+
+        char buf[1024];
+        la.ptr = lora.get();
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.task_name", buf, sizeof(buf));
+        la.task_name = buf;
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
+        la.prompt_prefix = buf;
+        pimpl->lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
+    }
+
     // updates params.sampling
     // TODO: fix naming
     common_init_sampler_from_model(model, params.sampling);
@@ -1243,24 +1264,6 @@ common_init_result_ptr common_init_from_params(common_params & params) {
         }
     }
 
-    // load and optionally apply lora adapters
-    for (auto & la : params.lora_adapters) {
-        llama_adapter_lora_ptr lora;
-        lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
-        if (lora == nullptr) {
-            LOG_ERR("%s: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
-            return res;
-        }
-
-        char buf[1024];
-        la.ptr = lora.get();
-        llama_adapter_meta_val_str(la.ptr, "adapter.lora.task_name", buf, sizeof(buf));
-        la.task_name = buf;
-        llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
-        la.prompt_prefix = buf;
-        res->lora().emplace_back(std::move(lora)); // copy to list of loaded adapters
-    }
-
     if (!params.lora_init_without_apply) {
         common_set_adapter_lora(lctx, params.lora_adapters);
     }
@@ -1339,10 +1342,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
         mparams.devices = params.devices.data();
     }
 
-    if (params.n_gpu_layers != -1) {
-        mparams.n_gpu_layers = params.n_gpu_layers;
-    }
-
+    mparams.n_gpu_layers    = params.n_gpu_layers;
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;

common/common.h

@@ -329,7 +329,7 @@ struct common_params {
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
-    int32_t n_gpu_layers       = -1;               // number of layers to store in VRAM (-1 - use default)
+    int32_t n_gpu_layers       = -1;               // number of layers to store in VRAM, -1 is auto, <= -2 is all
     int32_t main_gpu           = 0;                // the GPU that is used for scratch and small tensors
     float   tensor_split[128]  = {0};              // how split tensors should be distributed across GPUs
     bool    fit_params         = true;             // whether to fit unset model/context parameters to free device memory

convert_hf_to_gguf.py

@@ -141,16 +141,24 @@ class ModelBase:
         self.model_name = model_name
         self.dir_model_card = dir_model  # overridden in convert_lora_to_gguf.py
 
-        # Apply heuristics to figure out typical tensor encoding based on first layer tensor encoding type
+        # Apply heuristics to figure out typical tensor encoding based on first tensor's dtype
+        # NOTE: can't use field "torch_dtype" in config.json, because some finetunes lie.
         if self.ftype == gguf.LlamaFileType.GUESSED:
-            # NOTE: can't use field "torch_dtype" in config.json, because some finetunes lie.
-            _, first_tensor = next(self.get_tensors())
-            if first_tensor.dtype == torch.float16:
-                logger.info(f"choosing --outtype f16 from first tensor type ({first_tensor.dtype})")
-                self.ftype = gguf.LlamaFileType.MOSTLY_F16
+            for _, tensor in self.get_tensors():
+                if tensor.dim() < 2:
+                    continue
+
+                if tensor.dtype == torch.bfloat16:
+                    self.ftype = gguf.LlamaFileType.MOSTLY_BF16
+                    logger.info("heuristics detected bfloat16 tensor dtype, setting --outtype bf16")
+                    break
+                elif tensor.dtype == torch.float16:
+                    self.ftype = gguf.LlamaFileType.MOSTLY_F16
+                    logger.info("heuristics detected float16 tensor dtype, setting --outtype f16")
+                    break
             else:
-                logger.info(f"choosing --outtype bf16 from first tensor type ({first_tensor.dtype})")
-                self.ftype = gguf.LlamaFileType.MOSTLY_BF16
+                self.ftype = gguf.LlamaFileType.MOSTLY_F16
+                logger.info("heuristics unable to detect tensor dtype, defaulting to --outtype f16")
 
         self.dequant_model()
 
@@ -1204,6 +1212,9 @@ class TextModel(ModelBase):
         if chkhsh == "a1e163ecab2e718a4c829d1148b6e86824ec36163bb71941c3dca9cd5ac25756":
             # ref: https://huggingface.co/JetBrains/Mellum-4b-base
             res = "mellum"
+        if chkhsh == "a0b64b4385f123663873756336c085744376d015ff328bb1d901598f63c44152":
+            # ref: https://huggingface.co/answerdotai/ModernBERT-base
+            res = "modern-bert"
         if chkhsh == "49fc0303c9e0d2c2c565c510f64b2d9b271276acdcdadff733249eda9f7d59df":
             # ref: https://huggingface.co/arcee-ai/Trinity-Tokenizer
             res = "afmoe"
@@ -1685,6 +1696,84 @@ class TextModel(ModelBase):
         if template is not None:
             self.gguf_writer.add_chat_template(template)
 
+    def _set_vocab_plamo(self):
+        # PLaMo models use a custom tokenizer with a .jsonl file
+        tokenizer_jsonl_path = self.dir_model / "tokenizer.jsonl"
+        tokenizer_config_path = self.dir_model / "tokenizer_config.json"
+
+        if not tokenizer_jsonl_path.is_file():
+            raise FileNotFoundError(f"PLaMo tokenizer file not found: {tokenizer_jsonl_path}")
+
+        # Load tokenizer config
+        with open(tokenizer_config_path, "r", encoding="utf-8") as f:
+            tokenizer_config = json.load(f)
+
+        # Load tokens from JSONL file (actually a list format)
+        tokens = []
+        scores = []
+        toktypes = []
+
+        with open(tokenizer_jsonl_path, "r", encoding="utf-8") as f:
+            for line_num, line in enumerate(f):
+                if line.strip():
+                    token_data = json.loads(line)
+                    # Format: [token, score, type, ?, ?, ?, ?]
+                    token = token_data[0].encode("utf-8")
+                    score = float(token_data[1])
+                    token_type_str = token_data[2] if len(token_data) > 2 else "NORMAL"
+
+                    tokens.append(token)
+                    scores.append(score)
+
+                    if token_type_str == "UNKNOWN":
+                        toktypes.append(gguf.TokenType.UNKNOWN)
+                    elif token_type_str == "CONTROL":
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    elif token_type_str == "BYTE":
+                        toktypes.append(gguf.TokenType.BYTE)
+                    else:
+                        token_str = token_data[0]
+                        if token_str.startswith("<|plamo:") and token_str.endswith("|>"):
+                            toktypes.append(gguf.TokenType.CONTROL)
+                        else:
+                            toktypes.append(gguf.TokenType.NORMAL)
+
+        vocab_size = self.hparams["vocab_size"]
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(gguf.TokenType.UNUSED)
+
+        self.gguf_writer.add_tokenizer_model("plamo2")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] is not None:
+            token_id = tokens.index(tokenizer_config["bos_token"].encode("utf-8"))
+            self.gguf_writer.add_bos_token_id(token_id)
+        if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] is not None:
+            token_id = tokens.index(tokenizer_config["eos_token"].encode("utf-8"))
+            self.gguf_writer.add_eos_token_id(token_id)
+        if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] is not None:
+            token_id = tokens.index(tokenizer_config["pad_token"].encode("utf-8"))
+            self.gguf_writer.add_pad_token_id(token_id)
+        if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] is not None:
+            token_id = tokens.index(tokenizer_config["sep_token"].encode("utf-8"))
+            self.gguf_writer.add_sep_token_id(token_id)
+        if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] is not None:
+            token_id = tokens.index(tokenizer_config["unk_token"].encode("utf-8"))
+            self.gguf_writer.add_unk_token_id(token_id)
+
+        # Add <|plamo:op|> as EOT to ensure appropriate end of generation
+        self.gguf_writer.add_eot_token_id(4)
+
+        self.gguf_writer.add_add_space_prefix(False)
+
 
 class MmprojModel(ModelBase):
     model_type = ModelType.MMPROJ
@@ -4787,87 +4876,7 @@ class Plamo2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.PLAMO2
 
     def set_vocab(self):
-        # PLaMo 2 uses a custom tokenizer with a .jsonl file
-        # We need to handle this specially
-        tokenizer_jsonl_path = self.dir_model / "tokenizer.jsonl"
-        tokenizer_config_path = self.dir_model / "tokenizer_config.json"
-
-        if not tokenizer_jsonl_path.is_file():
-            raise FileNotFoundError(f"PLaMo 2 tokenizer file not found: {tokenizer_jsonl_path}")
-
-        # Load tokenizer config
-        with open(tokenizer_config_path, 'r', encoding='utf-8') as f:
-            tokenizer_config = json.load(f)
-
-        # Load tokens from JSONL file (actually a list format)
-        tokens = []
-        scores = []
-        toktypes = []
-
-        with open(tokenizer_jsonl_path, 'r', encoding='utf-8') as f:
-            for line_num, line in enumerate(f):
-                if line.strip():
-                    token_data = json.loads(line)
-                    # Format: [token, score, type, ?, ?, ?, ?]
-                    token = token_data[0].encode("utf-8")
-                    score = float(token_data[1])
-                    token_type_str = token_data[2] if len(token_data) > 2 else "NORMAL"
-
-                    tokens.append(token)
-                    scores.append(score)
-
-                    # Map token type strings to GGUF token types
-                    if token_type_str == "UNKNOWN":
-                        toktypes.append(gguf.TokenType.UNKNOWN)
-                    elif token_type_str == "CONTROL":
-                        toktypes.append(gguf.TokenType.CONTROL)
-                    elif token_type_str == "BYTE":
-                        toktypes.append(gguf.TokenType.BYTE)
-                    else:
-                        # Check for PLaMo-2 special tokens
-                        token_str = token_data[0]
-                        if token_str.startswith("<|plamo:") and token_str.endswith("|>"):
-                            toktypes.append(gguf.TokenType.CONTROL)
-                        else:
-                            toktypes.append(gguf.TokenType.NORMAL)
-
-        vocab_size = self.hparams["vocab_size"]
-        if vocab_size > len(tokens):
-            pad_count = vocab_size - len(tokens)
-            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
-            for i in range(1, pad_count + 1):
-                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
-                scores.append(-1000.0)
-                toktypes.append(gguf.TokenType.UNUSED)
-
-        # Use "plamo2" tokenizer type for PLaMo-2's custom Aho-Corasick tokenizer
-        self.gguf_writer.add_tokenizer_model("plamo2")
-        self.gguf_writer.add_tokenizer_pre("default")
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_scores(scores)
-        self.gguf_writer.add_token_types(toktypes)
-
-        # Add special tokens from config
-        if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] is not None:
-            token_id = tokens.index(tokenizer_config["bos_token"].encode("utf-8"))
-            self.gguf_writer.add_bos_token_id(token_id)
-        if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] is not None:
-            token_id = tokens.index(tokenizer_config["eos_token"].encode("utf-8"))
-            self.gguf_writer.add_eos_token_id(token_id)
-        if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] is not None:
-            token_id = tokens.index(tokenizer_config["pad_token"].encode("utf-8"))
-            self.gguf_writer.add_pad_token_id(token_id)
-        if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] is not None:
-            token_id = tokens.index(tokenizer_config["sep_token"].encode("utf-8"))
-            self.gguf_writer.add_sep_token_id(token_id)
-        if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] is not None:
-            token_id = tokens.index(tokenizer_config["unk_token"].encode("utf-8"))
-            self.gguf_writer.add_unk_token_id(token_id)
-
-        # Add <|plamo:op|> as EOT to ensure appropriate end of generation
-        self.gguf_writer.add_eot_token_id(4)
-
-        self.gguf_writer.add_add_space_prefix(False)
+        self._set_vocab_plamo()
 
     def set_gguf_parameters(self):
         hparams = self.hparams
@@ -4955,6 +4964,56 @@ class Plamo2Model(TextModel):
         return [(new_name, data_torch)]
 
 
+@ModelBase.register("Plamo3ForCausalLM", "PLaMo3ForCausalLM")
+class Plamo3Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.PLAMO3
+
+    def set_vocab(self):
+        self._set_vocab_plamo()
+
+        tokenizer_config_path = self.dir_model / "tokenizer_config.json"
+        tokenizer_config = {}
+
+        if tokenizer_config_path.is_file():
+            with open(tokenizer_config_path, encoding="utf-8") as f:
+                tokenizer_config = json.load(f)
+
+        chat_template = tokenizer_config.get("chat_template")
+        chat_template_jinja = self.dir_model / "chat_template.jinja"
+
+        if chat_template_jinja.is_file():
+            with open(chat_template_jinja, encoding="utf-8") as f:
+                chat_template = f.read()
+
+        if chat_template:
+            self.gguf_writer.add_chat_template(chat_template)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
+        if (sliding_window := self.find_hparam(["window_size", "sliding_window"], optional=True)) is not None:
+            self.gguf_writer.add_sliding_window(sliding_window)
+            self.gguf_writer.add_sliding_window_pattern(self.hparams["sliding_window_pattern"])
+            self.gguf_writer.add_rope_freq_base_swa(self.rope_parameters.get("sliding_attention", {"rope_theta": self.hparams.get("rope_local_theta")})["rope_theta"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+
+        if name.endswith(".pre_mixer_norm.weight"):
+            data_torch = data_torch + 1.0
+        elif name.endswith(".post_mixer_norm.weight"):
+            data_torch = data_torch + 1.0 / 5
+        elif name.endswith(".pre_mlp_norm.weight"):
+            data_torch = data_torch + 1.0
+        elif name.endswith(".post_mlp_norm.weight"):
+            data_torch = data_torch + 1.0 / (5**1.5)
+        elif name.endswith((".mixer.q_norm.weight", ".mixer.k_norm.weight")):
+            data_torch = data_torch + 1.0
+        elif name.endswith(".norm.weight"):
+            data_torch = data_torch + 1.0
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 @ModelBase.register("CodeShellForCausalLM")
 class CodeShellModel(TextModel):
     model_arch = gguf.MODEL_ARCH.CODESHELL
@@ -7351,6 +7410,90 @@ class MiniMaxM2Model(TextModel):
         return super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("MiMoV2FlashForCausalLM")
+class MimoV2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.MIMO2
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        assert self.hparams["swa_head_dim"] == self.hparams["head_dim"]
+        assert self.hparams["swa_num_attention_heads"] == self.hparams["num_attention_heads"]
+        assert self.hparams["swa_v_head_dim"] == self.hparams["v_head_dim"]
+        assert self.hparams["topk_method"] == "noaux_tc"
+
+        n_head_kv = self.hparams["num_key_value_heads"]
+        n_head_kv_swa = self.hparams["swa_num_key_value_heads"]
+        n_head_kv_arr = [n_head_kv_swa if use_swa == 1 else n_head_kv for use_swa in self.hparams["hybrid_layer_pattern"]]
+        self.gguf_writer.add_head_count_kv(n_head_kv_arr)
+
+        self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
+        self.gguf_writer.add_sliding_window_pattern(self.hparams["hybrid_layer_pattern"])
+        self.gguf_writer.add_rope_freq_base_swa(self.hparams["swa_rope_theta"])
+        self.gguf_writer.add_value_length(self.hparams["v_head_dim"])
+        self.gguf_writer.add_expert_count(self.hparams["n_routed_experts"])
+        self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
+
+        rope_dim = int(self.hparams["head_dim"] * self.hparams["partial_rotary_factor"])
+        self.gguf_writer.add_rope_dimension_count(rope_dim)
+
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("layernorm_epsilon", 1e-5))
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch, name, bid):
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+
+        if "attention_sink" in name and not name.endswith(".weight"):
+            name += ".weight"
+
+        # TODO: mimo v2 does not indicate the number of next-token-prediction layers, therefore we cannot do the same way as GLM4_MOE
+        if "model.mtp." in name:
+            return []
+
+        # process the experts separately
+        if name.find("mlp.experts") != -1:
+            n_experts = self.hparams["n_routed_experts"]
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["gate_proj", "up_proj", "down_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename_to_retrieve = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename_to_retrieve])
+                        del self._experts[bid][ename_to_retrieve]
+
+                    data_torch = torch.stack(datas, dim=0)
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+                    new_name = self.map_tensor_name(merged_name)
+                    tensors.append((new_name, data_torch))
+
+                return tensors
+            else:
+                return []
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 @ModelBase.register("PanguEmbeddedForCausalLM")
 class PanguEmbeddedModel(TextModel):
     model_arch = gguf.MODEL_ARCH.PANGU_EMBED
@@ -8684,6 +8827,11 @@ class NemotronHModel(GraniteHybridModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("LlamaBidirectionalModel")
+class LlamaEmbedNemotronModel(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.LLAMA_EMBED
+
+
 @ModelBase.register("BailingMoeForCausalLM")
 class BailingMoeModel(TextModel):
     model_arch = gguf.MODEL_ARCH.BAILINGMOE
@@ -9991,6 +10139,36 @@ class SmallThinkerModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("ModernBertModel", "ModernBertForMaskedLM", "ModernBertForSequenceClassification")
+class ModernBertModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.MODERN_BERT
+
+    def set_vocab(self):
+        self.gguf_writer.add_add_bos_token(True)
+        self.gguf_writer.add_add_eos_token(True)
+        self.gguf_writer.add_add_sep_token(True)
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_sliding_window(self.hparams["local_attention"])
+        if (sliding_window_pattern := self.hparams.get("global_attn_every_n_layers")) is not None:
+            self.gguf_writer.add_sliding_window_pattern(sliding_window_pattern)
+        self.gguf_writer.add_rope_freq_base_swa(self.rope_parameters.get("sliding_attention", {"rope_theta": self.hparams.get("local_rope_theta")})["rope_theta"])
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # these layers act as MLM head, so we don't need them
+        if name.startswith("decoder."):
+            return []
+
+        if name.startswith("model."):
+            name = name[6:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("ApertusForCausalLM")
 class ApertusModel(LlamaModel):
     model_arch = gguf.MODEL_ARCH.APERTUS
@@ -10557,8 +10735,8 @@ def parse_args() -> argparse.Namespace:
         help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
     )
     parser.add_argument(
-        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "tq1_0", "tq2_0", "auto"], default="f16",
-        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, tq1_0 or tq2_0 for ternary, and auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "tq1_0", "tq2_0", "auto"], default="auto",
+        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, tq1_0 or tq2_0 for ternary, and auto for the highest-fidelity 16-bit float type",
     )
     parser.add_argument(
         "--bigendian", action="store_true",

convert_hf_to_gguf_update.py

@@ -139,6 +139,7 @@ models = [
     {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
     {"name": "exaone4",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },
     {"name": "mellum",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },
+    {"name": "modern-bert",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/answerdotai/ModernBERT-base", },
     {"name": "afmoe",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/arcee-ai/Trinity-Tokenizer", },
     {"name": "bailingmoe2",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-mini-base-2.0", },
     {"name": "granite-docling",  "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ibm-granite/granite-docling-258M", },

docs/backend/OPENCL.md

@@ -17,7 +17,7 @@ OpenCL (Open Computing Language) is an open, royalty-free standard for cross-pla
 
 ### Llama.cpp + OpenCL
 
-The llama.cpp OpenCL backend is designed to enable llama.cpp on **Qualcomm Adreno GPU** firstly via OpenCL. Thanks to the portabilty of OpenCL, the OpenCL backend can also run on certain Intel GPUs although the performance is not optimal.
+The llama.cpp OpenCL backend is designed to enable llama.cpp on **Qualcomm Adreno GPU** firstly via OpenCL. Thanks to the portabilty of OpenCL, the OpenCL backend can also run on certain Intel GPUs such as those that do not have [SYCL](/docs/backend/SYCL.md) support although the performance is not optimal.
 
 ## OS
 

docs/backend/SYCL.md

@@ -829,7 +829,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
   No. We can't support Ollama issue directly, because we aren't familiar with Ollama.
 
-  Sugguest reproducing on llama.cpp and report similar issue to llama.cpp. We will surpport it.
+  Suggest reproducing on llama.cpp and report similar issue to llama.cpp. We will support it.
 
   It's same for other projects including llama.cpp SYCL backend.
 

docs/backend/hexagon/README.md

@@ -106,7 +106,7 @@ Here are some examples of running various llama.cpp tools via ADB.
 Simple question for Llama-3.2-1B
 
 ```
-~/src/llama.cpp$ M=Llama-3.2-1B-Instruct-Q4_0.gguf D=HTP0 ./scripts/snapdragon/adb/run-cli.sh -no-cnv -p "what is the most popular cookie in the world?"
+~/src/llama.cpp$ M=Llama-3.2-1B-Instruct-Q4_0.gguf D=HTP0 ./scripts/snapdragon/adb/run-completion.sh -p "what is the most popular cookie in the world?"
 ...
 ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
 ggml-hex: Hexagon Arch version v79
@@ -136,7 +136,7 @@ llama_memory_breakdown_print: |   - HTP0-REPACK        |                  504 =
 Summary request for OLMoE-1B-7B. This is a large model that requires two HTP sessions/devices
 
 ```
-~/src/llama.cpp$ M=OLMoE-1B-7B-0125-Instruct-Q4_0.gguf NDEV=2 D=HTP0,HTP1 ./scripts/snapdragon/adb/run-cli.sh -f surfing.txt -no-cnv
+~/src/llama.cpp$ M=OLMoE-1B-7B-0125-Instruct-Q4_0.gguf NDEV=2 D=HTP0,HTP1 ./scripts/snapdragon/adb/run-completion.sh -f surfing.txt
 ...
 ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
 ggml-hex: Hexagon Arch version v81
@@ -234,6 +234,6 @@ build: 6a8cf8914 (6733)
 
   Examples:
 
-      `GGML_HEXAGON_OPMASK=0x1 llama-cli ...` - Ops are enqueued but NPU-side processing is stubbed out
-      `GGML_HEXAGON_OPMASK=0x3 llama-cli ...` - NPU performs dynamic quantization and skips the rest
-      `GGML_HEXAGON_OPMASK=0x7 llama-cli ...` - Full queuing and processing of Ops (default)
+      `GGML_HEXAGON_OPMASK=0x1 llama-completion ...` - Ops are enqueued but NPU-side processing is stubbed out
+      `GGML_HEXAGON_OPMASK=0x3 llama-completion ...` - NPU performs dynamic quantization and skips the rest
+      `GGML_HEXAGON_OPMASK=0x7 llama-completion ...` - Full queuing and processing of Ops (default)

docs/backend/hexagon/developer.md

@@ -49,7 +49,7 @@ Each Hexagon device behaves like a GPU from the offload and model splitting pers
 Here is an example of running GPT-OSS-20B model on a newer Snapdragon device with 16GB of DDR.
 
 ```
-M=gpt-oss-20b-Q4_0.gguf NDEV=4 D=HTP0,HTP1,HTP2,HTP3 P=surfing.txt scripts/snapdragon/adb/run-cli.sh -no-cnv -f surfing.txt -n 32
+M=gpt-oss-20b-Q4_0.gguf NDEV=4 D=HTP0,HTP1,HTP2,HTP3 P=surfing.txt scripts/snapdragon/adb/run-completion.sh -f surfing.txt -n 32
 ...
 LD_LIBRARY_PATH=/data/local/tmp/llama.cpp/lib
 ADSP_LIBRARY_PATH=/data/local/tmp/llama.cpp/lib

docs/build.md

@@ -150,19 +150,38 @@ We also have a [guide](./backend/CUDA-FEDORA.md) for setting up CUDA toolkit in
 
 
 ### Compilation
+
+Make sure to read the notes about the CPU build for general instructions for e.g. speeding up the compilation.
+
 ```bash
 cmake -B build -DGGML_CUDA=ON
 cmake --build build --config Release
 ```
 
+### Non-Native Builds
+
+By default llama.cpp will be built for the hardware that is connected to the system at that time.
+For a build covering all CUDA GPUs, disable `GGML_NATIVE`:
+
+```bash
+cmake -B build -DGGML_CUDA=ON -DGGML_NATIVE=OFF
+```
+
+The resulting binary should run on all CUDA GPUs with optimal performance, though some just-in-time compilation may be required.
+
 ### Override Compute Capability Specifications
 
-If `nvcc` cannot detect your gpu, you may get compile-warnings such as:
+If `nvcc` cannot detect your gpu, you may get compile warnings such as:
  ```text
 nvcc warning : Cannot find valid GPU for '-arch=native', default arch is used
 ```
 
-To override the `native` GPU detection:
+One option is to do a non-native build as described above.
+However, this will result in a large binary that takes a long time to compile.
+Alternatively it is also possible to explicitly specify CUDA architectures.
+This may also make sense for a non-native build, for that one should look at the logic in `ggml/src/ggml-cuda/CMakeLists.txt` as a starting point.
+
+To override the default CUDA architectures:
 
 #### 1. Take note of the `Compute Capability` of your NVIDIA devices: ["CUDA: Your GPU Compute > Capability"](https://developer.nvidia.com/cuda-gpus).
 

docs/development/parsing.md

@@ -55,7 +55,7 @@ auto parser = build_chat_peg_native_parser([&](common_chat_peg_native_builder &
 ```
 
 For a more complete example, see `test_example_native()` in
-[tests/test-chat-peg-parser.cpp](tests/test-chat-peg-parser.cpp).
+[tests/test-chat-peg-parser.cpp](/tests/test-chat-peg-parser.cpp).
 
 ## Parsers/Combinators
 
@@ -175,7 +175,7 @@ Most model output can be placed in one of the following categories:
   (Qwen3-Coder, MiniMax M2) or pseudo-function calls (LFM2)
 
 To provide broad coverage,
-[`common/chat-peg-parser.h`](common/chat-peg-parser.h) contains builders and
+[`common/chat-peg-parser.h`](/common/chat-peg-parser.h) contains builders and
 mappers that help create parsers and visitors/extractors for these types. They
 require parsers to tag nodes to conform to an AST "shape". This normalization
 makes it easy to extract information and generalize parsing.

examples/gen-docs/gen-docs.cpp

@@ -2,57 +2,74 @@
 #include "common.h"
 
 #include <fstream>
+#include <sstream>
 #include <string>
 
 // Export usage message (-h) to markdown format
+// Automatically update the markdown docs
 
-static void write_table_header(std::ofstream & file) {
-    file << "| Argument | Explanation |\n";
-    file << "| -------- | ----------- |\n";
+#define HELP_START_MARKER "<!-- HELP_START -->"
+#define HELP_END_MARKER   "<!-- HELP_END -->"
+#define NOTE_MESSAGE      "<!-- IMPORTANT: The list below is auto-generated by llama-gen-docs; do NOT modify it manually -->"
+
+struct md_file {
+    llama_example ex;
+    std::string fname;
+    std::string specific_section_header;
+};
+
+std::vector<md_file> md_files = {
+    {LLAMA_EXAMPLE_CLI,        "tools/cli/README.md",        "CLI-specific params"},
+    {LLAMA_EXAMPLE_COMPLETION, "tools/completion/README.md", "Completion-specific params"},
+    {LLAMA_EXAMPLE_SERVER,     "tools/server/README.md",     "Server-specific params"},
+};
+
+static void write_table_header(std::ostringstream & ss) {
+    ss << "| Argument | Explanation |\n";
+    ss << "| -------- | ----------- |\n";
 }
 
-static void write_table_entry(std::ofstream & file, const common_arg & opt) {
-    file << "| `";
+static void write_table_entry(std::ostringstream & ss, const common_arg & opt) {
+    ss << "| `";
     // args
     auto all_args = opt.get_args();
     for (const auto & arg : all_args) {
     if (arg == all_args.front()) {
-            file << arg;
-            if (all_args.size() > 1) file << ", ";
+            ss << arg;
+            if (all_args.size() > 1) ss << ", ";
         } else {
-            file << arg << (arg != all_args.back() ? ", " : "");
+            ss << arg << (arg != all_args.back() ? ", " : "");
         }
     }
     // value hint
     if (opt.value_hint) {
         std::string md_value_hint(opt.value_hint);
         string_replace_all(md_value_hint, "|", "\\|");
-        file << " " << md_value_hint;
+        ss << " " << md_value_hint;
     }
     if (opt.value_hint_2) {
         std::string md_value_hint_2(opt.value_hint_2);
         string_replace_all(md_value_hint_2, "|", "\\|");
-        file << " " << md_value_hint_2;
+        ss << " " << md_value_hint_2;
     }
     // help text
     std::string md_help(opt.help);
+    md_help = string_strip(md_help);
     string_replace_all(md_help, "\n", "<br/>");
     string_replace_all(md_help, "|", "\\|");
-    file << "` | " << md_help << " |\n";
+    ss << "` | " << md_help << " |\n";
 }
 
-static void write_table(std::ofstream & file, std::vector<common_arg *> & opts) {
-    write_table_header(file);
+static void write_table(std::ostringstream & ss, std::vector<common_arg *> & opts) {
+    write_table_header(ss);
     for (const auto & opt : opts) {
-        write_table_entry(file, *opt);
+        write_table_entry(ss, *opt);
     }
 }
 
-static void export_md(std::string fname, llama_example ex, std::string name) {
-    std::ofstream file(fname, std::ofstream::out | std::ofstream::trunc);
-
+static void write_help(std::ostringstream & ss, const md_file & md) {
     common_params params;
-    auto ctx_arg = common_params_parser_init(params, ex);
+    auto ctx_arg = common_params_parser_init(params, md.ex);
 
     std::vector<common_arg *> common_options;
     std::vector<common_arg *> sparam_options;
@@ -68,18 +85,58 @@ static void export_md(std::string fname, llama_example ex, std::string name) {
         }
     }
 
-    file << "**Common params**\n\n";
-    write_table(file, common_options);
-    file << "\n\n**Sampling params**\n\n";
-    write_table(file, sparam_options);
-    file << "\n\n**" << name << "-specific params**\n\n";
-    write_table(file, specific_options);
+    ss << HELP_START_MARKER << "\n\n";
+
+    ss << NOTE_MESSAGE << "\n\n";
+
+    ss << "### Common params\n\n";
+    write_table(ss, common_options);
+    ss << "\n\n### Sampling params\n\n";
+    write_table(ss, sparam_options);
+    ss << "\n\n### " << md.specific_section_header << "\n\n";
+    write_table(ss, specific_options);
+
+    ss << "\n" << HELP_END_MARKER;
 }
 
 int main(int, char **) {
-    // TODO: add CLI
-    export_md("autogen-completion.md", LLAMA_EXAMPLE_COMPLETION, "Tool");
-    export_md("autogen-server.md", LLAMA_EXAMPLE_SERVER, "Server");
+    for (const auto & md : md_files) {
+        std::ifstream infile(md.fname);
+        if (!infile.is_open()) {
+            fprintf(stderr, "failed to open file '%s' for reading\n", md.fname.c_str());
+            return 1;
+        }
+
+        std::ostringstream ss;
+        ss << infile.rdbuf();
+        infile.close();
+
+        std::string content = ss.str();
+
+        size_t help_start = content.find(HELP_START_MARKER);
+        size_t help_end   = content.find(HELP_END_MARKER);
+
+        if (help_start == std::string::npos || help_end == std::string::npos || help_end <= help_start) {
+            fprintf(stderr, "failed to find help markers in file '%s'\n", md.fname.c_str());
+            return 1;
+        }
+
+        std::ostringstream new_help_ss;
+        write_help(new_help_ss, md);
+        std::string new_help = new_help_ss.str();
+
+        content = content.substr(0, help_start) + new_help + content.substr(help_end + strlen(HELP_END_MARKER));
+
+        std::ofstream outfile(md.fname);
+        if (!outfile.is_open()) {
+            fprintf(stderr, "failed to open file '%s' for writing\n", md.fname.c_str());
+            return 1;
+        }
+        outfile << content;
+        outfile.close();
+
+        printf("Updated help in '%s'\n", md.fname.c_str());
+    }
 
     return 0;
 }

examples/llama.android/app/build.gradle.kts

@@ -41,11 +41,8 @@ android {
         }
     }
     compileOptions {
-        sourceCompatibility = JavaVersion.VERSION_1_8
-        targetCompatibility = JavaVersion.VERSION_1_8
-    }
-    kotlinOptions {
-        jvmTarget = "1.8"
+        sourceCompatibility = JavaVersion.VERSION_17
+        targetCompatibility = JavaVersion.VERSION_17
     }
 }
 

examples/llama.android/app/src/main/java/com/example/llama/MainActivity.kt

@@ -6,6 +6,7 @@ import android.util.Log
 import android.widget.EditText
 import android.widget.TextView
 import android.widget.Toast
+import androidx.activity.addCallback
 import androidx.activity.enableEdgeToEdge
 import androidx.activity.result.contract.ActivityResultContracts
 import androidx.appcompat.app.AppCompatActivity
@@ -18,6 +19,7 @@ import com.arm.aichat.gguf.GgufMetadata
 import com.arm.aichat.gguf.GgufMetadataReader
 import com.google.android.material.floatingactionbutton.FloatingActionButton
 import kotlinx.coroutines.Dispatchers
+import kotlinx.coroutines.Job
 import kotlinx.coroutines.flow.onCompletion
 import kotlinx.coroutines.launch
 import kotlinx.coroutines.withContext
@@ -36,6 +38,7 @@ class MainActivity : AppCompatActivity() {
 
     // Arm AI Chat inference engine
     private lateinit var engine: InferenceEngine
+    private var generationJob: Job? = null
 
     // Conversation states
     private var isModelReady = false
@@ -47,11 +50,13 @@ class MainActivity : AppCompatActivity() {
         super.onCreate(savedInstanceState)
         enableEdgeToEdge()
         setContentView(R.layout.activity_main)
+        // View model boilerplate and state management is out of this basic sample's scope
+        onBackPressedDispatcher.addCallback { Log.w(TAG, "Ignore back press for simplicity") }
 
         // Find views
         ggufTv = findViewById(R.id.gguf)
         messagesRv = findViewById(R.id.messages)
-        messagesRv.layoutManager = LinearLayoutManager(this)
+        messagesRv.layoutManager = LinearLayoutManager(this).apply { stackFromEnd = true }
         messagesRv.adapter = messageAdapter
         userInputEt = findViewById(R.id.user_input)
         userActionFab = findViewById(R.id.fab)
@@ -157,33 +162,35 @@ class MainActivity : AppCompatActivity() {
      * Validate and send the user message into [InferenceEngine]
      */
     private fun handleUserInput() {
-        userInputEt.text.toString().also { userSsg ->
-            if (userSsg.isEmpty()) {
+        userInputEt.text.toString().also { userMsg ->
+            if (userMsg.isEmpty()) {
                 Toast.makeText(this, "Input message is empty!", Toast.LENGTH_SHORT).show()
             } else {
                 userInputEt.text = null
+                userInputEt.isEnabled = false
                 userActionFab.isEnabled = false
 
                 // Update message states
-                messages.add(Message(UUID.randomUUID().toString(), userSsg, true))
+                messages.add(Message(UUID.randomUUID().toString(), userMsg, true))
                 lastAssistantMsg.clear()
                 messages.add(Message(UUID.randomUUID().toString(), lastAssistantMsg.toString(), false))
 
-                lifecycleScope.launch(Dispatchers.Default) {
-                    engine.sendUserPrompt(userSsg)
+                generationJob = lifecycleScope.launch(Dispatchers.Default) {
+                    engine.sendUserPrompt(userMsg)
                         .onCompletion {
                             withContext(Dispatchers.Main) {
+                                userInputEt.isEnabled = true
                                 userActionFab.isEnabled = true
                             }
                         }.collect { token ->
-                            val messageCount = messages.size
-                            check(messageCount > 0 && !messages[messageCount - 1].isUser)
-
-                            messages.removeAt(messageCount - 1).copy(
-                                content = lastAssistantMsg.append(token).toString()
-                            ).let { messages.add(it) }
-
                             withContext(Dispatchers.Main) {
+                                val messageCount = messages.size
+                                check(messageCount > 0 && !messages[messageCount - 1].isUser)
+
+                                messages.removeAt(messageCount - 1).copy(
+                                    content = lastAssistantMsg.append(token).toString()
+                                ).let { messages.add(it) }
+
                                 messageAdapter.notifyItemChanged(messages.size - 1)
                             }
                         }
@@ -195,6 +202,7 @@ class MainActivity : AppCompatActivity() {
     /**
      * Run a benchmark with the model file
      */
+    @Deprecated("This benchmark doesn't accurately indicate GUI performance expected by app developers")
     private suspend fun runBenchmark(modelName: String, modelFile: File) =
         withContext(Dispatchers.Default) {
             Log.i(TAG, "Starts benchmarking $modelName")
@@ -223,6 +231,16 @@ class MainActivity : AppCompatActivity() {
             if (!it.exists()) { it.mkdir() }
         }
 
+    override fun onStop() {
+        generationJob?.cancel()
+        super.onStop()
+    }
+
+    override fun onDestroy() {
+        engine.destroy()
+        super.onDestroy()
+    }
+
     companion object {
         private val TAG = MainActivity::class.java.simpleName
 

examples/llama.android/app/src/main/res/layout/activity_main.xml

@@ -24,7 +24,7 @@
                 android:id="@+id/gguf"
                 android:layout_width="match_parent"
                 android:layout_height="wrap_content"
-                android:layout_margin="16dp"
+                android:padding="16dp"
                 android:text="Selected GGUF model's metadata will show here."
                 style="@style/TextAppearance.MaterialComponents.Body2" />
 
@@ -33,8 +33,7 @@
         <com.google.android.material.divider.MaterialDivider
             android:layout_width="match_parent"
             android:layout_height="2dp"
-            android:layout_marginHorizontal="16dp"
-            android:layout_marginVertical="8dp" />
+            android:layout_marginHorizontal="16dp" />
 
         <androidx.recyclerview.widget.RecyclerView
             android:id="@+id/messages"

examples/llama.android/gradle/libs.versions.toml

@@ -1,15 +1,15 @@
 [versions]
 
 # Plugins
-agp = "8.13.0"
-kotlin = "2.2.20"
+agp = "8.13.2"
+kotlin = "2.3.0"
 
 # AndroidX
-activity = "1.11.0"
+activity = "1.12.2"
 appcompat = "1.7.1"
 core-ktx = "1.17.0"
 constraint-layout = "2.2.1"
-datastore-preferences = "1.1.7"
+datastore-preferences = "1.2.0"
 
 # Material
 material = "1.13.0"

examples/llama.android/lib/src/main/cpp/ai_chat.cpp

@@ -560,6 +560,6 @@ Java_com_arm_aichat_internal_InferenceEngineImpl_unload(JNIEnv * /*unused*/, job
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_arm_aichat_internal_InferenceEngineImpl_shutdown(JNIEnv *env, jobject /*unused*/) {
+Java_com_arm_aichat_internal_InferenceEngineImpl_shutdown(JNIEnv *, jobject /*unused*/) {
     llama_backend_free();
 }

examples/llama.android/lib/src/main/java/com/arm/aichat/InferenceEngine.kt

@@ -38,7 +38,7 @@ interface InferenceEngine {
     /**
      * Unloads the currently loaded model.
      */
-    suspend fun cleanUp()
+    fun cleanUp()
 
     /**
      * Cleans up resources when the engine is no longer needed.

examples/llama.android/lib/src/main/java/com/arm/aichat/internal/InferenceEngineImpl.kt

@@ -15,9 +15,11 @@ import kotlinx.coroutines.cancel
 import kotlinx.coroutines.flow.Flow
 import kotlinx.coroutines.flow.MutableStateFlow
 import kotlinx.coroutines.flow.StateFlow
+import kotlinx.coroutines.flow.asStateFlow
 import kotlinx.coroutines.flow.flow
 import kotlinx.coroutines.flow.flowOn
 import kotlinx.coroutines.launch
+import kotlinx.coroutines.runBlocking
 import kotlinx.coroutines.withContext
 import java.io.File
 import java.io.IOException
@@ -109,9 +111,11 @@ internal class InferenceEngineImpl private constructor(
 
     private val _state =
         MutableStateFlow<InferenceEngine.State>(InferenceEngine.State.Uninitialized)
-    override val state: StateFlow<InferenceEngine.State> = _state
+    override val state: StateFlow<InferenceEngine.State> = _state.asStateFlow()
 
     private var _readyForSystemPrompt = false
+    @Volatile
+    private var _cancelGeneration = false
 
     /**
      * Single-threaded coroutine dispatcher & scope for LLama asynchronous operations
@@ -169,6 +173,8 @@ internal class InferenceEngineImpl private constructor(
                 }
                 Log.i(TAG, "Model loaded!")
                 _readyForSystemPrompt = true
+
+                _cancelGeneration = false
                 _state.value = InferenceEngine.State.ModelReady
             } catch (e: Exception) {
                 Log.e(TAG, (e.message ?: "Error loading model") + "\n" + pathToModel, e)
@@ -231,15 +237,19 @@ internal class InferenceEngineImpl private constructor(
 
             Log.i(TAG, "User prompt processed. Generating assistant prompt...")
             _state.value = InferenceEngine.State.Generating
-            while (true) {
+            while (!_cancelGeneration) {
                 generateNextToken()?.let { utf8token ->
                     if (utf8token.isNotEmpty()) emit(utf8token)
                 } ?: break
             }
-            Log.i(TAG, "Assistant generation complete. Awaiting user prompt...")
+            if (_cancelGeneration) {
+                Log.i(TAG, "Assistant generation aborted per requested.")
+            } else {
+                Log.i(TAG, "Assistant generation complete. Awaiting user prompt...")
+            }
             _state.value = InferenceEngine.State.ModelReady
         } catch (e: CancellationException) {
-            Log.i(TAG, "Generation cancelled by user.")
+            Log.i(TAG, "Assistant generation's flow collection cancelled.")
             _state.value = InferenceEngine.State.ModelReady
             throw e
         } catch (e: Exception) {
@@ -268,8 +278,9 @@ internal class InferenceEngineImpl private constructor(
     /**
      * Unloads the model and frees resources, or reset error states
      */
-    override suspend fun cleanUp() =
-        withContext(llamaDispatcher) {
+    override fun cleanUp() {
+        _cancelGeneration = true
+        runBlocking(llamaDispatcher) {
             when (val state = _state.value) {
                 is InferenceEngine.State.ModelReady -> {
                     Log.i(TAG, "Unloading model and free resources...")
@@ -293,17 +304,21 @@ internal class InferenceEngineImpl private constructor(
                 else -> throw IllegalStateException("Cannot unload model in ${state.javaClass.simpleName}")
             }
         }
+    }
 
     /**
      * Cancel all ongoing coroutines and free GGML backends
      */
     override fun destroy() {
-        _readyForSystemPrompt = false
-        llamaScope.cancel()
-        when(_state.value) {
-            is InferenceEngine.State.Uninitialized -> {}
-            is InferenceEngine.State.Initialized -> shutdown()
-            else -> { unload(); shutdown() }
+        _cancelGeneration = true
+        runBlocking(llamaDispatcher) {
+            _readyForSystemPrompt = false
+            when(_state.value) {
+                is InferenceEngine.State.Uninitialized -> {}
+                is InferenceEngine.State.Initialized -> shutdown()
+                else -> { unload(); shutdown() }
+            }
         }
+        llamaScope.cancel()
     }
 }

examples/model-conversion/Makefile

@@ -25,6 +25,8 @@ define quantize_model
 	@echo "Export the quantized model path to $(2) variable in your environment"
 endef
 
+DEVICE ?= auto
+
 ###
 ### Casual Model targets/recipes
 ###
@@ -53,7 +55,7 @@ causal-convert-mm-model:
 
 causal-run-original-model:
 	$(call validate_model_path,causal-run-original-model)
-	@MODEL_PATH="$(MODEL_PATH)" ./scripts/causal/run-org-model.py
+	@MODEL_PATH="$(MODEL_PATH)" ./scripts/causal/run-org-model.py --device "$(DEVICE)"
 
 causal-run-converted-model:
 	@CONVERTED_MODEL="$(CONVERTED_MODEL)" ./scripts/causal/run-converted-model.sh

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

@@ -5,8 +5,11 @@ set -e
 MODEL_PATH="${1:-"$MODEL_PATH"}"
 MODEL_NAME="${2:-$(basename "$MODEL_PATH")}"
 
+CONVERTED_MODEL_PATH="${1:-"$CONVERTED_MODEL"}"
+CONVERTED_MODEL_NAME="${2:-$(basename "$CONVERTED_MODEL_PATH" ".gguf")}"
+
 if [ -t 0 ]; then
-    CPP_EMBEDDINGS="data/llamacpp-${MODEL_NAME}-embeddings.bin"
+    CPP_EMBEDDINGS="data/llamacpp-${CONVERTED_MODEL_NAME}-embeddings.bin"
 else
     # Process piped JSON data and convert to binary (matching logits.cpp format)
     TEMP_FILE=$(mktemp /tmp/tmp.XXXXXX.binn)

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

@@ -4,149 +4,179 @@ import argparse
 import os
 import sys
 import importlib
+import torch
+import numpy as np
+
 from pathlib import Path
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModelForImageTextToText, AutoConfig
 
 # Add parent directory to path for imports
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
-
-from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModelForImageTextToText, AutoConfig
-import torch
-import numpy as np
 from utils.common import debug_hook
 
-parser = argparse.ArgumentParser(description="Process model with specified path")
-parser.add_argument("--model-path", "-m", help="Path to the model")
-parser.add_argument("--prompt-file", "-f", help="Optional prompt file", required=False)
-parser.add_argument("--verbose", "-v", action="store_true", help="Enable verbose debug output")
-args = parser.parse_args()
+def parse_arguments():
+    parser = argparse.ArgumentParser(description="Process model with specified path")
+    parser.add_argument("--model-path", "-m", help="Path to the model")
+    parser.add_argument("--prompt-file", "-f", help="Optional prompt file", required=False)
+    parser.add_argument("--verbose", "-v", action="store_true", help="Enable verbose debug output")
+    parser.add_argument("--device", "-d", help="Device to use (cpu, cuda, mps, auto)", default="auto")
+    return 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"
-    )
+def load_model_and_tokenizer(model_path, device="auto"):
+    print("Loading model and tokenizer using AutoTokenizer:", model_path)
+    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
+    multimodal = False
+    full_config = config
 
-### If you want to dump RoPE activations, uncomment the following lines:
-### === START ROPE DEBUG ===
-# from utils.common import setup_rope_debug
-# setup_rope_debug("transformers.models.apertus.modeling_apertus")
-### == END ROPE DEBUG ===
-
-
-print("Loading model and tokenizer using AutoTokenizer:", model_path)
-tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
-config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
-multimodal = False
-full_config = config
-
-print("Model type:       ", config.model_type)
-if "vocab_size" not in config and "text_config" in config:
-    config = config.text_config
-    multimodal = True
-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)
-
-unreleased_model_name = os.getenv("UNRELEASED_MODEL_NAME")
-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:
-    if multimodal:
-        model = AutoModelForImageTextToText.from_pretrained(
-            model_path, device_map="auto", offload_folder="offload", trust_remote_code=True, config=full_config
-        )
+    # Determine device_map based on device argument
+    if device == "cpu":
+        device_map = {"": "cpu"}
+        print("Forcing CPU usage")
+    elif device == "auto":
+        device_map = "auto"
     else:
-        model = AutoModelForCausalLM.from_pretrained(
-            model_path, device_map="auto", offload_folder="offload", trust_remote_code=True, config=config
+        device_map = {"": device}
+
+    print("Model type:       ", config.model_type)
+    if "vocab_size" not in config and "text_config" in config:
+        config = config.text_config
+        multimodal = True
+
+    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)
+
+    unreleased_model_name = os.getenv("UNRELEASED_MODEL_NAME")
+    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}")
 
-if args.verbose:
-    for name, module in model.named_modules():
-        if len(list(module.children())) == 0:  # only leaf modules
-            module.register_forward_hook(debug_hook(name))
+        try:
+            model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+            model = model_class.from_pretrained(
+                    model_path,
+                    device_map=device_map,
+                    offload_folder="offload",
+                    trust_remote_code=True,
+                    config=config
+            )
+        except (ImportError, AttributeError) as e:
+            print(f"Failed to import or load model: {e}")
+            exit(1)
+    else:
+        if multimodal:
+            model = AutoModelForImageTextToText.from_pretrained(
+                    model_path,
+                    device_map=device_map,
+                    offload_folder="offload",
+                    trust_remote_code=True,
+                    config=full_config
+            )
+        else:
+            model = AutoModelForCausalLM.from_pretrained(
+                    model_path,
+                    device_map=device_map,
+                    offload_folder="offload",
+                    trust_remote_code=True,
+                    config=config
+            )
 
-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__}")
+    print(f"Model class: {model.__class__.__name__}")
 
-device = next(model.parameters()).device
-if args.prompt_file:
-    with open(args.prompt_file, encoding='utf-8') as f:
-        prompt = f.read()
-elif os.getenv("MODEL_TESTING_PROMPT"):
-    prompt = os.getenv("MODEL_TESTING_PROMPT")
-else:
-    prompt = "Hello, my name is"
-input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
+    return model, tokenizer, config
 
-print(f"Input tokens: {input_ids}")
-print(f"Input text: {repr(prompt)}")
-print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")
+def enable_torch_debugging(model):
+        for name, module in model.named_modules():
+            if len(list(module.children())) == 0:  # only leaf modules
+                module.register_forward_hook(debug_hook(name))
 
-batch_size = 512
+def get_prompt(args):
+    if args.prompt_file:
+        with open(args.prompt_file, encoding='utf-8') as f:
+            return f.read()
+    elif os.getenv("MODEL_TESTING_PROMPT"):
+        return os.getenv("MODEL_TESTING_PROMPT")
+    else:
+        return "Hello, my name is"
 
-with torch.no_grad():
-    past = None
-    outputs = None
-    for i in range(0, input_ids.size(1), batch_size):
-        print(f"Processing chunk with tokens {i} to {i + batch_size}")
-        chunk = input_ids[:, i:i + batch_size]
-        outputs = model(chunk.to(model.device), past_key_values=past, use_cache=True)
-        past = outputs.past_key_values
+def main():
+    args = parse_arguments()
+    model_path = os.environ.get("MODEL_PATH", args.model_path)
+    if model_path is None:
+        print("Error: Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+        sys.exit(1)
 
-    logits = outputs.logits # type: ignore
 
-    # Extract logits for the last token (next token prediction)
-    last_logits = logits[0, -1, :].float().cpu().numpy()
+    model, tokenizer, config = load_model_and_tokenizer(model_path, args.device)
 
-    print(f"Logits shape: {logits.shape}")
-    print(f"Last token logits shape: {last_logits.shape}")
-    print(f"Vocab size: {len(last_logits)}")
+    if args.verbose:
+        enable_torch_debugging(model)
 
-    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"
+    model_name = os.path.basename(model_path)
 
-    # Save to file for comparison
-    last_logits.astype(np.float32).tofile(bin_filename)
+    # Iterate over the model parameters (the tensors) and get the first one
+    # and use it to get the device the model is on.
+    device = next(model.parameters()).device
+    prompt = get_prompt(args)
+    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
 
-    # 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(f"Input tokens: {input_ids}")
+    print(f"Input text: {repr(prompt)}")
+    print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")
 
-    # Print some sample logits for quick verification
-    print(f"First 10 logits: {last_logits[:10]}")
-    print(f"Last 10 logits: {last_logits[-10:]}")
+    batch_size = 512
 
-    # 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}")
+    with torch.no_grad():
+        past = None
+        outputs = None
+        for i in range(0, input_ids.size(1), batch_size):
+            print(f"Processing chunk with tokens {i} to {i + batch_size}")
+            chunk = input_ids[:, i:i + batch_size]
+            outputs = model(chunk.to(model.device), past_key_values=past, use_cache=True)
+            past = outputs.past_key_values
 
-    print(f"Saved bin logits to: {bin_filename}")
-    print(f"Saved txt logist to: {txt_filename}")
+        logits = outputs.logits # type: ignore
+
+        # Extract logits for the last token (next token prediction)
+        last_logits = logits[0, -1, :].float().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}")
+
+if __name__ == "__main__":
+    main()

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

@@ -2,6 +2,7 @@
 
 import argparse
 import os
+import sys
 import numpy as np
 import importlib
 from pathlib import Path
@@ -9,169 +10,243 @@ 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')
-parser.add_argument('--prompts-file', '-p', help='Path to file containing prompts (one per line)')
-parser.add_argument('--use-sentence-transformers', action='store_true',
-                    help='Use SentenceTransformer to apply all numbered layers (01_Pooling, 02_Dense, 03_Dense, 04_Normalize)')
-args = parser.parse_args()
+def parse_arguments():
+    parser = argparse.ArgumentParser(description='Run original embedding model')
+    parser.add_argument(
+        '--model-path',
+        '-m',
+        help='Path to the model'
+    )
+    parser.add_argument(
+        '--prompts-file',
+        '-p',
+        help='Path to file containing prompts (one per line)'
+    )
+    parser.add_argument(
+        '--use-sentence-transformers',
+        action='store_true',
+        help=('Use SentenceTransformer to apply all numbered layers '
+              '(01_Pooling, 02_Dense, 03_Dense, 04_Normalize)')
+    )
+    parser.add_argument(
+        '--device',
+        '-d',
+        help='Device to use (cpu, cuda, mps, auto)',
+        default='auto'
+    )
+    return parser.parse_args()
 
-def read_prompt_from_file(file_path):
-    try:
-        with open(file_path, 'r', encoding='utf-8') as f:
-            return f.read().strip()
-    except FileNotFoundError:
-        print(f"Error: Prompts file '{file_path}' not found")
-        exit(1)
-    except Exception as e:
-        print(f"Error reading prompts file: {e}")
-        exit(1)
 
-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")
-
-# Determine if we should use SentenceTransformer
-use_sentence_transformers = args.use_sentence_transformers or os.environ.get('USE_SENTENCE_TRANSFORMERS', '').lower() in ('1', 'true', 'yes')
-
-if use_sentence_transformers:
-    from sentence_transformers import SentenceTransformer
-    print("Using SentenceTransformer to apply all numbered layers")
-    model = SentenceTransformer(model_path)
-    tokenizer = model.tokenizer
-    config = model[0].auto_model.config  # type: ignore
-else:
-    tokenizer = AutoTokenizer.from_pretrained(model_path)
-
-    config = AutoConfig.from_pretrained(model_path)
-
-    # This can be used to override the sliding window size for manual testing. This
-    # can be useful to verify the sliding window attention mask in the original model
-    # and compare it with the converted .gguf model.
-    if hasattr(config, 'sliding_window'):
-        original_sliding_window = config.sliding_window
-        #original_sliding_window = 6
-        print(f"Modified sliding window: {original_sliding_window} -> {config.sliding_window}")
-
-    print(f"Using unreleased model: {unreleased_model_name}")
-    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, config=config)
-        except (ImportError, AttributeError) as e:
-            print(f"Failed to import or load model: {e}")
-            exit(1)
+def load_model_and_tokenizer(model_path, use_sentence_transformers=False, device="auto"):
+    if device == "cpu":
+        device_map = {"": "cpu"}
+        print("Forcing CPU usage")
+    elif device == "auto":
+        # On Mac, "auto" device_map can cause issues with accelerate
+        # So we detect the best device manually
+        if torch.cuda.is_available():
+            device_map = {"": "cuda"}
+            print("Using CUDA")
+        elif torch.backends.mps.is_available():
+            device_map = {"": "mps"}
+            print("Using MPS (Apple Metal)")
+        else:
+            device_map = {"": "cpu"}
+            print("Using CPU")
     else:
-        model = AutoModel.from_pretrained(model_path, config=config)
-    print(f"Model class: {type(model)}")
-    print(f"Model file: {type(model).__module__}")
+        device_map = {"": device}
 
-# Verify the model is using the correct sliding window
-if not use_sentence_transformers:
-    if hasattr(model.config, 'sliding_window'):  # type: ignore
-        print(f"Model's sliding_window: {model.config.sliding_window}")  # type: ignore
-    else:
-        print("Model config does not have sliding_window attribute")
-
-model_name = os.path.basename(model_path)
-
-if args.prompts_file:
-    prompt_text = read_prompt_from_file(args.prompts_file)
-    texts = [prompt_text]
-else:
-    texts = ["Hello world today"]
-
-with torch.no_grad():
     if use_sentence_transformers:
-        embeddings = model.encode(texts, convert_to_numpy=True)
-        all_embeddings = embeddings  # Shape: [batch_size, hidden_size]
-
-        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}'")
-
-        print(f"Embeddings shape (after all SentenceTransformer layers): {all_embeddings.shape}")
-        print(f"Embedding dimension: {all_embeddings.shape[1] if len(all_embeddings.shape) > 1 else all_embeddings.shape[0]}")  # type: ignore
+        from sentence_transformers import SentenceTransformer
+        print("Using SentenceTransformer to apply all numbered layers")
+        model = SentenceTransformer(model_path)
+        tokenizer = model.tokenizer
+        config = model[0].auto_model.config  # type: ignore
     else:
-        # Standard approach: use base model output only
-        encoded = tokenizer(
-            texts,
-            padding=True,
-            truncation=True,
-            return_tensors="pt"
-        )
+        tokenizer = AutoTokenizer.from_pretrained(model_path)
+        config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
 
-        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}'")
+        # This can be used to override the sliding window size for manual testing. This
+        # can be useful to verify the sliding window attention mask in the original model
+        # and compare it with the converted .gguf model.
+        if hasattr(config, 'sliding_window'):
+            original_sliding_window = config.sliding_window
+            print(f"Modified sliding window: {original_sliding_window} -> {config.sliding_window}")
 
-        outputs = model(**encoded)
-        hidden_states = outputs.last_hidden_state  # Shape: [batch_size, seq_len, hidden_size]
+        unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+        print(f"Using unreleased model: {unreleased_model_name}")
+        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}")
 
-        all_embeddings = hidden_states[0].cpu().numpy()  # Shape: [seq_len, hidden_size]
+            try:
+                model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+                model = model_class.from_pretrained(
+                    model_path,
+                    device_map=device_map,
+                    offload_folder="offload",
+                    trust_remote_code=True,
+                    config=config
+                )
+            except (ImportError, AttributeError) as e:
+                print(f"Failed to import or load model: {e}")
+                sys.exit(1)
+        else:
+            model = AutoModel.from_pretrained(
+                model_path,
+                device_map=device_map,
+                offload_folder="offload",
+                trust_remote_code=True,
+                config=config
+            )
+        print(f"Model class: {type(model)}")
+        print(f"Model file: {type(model).__module__}")
 
-        print(f"Hidden states shape: {hidden_states.shape}")
-        print(f"All embeddings shape: {all_embeddings.shape}")
-        print(f"Embedding dimension: {all_embeddings.shape[1]}")
+        # Verify the model is using the correct sliding window
+        if hasattr(model.config, 'sliding_window'):  # type: ignore
+            print(f"Model's sliding_window: {model.config.sliding_window}")  # type: ignore
+        else:
+            print("Model config does not have sliding_window attribute")
 
-    if len(all_embeddings.shape) == 1:
-        n_embd = all_embeddings.shape[0]  # type: ignore
-        n_embd_count = 1
-        all_embeddings = all_embeddings.reshape(1, -1)
+    return model, tokenizer, config
+
+
+def get_prompt(args):
+    if args.prompts_file:
+        try:
+            with open(args.prompts_file, 'r', encoding='utf-8') as f:
+                return f.read().strip()
+        except FileNotFoundError:
+            print(f"Error: Prompts file '{args.prompts_file}' not found")
+            sys.exit(1)
+        except Exception as e:
+            print(f"Error reading prompts file: {e}")
+            sys.exit(1)
     else:
-        n_embd = all_embeddings.shape[1]  # type: ignore
-        n_embd_count = all_embeddings.shape[0]  # type: ignore
+        return "Hello world today"
 
-    print()
 
-    for j in range(n_embd_count):
-        embedding = all_embeddings[j]
-        print(f"embedding {j}: ", end="")
+def main():
+    args = parse_arguments()
 
-        # Print first 3 values
-        for i in range(min(3, n_embd)):
-            print(f"{embedding[i]:9.6f} ", end="")
+    model_path = os.environ.get('EMBEDDING_MODEL_PATH', args.model_path)
+    if model_path is None:
+        print("Error: Model path must be specified either via --model-path argument "
+              "or EMBEDDING_MODEL_PATH environment variable")
+        sys.exit(1)
 
-        print(" ... ", end="")
+    # Determine if we should use SentenceTransformer
+    use_st = (
+        args.use_sentence_transformers or os.environ.get('USE_SENTENCE_TRANSFORMERS', '').lower() in ('1', 'true', 'yes')
+    )
 
-        # Print last 3 values
-        for i in range(n_embd - 3, n_embd):
-            print(f"{embedding[i]:9.6f} ", end="")
+    model, tokenizer, config = load_model_and_tokenizer(model_path, use_st, args.device)
 
-        print()  # New line
+    # Get the device the model is on
+    if not use_st:
+        device = next(model.parameters()).device
+    else:
+        # For SentenceTransformer, get device from the underlying model
+        device = next(model[0].auto_model.parameters()).device  # type: ignore
 
-    print()
+    model_name = os.path.basename(model_path)
 
-    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"
+    prompt_text = get_prompt(args)
+    texts = [prompt_text]
 
-    flattened_embeddings = all_embeddings.flatten()
-    flattened_embeddings.astype(np.float32).tofile(bin_filename)
+    with torch.no_grad():
+        if use_st:
+            embeddings = model.encode(texts, convert_to_numpy=True)
+            all_embeddings = embeddings  # Shape: [batch_size, hidden_size]
+
+            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}'")
+
+            print(f"Embeddings shape (after all SentenceTransformer layers): {all_embeddings.shape}")
+            print(f"Embedding dimension: {all_embeddings.shape[1] if len(all_embeddings.shape) > 1 else all_embeddings.shape[0]}")  # type: ignore
+        else:
+            # Standard approach: use base model output only
+            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}'")
+
+            # Move inputs to the same device as the model
+            encoded = {k: v.to(device) for k, v in encoded.items()}
+            outputs = model(**encoded)
+            hidden_states = outputs.last_hidden_state  # Shape: [batch_size, seq_len, hidden_size]
+
+            all_embeddings = hidden_states[0].float().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]}")
+
+        if len(all_embeddings.shape) == 1:
+            n_embd = all_embeddings.shape[0]  # type: ignore
+            n_embd_count = 1
+            all_embeddings = all_embeddings.reshape(1, -1)
+        else:
+            n_embd = all_embeddings.shape[1]  # type: ignore
+            n_embd_count = all_embeddings.shape[0]  # type: ignore
+
+        print()
 
-    with open(txt_filename, "w") as f:
-        idx = 0
         for j in range(n_embd_count):
-            for value in all_embeddings[j]:
-                f.write(f"{idx}: {value:.6f}\n")
-                idx += 1
-    print(f"Total values: {len(flattened_embeddings)} ({n_embd_count} embeddings × {n_embd} dimensions)")
-    print("")
-    print(f"Saved bin embeddings to: {bin_filename}")
-    print(f"Saved txt embeddings to: {txt_filename}")
+            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()
+
+        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"
+
+        flattened_embeddings = all_embeddings.flatten()
+        flattened_embeddings.astype(np.float32).tofile(bin_filename)
+
+        with open(txt_filename, "w") as f:
+            idx = 0
+            for j in range(n_embd_count):
+                for value in all_embeddings[j]:
+                    f.write(f"{idx}: {value:.6f}\n")
+                    idx += 1
+        print(f"Total values: {len(flattened_embeddings)} ({n_embd_count} embeddings × {n_embd} dimensions)")
+        print("")
+        print(f"Saved bin embeddings to: {bin_filename}")
+        print(f"Saved txt embeddings to: {txt_filename}")
+
+
+if __name__ == "__main__":
+    main()

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

@@ -166,7 +166,7 @@ def main():
     # 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)
+    config = AutoConfig.from_pretrained(args.model_path, trust_remote_code=True)
 
     if unreleased_model_name:
         model_name_lower = unreleased_model_name.lower()
@@ -186,9 +186,9 @@ def main():
             exit(1)
     else:
         if args.causal:
-            model = AutoModelForCausalLM.from_pretrained(args.model_path)
+            model = AutoModelForCausalLM.from_pretrained(args.model_path, trust_remote_code=True)
         else:
-            model = AutoModel.from_pretrained(args.model_path)
+            model = AutoModel.from_pretrained(args.model_path, trust_remote_code=True)
 
     encoded = tokenizer(prompt, return_tensors="pt")
     tokens = tokenizer.convert_ids_to_tokens(encoded['input_ids'][0])

examples/retrieval/retrieval.cpp

@@ -222,8 +222,8 @@ int main(int argc, char ** argv) {
     float * emb = embeddings.data();
 
     // break into batches
-    int p = 0; // number of prompts processed already
-    int s = 0; // number of prompts in current batch
+    unsigned int p = 0; // number of prompts processed already
+    unsigned int s = 0; // number of prompts in current batch
     for (int k = 0; k < n_chunks; k++) {
         // clamp to n_batch tokens
         auto & inp = chunks[k].tokens;
@@ -231,7 +231,7 @@ int main(int argc, char ** argv) {
         const uint64_t n_toks = inp.size();
 
         // encode if at capacity
-        if (batch.n_tokens + n_toks > n_batch) {
+        if (batch.n_tokens + n_toks > n_batch || s >= llama_n_seq_max(ctx)) {
             float * out = emb + p * n_embd;
             batch_process(ctx, batch, out, s, n_embd);
             common_batch_clear(batch);

examples/sycl/run-llama2.sh

@@ -22,9 +22,9 @@ if [ $# -gt 0 ]; then
     GGML_SYCL_DEVICE=$1
     echo "use $GGML_SYCL_DEVICE as main GPU"
     #use signle GPU only
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none
 
 else
     #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT}
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT}
 fi

examples/sycl/run-llama3.sh

@@ -24,8 +24,8 @@ export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
 if [ $# -gt 0 ]; then
     GGML_SYCL_DEVICE=$1
     echo "Using $GGML_SYCL_DEVICE as the main GPU"
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none
 else
     #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT}
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT}
 fi

examples/sycl/win-run-llama2.bat

@@ -8,4 +8,4 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 :: support malloc device memory more than 4GB.
 set UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
 
-.\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 99 -s 0
+.\build\bin\llama-completion.exe -m models\llama-2-7b.Q4_0.gguf -no-cnv -p %INPUT2% -n 400 -e -ngl 99 -s 0

examples/sycl/win-run-llama3.bat

@@ -8,4 +8,4 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 :: support malloc device memory more than 4GB.
 set UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
 
-.\build\bin\llama-cli.exe -m models\Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf -p %INPUT2% -n 400 -s 0 -e -ngl 99
+.\build\bin\llama-completion.exe -m models\Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf -no-cnv -p %INPUT2% -n 400 -s 0 -e -ngl 99

ggml/CMakeLists.txt

@@ -430,10 +430,22 @@ if (MSVC)
     configure_msvc_target(ggml-cpu-x64)
     configure_msvc_target(ggml-cpu-sse42)
     configure_msvc_target(ggml-cpu-sandybridge)
+    # __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+    # skipping            ggml-cpu-ivybridge
+    # skipping            ggml-cpu-piledriver
     configure_msvc_target(ggml-cpu-haswell)
     configure_msvc_target(ggml-cpu-skylakex)
+    configure_msvc_target(ggml-cpu-cannonlake)
+    configure_msvc_target(ggml-cpu-cascadelake)
     configure_msvc_target(ggml-cpu-icelake)
+    # MSVC 2022 doesn't support BF16 intrinsics without `/arch:AVX10.1` ?!
+    # https://learn.microsoft.com/en-us/cpp/intrinsics/x64-amd64-intrinsics-list?view=msvc-170
+    # https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170
+    # skipping            ggml-cpu-cooperlake
+    # skipping            ggml-cpu-zen4
     configure_msvc_target(ggml-cpu-alderlake)
+    # MSVC doesn't support AMX
+    # skipping            ggml-cpu-sapphirerapids
 
     if (GGML_BUILD_EXAMPLES)
         configure_msvc_target(common-ggml)

ggml/src/CMakeLists.txt

@@ -357,15 +357,29 @@ if (GGML_CPU_ALL_VARIANTS)
     endif()
     if (GGML_SYSTEM_ARCH STREQUAL "x86")
         ggml_add_cpu_backend_variant(x64)
-        ggml_add_cpu_backend_variant(sse42        SSE42)
-        ggml_add_cpu_backend_variant(sandybridge  SSE42 AVX)
-        ggml_add_cpu_backend_variant(haswell      SSE42 AVX F16C AVX2 BMI2 FMA)
-        ggml_add_cpu_backend_variant(skylakex     SSE42 AVX F16C AVX2 BMI2 FMA AVX512)
-        ggml_add_cpu_backend_variant(icelake      SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
-        ggml_add_cpu_backend_variant(alderlake    SSE42 AVX F16C AVX2 BMI2 FMA AVX_VNNI)
+        ggml_add_cpu_backend_variant(sse42              SSE42)
+        ggml_add_cpu_backend_variant(sandybridge        SSE42 AVX)
+        if (NOT MSVC)
+            # __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+            ggml_add_cpu_backend_variant(ivybridge      SSE42 AVX F16C)
+            ggml_add_cpu_backend_variant(piledriver     SSE42 AVX F16C FMA)
+        endif()
+        ggml_add_cpu_backend_variant(haswell            SSE42 AVX F16C FMA AVX2 BMI2)
+        ggml_add_cpu_backend_variant(skylakex           SSE42 AVX F16C FMA AVX2 BMI2 AVX512)
+        ggml_add_cpu_backend_variant(cannonlake         SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VBMI)
+        ggml_add_cpu_backend_variant(cascadelake        SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VNNI)
+        ggml_add_cpu_backend_variant(icelake            SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VBMI AVX512_VNNI)
+        if (NOT MSVC)
+            # MSVC 2022 doesn't support BF16 intrinsics without `/arch:AVX10.1` ?!
+            # https://learn.microsoft.com/en-us/cpp/intrinsics/x64-amd64-intrinsics-list?view=msvc-170
+            # https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170
+            ggml_add_cpu_backend_variant(cooperlake     SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VNNI AVX512_BF16)
+            ggml_add_cpu_backend_variant(zen4           SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16)
+        endif()
+        ggml_add_cpu_backend_variant(alderlake          SSE42 AVX F16C FMA AVX2 BMI2 AVX_VNNI)
         if (NOT MSVC)
             # MSVC doesn't support AMX
-            ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
+            ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
         endif()
     elseif(GGML_SYSTEM_ARCH STREQUAL "ARM")
         if (CMAKE_SYSTEM_NAME MATCHES "Linux")
@@ -387,8 +401,8 @@ if (GGML_CPU_ALL_VARIANTS)
             ggml_add_cpu_backend_variant(android_armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
             ggml_add_cpu_backend_variant(android_armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC MATMUL_INT8)
             ggml_add_cpu_backend_variant(android_armv9.0_1    DOTPROD MATMUL_INT8 FP16_VECTOR_ARITHMETIC SVE2)
-            ggml_add_cpu_backend_variant(android_armv9.2_1    DOTPROD MATMUL_INT8 FP16_VECTOR_ARITHMETIC SME)
-            ggml_add_cpu_backend_variant(android_armv9.2_2    DOTPROD MATMUL_INT8 FP16_VECTOR_ARITHMETIC SVE SME)
+            ggml_add_cpu_backend_variant(android_armv9.2_1    DOTPROD MATMUL_INT8 FP16_VECTOR_ARITHMETIC SVE SME)
+            ggml_add_cpu_backend_variant(android_armv9.2_2    DOTPROD MATMUL_INT8 FP16_VECTOR_ARITHMETIC SVE SVE2 SME)
         elseif (APPLE)
             ggml_add_cpu_backend_variant(apple_m1             DOTPROD)
             ggml_add_cpu_backend_variant(apple_m2_m3          DOTPROD MATMUL_INT8)

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -2338,19 +2338,19 @@ static void aclnn_rope_cache_init(ggml_backend_cann_context & ctx,
     // Step1.2: prepare rope_yarn_ramp, if this part updated, should update theta_scale_tensor.
     // TODO: acl_yarn_ramp_tensor use rope cache.
     bool                 yarn_ramp_tensor_updated = false;
-    ggml_cann_pool_alloc yarn_ramp_allocator(ctx.pool());
     acl_tensor_ptr       acl_yarn_ramp_tensor;
     if (ext_factor != 0 && (theta_scale_updated || ctx.rope_cache.theta_scale_length != theta_scale_length ||
                             ctx.rope_cache.freq_scale != freq_scale)) {
         yarn_ramp_tensor_updated = true;
-
+        if (ctx.rope_cache.yarn_ramp_cache != nullptr) {
+            ACL_CHECK(aclrtFree(ctx.rope_cache.yarn_ramp_cache));
+        }
+        ACL_CHECK(aclrtMalloc(&ctx.rope_cache.yarn_ramp_cache, theta_scale_length * sizeof(float), ACL_MEM_MALLOC_HUGE_FIRST));
         // -rope_yarn_ramp
         // const float y = (i0 / 2 - low) / MAX(0.001f, high - low);
         // return MIN(1, MAX(0, y)) - 1;
-        yarn_ramp_allocator.alloc(theta_scale_length * sizeof(float));
-        void * yarn_ramp_buffer = yarn_ramp_allocator.get();
         acl_yarn_ramp_tensor =
-            ggml_cann_create_tensor(yarn_ramp_buffer, ACL_FLOAT, sizeof(float), theta_scale_ne, theta_scale_nb, 1);
+            ggml_cann_create_tensor(ctx.rope_cache.yarn_ramp_cache, ACL_FLOAT, sizeof(float), theta_scale_ne, theta_scale_nb, 1);
         float          zero_value = 0, one_value = 1;
         float          denom_safe_value = MAX(0.001f, corr_dims[1] - corr_dims[0]);
         acl_scalar_ptr low              = ggml_cann_create_scalar(&corr_dims[0], aclDataType::ACL_FLOAT);
@@ -2380,8 +2380,10 @@ static void aclnn_rope_cache_init(ggml_backend_cann_context & ctx,
         acl_scalar_ptr freq_scale_1_sc = ggml_cann_create_scalar(&freq_scale_1, aclDataType::ACL_FLOAT);
         GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMuls, acl_yarn_ramp_tensor.get(), freq_scale_1_sc.get());
         GGML_CANN_CALL_ACLNN_OP(ctx, InplaceAdds, acl_yarn_ramp_tensor.get(), freq_scale_sc.get(), one.get());
+    } else {
+        acl_yarn_ramp_tensor =
+            ggml_cann_create_tensor(ctx.rope_cache.yarn_ramp_cache, ACL_FLOAT, sizeof(float), theta_scale_ne, theta_scale_nb, 1);
     }
-
     // Step 1.3: update theta_scale_tensor according to ext_factor or freq_scale.
     if (ext_factor != 0) {
         if (theta_scale_updated || yarn_ramp_tensor_updated) {
@@ -2988,32 +2990,156 @@ void ggml_cann_argmax(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     GGML_CANN_CALL_ACLNN_OP(ctx, ArgMax, acl_src.get(), 3, false, acl_dst.get());
 }
 
-void ggml_cann_conv_transpose_1d(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst){
     ggml_tensor * src0 = dst->src[0];
     ggml_tensor * src1 = dst->src[1];
 
     // stride
-    int64_t s0 = ((const int32_t *) (dst->op_params))[0];
+    int64_t s0 = ((const int32_t*)(dst->op_params))[0];
 
-    acl_tensor_ptr acl_input  = ggml_cann_create_tensor(src1, src1->ne, src1->nb, 3, ACL_FORMAT_NCL);
+    acl_tensor_ptr acl_input = ggml_cann_create_tensor(src1, src1->ne, src1->nb, 3, ACL_FORMAT_NCL);
     acl_tensor_ptr acl_weight = ggml_cann_create_tensor(src0, src0->ne, src0->nb, 3, ACL_FORMAT_NCL);
-    acl_tensor_ptr acl_dst    = ggml_cann_create_tensor(dst, dst->ne, dst->nb, 3, ACL_FORMAT_NCL);
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst, dst->ne, dst->nb, 3, ACL_FORMAT_NCL);
+
+    // get base information of input and kernel
+    int64_t input_len = *(src1->ne);
+    int64_t dst_len = *(dst->ne);
+    int64_t kernel_size = *(src0->ne);
+
+    // set the max kernel size for each conv
+    int64_t max_kernel_size = 255;
+
+    // compute the partition of kernel
+    int64_t part_num = 1;
+    part_num = (kernel_size + max_kernel_size - 1) / max_kernel_size;
 
     int64_t strideVal[1];
-    strideVal[0]                    = s0;
-    acl_int_array_ptr stride        = ggml_cann_create_int_array(strideVal, 1);
-    int64_t           paddingVal[]  = { 0 };
-    acl_int_array_ptr padding       = ggml_cann_create_int_array(paddingVal, 1);
-    int64_t           dilationVal[] = { 1 };
-    acl_int_array_ptr dilation      = ggml_cann_create_int_array(dilationVal, 1);
-    int8_t            cubeMathType  = 0;
+    strideVal[0] = s0;
+    acl_int_array_ptr stride = ggml_cann_create_int_array(strideVal, 1);
+    int64_t paddingVal[] = {0};
+    acl_int_array_ptr padding = ggml_cann_create_int_array(paddingVal, 1);
+    int64_t dilationVal[] = {1};
+    acl_int_array_ptr dilation = ggml_cann_create_int_array(dilationVal, 1);
+    bool transposed = true;
+    int64_t groups = 1;
+    int8_t cubeMathType = 0;
 
 #ifdef ASCEND_310P
     cubeMathType = 1;
 #endif
 
-    GGML_CANN_CALL_ACLNN_OP(ctx, Convolution, acl_input.get(), acl_weight.get(), nullptr, stride.get(), padding.get(),
-                            dilation.get(), true, padding.get(), 1, acl_dst.get(), cubeMathType);
+    auto weight_type = ggml_cann_type_mapping(src0->type);
+    auto dst_type = ggml_cann_type_mapping(dst->type);
+
+    // slice the kernel to make each conv available
+    int64_t slice_dim = -1;
+    int64_t slice_start = 0;
+    int64_t slice_end = max_kernel_size;
+    int64_t slice_step = 1;
+    int64_t interval = max_kernel_size;
+
+    int64_t left_pad_len = dilationVal[0] * (max_kernel_size - 1) + 1 - 2 * paddingVal[0];
+    int64_t right_pad_len = 0;
+
+    acl_scalar_ptr alpha = nullptr;
+    float alphaValue = 1.0;
+    alpha = ggml_cann_create_scalar(&alphaValue, aclDataType::ACL_FLOAT);
+
+    // set zero to destination
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, acl_dst.get());
+
+    for(int k = 0; k < part_num; k++){
+
+        // create part kernel tensor and slice from big kernel
+        slice_start = max_kernel_size * k;
+        if(k == part_num - 1){
+            slice_end = kernel_size;
+            interval = kernel_size - max_kernel_size * k;
+        }else{
+            slice_end = max_kernel_size * (k+1);
+        }
+
+        int64_t part_ne[4];
+        for(int i = 0; i < 4; i++) {
+            part_ne[i] = *(src0->ne + i);
+        }
+        part_ne[0] = interval;
+
+        size_t part_nb[4];
+        part_nb[0] = sizeof(weight_type);
+        for (int i = 1; i < 4; i++) {
+            part_nb[i] = part_nb[i - 1] * part_ne[i - 1];
+        }
+
+        ggml_cann_pool_alloc part_kernel_allocator;
+        part_kernel_allocator.alloc(ctx.pool(), part_nb[3]);
+        void* part_kernel_buf = part_kernel_allocator.get();
+
+        acl_tensor_ptr part_kernel = ggml_cann_create_tensor(part_kernel_buf, weight_type,
+                                ggml_element_size(src0), part_ne, part_nb, 3, ACL_FORMAT_NCL);
+
+        GGML_CANN_CALL_ACLNN_OP(ctx, Slice, acl_weight.get(), slice_dim, slice_start, slice_end, slice_step, part_kernel.get());
+
+        // create the part conv result tensor
+        int64_t part_dst_ne[4];
+        for(int i = 0; i < 4; i++){
+            part_dst_ne[i] = *(dst->ne + i);
+        }
+        part_dst_ne[0] = (input_len - 1) * strideVal[0] - 2 * paddingVal[0] + dilationVal[0] * (part_ne[0] - 1) + 1;
+
+        size_t part_dst_nb[4];
+        part_dst_nb[0] = sizeof(weight_type);
+        for (int i = 1; i < 4; i++) {
+            part_dst_nb[i] = part_dst_nb[i - 1] * part_dst_ne[i - 1];
+        }
+        ggml_cann_pool_alloc part_dst_allocator;
+        part_dst_allocator.alloc(ctx.pool(), part_dst_nb[3]);
+        void* part_dst_buf = part_dst_allocator.get();
+
+        acl_tensor_ptr acl_part_dst = ggml_cann_create_tensor(part_dst_buf, dst_type, ggml_element_size(dst),
+                                    part_dst_ne, part_dst_nb, 3, ACL_FORMAT_NCL);
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, acl_part_dst.get());
+
+        // compute part conv transpose 1d
+        GGML_CANN_CALL_ACLNN_OP(ctx, Convolution, acl_input.get(), part_kernel.get(), nullptr, stride.get(),
+        padding.get(), dilation.get(), transposed, padding.get(), groups, acl_part_dst.get(), cubeMathType);
+
+        // compute the position of part result in final result
+        int64_t global_start = slice_start;
+        int64_t global_end = std::min((input_len - 1) * strideVal[0] + slice_end, dst_len);
+
+        left_pad_len = global_start;
+        right_pad_len = dst_len - global_end;
+
+        std::vector<int64_t> padDataVal = {left_pad_len,right_pad_len};
+        acl_int_array_ptr padData = ggml_cann_create_int_array(padDataVal.data(), 2);
+
+        acl_scalar_ptr pad_value = nullptr;
+        float pad_valueVal = 0.0;
+        pad_value = ggml_cann_create_scalar(&pad_valueVal, aclDataType::ACL_FLOAT);
+
+        int64_t conv_result_ne[4];
+        for(int i = 0; i < 4; i++){
+            conv_result_ne[i] = *(dst->ne + i);
+        }
+
+        size_t conv_result_nb[4];
+        conv_result_nb[0] = sizeof(weight_type);
+        for (int i = 1; i < 4; i++) {
+            conv_result_nb[i] = conv_result_nb[i - 1] * conv_result_ne[i - 1];
+        }
+
+        ggml_cann_pool_alloc conv_result_allocator;
+        conv_result_allocator.alloc(ctx.pool(), conv_result_nb[3]);
+        void* conv_result_buf = conv_result_allocator.get();
+
+        acl_tensor_ptr conv_result = ggml_cann_create_tensor(conv_result_buf, dst_type, ggml_element_size(dst),
+                                    conv_result_ne, conv_result_nb, 3, ACL_FORMAT_NCL);
+
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, conv_result.get());
+        GGML_CANN_CALL_ACLNN_OP(ctx, ConstantPadNd, acl_part_dst.get(), padData.get(), pad_value.get(), conv_result.get());
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceAdd, acl_dst.get(), conv_result.get(), alpha.get());
+    }
 }
 
 void ggml_cann_elu(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
@@ -3576,3 +3702,106 @@ void ggml_cann_out_prod(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
             break;
     }
 }
+
+void ggml_cann_ssm_conv(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src0 = dst->src[0];  // conv_x
+    ggml_tensor * src1 = dst->src[1];  // conv1d.weight
+
+    // This op is currently defined only for F32 in ggml_cpu
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    // Shapes follow ggml_compute_forward_ssm_conv_f32
+    const int64_t nc  = src1->ne[0];   // d_conv
+    const int64_t ncs = src0->ne[0];   // d_conv - 1 + n_t
+    const int64_t nr  = src0->ne[1];   // d_inner
+    const int64_t n_s = src0->ne[2];   // n_seqs
+
+    const int64_t n_t = dst->ne[1];    // tokens per sequence
+
+    GGML_ASSERT(dst->ne[0] == nr);     // dst: {d_inner, n_t, n_s}
+    GGML_ASSERT(src1->ne[1] == nr);    // weight: {d_conv, d_inner}
+    GGML_ASSERT(ncs == nc - 1 + n_t);  // conv_x: {d_conv - 1 + n_t, d_inner, n_s}
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+    GGML_ASSERT(src1->nb[0] == sizeof(float));
+
+    // --- Build CANN tensors ---
+
+    // 1) Input: conv_x as NCL
+    //
+    // src0->ne = { ncs, nr, n_s, 1 }  // {L_in, C, N}
+    // Passing ACL_FORMAT_NCL here means:
+    //   reversed dims -> [N, C, L_in] = [n_s, nr, ncs]
+    acl_tensor_ptr acl_x = ggml_cann_create_tensor(src0, src0->ne, src0->nb, 3, ACL_FORMAT_NCL);
+
+    // 2) Weights: depthwise conv kernel, view src1 as {K, 1, C}
+    //
+    // src1 original:   ne = { nc, nr, 1, 1 }  // [K, C, 1, 1]
+    // we want a view:  ne_w = { nc, 1, nr }   // [K, 1, C]
+    // so that reversed dims -> [C, 1, K] which matches
+    //   [out_channels, in_channels/groups, kernel_size]
+    int64_t w_ne[GGML_MAX_DIMS] = { nc, 1, nr, 1 }; // [K, 1 input ch. per group, C groups]
+    // Layout: src1 data is [K, C] with
+    //   offset(k, c) = k*nb0 + c*nb1
+    // We want offset_w(k, 0, c) = k*nb0 + c*nb1,
+    // so we can reuse nb0 and nb1, and set nb2 = nb1.
+    size_t  w_nb[GGML_MAX_DIMS] = { src1->nb[0], src1->nb[1], src1->nb[1], src1->nb[3] }; // same as src1
+
+    acl_tensor_ptr acl_w = ggml_cann_create_tensor(
+        src1->data, ggml_cann_type_mapping(src1->type), ggml_type_size(src1->type), w_ne, w_nb, 3, ACL_FORMAT_NCL);
+
+    // 3) Output: dst is { d_inner, n_t, n_s } (CLN)
+    //
+    // We need an NCL view of the same buffer:
+    //   desired NCL logical shape: { L_out = n_t, C = nr, N = n_s }
+    //
+    // Original CLN layout:
+    //   dst->ne = { nr, n_t, n_s }
+    //   dst->nb[0] = sizeof(float)
+    //   dst->nb[1] = nr * sizeof(float)
+    //   dst->nb[2] = nr * n_t * sizeof(float)
+    //
+    // We want offset_new(L, C, N) = offset_orig(C, L, N).
+    // Choose:
+    //   nb_y[0] = nr * sizeof(float);           // step in L
+    //   nb_y[1] = sizeof(float);                // step in C
+    //   nb_y[2] = nr * n_t * sizeof(float);     // step in N
+    int64_t y_ne[GGML_MAX_DIMS] = { n_t, nr, n_s, 1 }; // [L_out, C, N]
+    size_t  y_nb[GGML_MAX_DIMS] = { dst->ne[0] * sizeof(float), sizeof(float), dst->ne[0] * dst->ne[1] * sizeof(float), dst->nb[3] }; // [nr, 1, nr * n_t]
+
+    acl_tensor_ptr acl_y = ggml_cann_create_tensor(
+        dst->data, ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type), y_ne, y_nb, 3, ACL_FORMAT_NCL);
+
+    // --- Conv1d parameters: depthwise, stride 1, no padding ("valid") ---
+    int64_t strideVal[1]   = { 1 };
+    int64_t paddingVal[1]  = { 0 };
+    int64_t dilationVal[1] = { 1 };
+
+    acl_int_array_ptr stride   = ggml_cann_create_int_array(strideVal, 1);
+    acl_int_array_ptr padding  = ggml_cann_create_int_array(paddingVal, 1);
+    acl_int_array_ptr dilation = ggml_cann_create_int_array(dilationVal, 1);
+
+    const bool    transposed   = false;
+    const int64_t groups       = nr;  // depthwise: one group per inner dim
+    int8_t        cubeMathType = 0;
+
+#ifdef ASCEND_310P
+    cubeMathType = 1;
+#endif
+
+    GGML_CANN_CALL_ACLNN_OP(ctx,
+                            Convolution,
+                            acl_x.get(),    // input:  N, C, L_in = ncs
+                            acl_w.get(),    // weight: [C, 1, K] with groups=nr
+                            nullptr,        // bias
+                            stride.get(),
+                            padding.get(),
+                            dilation.get(),
+                            transposed,
+                            padding.get(),   // output padding (unused for non-transposed)
+                            groups,
+                            acl_y.get(),
+                            cubeMathType);
+}
+

ggml/src/ggml-cann/aclnn_ops.h

@@ -47,6 +47,7 @@
 #include <aclnnop/aclnn_sign.h>
 #include <aclnnop/aclnn_silu.h>
 #include <aclnnop/aclnn_sin.h>
+#include <aclnnop/aclnn_slice.h>
 #include <aclnnop/aclnn_sqrt.h>
 #include <aclnnop/aclnn_tanh.h>
 
@@ -1032,6 +1033,8 @@ void ggml_cann_op_unary(std::function<void(ggml_backend_cann_context &, aclTenso
                         ggml_backend_cann_context &                                                ctx,
                         ggml_tensor *                                                              dst);
 
+void ggml_cann_ssm_conv(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
 /**
  * @brief Applies a gated (GLU-style) unary operation using the CANN backend.
  *

ggml/src/ggml-cann/common.h

@@ -229,6 +229,60 @@ struct ggml_graph_node_properties {
     // op
     ggml_op node_op;
     int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
+
+    /**
+     * @brief Check if a ggml tensor node matches this property set.
+     *
+     * This function compares all relevant fields (address, op type, shape, source inputs, op params)
+     * to determine whether the current node matches these previously recorded properties.
+     *
+     * @param node The current ggml tensor node.
+     * @return true if all fields match (excluding GGML_OP_VIEW); false otherwise.
+     */
+    bool has_matching_properties(ggml_tensor * node) {
+        if (node->data != this->node_address && node->op != GGML_OP_VIEW) {
+            return false;
+        }
+
+        if (node->op != this->node_op) {
+            return false;
+        }
+
+        for (int i = 0; i < GGML_MAX_DIMS; i++) {
+            if (node->ne[i] != this->ne[i]) {
+                return false;
+            }
+            if (node->nb[i] != this->nb[i]) {
+                return false;
+            }
+        }
+
+        for (int i = 0; i < GGML_MAX_SRC; i++) {
+            if (node->src[i]) {
+                if (node->src[i]->data != this->src_address[i] && node->op != GGML_OP_VIEW) {
+                    return false;
+                }
+
+                for (int d = 0; d < GGML_MAX_DIMS; d++) {
+                    if (node->src[i]->ne[d] != this->src_ne[i][d]) {
+                        return false;
+                    }
+                    if (node->src[i]->nb[d] != this->src_nb[i][d]) {
+                        return false;
+                    }
+                }
+            } else {
+                if (this->src_address[i] != nullptr) {
+                    return false;
+                }
+            }
+        }
+
+        if (node->op == GGML_OP_SCALE || node->op == GGML_OP_UNARY || node->op == GGML_OP_GLU) {
+            return memcmp(this->op_params, node->op_params, GGML_MAX_OP_PARAMS) == 0;
+        }
+        return true;
+    }
 };
 
 struct ggml_cann_graph {
@@ -241,6 +295,79 @@ struct ggml_cann_graph {
     aclmdlRI graph = nullptr;
 
     std::vector<ggml_graph_node_properties> ggml_graph_properties;
+
+    /**
+     * @brief Create a new CANN graph from a ggml computation graph.
+     *
+     * This function creates a new ggml_cann_graph object and fills its node properties
+     * (operation type, dimensions, strides, input sources, and operation parameters)
+     * based on the current ggml computation graph.
+     *
+     * Each node in the ggml graph is mapped to a property entry in the new CANN graph:
+     * - node address
+     * - operation type
+     * - shape (ne) and strides (nb)
+     * - source tensor addresses
+     * - operation parameters
+     *
+     * @param cgraph The current ggml computation graph.
+     * @return Pointer to the newly created ggml_cann_graph object.
+     */
+    static ggml_cann_graph * create_from_cgraph(ggml_cgraph * cgraph) {
+        ggml_cann_graph * new_graph = new ggml_cann_graph();
+        new_graph->ggml_graph_properties.resize(cgraph->n_nodes);
+
+        for (int node_idx = 0; node_idx < cgraph->n_nodes; ++node_idx) {
+            ggml_tensor * node = cgraph->nodes[node_idx];
+            auto &        prop = new_graph->ggml_graph_properties[node_idx];
+
+            prop.node_address = node->data;
+            prop.node_op      = node->op;
+
+            std::copy_n(node->ne, GGML_MAX_DIMS, prop.ne);
+            std::copy_n(node->nb, GGML_MAX_DIMS, prop.nb);
+
+            for (int src = 0; src < GGML_MAX_SRC; ++src) {
+                if (node->src[src]) {
+                    prop.src_address[src] = node->src[src]->data;
+                    std::copy_n(node->src[src]->ne, GGML_MAX_DIMS, prop.src_ne[src]);
+                    std::copy_n(node->src[src]->nb, GGML_MAX_DIMS, prop.src_nb[src]);
+                } else {
+                    prop.src_address[src] = nullptr;
+                    std::fill_n(prop.src_ne[src], GGML_MAX_DIMS, 0);
+                    std::fill_n(prop.src_nb[src], GGML_MAX_DIMS, 0);
+                }
+            }
+
+            memcpy(prop.op_params, node->op_params, GGML_MAX_OP_PARAMS);
+        }
+
+        return new_graph;
+    }
+
+    /**
+     * @brief Check whether this CANN graph matches the given ggml computation graph.
+     *
+     * This function compares the number of nodes and each node's properties
+     * (operation type, dimensions, strides, inputs, and operation parameters)
+     * to determine whether this CANN graph matches the given ggml graph.
+     *
+     * @param cgraph The current ggml computation graph.
+     * @return true if this CANN graph matches the ggml graph; false otherwise.
+     */
+    bool matches_cgraph(ggml_cgraph * cgraph) {
+        if (this->ggml_graph_properties.size() != static_cast<size_t>(cgraph->n_nodes)) {
+            return false;
+        }
+
+        for (int i = 0; i < cgraph->n_nodes; ++i) {
+            if (!this->ggml_graph_properties[i].has_matching_properties(cgraph->nodes[i])) {
+                return false;
+            }
+        }
+
+        return true;
+    }
 };
 
 /**
@@ -272,15 +399,6 @@ struct ggml_cann_graph_lru_cache {
         cache_list.push_front(new_node);
     }
 
-    /**
-     * @brief Move an existing graph to the front of the cache.
-     * @param node Pointer to the ggml_cann_graph to move.
-     */
-    void move_to_front(ggml_cann_graph * node) {
-        cache_list.remove(node);
-        cache_list.push_front(node);
-    }
-
     /**
      * @brief Clear all graphs from the cache (also frees memory).
      */
@@ -295,6 +413,28 @@ struct ggml_cann_graph_lru_cache {
      * @brief Destructor that clears the cache and frees all cached graphs.
      */
     ~ggml_cann_graph_lru_cache() { clear(); }
+
+    /**
+     * @brief Find a cached CANN graph that matches the given ggml graph and move it to front.
+     *
+     * This function iterates through the cached CANN graphs stored in the LRU cache and
+     * compares them against the given ggml computation graph. If a matching graph is found,
+     * it is promoted to the front of the LRU cache and returned. Otherwise, the function
+     * returns nullptr.
+     *
+     * @param cgraph The current ggml computation graph.
+     * @return true if found; false otherwise.
+     */
+    bool find_and_move_to_front(ggml_cgraph * cgraph) {
+        for (auto & graph_ptr : this->cache_list) {
+            if (graph_ptr->matches_cgraph(cgraph)) {
+                cache_list.remove(graph_ptr);
+                cache_list.push_front(graph_ptr);
+                return true;
+            }
+        }
+        return false;
+    }
 };
 #endif  // USE_ACL_GRAPH
 
@@ -318,6 +458,9 @@ struct ggml_cann_rope_cache {
         if (position_select_index_host) {
             free(position_select_index_host);
         }
+        if (yarn_ramp_cache) {
+            ACL_CHECK(aclrtFree(yarn_ramp_cache));
+        }
     }
 
     bool equal(int64_t theta_scale_length,
@@ -370,6 +513,7 @@ struct ggml_cann_rope_cache {
     float * theta_scale_exp_host       = nullptr;
     int *   position_select_index_host = nullptr;
     void *  position_select_index      = nullptr;
+    void *  yarn_ramp_cache            = nullptr;
     // sin/cos cache, used only to accelerate first layer on each device
     void *  sin_cache                  = nullptr;
     void *  cos_cache                  = nullptr;

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

@@ -1888,6 +1888,8 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context & ctx, struct gg
             break;
         case GGML_OP_OUT_PROD:
             ggml_cann_out_prod(ctx, dst);
+        case GGML_OP_SSM_CONV:
+            ggml_cann_ssm_conv(ctx, dst);
             break;
         default:
             return false;
@@ -2075,162 +2077,6 @@ static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
     ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
 }
 
-#ifdef USE_ACL_GRAPH
-/**
- * @brief Add a new CANN graph to the LRU cache by populating node properties from the ggml graph.
- *
- * This function creates a new ggml_cann_graph object and fills its node properties
- * (operation type, dimensions, strides, input sources, and operation parameters)
- * based on the current ggml computation graph.
- *
- * Each node in the ggml graph is mapped to a property entry in the new CANN graph:
- * - node address
- * - operation type
- * - shape (ne) and strides (nb)
- * - source tensor addresses
- * - operation parameters
- *
- * After initialization, the new graph is pushed into the LRU cache owned by the
- * CANN backend context. The cache takes ownership of the graph and manages its
- * lifetime (including deletion upon eviction).
- *
- * @param cann_ctx  The CANN backend context containing the graph cache.
- * @param cgraph    The current ggml computation graph.
- */
-static void add_lru_matched_graph_node_properties(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph) {
-    // Create a new ggml_cann_graph object on the heap (its lifetime is managed by the cache).
-    ggml_cann_graph * new_graph = new ggml_cann_graph();
-    new_graph->ggml_graph_properties.resize(cgraph->n_nodes);
-
-    for (int node_idx = 0; node_idx < cgraph->n_nodes; ++node_idx) {
-        ggml_tensor * node = cgraph->nodes[node_idx];
-        auto &        prop = new_graph->ggml_graph_properties[node_idx];
-
-        prop.node_address = node->data;
-        prop.node_op      = node->op;
-
-        std::copy_n(node->ne, GGML_MAX_DIMS, prop.ne);
-        std::copy_n(node->nb, GGML_MAX_DIMS, prop.nb);
-
-        for (int src = 0; src < GGML_MAX_SRC; ++src) {
-            if (node->src[src]) {
-                prop.src_address[src] = node->src[src]->data;
-                std::copy_n(node->src[src]->ne, GGML_MAX_DIMS, prop.src_ne[src]);
-                std::copy_n(node->src[src]->nb, GGML_MAX_DIMS, prop.src_nb[src]);
-            } else {
-                prop.src_address[src] = nullptr;
-                std::fill_n(prop.src_ne[src], GGML_MAX_DIMS, 0);
-                std::fill_n(prop.src_nb[src], GGML_MAX_DIMS, 0);
-            }
-        }
-
-        memcpy(prop.op_params, node->op_params, GGML_MAX_OP_PARAMS);
-    }
-
-    // Insert into the LRU cache (cache takes ownership and will delete it when evicted).
-    cann_ctx->graph_lru_cache.push(new_graph);
-}
-
-/**
- * @brief Check if a ggml tensor node matches a previously captured CANN graph node.
- *
- * This function compares all relevant fields (address, op type, shape, source inputs, op params)
- * to determine whether the current node matches a previously recorded version.
- *
- * @param node                  The current ggml tensor node.
- * @param graph_node_properties The stored properties of a CANN graph node.
- * @return true if all fields match (excluding GGML_OP_VIEW); false otherwise.
- */
-static bool ggml_graph_node_has_matching_properties(ggml_tensor *                node,
-                                                    ggml_graph_node_properties * graph_node_properties) {
-    if (node->data != graph_node_properties->node_address && node->op != GGML_OP_VIEW) {
-        return false;
-    }
-
-    if (node->op != graph_node_properties->node_op) {
-        return false;
-    }
-
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        if (node->ne[i] != graph_node_properties->ne[i]) {
-            return false;
-        }
-        if (node->nb[i] != graph_node_properties->nb[i]) {
-            return false;
-        }
-    }
-
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        if (node->src[i]) {
-            if (node->src[i]->data != graph_node_properties->src_address[i] && node->op != GGML_OP_VIEW) {
-                return false;
-            }
-
-            for (int d = 0; d < GGML_MAX_DIMS; d++) {
-                if (node->src[i]->ne[d] != graph_node_properties->src_ne[i][d]) {
-                    return false;
-                }
-                if (node->src[i]->nb[d] != graph_node_properties->src_nb[i][d]) {
-                    return false;
-                }
-            }
-        } else {
-            if (graph_node_properties->src_address[i] != nullptr) {
-                return false;
-            }
-        }
-    }
-
-    if (node->op == GGML_OP_SCALE || node->op == GGML_OP_UNARY || node->op == GGML_OP_GLU) {
-        return memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) == 0;
-    }
-    return true;
-}
-
-/**
- * @brief Check whether there is a cached CANN graph that matches the current ggml graph.
- *
- * This function iterates through the cached CANN graphs stored in the LRU cache and
- * compares them against the given ggml computation graph. A match requires that the
- * number of nodes is the same and that each node’s properties (operation type,
- * dimensions, strides, inputs, and operation parameters) are identical.
- *
- * If a matching graph is found, it is promoted to the front of the LRU cache and the
- * function returns true. Otherwise, the function returns false, indicating that a new
- * CANN graph needs to be captured.
- *
- * @param cann_ctx  The CANN backend context containing the graph cache.
- * @param cgraph    The current ggml computation graph.
- * @return true if a matching cached graph exists; false otherwise.
- */
-static bool is_matched_graph(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph) {
-    ggml_cann_graph_lru_cache & lru_cache = cann_ctx->graph_lru_cache;
-    for (auto & graph_ptr : lru_cache.cache_list) {
-        // Skip graphs with a different number of nodes.
-        if (graph_ptr->ggml_graph_properties.size() != static_cast<size_t>(cgraph->n_nodes)) {
-            continue;
-        }
-
-        // Check if all nodes match.
-        bool all_match = true;
-        for (int i = 0; i < cgraph->n_nodes; ++i) {
-            if (!ggml_graph_node_has_matching_properties(cgraph->nodes[i], &graph_ptr->ggml_graph_properties[i])) {
-                all_match = false;
-                break;
-            }
-        }
-
-        if (all_match) {
-            // update cache_list && renturn graph_ptr
-            lru_cache.move_to_front(graph_ptr);
-            return true;
-        }
-    }
-
-    return false;
-}
-#endif  // USE_ACL_GRAPH
-
 /**
  * @brief Evaluate the computation graph and optionally capture or execute it using CANN graph API.
  *
@@ -2239,23 +2085,23 @@ static bool is_matched_graph(ggml_backend_cann_context * cann_ctx, ggml_cgraph *
  *
  * Otherwise, it falls back to op-by-op execution using the CANN compute kernel dispatcher.
  *
- * @param cann_ctx                 The CANN backend context.
- * @param cgraph                   The ggml computation graph.
- * @param use_cann_graph           Whether to use CANN graph execution.
- * @param cann_graph_update_required Whether graph capture is needed due to graph changes.
+ * @param cann_ctx                     The CANN backend context.
+ * @param cgraph                       The ggml computation graph.
+ * @param use_cann_graph               Whether to use CANN graph execution.
+ * @param cann_graph_capture_required  Whether graph capture is needed due to graph changes.
  */
 static void evaluate_and_capture_cann_graph(ggml_backend_cann_context * cann_ctx,
                                             ggml_cgraph *               cgraph,
-                                            bool &                      use_cann_graph,
-                                            bool &                      cann_graph_update_required) {
+                                            bool                        use_cann_graph,
+                                            bool                        cann_graph_capture_required) {
 #ifdef USE_ACL_GRAPH
-    if (use_cann_graph && cann_graph_update_required) {  // Begin CANN graph capture
+    if (use_cann_graph && cann_graph_capture_required) {  // Begin CANN graph capture
         ACL_CHECK(aclmdlRICaptureBegin(cann_ctx->stream(), ACL_MODEL_RI_CAPTURE_MODE_GLOBAL));
     }
 #endif  // USE_ACL_GRAPH
     // Only perform the graph execution if CANN graphs are not enabled, or we are capturing the graph.
     // With the use of CANN graphs, the execution will be performed by the graph launch.
-    if (!use_cann_graph || cann_graph_update_required) {
+    if (!use_cann_graph || cann_graph_capture_required) {
         for (int i = 0; i < cgraph->n_nodes; i++) {
             ggml_tensor * node = cgraph->nodes[i];
 
@@ -2274,9 +2120,10 @@ static void evaluate_and_capture_cann_graph(ggml_backend_cann_context * cann_ctx
 
 #ifdef USE_ACL_GRAPH
     if (use_cann_graph) {
+        GGML_ASSERT(!cann_ctx->graph_lru_cache.cache_list.empty());
         ggml_cann_graph * matched_graph = cann_ctx->graph_lru_cache.cache_list.front();
 
-        if (cann_graph_update_required) {  // End CANN graph capture
+        if (cann_graph_capture_required) {  // End CANN graph capture
             ACL_CHECK(aclmdlRICaptureEnd(cann_ctx->stream(), &matched_graph->graph));
         }
 
@@ -2306,7 +2153,7 @@ static enum ggml_status ggml_backend_cann_graph_compute(ggml_backend_t backend,
     // calculate rope cache for fist layer in current device.
     cann_ctx->rope_cache.cached = false;
 
-    bool cann_graph_update_required = false;
+    bool graph_capture_required = false;
 #ifdef USE_ACL_GRAPH
     bool use_cann_graph = true;
 
@@ -2331,16 +2178,17 @@ static enum ggml_status ggml_backend_cann_graph_compute(ggml_backend_t backend,
 
     if (use_cann_graph) {
         // If no matching graph is found, the graph needs to be recaptured.
-        cann_graph_update_required = !is_matched_graph(cann_ctx, cgraph);
-        if (cann_graph_update_required) {
+        graph_capture_required = !cann_ctx->graph_lru_cache.find_and_move_to_front(cgraph);
+        if (graph_capture_required) {
             // If no matching graph is found, add a new ACL graph.
-            add_lru_matched_graph_node_properties(cann_ctx, cgraph);
+            ggml_cann_graph * new_graph = ggml_cann_graph::create_from_cgraph(cgraph);
+            cann_ctx->graph_lru_cache.push(new_graph);
         }
     }
 #else
     bool use_cann_graph = false;
 #endif  // USE_ACL_GRAPH
-    evaluate_and_capture_cann_graph(cann_ctx, cgraph, use_cann_graph, cann_graph_update_required);
+    evaluate_and_capture_cann_graph(cann_ctx, cgraph, use_cann_graph, graph_capture_required);
 
     return GGML_STATUS_SUCCESS;
 }
@@ -2578,8 +2426,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
                 }
             }
         case GGML_OP_CONV_TRANSPOSE_1D:
-            // TODO: ((weightL - 1) * dilationW - padLeft)=1336 should not be larger than 255.
-            return (op->src[0]->ne[0] - 1) <= 255;
+            return true;
         case GGML_OP_SCALE:
             float bias;
             memcpy(&bias, (const float *) (op->op_params) + 1, sizeof(float));
@@ -2626,6 +2473,8 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
                 }
                 return true;
             }
+        case GGML_OP_SSM_CONV:
+            return true;
         default:
             return false;
     }

ggml/src/ggml-cpu/CMakeLists.txt

@@ -561,9 +561,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.14.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.16.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "45e110675d93f99f82c23a1afcca76bc")
+        set(KLEIDIAI_ARCHIVE_MD5  "0a9e9008adb6031f9e8cf70dff4a3321")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)
@@ -615,6 +615,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         string(FIND "${ARCH_FLAGS_TEMP}" "+dotprod" DOTPROD_ENABLED)
         string(FIND "${ARCH_FLAGS_TEMP}" "+i8mm" I8MM_ENABLED)
         string(FIND "${ARCH_FLAGS_TEMP}" "+sme" SME_ENABLED)
+        string(FIND "${ARCH_FLAGS_TEMP}" "+sve" SVE_ENABLED)
 
         set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS_TEMP})
 
@@ -659,6 +660,15 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2")
         endif()
 
+        if (NOT SVE_ENABLED MATCHES -1)
+            list(APPEND GGML_KLEIDIAI_SOURCES
+                ${KLEIDIAI_SRC}/kai/kai_common_sve_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm.c)
+        endif()
+
         set_source_files_properties(${GGML_KLEIDIAI_SOURCES} PROPERTIES COMPILE_OPTIONS "${PRIVATE_ARCH_FLAGS}")
         list(APPEND GGML_CPU_SOURCES ${GGML_KLEIDIAI_SOURCES})
     endif()

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

@@ -328,7 +328,7 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <intrin.h>
-#elif defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
+#elif defined(__SSE__) || defined(__SSE3__) || defined(__SSSE3__) || defined(__AVX__) || defined(__F16C__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX512BF16__)
 #include <immintrin.h>
 #endif
 

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

@@ -18,6 +18,8 @@
 #include "kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.h"
 #include "kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.h"
 #include "kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.h"
+#include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm.h"
+#include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.h"
 
 #include "kai_lhs_pack_bf16p2vlx2_f32_sme.h"
 #include "kai_lhs_quant_pack_qsi8d32p_f32.h"
@@ -69,9 +71,9 @@ static inline void kernel_run_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
 
 template<void(*Fn)(size_t,size_t,size_t,const void*,const void*,float*,size_t,size_t,float,float)>
 static inline void kernel_run_float_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
-                                     const void* lhs, const void* rhs, void* dst,
-                                     size_t dst_stride_row, size_t dst_stride_col,
-                                     float clamp_min, float clamp_max) {
+                                         const void* lhs, const void* rhs, void* dst,
+                                         size_t dst_stride_row, size_t dst_stride_col,
+                                         float clamp_min, float clamp_max) {
     Fn(m, n, k, lhs, rhs, static_cast<float*>(dst), dst_stride_row, dst_stride_col, clamp_min, clamp_max);
 }
 
@@ -152,8 +154,8 @@ static inline void rhs_pack_fn12(size_t num_groups, size_t n, size_t k, size_t n
 
 template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,const int8_t*,const float*,const float*,void*,size_t,const struct kai_rhs_pack_qsi8cx_params*)>
 static inline void rhs_pack_scale_fn12(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t /*bl*/,
-                                               size_t /*rhs_stride*/, const void* rhs, const void* bias, const void* scale,
-                                               void* rhs_packed, size_t extra_bytes, const void* params) {
+                                       size_t /*rhs_stride*/, const void* rhs, const void* bias, const void* scale,
+                                       void* rhs_packed, size_t extra_bytes, const void* params) {
     Fn(num_groups, n, k, nr, kr, sr,
        static_cast<const int8_t*>(rhs),
        static_cast<const float*>(bias),
@@ -524,6 +526,61 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
     },
 #endif
 #else
+#if defined(__ARM_FEATURE_SVE)
+    {
+        /* SVE i8mm GEMM */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm>,
+        },
+        /* .gemm_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+        },
+        /* SVE dotprod GEMV */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod>,
+        },
+        /* .gemv_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
+        },
+        /* .rhs_info = */ {
+            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
+            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+        },
+        /* .required_cpu       = */ CPU_FEATURE_SVE | CPU_FEATURE_I8MM | CPU_FEATURE_DOTPROD,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
+    },
+#endif
 #if defined(__ARM_FEATURE_MATMUL_INT8)
     {
         /* i8mm GEMM */
@@ -578,7 +635,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
         /* .rhs_type           = */ GGML_TYPE_Q4_0,
         /* .op_type            = */ GGML_TYPE_F32,
     },
-#endif
+#endif // __ARM_FEATURE_MATMUL_INT8
 #if defined(__ARM_FEATURE_DOTPROD)
     {
         /* DOTPROD GEMM */
@@ -811,26 +868,27 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
     ggml_kleidiai_kernels * kernel = nullptr;
 
     if (tensor->op == GGML_OP_MUL_MAT && tensor->src[0] != nullptr && tensor->src[1] != nullptr) {
-#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_DOTPROD) || defined(__ARM_FEATURE_MATMUL_INT8)
-        for (size_t i = 0; i < NELEMS(gemm_gemv_kernels) - 1; ++i) {
-            if ((cpu_features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu &&
-                gemm_gemv_kernels[i].lhs_type == tensor->src[1]->type &&
-                gemm_gemv_kernels[i].rhs_type == tensor->src[0]->type &&
-                gemm_gemv_kernels[i].op_type  == tensor->type) {
-                kernel = &gemm_gemv_kernels[i];
-                break;
-            }
-        }
-        if (!kernel) {
-            for (size_t i = 0; i < NELEMS(gemm_gemv_kernels_q8) - 1; ++i) {
-                if ((cpu_features & gemm_gemv_kernels_q8[i].required_cpu) == gemm_gemv_kernels_q8[i].required_cpu &&
-                    gemm_gemv_kernels_q8[i].lhs_type == tensor->src[1]->type &&
-                    gemm_gemv_kernels_q8[i].rhs_type == tensor->src[0]->type &&
-                    gemm_gemv_kernels_q8[i].op_type  == tensor->type) {
-                    kernel = &gemm_gemv_kernels_q8[i];
-                    break;
+#if defined(__ARM_FEATURE_SME)          ||  \
+    defined(__ARM_FEATURE_DOTPROD)      ||  \
+    defined(__ARM_FEATURE_MATMUL_INT8)  ||  \
+    defined(__ARM_FEATURE_SVE)
+        auto try_table = [&](auto & table) {
+            for (size_t i = 0; i < NELEMS(table) - 1; ++i) {
+                if ((cpu_features & table[i].required_cpu) == table[i].required_cpu &&
+                    table[i].lhs_type == tensor->src[1]->type &&
+                    table[i].rhs_type == tensor->src[0]->type &&
+                    table[i].op_type  == tensor->type) {
+                    kernel = &table[i];
+                    return true;
                 }
             }
+            return false;
+        };
+
+        if (tensor->src[0]->type == GGML_TYPE_Q8_0) {
+            try_table(gemm_gemv_kernels_q8);
+        } else {
+            try_table(gemm_gemv_kernels);
         }
 #else
     GGML_UNUSED(gemm_gemv_kernels);
@@ -845,7 +903,10 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
 ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features) {
     ggml_kleidiai_kernels * kernels = nullptr;
 
-#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_DOTPROD) || defined(__ARM_FEATURE_MATMUL_INT8)
+#if defined(__ARM_FEATURE_SME)          ||  \
+    defined(__ARM_FEATURE_DOTPROD)      ||  \
+    defined(__ARM_FEATURE_MATMUL_INT8)  ||  \
+    defined(__ARM_FEATURE_SVE)
     for (size_t i = 0; i < NELEMS(gemm_gemv_kernels) - 1; ++i) {
         if ((features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu) {
             kernels = &gemm_gemv_kernels[i];

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

@@ -46,13 +46,20 @@ struct ggml_kleidiai_context {
 } static ctx = { CPU_FEATURE_NONE, NULL, NULL };
 
 static const char* cpu_feature_to_string(cpu_feature f) {
-    switch (f) {
-        case CPU_FEATURE_NONE:    return "NONE";
-        case CPU_FEATURE_DOTPROD: return "DOTPROD";
-        case CPU_FEATURE_I8MM:    return "I8MM";
-        case CPU_FEATURE_SVE:     return "SVE";
-        case CPU_FEATURE_SME:     return "SME";
-        default:                  return "UNKNOWN";
+    if (f == CPU_FEATURE_NONE) {
+        return "NONE";
+    } else if ((f & CPU_FEATURE_SME) == CPU_FEATURE_SME) {
+        return "SME";
+    } else if ((f & CPU_FEATURE_SVE) == CPU_FEATURE_SVE) {
+        return "SVE";
+    }
+    else if ((f & CPU_FEATURE_I8MM) == CPU_FEATURE_I8MM) {
+        return "I8MM";
+    } else if ((f & CPU_FEATURE_DOTPROD) == CPU_FEATURE_DOTPROD) {
+        return "DOTPROD";
+    }
+    else {
+        return "UNKNOWN";
     }
 }
 
@@ -68,7 +75,7 @@ static void init_kleidiai_context(void) {
 
         ctx.features  = (ggml_cpu_has_dotprod()     ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) |
                         (ggml_cpu_has_matmul_int8() ? CPU_FEATURE_I8MM    : CPU_FEATURE_NONE) |
-                        (ggml_cpu_has_sve()         ? CPU_FEATURE_SVE     : CPU_FEATURE_NONE);
+                        ((ggml_cpu_has_sve() && ggml_cpu_get_sve_cnt() == QK8_0) ? CPU_FEATURE_SVE : CPU_FEATURE_NONE);
 
         if (env_var) {
             sme_enabled = atoi(env_var);

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

@@ -69,6 +69,10 @@
 #define VECTOR_REGISTERS 16
 #endif
 
+#if defined(__riscv_v_intrinsic)
+#define LMUL 4
+#endif
+
 #define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
 
 namespace {
@@ -175,6 +179,46 @@ inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
 }
 #endif
 
+#if defined(__riscv_zvfh)
+template <>
+inline vfloat32m1_t madd(vfloat16mf2_t a, vfloat16mf2_t b, vfloat32m1_t c) {
+    return __riscv_vfwmacc_vv_f32m1(c, a, b, __riscv_vsetvlmax_e32m1());
+}
+inline vfloat32m2_t madd(vfloat16m1_t a, vfloat16m1_t b, vfloat32m2_t c) {
+    return __riscv_vfwmacc_vv_f32m2(c, a, b, __riscv_vsetvlmax_e32m2());
+}
+inline vfloat32m4_t madd(vfloat16m2_t a, vfloat16m2_t b, vfloat32m4_t c) {
+    return __riscv_vfwmacc_vv_f32m4(c, a, b, __riscv_vsetvlmax_e32m4());
+}
+inline vfloat32m8_t madd(vfloat16m4_t a, vfloat16m4_t b, vfloat32m8_t c) {
+    return __riscv_vfwmacc_vv_f32m8(c, a, b, __riscv_vsetvlmax_e32m8());
+}
+inline vfloat32m1_t madd(vfloat32m1_t a, vfloat32m1_t b, vfloat32m1_t c) {
+    return __riscv_vfmacc_vv_f32m1(c, a, b, __riscv_vsetvlmax_e32m1());
+}
+inline vfloat32m2_t madd(vfloat32m2_t a, vfloat32m2_t b, vfloat32m2_t c) {
+    return __riscv_vfmacc_vv_f32m2(c, a, b, __riscv_vsetvlmax_e32m2());
+}
+inline vfloat32m4_t madd(vfloat32m4_t a, vfloat32m4_t b, vfloat32m4_t c) {
+    return __riscv_vfmacc_vv_f32m4(c, a, b, __riscv_vsetvlmax_e32m4());
+}
+inline vfloat32m8_t madd(vfloat32m8_t a, vfloat32m8_t b, vfloat32m8_t c) {
+    return __riscv_vfmacc_vv_f32m8(c, a, b, __riscv_vsetvlmax_e32m8());
+}
+#endif
+
+#if defined(__riscv_zvfbfwma)
+inline vfloat32m1_t madd(vbfloat16mf2_t a, vbfloat16mf2_t b, vfloat32m1_t c) {
+    return __riscv_vfwmaccbf16_vv_f32m1(c, a, b, __riscv_vsetvlmax_e32m1());
+}
+inline vfloat32m2_t madd(vbfloat16m1_t a, vbfloat16m1_t b, vfloat32m2_t c) {
+    return __riscv_vfwmaccbf16_vv_f32m2(c, a, b, __riscv_vsetvlmax_e32m2());
+}
+inline vfloat32m4_t madd(vbfloat16m2_t a, vbfloat16m2_t b, vfloat32m4_t c) {
+    return __riscv_vfwmaccbf16_vv_f32m4(c, a, b, __riscv_vsetvlmax_e32m4());
+}
+#endif
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // VECTORIZED HORIZONTAL SUM
 
@@ -227,6 +271,25 @@ inline float hsum(__m512 x) {
 }
 #endif // __AVX512F__
 
+#if defined(__riscv_zvfh)
+inline float hsum(vfloat32m1_t x) {
+    return __riscv_vfmv_f_s_f32m1_f32(
+        __riscv_vfredusum_vs_f32m1_f32m1(x, __riscv_vfmv_v_f_f32m1(0, 1), __riscv_vsetvlmax_e32m1()));
+}
+inline float hsum(vfloat32m2_t x) {
+    return __riscv_vfmv_f_s_f32m1_f32(
+        __riscv_vfredusum_vs_f32m2_f32m1(x, __riscv_vfmv_v_f_f32m1(0, 1), __riscv_vsetvlmax_e32m2()));
+}
+inline float hsum(vfloat32m4_t x) {
+    return __riscv_vfmv_f_s_f32m1_f32(
+        __riscv_vfredusum_vs_f32m4_f32m1(x, __riscv_vfmv_v_f_f32m1(0, 1), __riscv_vsetvlmax_e32m4()));
+}
+inline float hsum(vfloat32m8_t x) {
+    return __riscv_vfmv_f_s_f32m1_f32(
+        __riscv_vfredusum_vs_f32m8_f32m1(x, __riscv_vfmv_v_f_f32m1(0, 1), __riscv_vsetvlmax_e32m8()));
+}
+#endif
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // VECTORIZED MEMORY LOADING
 
@@ -315,6 +378,88 @@ template <> inline __m256bh load(const float *p) {
 }
 #endif
 
+#if defined(__riscv_zvfh)
+template <> inline vfloat16mf2_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16mf2(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16mf2());
+}
+template <> inline vfloat16m1_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16m1(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m1());
+}
+template <> inline vfloat16m2_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16m2(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m2());
+}
+template <> inline vfloat16m4_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16m4(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m4());
+}
+template <> inline vfloat32m1_t load(const float *p) {
+    return __riscv_vle32_v_f32m1(p, __riscv_vsetvlmax_e32m1());
+}
+template <> inline vfloat32m2_t load(const float *p) {
+    return __riscv_vle32_v_f32m2(p, __riscv_vsetvlmax_e32m2());
+}
+template <> inline vfloat32m4_t load(const float *p) {
+    return __riscv_vle32_v_f32m4(p, __riscv_vsetvlmax_e32m4());
+}
+template <> inline vfloat32m8_t load(const float *p) {
+    return __riscv_vle32_v_f32m8(p, __riscv_vsetvlmax_e32m8());
+}
+#endif
+
+#if defined(__riscv_zvfbfwma)
+template <> inline vbfloat16mf2_t load(const ggml_bf16_t *p) {
+    return __riscv_vle16_v_bf16mf2(reinterpret_cast<const __bf16*>(p), __riscv_vsetvlmax_e16mf2());
+}
+template <> inline vbfloat16m1_t load(const ggml_bf16_t *p) {
+    return __riscv_vle16_v_bf16m1(reinterpret_cast<const __bf16*>(p), __riscv_vsetvlmax_e16m1());
+}
+template <> inline vbfloat16m2_t load(const ggml_bf16_t *p) {
+    return __riscv_vle16_v_bf16m2(reinterpret_cast<const __bf16*>(p), __riscv_vsetvlmax_e16m2());
+}
+#endif
+
+#if defined(__riscv_zvfh)
+template <typename T> T set_zero();
+
+template <> inline vfloat16mf2_t set_zero() {
+    return __riscv_vfmv_v_f_f16mf2(0, __riscv_vsetvlmax_e16mf2());
+}
+template <> inline vfloat16m1_t set_zero() {
+    return __riscv_vfmv_v_f_f16m1(0, __riscv_vsetvlmax_e16m1());
+}
+template <> inline vfloat16m2_t set_zero() {
+    return __riscv_vfmv_v_f_f16m2(0, __riscv_vsetvlmax_e16m2());
+}
+template <> inline vfloat16m4_t set_zero() {
+    return __riscv_vfmv_v_f_f16m4(0, __riscv_vsetvlmax_e16m4());
+}
+template <> inline vfloat32m1_t set_zero() {
+    return __riscv_vfmv_v_f_f32m1(0.0f, __riscv_vsetvlmax_e32m1());
+}
+template <> inline vfloat32m2_t set_zero() {
+    return __riscv_vfmv_v_f_f32m2(0, __riscv_vsetvlmax_e32m2());
+}
+template <> inline vfloat32m4_t set_zero() {
+    return __riscv_vfmv_v_f_f32m4(0, __riscv_vsetvlmax_e32m4());
+}
+template <> inline vfloat32m8_t set_zero() {
+    return __riscv_vfmv_v_f_f32m8(0, __riscv_vsetvlmax_e32m8());
+}
+#endif
+
+#if defined(__riscv_v_intrinsic)
+template <typename T> size_t vlmax() {
+    if constexpr (std::is_same_v<T, vfloat16mf2_t>) { return  __riscv_vsetvlmax_e16mf2(); }
+    else if constexpr (std::is_same_v<T, vfloat16m1_t>) { return  __riscv_vsetvlmax_e16m1(); }
+    else if constexpr (std::is_same_v<T, vfloat16m2_t>) { return  __riscv_vsetvlmax_e16m2(); }
+    else if constexpr (std::is_same_v<T, vfloat16m4_t>) { return  __riscv_vsetvlmax_e16m4(); }
+    else if constexpr (std::is_same_v<T, vfloat32m1_t>) { return  __riscv_vsetvlmax_e32m1(); }
+    else if constexpr (std::is_same_v<T, vfloat32m2_t>) { return  __riscv_vsetvlmax_e32m2(); }
+    else if constexpr (std::is_same_v<T, vfloat32m4_t>) { return  __riscv_vsetvlmax_e32m4(); }
+    else if constexpr (std::is_same_v<T, vfloat32m8_t>) { return  __riscv_vsetvlmax_e32m8(); }
+    return 0;
+}
+#endif
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // FLOATING POINT MATRIX MULTIPLICATION
 
@@ -488,6 +633,573 @@ class tinyBLAS {
     const int64_t ldc;
 };
 
+#if defined(__riscv_v_intrinsic)
+template <typename D, typename V, typename TA, typename TB, typename TC>
+class tinyBLAS_RVV {
+  public:
+    tinyBLAS_RVV(const ggml_compute_params * params, int64_t k,
+             const TA *A, int64_t lda,
+             const TB *B, int64_t ldb,
+             TC *C, int64_t ldc)
+        : params(params), A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc) {
+    }
+
+    bool matmul(int64_t m, int64_t n) {
+        if (k % vlmax<V>() != 0) {
+            return false;
+        }
+
+#if LMUL == 1
+        if (m % 16 == 0 && (m/16 >= params->nth)) {
+            const int64_t SIZE_N = BLOCK_SIZE<6>(n);
+            mnpack<4, 6, 4>(m, n, SIZE_N, 12);
+            return true;
+        }
+        if (m % 8 == 0 ) {
+            const int64_t SIZE_N = BLOCK_SIZE<6>(n);
+            mnpack<4, 6, 2>(m, n, SIZE_N, 12);
+            return true;
+        }
+        if (m % 4 == 0) {
+            const int64_t SIZE_N = BLOCK_SIZE<6>(n);
+            mnpack<4, 6, 1>(m, n, SIZE_N, 12);
+            return true;
+        }
+#elif LMUL == 2
+        if (m % 16 == 0 && (m/16 >= params->nth)) {
+            const int64_t SIZE_N = BLOCK_SIZE<3>(n);
+            mnpack<4, 3, 4>(m, n, SIZE_N, 24);
+            return true;
+        }
+        if (m % 8 == 0 ) {
+            const int64_t SIZE_N = BLOCK_SIZE<3>(n);
+            mnpack<4, 3, 2>(m, n, SIZE_N, 24);
+            return true;
+        }
+        if (m % 4 == 0) {
+            const int64_t SIZE_N = BLOCK_SIZE<3>(n);
+            mnpack<4, 3, 1>(m, n, SIZE_N, 24);
+            return true;
+        }
+#else // LMUL = 4
+        if (m % 16 == 0 && (m/16 >= params->nth)) {
+            const int64_t SIZE_N = BLOCK_SIZE<2>(n);
+            mnpack<2, 2, 8>(m, n, SIZE_N, 36);
+            return true;
+        }
+        if (m % 8 == 0 ) {
+            const int64_t SIZE_N = BLOCK_SIZE<2>(n);
+            mnpack<2, 2, 4>(m, n, SIZE_N, 36);
+            return true;
+        }
+        if (m % 4 == 0) {
+            const int64_t SIZE_N = BLOCK_SIZE<2>(n);
+            mnpack<2, 2, 2>(m, n, SIZE_N, 36);
+            return true;
+        }
+#endif
+        return false;
+    }
+
+  private:
+    template<int RM, int RN, int BM>
+    inline void mnpack(int64_t m, int64_t n, int64_t SIZE_N, int64_t BN) {
+        if (SIZE_N == RN) {
+            return gemm<RM, RN, BM>(m, n, BN);
+        }
+        if constexpr (RN > 1) {
+            return mnpack<RM, RN-1, BM>(m, n, SIZE_N, BN);
+        } else {
+            GGML_LOG_ERROR("mnpack<%d, %d> bloc size not supported\n", RM, (int)SIZE_N);
+            GGML_ASSERT(false); // we have miss something.
+        }
+    }
+
+    inline void gemm_bloc_4x6(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+        D Cv02 = set_zero<D>();
+        D Cv03 = set_zero<D>();
+        D Cv10 = set_zero<D>();
+        D Cv11 = set_zero<D>();
+        D Cv12 = set_zero<D>();
+        D Cv13 = set_zero<D>();
+        D Cv20 = set_zero<D>();
+        D Cv21 = set_zero<D>();
+        D Cv22 = set_zero<D>();
+        D Cv23 = set_zero<D>();
+        D Cv30 = set_zero<D>();
+        D Cv31 = set_zero<D>();
+        D Cv32 = set_zero<D>();
+        D Cv33 = set_zero<D>();
+        D Cv40 = set_zero<D>();
+        D Cv41 = set_zero<D>();
+        D Cv42 = set_zero<D>();
+        D Cv43 = set_zero<D>();
+        D Cv50 = set_zero<D>();
+        D Cv51 = set_zero<D>();
+        D Cv52 = set_zero<D>();
+        D Cv53 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            V Bv1 = load<V>(B + ldb * (jj + 1) + l);
+            V Bv2 = load<V>(B + ldb * (jj + 2) + l);
+            V Bv3 = load<V>(B + ldb * (jj + 3) + l);
+            V Bv4 = load<V>(B + ldb * (jj + 4) + l);
+            V Bv5 = load<V>(B + ldb * (jj + 5) + l);
+
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv10 = madd(Av0, Bv1, Cv10);
+            Cv20 = madd(Av0, Bv2, Cv20);
+            Cv30 = madd(Av0, Bv3, Cv30);
+            Cv40 = madd(Av0, Bv4, Cv40);
+            Cv50 = madd(Av0, Bv5, Cv50);
+
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+            Cv01 = madd(Av1, Bv0, Cv01);
+            Cv11 = madd(Av1, Bv1, Cv11);
+            Cv21 = madd(Av1, Bv2, Cv21);
+            Cv31 = madd(Av1, Bv3, Cv31);
+            Cv41 = madd(Av1, Bv4, Cv41);
+            Cv51 = madd(Av1, Bv5, Cv51);
+
+            V Av2 = load<V>(A + lda * (ii + 2) + l);
+            Cv02 = madd(Av2, Bv0, Cv02);
+            Cv12 = madd(Av2, Bv1, Cv12);
+            Cv22 = madd(Av2, Bv2, Cv22);
+            Cv32 = madd(Av2, Bv3, Cv32);
+            Cv42 = madd(Av2, Bv4, Cv42);
+            Cv52 = madd(Av2, Bv5, Cv52);
+
+            V Av3 = load<V>(A + lda * (ii + 3) + l);
+            Cv03 = madd(Av3, Bv0, Cv03);
+            Cv13 = madd(Av3, Bv1, Cv13);
+            Cv23 = madd(Av3, Bv2, Cv23);
+            Cv33 = madd(Av3, Bv3, Cv33);
+            Cv43 = madd(Av3, Bv4, Cv43);
+            Cv53 = madd(Av3, Bv5, Cv53);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+        C[ldc * (jj + 0) + (ii + 2)] = hsum(Cv02);
+        C[ldc * (jj + 0) + (ii + 3)] = hsum(Cv03);
+        C[ldc * (jj + 1) + (ii + 0)] = hsum(Cv10);
+        C[ldc * (jj + 1) + (ii + 1)] = hsum(Cv11);
+        C[ldc * (jj + 1) + (ii + 2)] = hsum(Cv12);
+        C[ldc * (jj + 1) + (ii + 3)] = hsum(Cv13);
+        C[ldc * (jj + 2) + (ii + 0)] = hsum(Cv20);
+        C[ldc * (jj + 2) + (ii + 1)] = hsum(Cv21);
+        C[ldc * (jj + 2) + (ii + 2)] = hsum(Cv22);
+        C[ldc * (jj + 2) + (ii + 3)] = hsum(Cv23);
+        C[ldc * (jj + 3) + (ii + 0)] = hsum(Cv30);
+        C[ldc * (jj + 3) + (ii + 1)] = hsum(Cv31);
+        C[ldc * (jj + 3) + (ii + 2)] = hsum(Cv32);
+        C[ldc * (jj + 3) + (ii + 3)] = hsum(Cv33);
+        C[ldc * (jj + 4) + (ii + 0)] = hsum(Cv40);
+        C[ldc * (jj + 4) + (ii + 1)] = hsum(Cv41);
+        C[ldc * (jj + 4) + (ii + 2)] = hsum(Cv42);
+        C[ldc * (jj + 4) + (ii + 3)] = hsum(Cv43);
+        C[ldc * (jj + 5) + (ii + 0)] = hsum(Cv50);
+        C[ldc * (jj + 5) + (ii + 1)] = hsum(Cv51);
+        C[ldc * (jj + 5) + (ii + 2)] = hsum(Cv52);
+        C[ldc * (jj + 5) + (ii + 3)] = hsum(Cv53);
+    }
+
+    inline void gemm_bloc_4x5(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+        D Cv02 = set_zero<D>();
+        D Cv03 = set_zero<D>();
+        D Cv10 = set_zero<D>();
+        D Cv11 = set_zero<D>();
+        D Cv12 = set_zero<D>();
+        D Cv13 = set_zero<D>();
+        D Cv20 = set_zero<D>();
+        D Cv21 = set_zero<D>();
+        D Cv22 = set_zero<D>();
+        D Cv23 = set_zero<D>();
+        D Cv30 = set_zero<D>();
+        D Cv31 = set_zero<D>();
+        D Cv32 = set_zero<D>();
+        D Cv33 = set_zero<D>();
+        D Cv40 = set_zero<D>();
+        D Cv41 = set_zero<D>();
+        D Cv42 = set_zero<D>();
+        D Cv43 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            V Bv1 = load<V>(B + ldb * (jj + 1) + l);
+            V Bv2 = load<V>(B + ldb * (jj + 2) + l);
+            V Bv3 = load<V>(B + ldb * (jj + 3) + l);
+            V Bv4 = load<V>(B + ldb * (jj + 4) + l);
+
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv10 = madd(Av0, Bv1, Cv10);
+            Cv20 = madd(Av0, Bv2, Cv20);
+            Cv30 = madd(Av0, Bv3, Cv30);
+            Cv40 = madd(Av0, Bv4, Cv40);
+
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+            Cv01 = madd(Av1, Bv0, Cv01);
+            Cv11 = madd(Av1, Bv1, Cv11);
+            Cv21 = madd(Av1, Bv2, Cv21);
+            Cv31 = madd(Av1, Bv3, Cv31);
+            Cv41 = madd(Av1, Bv4, Cv41);
+
+            V Av2 = load<V>(A + lda * (ii + 2) + l);
+            Cv02 = madd(Av2, Bv0, Cv02);
+            Cv12 = madd(Av2, Bv1, Cv12);
+            Cv22 = madd(Av2, Bv2, Cv22);
+            Cv32 = madd(Av2, Bv3, Cv32);
+            Cv42 = madd(Av2, Bv4, Cv42);
+
+            V Av3 = load<V>(A + lda * (ii + 3) + l);
+            Cv03 = madd(Av3, Bv0, Cv03);
+            Cv13 = madd(Av3, Bv1, Cv13);
+            Cv23 = madd(Av3, Bv2, Cv23);
+            Cv33 = madd(Av3, Bv3, Cv33);
+            Cv43 = madd(Av3, Bv4, Cv43);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+        C[ldc * (jj + 0) + (ii + 2)] = hsum(Cv02);
+        C[ldc * (jj + 0) + (ii + 3)] = hsum(Cv03);
+        C[ldc * (jj + 1) + (ii + 0)] = hsum(Cv10);
+        C[ldc * (jj + 1) + (ii + 1)] = hsum(Cv11);
+        C[ldc * (jj + 1) + (ii + 2)] = hsum(Cv12);
+        C[ldc * (jj + 1) + (ii + 3)] = hsum(Cv13);
+        C[ldc * (jj + 2) + (ii + 0)] = hsum(Cv20);
+        C[ldc * (jj + 2) + (ii + 1)] = hsum(Cv21);
+        C[ldc * (jj + 2) + (ii + 2)] = hsum(Cv22);
+        C[ldc * (jj + 2) + (ii + 3)] = hsum(Cv23);
+        C[ldc * (jj + 3) + (ii + 0)] = hsum(Cv30);
+        C[ldc * (jj + 3) + (ii + 1)] = hsum(Cv31);
+        C[ldc * (jj + 3) + (ii + 2)] = hsum(Cv32);
+        C[ldc * (jj + 3) + (ii + 3)] = hsum(Cv33);
+        C[ldc * (jj + 4) + (ii + 0)] = hsum(Cv40);
+        C[ldc * (jj + 4) + (ii + 1)] = hsum(Cv41);
+        C[ldc * (jj + 4) + (ii + 2)] = hsum(Cv42);
+        C[ldc * (jj + 4) + (ii + 3)] = hsum(Cv43);
+    }
+
+    inline void gemm_bloc_4x4(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+        D Cv02 = set_zero<D>();
+        D Cv03 = set_zero<D>();
+        D Cv10 = set_zero<D>();
+        D Cv11 = set_zero<D>();
+        D Cv12 = set_zero<D>();
+        D Cv13 = set_zero<D>();
+        D Cv20 = set_zero<D>();
+        D Cv21 = set_zero<D>();
+        D Cv22 = set_zero<D>();
+        D Cv23 = set_zero<D>();
+        D Cv30 = set_zero<D>();
+        D Cv31 = set_zero<D>();
+        D Cv32 = set_zero<D>();
+        D Cv33 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+            V Av2 = load<V>(A + lda * (ii + 2) + l);
+            V Av3 = load<V>(A + lda * (ii + 3) + l);
+
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv01 = madd(Av1, Bv0, Cv01);
+            Cv02 = madd(Av2, Bv0, Cv02);
+            Cv03 = madd(Av3, Bv0, Cv03);
+
+            V Bv1 = load<V>(B + ldb * (jj + 1) + l);
+            Cv10 = madd(Av0, Bv1, Cv10);
+            Cv11 = madd(Av1, Bv1, Cv11);
+            Cv12 = madd(Av2, Bv1, Cv12);
+            Cv13 = madd(Av3, Bv1, Cv13);
+
+            V Bv2 = load<V>(B + ldb * (jj + 2) + l);
+            Cv20 = madd(Av0, Bv2, Cv20);
+            Cv21 = madd(Av1, Bv2, Cv21);
+            Cv22 = madd(Av2, Bv2, Cv22);
+            Cv23 = madd(Av3, Bv2, Cv23);
+
+            V Bv3 = load<V>(B + ldb * (jj + 3) + l);
+            Cv30 = madd(Av0, Bv3, Cv30);
+            Cv31 = madd(Av1, Bv3, Cv31);
+            Cv32 = madd(Av2, Bv3, Cv32);
+            Cv33 = madd(Av3, Bv3, Cv33);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+        C[ldc * (jj + 0) + (ii + 2)] = hsum(Cv02);
+        C[ldc * (jj + 0) + (ii + 3)] = hsum(Cv03);
+        C[ldc * (jj + 1) + (ii + 0)] = hsum(Cv10);
+        C[ldc * (jj + 1) + (ii + 1)] = hsum(Cv11);
+        C[ldc * (jj + 1) + (ii + 2)] = hsum(Cv12);
+        C[ldc * (jj + 1) + (ii + 3)] = hsum(Cv13);
+        C[ldc * (jj + 2) + (ii + 0)] = hsum(Cv20);
+        C[ldc * (jj + 2) + (ii + 1)] = hsum(Cv21);
+        C[ldc * (jj + 2) + (ii + 2)] = hsum(Cv22);
+        C[ldc * (jj + 2) + (ii + 3)] = hsum(Cv23);
+        C[ldc * (jj + 3) + (ii + 0)] = hsum(Cv30);
+        C[ldc * (jj + 3) + (ii + 1)] = hsum(Cv31);
+        C[ldc * (jj + 3) + (ii + 2)] = hsum(Cv32);
+        C[ldc * (jj + 3) + (ii + 3)] = hsum(Cv33);
+    }
+
+    inline void gemm_bloc_4x3(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+        D Cv02 = set_zero<D>();
+        D Cv03 = set_zero<D>();
+        D Cv10 = set_zero<D>();
+        D Cv11 = set_zero<D>();
+        D Cv12 = set_zero<D>();
+        D Cv13 = set_zero<D>();
+        D Cv20 = set_zero<D>();
+        D Cv21 = set_zero<D>();
+        D Cv22 = set_zero<D>();
+        D Cv23 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+            V Av2 = load<V>(A + lda * (ii + 2) + l);
+            V Av3 = load<V>(A + lda * (ii + 3) + l);
+
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv01 = madd(Av1, Bv0, Cv01);
+            Cv02 = madd(Av2, Bv0, Cv02);
+            Cv03 = madd(Av3, Bv0, Cv03);
+
+            V Bv1 = load<V>(B + ldb * (jj + 1) + l);
+            Cv10 = madd(Av0, Bv1, Cv10);
+            Cv11 = madd(Av1, Bv1, Cv11);
+            Cv12 = madd(Av2, Bv1, Cv12);
+            Cv13 = madd(Av3, Bv1, Cv13);
+
+            V Bv2 = load<V>(B + ldb * (jj + 2) + l);
+            Cv20 = madd(Av0, Bv2, Cv20);
+            Cv21 = madd(Av1, Bv2, Cv21);
+            Cv22 = madd(Av2, Bv2, Cv22);
+            Cv23 = madd(Av3, Bv2, Cv23);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+        C[ldc * (jj + 0) + (ii + 2)] = hsum(Cv02);
+        C[ldc * (jj + 0) + (ii + 3)] = hsum(Cv03);
+        C[ldc * (jj + 1) + (ii + 0)] = hsum(Cv10);
+        C[ldc * (jj + 1) + (ii + 1)] = hsum(Cv11);
+        C[ldc * (jj + 1) + (ii + 2)] = hsum(Cv12);
+        C[ldc * (jj + 1) + (ii + 3)] = hsum(Cv13);
+        C[ldc * (jj + 2) + (ii + 0)] = hsum(Cv20);
+        C[ldc * (jj + 2) + (ii + 1)] = hsum(Cv21);
+        C[ldc * (jj + 2) + (ii + 2)] = hsum(Cv22);
+        C[ldc * (jj + 2) + (ii + 3)] = hsum(Cv23);
+    }
+
+    inline void gemm_bloc_4x2(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+        D Cv02 = set_zero<D>();
+        D Cv03 = set_zero<D>();
+        D Cv10 = set_zero<D>();
+        D Cv11 = set_zero<D>();
+        D Cv12 = set_zero<D>();
+        D Cv13 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+            V Av2 = load<V>(A + lda * (ii + 2) + l);
+            V Av3 = load<V>(A + lda * (ii + 3) + l);
+
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv01 = madd(Av1, Bv0, Cv01);
+            Cv02 = madd(Av2, Bv0, Cv02);
+            Cv03 = madd(Av3, Bv0, Cv03);
+
+            V Bv1 = load<V>(B + ldb * (jj + 1) + l);
+            Cv10 = madd(Av0, Bv1, Cv10);
+            Cv11 = madd(Av1, Bv1, Cv11);
+            Cv12 = madd(Av2, Bv1, Cv12);
+            Cv13 = madd(Av3, Bv1, Cv13);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+        C[ldc * (jj + 0) + (ii + 2)] = hsum(Cv02);
+        C[ldc * (jj + 0) + (ii + 3)] = hsum(Cv03);
+        C[ldc * (jj + 1) + (ii + 0)] = hsum(Cv10);
+        C[ldc * (jj + 1) + (ii + 1)] = hsum(Cv11);
+        C[ldc * (jj + 1) + (ii + 2)] = hsum(Cv12);
+        C[ldc * (jj + 1) + (ii + 3)] = hsum(Cv13);
+    }
+
+    inline void gemm_bloc_4x1(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+        D Cv02 = set_zero<D>();
+        D Cv03 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+            V Av2 = load<V>(A + lda * (ii + 2) + l);
+            V Av3 = load<V>(A + lda * (ii + 3) + l);
+
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv01 = madd(Av1, Bv0, Cv01);
+            Cv02 = madd(Av2, Bv0, Cv02);
+            Cv03 = madd(Av3, Bv0, Cv03);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+        C[ldc * (jj + 0) + (ii + 2)] = hsum(Cv02);
+        C[ldc * (jj + 0) + (ii + 3)] = hsum(Cv03);
+    }
+
+    inline void gemm_bloc_2x2(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+        D Cv10 = set_zero<D>();
+        D Cv11 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv01 = madd(Av1, Bv0, Cv01);
+
+            V Bv1 = load<V>(B + ldb * (jj + 1) + l);
+            Cv10 = madd(Av0, Bv1, Cv10);
+            Cv11 = madd(Av1, Bv1, Cv11);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+        C[ldc * (jj + 1) + (ii + 0)] = hsum(Cv10);
+        C[ldc * (jj + 1) + (ii + 1)] = hsum(Cv11);
+    }
+
+    inline void gemm_bloc_2x1(int64_t ii, int64_t jj) {
+        size_t vl = vlmax<V>();
+        D Cv00 = set_zero<D>();
+        D Cv01 = set_zero<D>();
+
+        for (int64_t l = 0; l < k; l += vl) {
+            V Av0 = load<V>(A + lda * (ii + 0) + l);
+            V Av1 = load<V>(A + lda * (ii + 1) + l);
+
+            V Bv0 = load<V>(B + ldb * (jj + 0) + l);
+            Cv00 = madd(Av0, Bv0, Cv00);
+            Cv01 = madd(Av1, Bv0, Cv01);
+        }
+
+        C[ldc * (jj + 0) + (ii + 0)] = hsum(Cv00);
+        C[ldc * (jj + 0) + (ii + 1)] = hsum(Cv01);
+    }
+
+    template <int RM, int RN>
+    inline void gemm_bloc(int64_t ii, int64_t jj) {
+        if constexpr (RM == 4) {
+            if constexpr (RN == 6) { return gemm_bloc_4x6(ii, jj); }
+            if constexpr (RN == 5) { return gemm_bloc_4x5(ii, jj); }
+            if constexpr (RN == 4) { return gemm_bloc_4x4(ii, jj); }
+            if constexpr (RN == 3) { return gemm_bloc_4x3(ii, jj); }
+            if constexpr (RN == 2) { return gemm_bloc_4x2(ii, jj); }
+            if constexpr (RN == 1) { return gemm_bloc_4x1(ii, jj); }
+        } else if constexpr (RM == 2) {
+            if constexpr (RN == 2) { return gemm_bloc_2x2(ii, jj); }
+            if constexpr (RN == 1) { return gemm_bloc_2x1(ii, jj); }
+        }
+    }
+
+    template <int RM, int RN, int BM>
+    NOINLINE void gemm(int64_t m, int64_t n, int64_t BN) {
+        GGML_ASSERT(m % (RM * BM) == 0);
+        const int64_t ytiles = m / (RM * BM);
+        const int64_t xtiles = (n + RN -1) / RN;
+        const int64_t jj_RN = (xtiles - (xtiles * RN - n));
+
+        // "round" bloc_size to "nearest" BN
+        const int64_t NB_BN = xtiles < BN ? 1 : (xtiles + BN / 2) / BN;
+        const int64_t SIZE_BN = xtiles % NB_BN == 0 ? xtiles / NB_BN : xtiles / NB_BN + 1;
+        const int64_t jj_BN = (NB_BN - (NB_BN * SIZE_BN - xtiles));
+        const int64_t nb_job = ytiles * NB_BN;
+
+        if (params->ith == 0) {
+            GGML_ASSERT( jj_BN * SIZE_BN + (NB_BN - jj_BN) * (SIZE_BN - 1) == xtiles);
+            // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
+            ggml_threadpool_chunk_set(params->threadpool, params->nth);
+        }
+
+        ggml_barrier(params->threadpool);
+
+        int64_t job = params->ith;
+        while (job < nb_job) {
+            const int64_t ii = (job % ytiles) * RM * BM;
+            const int64_t jb =  job / ytiles;
+            const int64_t jr0 = BLOC_POS(jb  , jj_BN, SIZE_BN);
+            const int64_t jrN = BLOC_POS(jb+1, jj_BN, SIZE_BN);
+
+            const int64_t jj0 = BLOC_POS(jr0, jj_RN, RN);
+            const int64_t jj2 = BLOC_POS(jrN, jj_RN, RN);
+            const int64_t jj1 = jj2 < jj_RN * RN ? jj2 : jj_RN * RN;
+
+            for (int64_t bi = 0; bi < BM * RM; bi += RM) {
+                int64_t jj = jj0;
+                for (; jj < jj1; jj += RN) {
+                    gemm_bloc<RM, RN>(ii + bi, jj);
+                }
+                if constexpr (RN > 1) {
+                    for (; jj < jj2; jj += RN - 1) {
+                        gemm_bloc<RM, RN-1>(ii + bi, jj);
+                    }
+                }
+                GGML_ASSERT(jj == jj2);
+            }
+
+            job = ggml_threadpool_chunk_add(params->threadpool, 1);
+        }
+
+        ggml_barrier(params->threadpool);
+        return;
+    }
+
+    const ggml_compute_params * params;
+    const TA *const A;
+    const TB *const B;
+    TC *const C;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+};
+#endif
+
 //////////////////////////////////////////////////////////////////////////////////////////
 // QUANT ZERO MATRIX MULTIPLICATION
 
@@ -2657,6 +3369,24 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             params->ith, params->nth};
         tb.matmul(m, n);
         return true;
+#elif defined(__riscv_zvfh)
+    #if LMUL == 1
+        tinyBLAS_RVV<vfloat32m1_t, vfloat32m1_t, float, float, float> tb{ params,
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc};
+    #elif LMUL == 2
+        tinyBLAS_RVV<vfloat32m2_t, vfloat32m2_t, float, float, float> tb{ params,
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc};
+    #else // LMUL = 4
+        tinyBLAS_RVV<vfloat32m4_t, vfloat32m4_t, float, float, float> tb{ params,
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc};
+    #endif
+        return tb.matmul(m, n);
 #else
         return false;
 #endif
@@ -2699,6 +3429,24 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
         tb.matmul(m, n);
         return true;
         }
+#elif defined(__riscv_zvfbfwma)
+        #if LMUL == 1
+            tinyBLAS_RVV<vfloat32m1_t, vbfloat16mf2_t, ggml_bf16_t, ggml_bf16_t, float> tb{ params,
+                k, (const ggml_bf16_t *)A, lda,
+                (const ggml_bf16_t *)B, ldb,
+                (float *)C, ldc};
+        #elif LMUL == 2
+            tinyBLAS_RVV<vfloat32m2_t, vbfloat16m1_t, ggml_bf16_t, ggml_bf16_t, float> tb{ params,
+                k, (const ggml_bf16_t *)A, lda,
+                (const ggml_bf16_t *)B, ldb,
+                (float *)C, ldc};
+        #else // LMUL = 4
+            tinyBLAS_RVV<vfloat32m4_t, vbfloat16m2_t, ggml_bf16_t, ggml_bf16_t, float> tb{ params,
+                k, (const ggml_bf16_t *)A, lda,
+                (const ggml_bf16_t *)B, ldb,
+                (float *)C, ldc};
+        #endif
+            return tb.matmul(m, n);
 #endif
         return false;
     }
@@ -2748,6 +3496,26 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
                 (float *)C, ldc};
             return tb.matmul(m, n);
         }
+#elif defined(__riscv_zvfh)
+        if (Btype == GGML_TYPE_F16) {
+        #if LMUL == 1
+            tinyBLAS_RVV<vfloat32m1_t, vfloat16mf2_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
+                k, (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc};
+        #elif LMUL == 2
+            tinyBLAS_RVV<vfloat32m2_t, vfloat16m1_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
+                k, (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc};
+        #else // LMUL = 4
+            tinyBLAS_RVV<vfloat32m4_t, vfloat16m2_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
+                k, (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc};
+        #endif
+            return tb.matmul(m, n);
+        }
 #endif
         return false;
     }

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

@@ -14,10 +14,6 @@
 #include <arm_neon.h>
 #endif
 
-#if defined(__F16C__)
-#include <immintrin.h>
-#endif
-
 #if defined(__riscv_v_intrinsic)
 #include <riscv_vector.h>
 #endif

ggml/src/ggml-cuda/CMakeLists.txt

@@ -15,6 +15,7 @@ if (CUDAToolkit_FOUND)
         # 80     == Ampere, asynchronous data loading, faster tensor core instructions
         # 86     == RTX 3000, needs CUDA v11.1
         # 89     == RTX 4000, needs CUDA v11.8
+        # 120    == Blackwell, needs CUDA v12.8, FP4 tensor cores
         #
         # XX-virtual == compile CUDA code as PTX, do JIT compilation to binary code on first run
         # XX-real    == compile CUDA code as device code for this specific architecture
@@ -34,12 +35,52 @@ if (CUDAToolkit_FOUND)
             if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8")
                 list(APPEND CMAKE_CUDA_ARCHITECTURES 89-real)
             endif()
+
+            if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")
+                # The CUDA architecture 120f-virtual would in principle work for Blackwell support
+                #     but the newly added "f" suffix conflicted with a preexising regex for validating CUDA architectures in CMake.
+                # So either a recent CMake version or one with the backported fix is needed.
+                # The following versions should work:
+                #   - CMake >= v3.31.8 && CMake < v4.0.0
+                #   - CMake >= v4.0.2
+                # This is NOT documented in the CMake release notes,
+                #     check Modules/Internal/CMakeCUDAArchitecturesValidate.cmake in the CMake git repository instead.
+                # However, the architectures 120a-real and 121a-real should work with basically any CMake version and
+                #     until the release of e.g. Rubin there is no benefit to shipping virtual architectures for Blackwell.
+                list(APPEND CMAKE_CUDA_ARCHITECTURES 120a-real 121a-real)
+            endif()
         endif()
     endif()
-    message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
 
     enable_language(CUDA)
 
+    # Replace any plain 12X CUDA architectures with their "architecture-specific" equivalents 12Xa.
+    # 12X is forwards-compatible, 12Xa is not.
+    # Notably the Blackwell FP4 tensor core instructions are not forwards compatible and therefore need 12Xa.
+    # But while 12X vs. 12Xa can be checked in device code there is (to my knowledge) no easy way to do the same check in host code.
+    # So for now just replace all instances of 12X with 12Xa, this should be fine until Rubin is released.
+    foreach(ARCHS IN ITEMS CMAKE_CUDA_ARCHITECTURES CMAKE_CUDA_ARCHITECTURES_NATIVE)
+        set(FIXED_ARCHS "")
+        foreach(ARCH IN LISTS ${ARCHS})
+            if (ARCH MATCHES "^12[0-9](-real|-virtual)?$")
+                string(REGEX REPLACE "^(12[0-9])((-real|-virtual)?)$" "\\1a\\2" FIXED_ARCH ${ARCH})
+                message(STATUS "Replacing ${ARCH} in ${ARCHS} with ${FIXED_ARCH}")
+                list(APPEND FIXED_ARCHS "${FIXED_ARCH}")
+            else()
+                list(APPEND FIXED_ARCHS "${ARCH}")
+            endif()
+        endforeach()
+        set(${ARCHS} ${FIXED_ARCHS})
+    endforeach()
+
+    # If we try to compile a "native" build it will use the 12X architectures and fail.
+    # So we should instead use the native architectures as determined by CMake after replacing 12X with 12Xa.
+    # But if at the time of the build no GPUs are connected at all CMAKE_CUDA_ARCHITECTURES will contain garbage that we should not use.
+    if (CMAKE_CUDA_ARCHITECTURES STREQUAL "native" AND CMAKE_CUDA_ARCHITECTURES_NATIVE MATCHES "^[0-9]+(a|f)?(-real|-virtual)?(;[0-9]+(a|f)?(-real|-virtual)?|;)*$")
+        set(CMAKE_CUDA_ARCHITECTURES ${CMAKE_CUDA_ARCHITECTURES_NATIVE})
+    endif()
+    message(STATUS "Using CMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} CMAKE_CUDA_ARCHITECTURES_NATIVE=${CMAKE_CUDA_ARCHITECTURES_NATIVE}")
+
     file(GLOB   GGML_HEADERS_CUDA "*.cuh")
     list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")
 

ggml/src/ggml-cuda/common.cuh

@@ -50,6 +50,10 @@
 #define GGML_CUDA_CC_TURING          750
 #define GGML_CUDA_CC_AMPERE          800
 #define GGML_CUDA_CC_ADA_LOVELACE    890
+// While BW spans CC 1000, 1100 & 1200, we are integrating Tensor Core instructions available to 1200 family, see
+// https://docs.nvidia.com/cutlass/media/docs/cpp/blackwell_functionality.html#blackwell-sm120-gemms
+#define GGML_CUDA_CC_BLACKWELL       1200
+#define GGML_CUDA_CC_RUBIN           1300
 #define GGML_CUDA_CC_OFFSET_AMD      0x1000000
 #define GGML_CUDA_CC_OFFSET_MTHREADS 0x0100000
 #define GGML_CUDA_CC_IS_NVIDIA(cc)   (cc < GGML_CUDA_CC_OFFSET_MTHREADS)
@@ -246,6 +250,10 @@ static const char * cu_get_error_str(CUresult err) {
 #define AMPERE_MMA_AVAILABLE
 #endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_BLACKWELL && __CUDA_ARCH__ < GGML_CUDA_CC_RUBIN
+#    define BLACKWELL_MMA_AVAILABLE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_BLACKWELL
+
 #if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #define CP_ASYNC_AVAILABLE
 #endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
@@ -316,6 +324,11 @@ static bool cp_async_available(const int cc) {
     return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_AMPERE;
 }
 
+static bool blackwell_mma_available(const int cc) {
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_BLACKWELL &&
+           ggml_cuda_highest_compiled_arch(cc) < GGML_CUDA_CC_RUBIN;
+}
+
 static constexpr __device__ int ggml_cuda_get_physical_warp_size() {
 #if defined(GGML_USE_HIP) && (defined(__GFX9__) || defined(__GFX8__))
     return 64;
@@ -701,6 +714,28 @@ static __device__ __forceinline__ float ggml_cuda_e8m0_to_fp32(uint8_t x) {
 #endif // CUDART_VERSION >= 12050
 }
 
+__device__ __forceinline__ uint8_t ggml_cuda_float_to_fp4_e2m1(float x, float e) {
+    const uint8_t sign_bit = (x < 0.0f) << 3;
+    float         ax       = fabsf(x) * e;
+
+    // Positive LUT
+    static constexpr float pos_lut[8] = { 0.0f, 0.5f, 1.0f, 1.5f, 2.0f, 3.0f, 4.0f, 6.0f };
+
+    int   best_i   = 0;
+    float best_err = fabsf(ax - pos_lut[0]);
+
+#pragma unroll
+    for (int i = 1; i < 8; ++i) {
+        const float err = fabsf(ax - pos_lut[i]);
+        if (err < best_err) {
+            best_err = err;
+            best_i   = i;
+        }
+    }
+
+    return static_cast<uint8_t>(best_i | sign_bit);
+}
+
 // See https://gmplib.org/~tege/divcnst-pldi94.pdf figure 4.1.
 // Precompute mp (m' in the paper) and L such that division
 // can be computed using a multiply (high 32b of 64b result)

ggml/src/ggml-cuda/cumsum.cu

@@ -5,7 +5,7 @@
 #include "ggml.h"
 
 #ifdef GGML_CUDA_USE_CUB
-#   include <cub/device/device_scan.cuh>
+#   include <cub/block/block_scan.cuh>
 #endif // GGML_CUDA_USE_CUB
 
 template<typename T, int BLOCK_SIZE>
@@ -16,12 +16,14 @@ static __global__ void cumsum_cub_kernel(
         const int64_t  s01, const int64_t  s02, const int64_t  s03,
         const int64_t   s1,  const int64_t   s2,  const int64_t   s3) {
 #ifdef GGML_CUDA_USE_CUB
-    using BlockScan = cub::BlockScan<T, BLOCK_SIZE>;
+    using BlockScanT = cub::BlockScan<T, BLOCK_SIZE>;
 
-    __shared__ typename BlockScan::TempStorage temp_storage;
-    __shared__ T block_carry;      // carry from previous tile
+    __shared__ typename BlockScanT::TempStorage temp_storage;
+    __shared__ T block_carry;
 
     const int tid = threadIdx.x;
+    constexpr int UNROLL_FACTOR = 4;
+    constexpr int TILE_SIZE = BLOCK_SIZE * UNROLL_FACTOR;
 
     const int64_t i1 = blockIdx.x;
     const int64_t i2 = blockIdx.y;
@@ -39,37 +41,47 @@ static __global__ void cumsum_cub_kernel(
     }
     __syncthreads();
 
-    for (int64_t start = 0; start < ne00; start += BLOCK_SIZE) {
-        int64_t idx = start + tid;
-        T x = (idx < ne00) ? src_row[idx] : T(0);
+    for (int64_t start = 0; start < ne00; start += TILE_SIZE) {
+        T items[UNROLL_FACTOR];
+        T thread_sum = T(0);
 
-        T inclusive;
+#pragma unroll
+        for (int i = 0; i < UNROLL_FACTOR; i++) {
+            int64_t idx = start + tid * UNROLL_FACTOR + i;
+            T val = (idx < ne00) ? src_row[idx] : T(0);
+            thread_sum += val;
+            items[i] = thread_sum;
+        }
+
+        // Block-wide scan on thread sums
+        T thread_prefix;
         T block_total;
-        BlockScan(temp_storage).InclusiveSum(x, inclusive, block_total);
-
+        BlockScanT(temp_storage).InclusiveSum(thread_sum, thread_prefix, block_total);
         __syncthreads();
 
-        T final_val = inclusive + block_carry;
-
-        // store result
-        if (idx < ne00) {
-            dst_row[idx] = final_val;
+        // Add offset to each item and store
+        T thread_offset = thread_prefix - thread_sum + block_carry;
+#pragma unroll
+        for (int i = 0; i < UNROLL_FACTOR; i++) {
+            int64_t idx = start + tid * UNROLL_FACTOR + i;
+            if (idx < ne00) {
+                dst_row[idx] = items[i] + thread_offset;
+            }
         }
 
         __syncthreads();
 
+        // Update carry for next tile
         if (tid == 0) {
             block_carry += block_total;
         }
-
-        __syncthreads();
     }
 #else
     NO_DEVICE_CODE;
 #endif // GGML_CUDA_USE_CUB
 }
 
-// Fallback kernel implementation (original)
+// Fallback kernel implementation
 template<typename T>
 static __global__ void cumsum_kernel(
         const T * src, T * dst,
@@ -86,10 +98,10 @@ static __global__ void cumsum_kernel(
     const int warps_per_block = blockDim.x / warp_size;
 
     extern __shared__ float smem[];
-    float * s_vals = smem;
-    float * s_warp_sums = smem + blockDim.x;
-    float * s_carry = smem + blockDim.x + warps_per_block;
-    float * s_chunk_total = s_carry + 1;
+    float *                 s_vals        = smem;
+    float *                 s_warp_sums   = smem + blockDim.x;
+    float *                 s_carry       = smem + blockDim.x + warps_per_block;
+    float *                 s_chunk_total = s_carry + 1;
 
     // Initialize carry
     if (tid == 0) {
@@ -107,21 +119,39 @@ static __global__ void cumsum_kernel(
     const T * src_row = src + i1 * s01 + i2 * s02 + i3 * s03;
     T       * dst_row = dst + i1 * s1  + i2 * s2  + i3 * s3;
 
-    for (int64_t start = 0; start < ne00; start += blockDim.x) {
-        int64_t idx = start + tid;
-        float val = (idx < ne00) ? ggml_cuda_cast<float, T>(src_row[idx]) : 0.0f;
+    // register blocking: process 4 elements per thread to hide latency
+    // and reduce synchronization overhead
+    constexpr int num_unroll = 4;
+    T             temp[num_unroll];
 
-        // 1. Warp inclusive scan
+    for (int64_t i = 0; i < ne00; i += num_unroll * blockDim.x) {
+        int64_t idx = i + tid * num_unroll;
+
+        // thread local sequential scan
+        temp[0] = (idx < ne00 ? src_row[idx] : T(0));
+#pragma unroll
+        for (int64_t j = 1; j < num_unroll; j++) {
+            temp[j] = temp[j - 1];
+            if (idx + j < ne00) {
+                temp[j] += src_row[idx + j];
+            } else {
+                temp[j] += 0;
+            }
+        }
+
+        // last emenent is sum of all values assigned to thread
+        float val = (idx < ne00) ? ggml_cuda_cast<float, T>(temp[num_unroll - 1]) : 0.0f;
+
+        // Warp inclusive scan
         val = warp_prefix_inclusive_sum<T, warp_size>(val);
         s_vals[tid] = val;
 
-        // Store warp total
         if (lane == warp_size - 1) {
             s_warp_sums[warp] = val;
         }
         __syncthreads();
 
-        // 2. Exclusive scan of warp sums (warp 0 only)
+        // Exclusive scan of warp sums (warp 0 only)
         if (warp == 0) {
             float w = (tid < warps_per_block) ? s_warp_sums[tid] : 0.0f;
             float inc = warp_prefix_inclusive_sum<T, warp_size>(w);
@@ -134,18 +164,24 @@ static __global__ void cumsum_kernel(
         }
         __syncthreads();
 
+        // write back results
         float carry = *s_carry;
-        float final_val = s_vals[tid] + s_warp_sums[warp] + carry;
-        if (idx < ne00) {
-            dst_row[idx] = ggml_cuda_cast<T, float>(final_val);
+        // calculate sum offset for this thread
+        float final_val_offset = s_vals[tid] + s_warp_sums[warp] + carry - temp[num_unroll - 1];
+
+#pragma unroll
+        for (int32_t j = 0; j < num_unroll; j++) {
+            if (idx + j < ne00) {
+                dst_row[idx + j] = temp[j] + ggml_cuda_cast<T, float>(final_val_offset);
+            }
         }
+
         __syncthreads();
 
         // Update carry for next chunk
         if (tid == 0) {
             *s_carry += *s_chunk_total;
         }
-        __syncthreads();
     }
 }
 
@@ -177,7 +213,7 @@ static void cumsum_cuda(
     const int warps_per_block = block_size / warp_size;
     const size_t shmem_size = (block_size + warps_per_block + 2) * sizeof(float);
 
-    if (use_cub) {
+    if (use_cub && ne00 >= 1024) {
         cumsum_cub_kernel<T, CUDA_CUMSUM_BLOCK_SIZE><<<grid_dims, CUDA_CUMSUM_BLOCK_SIZE, 0, stream>>>(
             src, dst,
             ne00, ne01, ne02, ne03,

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

@@ -2211,7 +2211,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
 
             const int cc            = ggml_cuda_info().devices[id].cc;
             const int warp_size     = ggml_cuda_info().devices[id].warp_size;
-            use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+            use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1], /*n_experts=*/0);
             use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src0->nb, src1->ne[1], /*mul_mat_id=*/false);
             use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, src1->ne[1]);
             any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
@@ -2219,7 +2219,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     } else {
         const int cc            = ggml_cuda_info().devices[ctx.device].cc;
         const int warp_size     = ggml_cuda_info().devices[ctx.device].warp_size;
-        use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+        use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1], /*n_experts=*/0);
         use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src0->nb, src1->ne[1], /*mul_mat_id=*/false);
         use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, src1->ne[1]);
         any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
@@ -2287,7 +2287,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
             return;
         }
 
-        if (ggml_cuda_should_use_mmq(src0->type, cc, ne12)) {
+        if (ggml_cuda_should_use_mmq(src0->type, cc, ne12, /*n_experts=*/ne02)) {
             ggml_cuda_mul_mat_q(ctx, src0, src1, ids, dst);
             return;
         }
@@ -3076,8 +3076,11 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 9 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx];
         ggml_tensor * weights = cgraph->nodes[node_idx + 9];
+        ggml_tensor * get_rows = cgraph->nodes[node_idx + 4];
+        ggml_tensor * argsort = cgraph->nodes[node_idx + 2];
+        int n_expert = cgraph->nodes[node_idx]->src[0]->ne[0];
 
-        if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
+        if (ggml_cuda_should_use_topk_moe(softmax, weights, get_rows, argsort, nullptr, n_expert)) {
             return true;
         }
     }
@@ -3085,7 +3088,11 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
     if (is_equal(topk_moe_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 4 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx];
         ggml_tensor * weights = cgraph->nodes[node_idx + 4];
-        if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
+        ggml_tensor * get_rows = cgraph->nodes[node_idx + 4];
+        ggml_tensor * argsort = cgraph->nodes[node_idx + 2];
+        int n_expert = cgraph->nodes[node_idx]->src[0]->ne[0];
+
+        if (ggml_cuda_should_use_topk_moe(softmax, weights, get_rows, argsort, nullptr, n_expert)) {
             return true;
         }
     }
@@ -3094,8 +3101,11 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 1, node_idx + 5 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx + 4];
         ggml_tensor * weights = cgraph->nodes[node_idx + 5];
+        ggml_tensor * get_rows = cgraph->nodes[node_idx + 2];
+        ggml_tensor * argsort = cgraph->nodes[node_idx + 0];
+        int n_expert = cgraph->nodes[node_idx]->src[0]->ne[0];
 
-        if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
+        if (ggml_cuda_should_use_topk_moe(softmax, weights, get_rows, argsort, nullptr, n_expert)) {
             return true;
         }
     }
@@ -4775,6 +4785,16 @@ static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t
         features.push_back({ "FA_ALL_QUANTS", "1" });
     #endif
 
+    {
+        const auto & info = ggml_cuda_info();
+        for (int id = 0; id < info.device_count; ++id) {
+            if (blackwell_mma_available(info.devices[id].cc)) {
+                features.push_back({ "BLACKWELL_NATIVE_FP4", "1"});
+                break;
+            }
+        }
+    }
+
     #undef _STRINGIFY
     #undef STRINGIFY
 

ggml/src/ggml-cuda/mma.cuh

@@ -900,6 +900,27 @@ namespace ggml_cuda_mma {
 #endif // AMPERE_MMA_AVAILABLE
     }
 
+    static __device__ __forceinline__ void mma_block_scaled(tile<16, 8, float> &     D,
+                                                            const tile<16, 8, int> & A,
+                                                            const tile<8, 8, int> &  B,
+                                                            uint32_t                 a_scale,
+                                                            uint32_t                 b_scale) {
+#ifdef BLACKWELL_MMA_AVAILABLE
+        const int * Axi = (const int *) A.x;
+        const int * Bxi = (const int *) B.x;
+        float *     Dxi = (float *) D.x;
+
+        asm volatile(
+            "mma.sync.aligned.kind::mxf4.block_scale.scale_vec::2X.m16n8k64.row.col.f32.e2m1.e2m1.f32.ue8m0 "
+            "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3}, "
+            "%10, {0, 0}, %11, {0, 0};"
+            : "+f"(Dxi[0]), "+f"(Dxi[1]), "+f"(Dxi[2]), "+f"(Dxi[3])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[1]), "r"(a_scale), "r"(b_scale));
+#else
+        GGML_UNUSED_VARS(D, A, B, a_scale, b_scale);
+#endif  // BLACKWELL_MMA_AVAILABLE
+    }
+
     static __device__ __forceinline__ void mma(
             tile<16, 8, float> & D, const tile<16, 8, half2> & A, const tile<8, 8, half2> & B) {
 #ifdef TURING_MMA_AVAILABLE

ggml/src/ggml-cuda/mmq.cu

@@ -1,3 +1,4 @@
+#include "common.cuh"
 #include "mmq.cuh"
 #include "quantize.cuh"
 #include "mmid.cuh"
@@ -114,6 +115,9 @@ void ggml_cuda_mul_mat_q(
     const bool use_stream_k = (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA)
                             || GGML_CUDA_CC_IS_CDNA(cc);
 
+    // TODO: tighter pool buffer size vs q8 path
+    const bool use_native_mxfp4 = blackwell_mma_available(cc) && src0->type == GGML_TYPE_MXFP4;
+
     if (!ids) {
         const size_t nbytes_src1_q8_1 = ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1 +
             get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq);
@@ -123,12 +127,24 @@ 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[3] / ts_src1;
-            quantize_mmq_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type,
-                ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
+            if (use_native_mxfp4) {
+                static_assert(sizeof(block_fp4_mmq) == 4 * sizeof(block_q8_1));
+                quantize_mmq_mxfp4_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded,
+                                        ne11, ne12, ne13, stream);
+
+            } else {
+                quantize_mmq_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded,
+                                       ne11, ne12, ne13, stream);
+            }
             CUDA_CHECK(cudaGetLastError());
         }
 
-        const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));
+        // Stride depends on quantization format
+        const int64_t s12 = use_native_mxfp4 ?
+                                ne11 * ne10_padded * sizeof(block_fp4_mmq) /
+                                    (8 * QK_MXFP4 * sizeof(int))  // block_fp4_mmq holds 256 values (8 blocks of 32)
+                                :
+                                ne11 * ne10_padded * sizeof(block_q8_1) / (QK8_1 * sizeof(int));
         const int64_t s13 = ne12*s12;
 
         const mmq_args args = {
@@ -175,12 +191,19 @@ 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.get(), src1_q8_1.get(), src0->type,
-            ne10, s11, s12, s13, ne10_padded, ne11_flat, ne12_flat, ne13_flat, stream);
+
+        if (use_native_mxfp4) {
+            quantize_mmq_mxfp4_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);
+        } else {
+            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());
     }
 
-    const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));
+    const int64_t s12 = use_native_mxfp4 ? ne11 * ne10_padded * sizeof(block_fp4_mmq) / (8 * QK_MXFP4 * sizeof(int)) :
+                                           ne11 * ne10_padded * sizeof(block_q8_1) / (QK8_1 * sizeof(int));
     const int64_t s13 = ne12*s12;
 
     // Note that ne02 is used instead of ne12 because the number of y channels determines the z dimension of the CUDA grid.
@@ -236,7 +259,7 @@ void ggml_cuda_op_mul_mat_q(
     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) {
+bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t n_experts) {
 #ifdef GGML_CUDA_FORCE_CUBLAS
     return false;
 #endif // GGML_CUDA_FORCE_CUBLAS
@@ -297,7 +320,10 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         if (GGML_CUDA_CC_IS_CDNA3(cc)) {
             return true;
         }
-        if (ne11 <= 128 || type == GGML_TYPE_Q4_0 || type == GGML_TYPE_Q4_1 || type == GGML_TYPE_Q5_0 || type == GGML_TYPE_Q5_1) {
+        if (n_experts > 64 || ne11 <= 128) {
+            return true;
+        }
+        if (type == GGML_TYPE_Q4_0 || type == GGML_TYPE_Q4_1 || type == GGML_TYPE_Q5_0 || type == GGML_TYPE_Q5_1) {
             return true;
         }
         if (ne11 <= 256 && (type == GGML_TYPE_Q4_K || type == GGML_TYPE_Q5_K)) {

ggml/src/ggml-cuda/mmq.cuh

@@ -11,6 +11,7 @@ using namespace ggml_cuda_mma;
 
 #define MMQ_DP4A_MAX_BATCH_SIZE 64 // Max. batch size to use for dp4a MMQ kernels when FP16 tensor cores are available.
 #define MMQ_ITER_K 256
+#define MMQ_ITER_K_MXFP4_FP4    512
 #define MMQ_NWARPS 8
 
 typedef void (*load_tiles_mmq_t)(const char * __restrict__ x, int * x_tile, const int kbx0, const int i_max, const int stride);
@@ -44,8 +45,15 @@ struct block_q8_1_mmq {
     };
     int8_t qs[4*QK8_1]; // 128 values quantized to 8 bit each
 };
+
+struct block_fp4_mmq {
+    uint32_t d4[4];       // 8 E8M0 scales (1 per 32 values), 2 packed per uint32: d4[0]={s0,s1}, d4[1]={s2,s3}, etc.
+    int8_t   qs[4 * 32];  // 256 FP4 values packed as 4-bit pairs (2 per byte), 8 blocks of 32 values
+};
+
 static_assert(sizeof(block_q8_1_mmq) == 4*QK8_1 + 4*sizeof(half2), "Unexpected block_q8_1_mmq size");
 static_assert(sizeof(block_q8_1_mmq) == 4*sizeof(block_q8_1),      "Unexpected block_q8_1_mmq size");
+static_assert(sizeof(block_fp4_mmq)  == sizeof(block_q8_1_mmq),    "Unexpected block_fp4_mmq size");
 
 static mmq_q8_1_ds_layout mmq_get_q8_1_ds_layout(const ggml_type type_x) {
     switch (type_x) {
@@ -129,6 +137,14 @@ static int get_mmq_y_host(const int cc) {
         ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ? 128 : 64);
 }
 
+static constexpr __device__ int get_iter_k([[maybe_unused]] const ggml_type type) {
+#if defined(BLACKWELL_MMA_AVAILABLE)
+    return type == GGML_TYPE_MXFP4 ? MMQ_ITER_K_MXFP4_FP4 : MMQ_ITER_K;
+#else
+    return MMQ_ITER_K;
+#endif // defined(BLACKWELL_MMA_AVAILABLE)
+}
+
 static constexpr __device__ int get_mmq_y_device() {
 #if defined(GGML_USE_HIP)
 #if defined(RDNA1)
@@ -191,6 +207,7 @@ static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml
 }
 
 #define MMQ_MMA_TILE_X_K_Q8_0 (2*MMQ_TILE_NE_K + 2*MMQ_TILE_NE_K/QI8_0                   + 4)
+#define MMQ_MMA_TILE_X_K_FP4  (2*MMQ_TILE_NE_K + 8                                       + 4)
 #define MMQ_MMA_TILE_X_K_Q8_1 (2*MMQ_TILE_NE_K + 2*MMQ_TILE_NE_K/QI8_0                   + 4)
 #define MMQ_MMA_TILE_X_K_Q2_K (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K                           + 4)
 #define MMQ_MMA_TILE_X_K_Q3_K (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/2                         + 4)
@@ -201,6 +218,8 @@ static_assert(MMQ_MMA_TILE_X_K_Q8_1 % 8 == 4, "Wrong padding.");
 static_assert(MMQ_MMA_TILE_X_K_Q2_K % 8 == 4, "Wrong padding.");
 static_assert(MMQ_MMA_TILE_X_K_Q3_K % 8 == 4, "Wrong padding.");
 static_assert(MMQ_MMA_TILE_X_K_Q6_K % 8 == 4, "Wrong padding.");
+static_assert(MMQ_MMA_TILE_X_K_FP4  % 8 == 4, "Wrong padding.");
+static_assert(MMQ_MMA_TILE_X_K_FP4 == MMQ_MMA_TILE_X_K_Q8_1, "Wrong tile size for MXFP4");
 
 static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
     switch (type) {
@@ -209,6 +228,7 @@ static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
         case GGML_TYPE_Q5_0:    return MMQ_MMA_TILE_X_K_Q8_0;
         case GGML_TYPE_Q5_1:    return MMQ_MMA_TILE_X_K_Q8_1;
         case GGML_TYPE_Q8_0:    return MMQ_MMA_TILE_X_K_Q8_0;
+        // tile sizes are the same for Q8_1 and FP4 for blackwell
         case GGML_TYPE_MXFP4:   return MMQ_MMA_TILE_X_K_Q8_1;
         case GGML_TYPE_Q2_K:    return MMQ_MMA_TILE_X_K_Q2_K;
         case GGML_TYPE_Q3_K:    return MMQ_MMA_TILE_X_K_Q3_K;
@@ -228,7 +248,8 @@ static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
 }
 
 // block_q8_1_mmq has (128 8-bit ints == 32 32-bit ints + 4 32-bit scales)
-#define MMQ_TILE_Y_K (MMQ_TILE_NE_K + MMQ_TILE_NE_K/QI8_1)
+#define MMQ_TILE_Y_K     (MMQ_TILE_NE_K + MMQ_TILE_NE_K / QI8_1)
+#define MMQ_TILE_Y_FP4_K MMQ_TILE_Y_K
 
 static int mmq_get_granularity_host(const int mmq_x, const int cc) {
     if (amd_mfma_available(cc) || amd_wmma_available(cc)) {
@@ -761,6 +782,50 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
     }
 }
 
+template <int mmq_y, bool need_check>
+static __device__ __forceinline__ void load_tiles_mxfp4_fp4(const char * __restrict__ x,
+                                                            int * __restrict__ x_tile,
+                                                            const int kbx0,
+                                                            const int i_max,
+                                                            const int stride) {
+    constexpr int nwarps = mmq_get_nwarps_device();
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+
+    int *      x_qs = (int *) x_tile;
+    uint32_t * x_sc = (uint32_t *) (x_qs + 2 * MMQ_TILE_NE_K);
+
+    const int txi = threadIdx.x;
+
+    constexpr int iter_k = get_iter_k(GGML_TYPE_MXFP4);
+
+    constexpr int threads_per_row = iter_k / QK_MXFP4;  // each thread processes 1 block
+    constexpr int rows_per_warp   = warp_size / threads_per_row;
+    const int     kbx             = txi % threads_per_row;
+    const int     row_in_warp     = txi / threads_per_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += rows_per_warp * nwarps) {
+        int i = i0 + threadIdx.y * rows_per_warp + row_in_warp;
+
+        if constexpr (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i * stride + kbx;
+
+        // quantize_mxfp4_mmq permutes nibbles to match the quantized format
+        const int k0 = kbx * 4;
+        memcpy(x_qs + i * MMQ_MMA_TILE_X_K_FP4 + k0, bxi->qs, 16);
+
+        // Load E8M0 scales: pack 2 consecutive scales into one uint32
+        if (kbx % 2 == 0) {
+            uint32_t e = bxi->e;
+            e |= ((bxi + 1)->e << 8);
+            x_sc[i * MMQ_MMA_TILE_X_K_FP4 + kbx / 2] = e;
+        }
+    }
+}
+
 template <int mmq_x, int mmq_y>
 static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
@@ -931,6 +996,78 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma(
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 }
 
+template <int mmq_x, int mmq_y>
+static __device__ __forceinline__ void vec_dot_mxfp4_mxfp4_mma(const int * __restrict__ x,
+                                                               const int * __restrict__ y,
+                                                               float * __restrict__ sum,
+                                                               const int k00) {
+    typedef tile<16, 8, int>   tile_A;
+    typedef tile<8, 8, int>    tile_B;
+    typedef tile<16, 8, float> tile_C;  // Output is float for native scaled MMA
+
+    constexpr int granularity   = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = 2 * granularity;
+    constexpr int ntx           = rows_per_warp / tile_C::I;  // Number of x minitiles per warp.
+
+    y += (threadIdx.y % ntx) * (tile_C::J * MMQ_TILE_Y_FP4_K);
+
+    // Match layout from load_tiles_mxfp4_fp4
+    const int *      x_qs = (const int *) x;
+    const uint32_t * x_sc = (const uint32_t *) (x_qs + 2 * MMQ_TILE_NE_K);
+    const int *      y_qs = (const int *) y + 4;
+    const uint32_t * y_sc = (const uint32_t *) y;
+
+    // tile_A has a length of 64 logical values vs. 32 values in block_mxfp4
+    tile_A   A[ntx][MMQ_TILE_NE_K / (2 * QI_MXFP4)];
+    uint32_t scaleA[ntx][MMQ_TILE_NE_K / (2 * QI_MXFP4)];
+
+    // Block scale
+    // Each thread has to point to a 4 byte scale value
+    // https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-block-scaling
+
+    const int i0 = (threadIdx.y / ntx) * rows_per_warp;
+
+#pragma unroll
+    for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+        for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 2 * QI_MXFP4) {
+            const int k0 = k00 + k01;
+
+            load_ldmatrix(A[n][k01 / (2 * QI_MXFP4)], x_qs + (i0 + n * tile_A::I) * MMQ_MMA_TILE_X_K_FP4 + k0,
+                          MMQ_MMA_TILE_X_K_FP4);
+
+            // based on block-scaling document, 2 threads in each quad need to supply to the scale value
+            const int tidx         = threadIdx.x / 4 + (threadIdx.x % 2) * 8;
+            scaleA[n][k01 / (2 * QI_MXFP4)] =
+                *(x_sc + (i0 + n * tile_A::I + tidx) * MMQ_MMA_TILE_X_K_FP4 + k0 / (2 * QI_MXFP4));
+        }
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += ntx * tile_C::J) {
+#pragma unroll
+        for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 2 * QI_MXFP4) {
+            tile_B   B;
+            uint32_t scaleB;  // 2xN scales
+
+            load_generic(B, y_qs + j0 * MMQ_TILE_Y_FP4_K + k01, MMQ_TILE_Y_FP4_K);
+
+            scaleB = y_sc[(j0 + threadIdx.x / 4) * MMQ_TILE_Y_FP4_K + k01 / (2 * QI_MXFP4)];
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+                tile_C C;
+
+                mma_block_scaled(C, A[n][k01 / (2 * QI_MXFP4)], B, scaleA[n][k01 / (2 * QI_MXFP4)], scaleB);
+#pragma unroll
+                for (int l = 0; l < tile_C::ne; ++l) {
+                    sum[(j0 / tile_C::J + n) * tile_C::ne + l] += C.x[l];
+                }
+            }
+        }
+    }
+}
+
 template <int mmq_x, int mmq_y>
 static __device__ __forceinline__ void vec_dot_q8_1_q8_1_dp4a(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
@@ -3109,8 +3246,13 @@ struct mmq_type_traits<mmq_x, mmq_y, need_check, GGML_TYPE_Q8_0> {
 template <int mmq_x, int mmq_y, bool need_check>
 struct mmq_type_traits<mmq_x, mmq_y, need_check, GGML_TYPE_MXFP4> {
     static constexpr int              vdr          = VDR_MXFP4_Q8_1_MMQ;
+#ifdef BLACKWELL_MMA_AVAILABLE
+    static constexpr load_tiles_mmq_t load_tiles  = load_tiles_mxfp4_fp4<mmq_y, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma = vec_dot_mxfp4_mxfp4_mma<mmq_x, mmq_y>;
+#else
     static constexpr load_tiles_mmq_t load_tiles   = load_tiles_mxfp4<mmq_y, need_check>;
     static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, MMQ_Q8_1_DS_LAYOUT_D4>;
+#endif // BLACKWELL_MMA_AVAILABLE
     static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y>;
 };
 
@@ -3243,17 +3385,26 @@ static __device__ __forceinline__ void mul_mat_q_process_tile(
     constexpr mmq_write_back_t write_back = mmq_write_back_dp4a<mmq_x, mmq_y, need_check>;
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
-    constexpr int blocks_per_iter = MMQ_ITER_K / qk;
+#if defined(BLACKWELL_MMA_AVAILABLE)
+    // FP4 tile stores 8 blocks
+    constexpr int ne_block = (type == GGML_TYPE_MXFP4) ? 8 * QK_MXFP4 : 4 * QK8_1;
+#else
+    constexpr int ne_block = 4 * QK8_1;
+#endif  // defined(BLACKWELL_MMA_AVAILABLE)
+
+    constexpr int ITER_K          = get_iter_k(type);
+    constexpr int blocks_per_iter = ITER_K / qk;
 
     float sum[mmq_x*mmq_y / (nwarps*warp_size)] = {0.0f};
 
+    constexpr int sz = sizeof(block_q8_1_mmq) / sizeof(int);
+
     for (int kb0 = kb0_start; kb0 < kb0_stop; kb0 += blocks_per_iter) {
         load_tiles(x, tile_x, offset_x + kb0, tile_x_max_i, stride_row_x);
-
         {
-            const int * by0 = y + ncols_y*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 0*sizeof(block_q8_1_mmq)/sizeof(int));
+            const int * by0 = y + ncols_y * (kb0 * qk / ne_block) * sz;
 #pragma unroll
-            for (int l0 = 0; l0 < mmq_x*MMQ_TILE_Y_K; l0 += nwarps*warp_size) {
+            for (int l0 = 0; l0 < mmq_x * MMQ_TILE_Y_K; l0 += nwarps * warp_size) {
                 int l = l0 + threadIdx.y*warp_size + threadIdx.x;
 
                 tile_y[l] = by0[l];
@@ -3267,9 +3418,9 @@ static __device__ __forceinline__ void mul_mat_q_process_tile(
         __syncthreads();
 
         {
-            const int * by0 = y + ncols_y*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 1*sizeof(block_q8_1_mmq)/sizeof(int));
+            const int * by0 = y + ncols_y * ((kb0 * qk / ne_block) * sz + sz);
 #pragma unroll
-            for (int l0 = 0; l0 < mmq_x*MMQ_TILE_Y_K; l0 += nwarps*warp_size) {
+            for (int l0 = 0; l0 < mmq_x * MMQ_TILE_Y_K; l0 += nwarps * warp_size) {
                 int l = l0 + threadIdx.y*warp_size + threadIdx.x;
 
                 tile_y[l] = by0[l];
@@ -3401,8 +3552,10 @@ static __global__ void mul_mat_q(
     }
 #endif // (defined(GGML_USE_HIP) && !defined(CDNA3)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
 
+    constexpr int ITER_K = get_iter_k(type);
+
     const     int64_t blocks_per_ne00 = ncols_x / qk;
-    constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
+    constexpr int     blocks_per_iter = ITER_K / qk;
 
     // kbc == k block continuous, current index in continuous ijk space.
     int64_t kbc      = (int64_t) blockIdx.x     *nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
@@ -3463,7 +3616,7 @@ static __global__ void mul_mat_q(
             __syncthreads();
         }
 
-        offset_y   += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int));
+        offset_y += (col_low + jt * mmq_x) * (sizeof(block_q8_1_mmq) / sizeof(int));
         offset_dst += it*mmq_y;
 
         const int tile_x_max_i = nrows_x  - it*mmq_y - 1;
@@ -3530,7 +3683,7 @@ static __global__ void mul_mat_q(
         __syncthreads();
     }
 
-    offset_y   += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int));
+    offset_y += (col_low + jt * mmq_x) * (sizeof(block_q8_1_mmq) / sizeof(int));
     offset_dst += it*mmq_y;
 
     const int tile_x_max_i = nrows_x  - it*mmq_y - 1;
@@ -3553,7 +3706,9 @@ static __global__ void mul_mat_q_stream_k_fixup(
         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;
+    constexpr int     ITER_K          = get_iter_k(type);
+
+    constexpr int     blocks_per_iter = ITER_K / qk;
     const     int64_t blocks_per_ne00 = ncols_x / qk;
 
     constexpr int nwarps = mmq_get_nwarps_device();
@@ -3711,7 +3866,7 @@ static size_t mmq_get_nbytes_shared(const int mmq_x, const int mmq_y, const int
     const int mmq_tile_x_k = mmq_get_mma_tile_x_k(type);
     const size_t nbs_ids = mmq_x*sizeof(int);
     const size_t nbs_x = (turing_mma_available(cc) || amd_mfma_available(cc) || amd_wmma_available(cc)) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
-    const size_t nbs_y = mmq_x*sizeof(block_q8_1_mmq);
+    const size_t nbs_y = mmq_x * (sizeof(block_q8_1_mmq));
     return nbs_ids + nbs_x + GGML_PAD(nbs_y, nwarps*warp_size*sizeof(int));
 }
 
@@ -3927,4 +4082,4 @@ void ggml_cuda_op_mul_mat_q(
     const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
     const int64_t src1_padded_row_size, cudaStream_t stream);
 
-bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11);
+bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t n_experts);

ggml/src/ggml-cuda/quantize.cu

@@ -47,6 +47,131 @@ static __global__ void quantize_q8_1(
     y[ib].ds = make_half2(d, sum);
 }
 
+__device__ __forceinline__ uint8_t compute_e8m0_scale(float amax) {
+    if (!(amax > 0.0f)) {
+        return 0;
+    }
+
+    // FP4 E2M1: max exponent (unbiased) is 2.
+    constexpr int FP4_E2M1_EMAX = 2;
+
+    const float e = log2f(amax);
+
+    // "even" -> round-to-nearest integer, ties-to-even
+    const int e_int = __float2int_rn(e);
+
+    const int shared_exp = e_int - FP4_E2M1_EMAX;
+
+    int biased = shared_exp + 127;
+
+    biased = max(biased, 0);
+    biased = min(biased, 254);
+
+    return static_cast<uint8_t>(biased);
+}
+
+// quantize values in the format mxfp4 is stored which is interleaved nibbles
+// i.e. a block a0-a31 is represented as a0a16,a1a17 ...a15a31
+static __global__ void quantize_mmq_mxfp4(const float * __restrict__ x,
+                                          const int32_t * __restrict__ ids,
+                                          void * __restrict__ vy,
+                                          const int64_t ne00,
+                                          const int64_t s01,
+                                          const int64_t s02,
+                                          const int64_t s03,
+                                          const int64_t ne0,
+                                          const int     ne1,
+                                          const int     ne2) {
+    constexpr int vals_per_scale = 32;
+    constexpr int vals_per_warp  = 2 * vals_per_scale;  // Each warp processes 2 blocks of 32 = 64 values
+
+    const int warp_id = threadIdx.y;
+    const int lane_id_32 = threadIdx.x;
+
+    const int nwarps = blockDim.y;
+
+    const int64_t warp_start_offset = (blockIdx.y * nwarps + warp_id) * vals_per_warp;
+
+    if (warp_start_offset >= ne0) {
+        return;
+    }
+
+    const int64_t i1 = blockIdx.x;
+    const int64_t i2 = blockIdx.z % ne2;
+    const int64_t i3 = blockIdx.z / ne2;
+
+    const int64_t i01 = ids ? ids[i1] : i1;
+    const int64_t i02 = i2;
+    const int64_t i03 = i3;
+
+    block_fp4_mmq * y = (block_fp4_mmq *) vy;
+
+    const int64_t block_fp4_mmq_size = 8 * QK_MXFP4;  // 256 values
+    const int64_t ib0                = blockIdx.z * ((int64_t) ne1 * (ne0 / block_fp4_mmq_size));
+    const int64_t ib = ib0 + (warp_start_offset / block_fp4_mmq_size) * ne1 + blockIdx.x;
+    const int64_t quad_idx_in_block  = (warp_start_offset % block_fp4_mmq_size) / vals_per_warp;
+
+    const int group_id = lane_id_32 / 4;
+    const int lane_in_group = lane_id_32 % 4;
+    const int base = group_id * 2;
+    char2 * yqs2 = (char2 *) y[ib].qs;
+
+    const int64_t base_pos = i03 * s03 + i02 * s02 + i01 * s01;
+
+    uint8_t scales[2];
+
+#pragma unroll
+    for (int b = 0; b < 2; ++b) {
+        const int64_t i0 = warp_start_offset + b * vals_per_scale + lane_id_32;
+        const float xi = (i0 < ne00) ? x[base_pos + i0] : 0.0f;
+
+        float amax = fabsf(xi);
+#pragma unroll
+        for (int mask = 16; mask > 0; mask >>= 1) {
+            amax = fmaxf(amax, __shfl_xor_sync(0xFFFFFFFF, amax, mask, WARP_SIZE));
+        }
+
+        const uint8_t e = compute_e8m0_scale(amax);
+        scales[b] = e;
+        const float inv_s = (amax == 0.0f) ? 0.0f : __frcp_rn(ggml_cuda_e8m0_to_fp32(e));
+
+#if CUDART_VERSION >= 12080
+        const float scaled_val = xi * inv_s;
+
+        const float val0 = __shfl_sync(0xFFFFFFFF, scaled_val, base, WARP_SIZE);
+        const float val1 = __shfl_sync(0xFFFFFFFF, scaled_val, base + 16, WARP_SIZE);
+        const float val2 = __shfl_sync(0xFFFFFFFF, scaled_val, base + 1, WARP_SIZE);
+        const float val3 = __shfl_sync(0xFFFFFFFF, scaled_val, base + 17, WARP_SIZE);
+
+        if (lane_in_group == 0) {
+            __nv_fp4x4_e2m1 fp4_packed(make_float4(val0, val1, val2, val3));
+
+            yqs2[quad_idx_in_block * 16 + b * 8 + group_id] = *(char2 *) &fp4_packed;
+        }
+#else
+        // Fallback: manual FP4 conversion using LUT
+        const uint8_t q_val = ggml_cuda_float_to_fp4_e2m1(xi, inv_s);
+
+        const uint8_t q_lo_0 = __shfl_sync(0xFFFFFFFF, q_val, base,      WARP_SIZE);
+        const uint8_t q_lo_1 = __shfl_sync(0xFFFFFFFF, q_val, base + 1,  WARP_SIZE);
+        const uint8_t q_hi_0 = __shfl_sync(0xFFFFFFFF, q_val, base + 16, WARP_SIZE);
+        const uint8_t q_hi_1 = __shfl_sync(0xFFFFFFFF, q_val, base + 17, WARP_SIZE);
+
+        if (lane_in_group == 0) {
+            char2 q;
+            q.x = (q_hi_0 << 4) | q_lo_0;
+            q.y = (q_hi_1 << 4) | q_lo_1;
+            yqs2[quad_idx_in_block * 16 + b * 8 + group_id] = q;
+        }
+#endif // CUDART_VERSION >= 12080
+    }
+
+    if (lane_id_32 == 0) {
+        // Store 2 scales packed into 1 uint32
+        y[ib].d4[quad_idx_in_block] = (scales[1] << 8) | scales[0];
+    }
+}
+
 template <mmq_q8_1_ds_layout ds_layout>
 static __global__ void quantize_mmq_q8_1(
         const float * __restrict__ x, const int32_t * __restrict__ ids, void * __restrict__ vy,
@@ -190,3 +315,29 @@ void quantize_mmq_q8_1_cuda(
             break;
     }
 }
+
+void quantize_mmq_mxfp4_cuda(const float *                    x,
+                             const int32_t *                  ids,
+                             void *                           vy,
+                             [[maybe_unused]] const ggml_type type_src0,
+                             const int64_t                    ne00,
+                             const int64_t                    s01,
+                             const int64_t                    s02,
+                             const int64_t                    s03,
+                             const int64_t                    ne0,
+                             const int64_t                    ne1,
+                             const int64_t                    ne2,
+                             const int64_t                    ne3,
+                             cudaStream_t                     stream) {
+    GGML_ASSERT(ne0 % (2 * QK_MXFP4) == 0);
+
+    constexpr int nwarps = 8;
+    constexpr int vals_per_warp  = 2 * QK_MXFP4;
+    constexpr int vals_per_block = nwarps * vals_per_warp;
+
+    const int64_t block_num_y = (ne0 + vals_per_block - 1) / vals_per_block;
+    const dim3    num_blocks(ne1, block_num_y, ne2 * ne3);
+    const dim3    block_size(WARP_SIZE, nwarps, 1);
+
+    quantize_mmq_mxfp4<<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
+}

ggml/src/ggml-cuda/quantize.cuh

@@ -25,3 +25,17 @@ void quantize_mmq_q8_1_cuda(
         const float * x, const int32_t * ids, void * vy,
         ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
         int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);
+
+void quantize_mmq_mxfp4_cuda(const float *   x,
+                             const int32_t * ids,
+                             void *          vy,
+                             ggml_type       type_src0,
+                             int64_t         ne00,
+                             int64_t         s01,
+                             int64_t         s02,
+                             int64_t         s03,
+                             int64_t         ne0,
+                             int64_t         ne1,
+                             int64_t         ne2,
+                             int64_t         ne3,
+                             cudaStream_t    stream);

ggml/src/ggml-cuda/topk-moe.cu

@@ -268,7 +268,23 @@ void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx,
     }
 }
 
-bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, const ggml_tensor * weights, const ggml_tensor * clamp) {
+bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax,
+                                   const ggml_tensor * weights,
+                                   const ggml_tensor * get_rows,
+                                   const ggml_tensor * argsort,
+                                   const ggml_tensor * clamp,
+                                   int n_expert) {
+    ggml_tensor * probs = get_rows->src[0];
+    if (probs->op != GGML_OP_RESHAPE) {
+        return false;
+    }
+    probs = probs->src[0];
+    ggml_tensor * selection_probs = argsort->src[0];
+
+    if (probs != selection_probs) {
+        return false;
+    }
+
     float scale    = 1.0f;
     float max_bias = 0.0f;
 
@@ -288,7 +304,6 @@ bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, const ggml_tenso
         return false;
     }
 
-    const int n_expert = softmax->ne[0];
     // n_expert must be a power of 2
     if ((n_expert & (n_expert - 1)) != 0 || n_expert > 512) {
         return false;

ggml/src/ggml-cuda/topk-moe.cuh

@@ -11,6 +11,11 @@ void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx,
                            const bool                  delayed_softmax = false,
                            ggml_tensor *               weight_clamp    = nullptr);
 
-bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, const ggml_tensor * weights, const ggml_tensor * clamp = nullptr);
+bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax,
+                                   const ggml_tensor * weights,
+                                   const ggml_tensor * get_rows,
+                                   const ggml_tensor * argsort,
+                                   const ggml_tensor * clamp,
+                                   int n_expert);
 
 std::initializer_list<enum ggml_op> ggml_cuda_topk_moe_ops(bool with_norm, bool delayed_softmax = false);

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

@@ -10,6 +10,10 @@
 #include <cuda_fp8.h>
 #endif // CUDART_VERSION >= 12050
 
+#if CUDART_VERSION >= 12080
+#include <cuda_fp4.h>
+#endif // CUDART_VERSION >= 12080
+
 #if CUDART_VERSION < 11020
 #define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
 #define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH

ggml/src/ggml-hexagon/htp-utils.h

@@ -8,6 +8,7 @@ extern "C" {
 #include <AEEStdErr.h>
 #include <inttypes.h>
 #include <remote.h>
+#include <rpcmem.h>
 #include <stdbool.h>
 
 /* Offset to differentiate HLOS and Hexagon error codes.

ggml/src/ggml-hexagon/htp/act-ops.c

@@ -263,7 +263,8 @@ static void unary_gelu_fp32_per_thread(const struct htp_tensor * src0,
                                        struct htp_spad *         dst_spad,
                                        uint32_t                  nth,
                                        uint32_t                  ith,
-                                       uint32_t                  src0_nrows_per_thread) {
+                                       uint32_t                  src0_nrows_per_thread,
+                                       dma_queue *               dma_queue) {
     htp_act_preamble2;
 
     uint64_t t1, t2;
@@ -271,6 +272,8 @@ static void unary_gelu_fp32_per_thread(const struct htp_tensor * src0,
 
     const size_t src0_row_size = nb01;
     const size_t dst_row_size  = nb1;
+    const size_t src0_row_size_aligned = htp_round_up(src0_row_size, VLEN);
+    const size_t dst_row_size_aligned  = htp_round_up(dst_row_size, VLEN);
 
     const uint32_t src0_nrows = ne01 * ne02 * ne03;
 
@@ -282,60 +285,81 @@ static void unary_gelu_fp32_per_thread(const struct htp_tensor * src0,
         return;
     }
 
-    int is_aligned = 1;
-    int opt_path   = 0;
-    if (!htp_is_aligned((void *) src0->data, VLEN) || !htp_is_aligned((void *) dst->data, VLEN)) {
-        is_aligned = 0;
-        FARF(HIGH, "silu-f32: unaligned addresses in elementwise op, possibly slower execution\n");
-    }
-    if ((1 == is_aligned) && !(nb01 & (VLEN - 1))) {
-        opt_path = 1;
+    const uint8_t * data_src0 = (const uint8_t *) src0->data;
+    uint8_t * data_dst        = (uint8_t *) dst->data;
+
+    uint8_t * src0_spad_data = src0_spad->data + (ith * src0_spad->size_per_thread);
+    uint8_t * dst_spad_data  = dst_spad->data  + (ith * dst_spad->size_per_thread);
+
+    // While given src0_spad->size_per_thread, divide it to two ping-pong buffer for src0
+    size_t src0_spad_half_size = src0_spad->size_per_thread / 2;
+    size_t dst_spad_half_size  = dst_spad->size_per_thread  / 2;
+
+    // In gelu = x*sigmoid(x*1.702)
+    const int BLOCK = src0_spad_half_size / src0_row_size_aligned; // How many rows can we process in one block
+
+    if (BLOCK == 0) {
+        FARF(ERROR, "gelu-f32 : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n",
+                src0_spad->size_per_thread, src0_row_size_aligned);
+        return;
     }
 
-    const uint8_t * restrict data_src0 = (const uint8_t *) src0->data;
-    uint8_t * restrict data_dst        = (uint8_t *) dst->data;
+    // See discussion: https://github.com/ggml-org/llama.cpp/pull/18151#issuecomment-3678235379
+    for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) {
+        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
 
-    uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_row_size);
-    uint8_t * restrict dst_spad_data  = dst_spad->data + (ith * dst_row_size);
+        // Dummy DMA transation for sequencing (interleaving dst,src,dst,...)
+        dma_queue_push_vtcm_to_ddr(dma_queue,
+            dma_make_ptr(data_dst, dst_spad_data + (spad_idx * dst_spad_half_size)),
+            dst_row_size, dst_row_size_aligned, 0);
+
+        dma_queue_push_ddr_to_vtcm(dma_queue,
+            dma_make_ptr(src0_spad_data + (spad_idx * src0_spad_half_size), data_src0 + (ir * src0_row_size)),
+            src0_row_size_aligned, src0_row_size, block_size);
+    }
 
-    const int BLOCK = 8;
     for (uint32_t ir = src0_start_row; ir < src0_end_row; ir += BLOCK) {
-        const uint32_t block_end = MIN(ir + BLOCK, src0_end_row);
+        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
 
-        // Prefetch next block
-        if (block_end < src0_end_row) {
-            const float * restrict prefetch_ptr = (float *) (data_src0 + (block_end * src0_row_size));
-            htp_l2fetch(prefetch_ptr, 1, block_end * src0_row_size, src0_row_size);
-        }
+        float* dst_spad  = (float *) dma_queue_pop(dma_queue).src;
+        float* src0_spad = (float *) dma_queue_pop(dma_queue).dst;
 
-        // Process rows in current block
-        for (uint32_t ib = ir; ib < block_end; ib++) {
-            const float * restrict src0 = (float *) (data_src0 + (ib * src0_row_size));
-            float * restrict dst        = (float *) (data_dst + (ib * dst_row_size));
+        for (uint32_t ib = 0; ib < block_size; ib++) {
+            const float* src0_spad_ptr = src0_spad + ib * (src0_row_size_aligned / sizeof(float));
+            float* dst_spad_ptr        = dst_spad  + ib * (dst_row_size_aligned  / sizeof(float));
 
             // gelu = x * sigmoid(1.702 * x) // current implementation
-            if (1 == opt_path) {
-                hvx_mul_scalar_f32((const uint8_t *) src0, (float) 1.702, (uint8_t *) src0_spad_data, ne0);
-                hvx_fast_sigmoid_f32((const uint8_t *) src0_spad_data, (uint8_t *) src0_spad_data, ne0);
-                hvx_mul_f32_opt((const uint8_t *) src0, src0_spad_data, (uint8_t *) dst, ne0);
-            } else {
-                hvx_mul_scalar_f32( (const uint8_t *) src0, (float)1.702, (uint8_t *) src0_spad_data, ne0);
-                hvx_sigmoid_f32((const uint8_t *) src0_spad_data, (uint8_t *) src0_spad_data, ne0);
-                hvx_mul_f32((const uint8_t *) src0, src0_spad_data, (uint8_t *) dst, ne0);
-            }
+            hvx_mul_scalar_f32((const uint8_t *) src0_spad_ptr, (float) 1.702, (uint8_t *) dst_spad_ptr, ne0);
+            hvx_fast_sigmoid_f32((const uint8_t *) dst_spad_ptr, (uint8_t *) dst_spad_ptr, ne0);
+            hvx_mul_f32_opt((const uint8_t *) src0_spad_ptr, (uint8_t *) dst_spad_ptr, (uint8_t *) dst_spad_ptr, ne0);
+        }
+
+        dma_queue_push_vtcm_to_ddr(dma_queue,
+            dma_make_ptr(data_dst + (ir * dst_row_size), dst_spad),
+            dst_row_size, dst_row_size_aligned, block_size);
+
+        // prefetch N+2 loop iteration if any
+        const uint32_t pref_block = (ir + BLOCK * 2);
+        if (pref_block < src0_end_row) {
+            const uint32_t pref_block_size = MIN(BLOCK, src0_end_row - pref_block);
+            dma_queue_push_ddr_to_vtcm(dma_queue,
+                dma_make_ptr(src0_spad, data_src0 + (pref_block * src0_row_size)),
+                src0_row_size_aligned, src0_row_size, pref_block_size);
         }
     }
 
+    dma_queue_flush(dma_queue);
+
     t2 = HAP_perf_get_qtimer_count();
 
-    FARF(HIGH, "gelu-f32 %d/%d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u usec %u\n", ith, nth, opt_path, ne00, ne01, ne02,
+    FARF(HIGH, "gelu-f32 %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u usec %u\n", ith, nth, ne00, ne01, ne02,
          ne03, src0_start_row, src0_end_row, ne0, ne1, ne2, ne3, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
 }
 
 static void unary_gelu_fp32(unsigned int n, unsigned int i, void * data) {
     struct htp_ops_context * octx = (struct htp_ops_context *) data;
     unary_gelu_fp32_per_thread(&octx->src0, &octx->dst, octx->op_params, &octx->src0_spad, &octx->dst_spad, n, i,
-                               octx->src0_nrows_per_thread);
+                               octx->src0_nrows_per_thread, octx->ctx->dma[i]);
 }
 
 
@@ -468,21 +492,45 @@ static int execute_op_activations_fp32(struct htp_ops_context * octx) {
     const uint32_t n_threads  = octx->n_threads;
     const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
 
-    const size_t src0_row_size = src0->nb[1];
-    const size_t src1_row_size = src1->ne[0] ? src1->nb[1] : src0->nb[1];
-    const size_t dst_row_size  = dst->nb[1];
+    size_t src0_row_size = src0->nb[1];
+    size_t src1_row_size = src1->nb[1]; // zero bytes if src1 is not used
+    size_t dst_row_size  = dst->nb[1];
 
+    const bool src1_valid = src1->ne[0];
+    if (!src1_valid) {
+        src1_row_size = src0_row_size;
+    }
+
+    const size_t src0_row_size_aligned = htp_round_up(src0_row_size, VLEN);
+    const size_t src1_row_size_aligned = htp_round_up(src1_row_size, VLEN);
+    const size_t dst_row_size_aligned  = htp_round_up(dst_row_size, VLEN);
     // VTCM scratchpads for all tensors
     // N rows per thread, padded to HVX vector size
-    octx->dst_spad.size  = htp_round_up(dst_row_size, 128) * octx->n_threads;
-    octx->src0_spad.size = htp_round_up(src0_row_size, 128) * octx->n_threads;
-    octx->src1_spad.size = htp_round_up(src1_row_size, 128) * octx->n_threads;
 
-    size_t spad_size = octx->src0_spad.size + octx->src1_spad.size + octx->dst_spad.size;
+    size_t spad_size_per_row   = (src0_row_size_aligned + src1_row_size_aligned) + dst_row_size_aligned;
+    size_t vtcm_row_per_thread = (octx->ctx->vtcm_size)/ (n_threads* spad_size_per_row);
+
+    // Make sure the reserved vtcm size is sufficient
+    if(vtcm_row_per_thread ==0){
+        FARF(ERROR, "act-%s : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n", op_type, octx->ctx->vtcm_size,
+             spad_size_per_row * n_threads);
+        return HTP_STATUS_VTCM_TOO_SMALL;
+    }
+
+    octx->src0_spad.size_per_thread = src0_row_size_aligned * vtcm_row_per_thread;
+    octx->src1_spad.size_per_thread = src1_row_size_aligned * vtcm_row_per_thread;
+    octx->dst_spad.size_per_thread  = dst_row_size_aligned * vtcm_row_per_thread;
+
+    octx->dst_spad.size  = n_threads* octx->dst_spad.size_per_thread;
+    octx->src0_spad.size = n_threads* octx->src0_spad.size_per_thread;
+    octx->src1_spad.size = n_threads* octx->src1_spad.size_per_thread;
+
+    octx->src0_spad.data = octx->ctx->vtcm_base;
+    octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
+    octx->dst_spad.data  = octx->src1_spad.data + octx->src1_spad.size;
 
     if (src1->ne[0]) {
-        FARF(HIGH,
-             "%s: %ux%ux%ux%u x %ux%ux%ux%u -> %ux%ux%ux%u : src0-spad-size %u src1-spad-size %u dst-spad-size %u\n",
+        FARF(HIGH, "%s: %ux%ux%ux%u x %ux%ux%ux%u -> %ux%ux%ux%u : src0-spad-size %u src1-spad-size %u dst-spad-size %u\n",
              op_type, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], src1->ne[0], src1->ne[1], src1->ne[2],
              src1->ne[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], octx->src0_spad.size, octx->src1_spad.size,
              octx->dst_spad.size);
@@ -492,20 +540,8 @@ static int execute_op_activations_fp32(struct htp_ops_context * octx) {
              octx->src0_spad.size, octx->src1_spad.size, octx->dst_spad.size);
     }
 
-    // Make sure the reserved vtcm size is sufficient
-    if (octx->ctx->vtcm_size < spad_size) {
-        FARF(ERROR, "act-%s : current VTCM reservation %zu is too small, needed %zu\n", op_type, octx->ctx->vtcm_size,
-             spad_size);
-        return HTP_STATUS_VTCM_TOO_SMALL;
-    }
-
-    octx->src0_spad.data = octx->ctx->vtcm_base;
-    octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
-    octx->dst_spad.data  = octx->src1_spad.data + octx->src1_spad.size;
-
     if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
         uint32_t n_jobs = MIN(n_threads, src0_nrows);
-
         octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
         worker_pool_run_func(octx->ctx->worker_pool, act_op_func, octx, n_jobs);
     }

ggml/src/ggml-hexagon/htp/htp-dma.c

@@ -34,12 +34,12 @@ dma_queue * dma_queue_create(size_t capacity) {
     q->desc = (hexagon_udma_descriptor_type1_t *) memalign(64, capacity * sizeof(hexagon_udma_descriptor_type1_t));
     memset(q->desc, 0, capacity * sizeof(hexagon_udma_descriptor_type1_t));
 
-    q->dst = (void **) memalign(4, capacity * sizeof(void *));
-    memset(q->dst, 0, capacity * sizeof(void *));
+    q->dptr = (dma_ptr *) memalign(4, capacity * sizeof(dma_ptr));
+    memset(q->dptr, 0, capacity * sizeof(dma_ptr));
 
     q->tail = &q->desc[capacity - 1];
 
-    if (!q->desc && !q->dst) {
+    if (!q->desc && !q->dptr) {
         FARF(ERROR, "%s: failed to allocate DMA queue items\n", __FUNCTION__);
         return NULL;
     }
@@ -54,16 +54,10 @@ void dma_queue_delete(dma_queue * q) {
         return;
     }
     free(q->desc);
-    free(q->dst);
+    free(q->dptr);
     free(q);
 }
 
 void dma_queue_flush(dma_queue * q) {
-    while (1) {
-        uint32_t s = dmwait() & 0x3;
-        if (s == HEXAGON_UDMA_DM0_STATUS_IDLE) {
-            break;
-        }
-    }
-    q->tail = NULL;
+    while (dma_queue_pop(q).dst != NULL) ;
 }

ggml/src/ggml-hexagon/htp/htp-dma.h

@@ -11,10 +11,15 @@
 extern "C" {
 #endif
 
+typedef struct {
+    void *dst;
+    const void *src;
+} dma_ptr;
+
 typedef struct {
     hexagon_udma_descriptor_type1_t * desc;  // descriptor pointers
     hexagon_udma_descriptor_type1_t * tail;  // tail pointer
-    void **                           dst;   // dst pointers
+    dma_ptr                         * dptr;  // dst/src pointers
     uint32_t                          push_idx;
     uint32_t                          pop_idx;
     uint32_t                          capacity;
@@ -49,13 +54,20 @@ static inline unsigned int dmwait(void) {
     return ret;
 }
 
-static inline bool dma_queue_push(dma_queue *  q,
-                                  void *       dst,
-                                  const void * src,
-                                  size_t       dst_row_size,
-                                  size_t       src_row_size,
-                                  size_t       nrows) {
+static inline dma_ptr dma_make_ptr(void *dst, const void *src)
+{
+    dma_ptr p = { dst, src };
+    return p;
+}
+
+static inline bool dma_queue_push(dma_queue * q,
+                                  dma_ptr     dptr,
+                                  size_t      dst_row_size,
+                                  size_t      src_row_size,
+                                  size_t      width, // width in bytes. number of bytes to transfer per row
+                                  size_t      nrows) {
     if (((q->push_idx + 1) & q->idx_mask) == q->pop_idx) {
+        FARF(ERROR, "dma-push: queue full\n");
         return false;
     }
 
@@ -75,18 +87,18 @@ static inline bool dma_queue_push(dma_queue *  q,
 #endif
     desc->order          = 0;
     desc->dstate         = HEXAGON_UDMA_DESC_DSTATE_INCOMPLETE;
-    desc->src            = (void *) src;
-    desc->dst            = (void *) dst;
+    desc->src            = (void *) dptr.src;
+    desc->dst            = (void *) dptr.dst;
     desc->allocation     = 0;
     desc->padding        = 0;
-    desc->roiwidth       = src_row_size;
+    desc->roiwidth       = width;
     desc->roiheight      = nrows;
     desc->srcstride      = src_row_size;
     desc->dststride      = dst_row_size;
     desc->srcwidthoffset = 0;
     desc->dstwidthoffset = 0;
 
-    q->dst[q->push_idx] = dst;
+    q->dptr[q->push_idx] = dptr;
 
     dmlink(q->tail, desc);
     q->tail = desc;
@@ -96,9 +108,28 @@ static inline bool dma_queue_push(dma_queue *  q,
     return true;
 }
 
-static inline uint8_t * dma_queue_pop(dma_queue * q) {
+static inline bool dma_queue_push_ddr_to_vtcm(dma_queue * q,
+                                              dma_ptr     dptr,
+                                              size_t      dst_row_size,
+                                              size_t      src_row_size,
+                                              size_t      nrows) {
+    return dma_queue_push(q, dptr, dst_row_size, src_row_size, src_row_size, nrows);
+}
+
+
+static inline bool dma_queue_push_vtcm_to_ddr(dma_queue * q,
+                                              dma_ptr     dptr,
+                                              size_t      dst_row_size,
+                                              size_t      src_row_size,
+                                              size_t      nrows) {
+    return dma_queue_push(q, dptr, dst_row_size, src_row_size, dst_row_size, nrows);
+}
+
+static inline dma_ptr dma_queue_pop(dma_queue * q) {
+    dma_ptr dptr  = { NULL };
+
     if (q->push_idx == q->pop_idx) {
-        return NULL;
+        return dptr;
     }
 
     hexagon_udma_descriptor_type1_t * desc = &q->desc[q->pop_idx];
@@ -112,11 +143,11 @@ static inline uint8_t * dma_queue_pop(dma_queue * q) {
         // FARF(ERROR, "dma-pop: waiting for DMA : %u\n", q->pop_idx);
     }
 
-    uint8_t * dst = (uint8_t *) q->dst[q->pop_idx];
+    dptr = q->dptr[q->pop_idx];
 
     // FARF(ERROR, "dma-pop: i %u dst %p\n", q->pop_idx, dst);
     q->pop_idx = (q->pop_idx + 1) & q->idx_mask;
-    return dst;
+    return dptr;
 }
 
 #ifdef __cplusplus

ggml/src/ggml-hexagon/htp/hvx-utils.h

@@ -980,8 +980,6 @@ static inline void hvx_fast_sigmoid_f32(const uint8_t * restrict src, uint8_t *
     int step_of_1 = num_elems >> 5;
     int remaining = num_elems - step_of_1 * VLEN_FP32;
 
-    assert(remaining == 0);
-
     const HVX_Vector * restrict v_src = (HVX_Vector *) src;
     HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
 
@@ -996,8 +994,16 @@ static inline void hvx_fast_sigmoid_f32(const uint8_t * restrict src, uint8_t *
     for (int i = 0; i < step_of_1; i++) {
         v_dst[i] = hvx_vec_fast_sigmoid_fp32_guard(v_src[i], one, max_exp, min_exp);
     }
-}
 
+    if (remaining > 0) {
+        const float * srcf = ((const float *) src) + step_of_1* VLEN_FP32;
+        float *       dstf = (float *) dst + step_of_1*VLEN_FP32;
+
+        HVX_Vector in  = *(HVX_UVector *) srcf;
+        HVX_Vector out = hvx_vec_fast_sigmoid_fp32_guard(in, one, max_exp, min_exp);
+        hvx_vec_store_u((void *) dstf, remaining * SIZEOF_FP32, out);
+    }
+}
 
 static inline void hvx_sigmoid_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int num_elems){
     int step_of_1 = num_elems >> 5;  // divby 32, because 32 float = 128 bytes per HVX vector

ggml/src/ggml-hexagon/htp/main.c

@@ -299,7 +299,8 @@ AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_que
 
     ctx->n_threads = n_hvx;
     for (int i = 0; i < ctx->n_threads; i++) {
-        ctx->dma[i] = dma_queue_create(HTP_SPAD_SRC0_NROWS * 2);
+        // see discussion https://github.com/ggml-org/llama.cpp/pull/18151#discussion_r2632388541
+        ctx->dma[i] = dma_queue_create(64);
     }
 
     // init worker pool

ggml/src/ggml-hexagon/htp/matmul-ops.c

@@ -1127,13 +1127,13 @@ static void matmul(struct htp_matmul_type * mt,
         if (is0 >= HTP_SPAD_SRC0_NROWS) {
             break;
         }
-        dma_queue_push(dma_queue, spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size,
+        dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size),
                        src0_row_size_padded, src0_row_size, 2);
     }
 
     // Process src0 rows
     for (uint32_t ir0 = src0_start_row; ir0 < src0_end_row_x2; ir0 += 2) {
-        const uint8_t * ss0 = dma_queue_pop(dma_queue);
+        const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
 
         #pragma unroll(2)
         for (uint32_t ir1 = 0; ir1 < src1_nrows; ++ir1) {
@@ -1146,7 +1146,7 @@ static void matmul(struct htp_matmul_type * mt,
         const int pr0 = (ir0 + HTP_SPAD_SRC0_NROWS);
         const int is0 = (pr0 - src0_start_row) % HTP_SPAD_SRC0_NROWS;
         if (pr0 < src0_end_row_x2) {
-            dma_queue_push(dma_queue, spad_src0 + is0 * src0_row_size_padded, src0_row + pr0 * src0_row_size,
+            dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_row_size_padded, src0_row + pr0 * src0_row_size),
                            src0_row_size_padded, src0_row_size, 2);
         }
     }
@@ -1155,9 +1155,9 @@ static void matmul(struct htp_matmul_type * mt,
     if (src0_end_row != src0_end_row_x2) {
         uint32_t  ir0 = src0_end_row_x2;
         const int is0 = (ir0 - src0_start_row);
-        dma_queue_push(dma_queue, spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size,
+        dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size),
                        src0_row_size_padded, src0_row_size, 1);
-        const uint8_t * ss0 = dma_queue_pop(dma_queue);
+        const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
 
         #pragma unroll(2)
         for (uint32_t ir1 = 0; ir1 < src1_nrows; ++ir1) {
@@ -1229,20 +1229,20 @@ static void matvec(struct htp_matmul_type * mt,
         if (is0 >= HTP_SPAD_SRC0_NROWS) {
             break;
         }
-        dma_queue_push(dma_queue, spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size,
+        dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size),
                        src0_row_size_padded, src0_row_size, 2);
     }
 
     // Process src0 rows
     for (uint32_t ir0 = src0_start_row; ir0 < src0_end_row_x2; ir0 += 2) {
-        const uint8_t * ss0 = dma_queue_pop(dma_queue);
+        const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
         mt->vec_dot_rx2(ne00, &tmp[ir0 - src0_start_row], ss0, src0_row_size_padded, src1_col);
 
         // Prefetch next (n + spad_nrows) row
         const uint32_t pr0 = (ir0 + HTP_SPAD_SRC0_NROWS);
         const uint32_t is0 = (pr0 - src0_start_row) % HTP_SPAD_SRC0_NROWS;
         if (pr0 < src0_end_row_x2) {
-            dma_queue_push(dma_queue, spad_src0 + is0 * src0_row_size_padded, src0_row + pr0 * src0_row_size,
+            dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_row_size_padded, src0_row + pr0 * src0_row_size),
                            src0_row_size_padded, src0_row_size, 2);
         }
     }
@@ -1251,9 +1251,9 @@ static void matvec(struct htp_matmul_type * mt,
     if (src0_end_row != src0_end_row_x2) {
         const uint32_t ir0 = src0_end_row_x2;
         const uint32_t is0 = (ir0 - src0_start_row);
-        dma_queue_push(dma_queue, spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size,
+        dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size),
                        src0_row_size_padded, src0_row_size, 1);
-        const uint8_t * ss0 = dma_queue_pop(dma_queue);
+        const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
         mt->vec_dot(ne00, &tmp[ir0 - src0_start_row], ss0, src1_col);
     }
 
@@ -1343,13 +1343,13 @@ static void matmul_id(struct htp_matmul_type * mt,
             if (is0 >= HTP_SPAD_SRC0_NROWS) {
                 break;
             }
-            dma_queue_push(dma_queue, spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size,
+            dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size),
                            src0_row_size_padded, src0_row_size, 2);
         }
 
         // Process src0 rows
         for (uint32_t ir0 = src0_start_row; ir0 < src0_end_row_x2; ir0 += 2) {
-            const uint8_t * ss0 = dma_queue_pop(dma_queue);
+            const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
 
             for (uint32_t cid = 0; cid < cne1; ++cid) {
                 struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, cid);
@@ -1368,7 +1368,7 @@ static void matmul_id(struct htp_matmul_type * mt,
             const int pr0 = (ir0 + HTP_SPAD_SRC0_NROWS);
             const int is0 = (pr0 - src0_start_row) % HTP_SPAD_SRC0_NROWS;
             if (pr0 < src0_end_row_x2) {
-                dma_queue_push(dma_queue, spad_src0 + is0 * src0_row_size_padded, src0_row + pr0 * src0_row_size,
+                dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_row_size_padded, src0_row + pr0 * src0_row_size),
                                src0_row_size_padded, src0_row_size, 2);
             }
         }
@@ -1377,9 +1377,9 @@ static void matmul_id(struct htp_matmul_type * mt,
         if (src0_end_row != src0_end_row_x2) {
             uint32_t       ir0 = src0_end_row_x2;
             const uint32_t is0 = (ir0 - src0_start_row);
-            dma_queue_push(dma_queue, spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size,
+            dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size),
                            src0_row_size_padded, src0_row_size, 1);
-            const uint8_t * ss0 = dma_queue_pop(dma_queue);
+            const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
 
             for (uint32_t cid = 0; cid < cne1; ++cid) {
                 struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, cid);
@@ -1467,20 +1467,20 @@ static void matvec_id(struct htp_matmul_type * mt,
             if (is0 >= HTP_SPAD_SRC0_NROWS) {
                 break;
             }
-            dma_queue_push(dma_queue, spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size,
+            dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size),
                            src0_row_size_padded, src0_row_size, 2);
         }
 
         // Process src0 rows
         for (uint32_t ir0 = src0_start_row; ir0 < src0_end_row_x2; ir0 += 2) {
-            const uint8_t * ss0 = dma_queue_pop(dma_queue);
+            const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
             mt->vec_dot_rx2(ne00, &dst_row[ir0], ss0, src0_row_size_padded, src1_col);
 
             // Prefetch next (n + spad_nrows) row
             const int pr0 = (ir0 + HTP_SPAD_SRC0_NROWS);
             const int is0 = (pr0 - src0_start_row) % HTP_SPAD_SRC0_NROWS;
             if (pr0 < src0_end_row_x2) {
-                dma_queue_push(dma_queue, spad_src0 + is0 * src0_row_size_padded, src0_row + pr0 * src0_row_size,
+                dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_row_size_padded, src0_row + pr0 * src0_row_size),
                                src0_row_size_padded, src0_row_size, 2);
             }
         }
@@ -1489,9 +1489,9 @@ static void matvec_id(struct htp_matmul_type * mt,
         if (src0_end_row != src0_end_row_x2) {
             uint32_t       ir0 = src0_end_row_x2;
             const uint32_t is0 = (ir0 - src0_start_row);
-            dma_queue_push(dma_queue, spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size,
+            dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size),
                            src0_row_size_padded, src0_row_size, 1);
-            const uint8_t * ss0 = dma_queue_pop(dma_queue);
+            const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
             mt->vec_dot(ne00, &dst_row[ir0], ss0, src1_col);
         }
     }

ggml/src/ggml-hexagon/op-desc.h

@@ -0,0 +1,153 @@
+#ifndef OP_DESC_H
+#define OP_DESC_H
+
+#define GGML_COMMON_IMPL_CPP
+#include "ggml-backend-impl.h"
+#include "ggml-common.h"
+
+#include <string>
+#include <stdio.h>
+
+struct op_desc {
+    char strides[64 * GGML_MAX_SRC];
+    char dims[64 * GGML_MAX_SRC];
+    char types[16 * GGML_MAX_SRC];
+    char buffs[64 * GGML_MAX_SRC];
+    char names[64 * GGML_MAX_SRC];
+
+    int format_tensor_dims(char * str, const struct ggml_tensor * t) {
+        if (t->ne[2] == 1 && t->ne[3] == 1) {
+            return sprintf(str, "%d:%d", (int) t->ne[0], (int) t->ne[1]);
+        } else {
+            return sprintf(str, "%d:%d:%d:%d", (int) t->ne[0], (int) t->ne[1], (int) t->ne[2], (int) t->ne[3]);
+        }
+    }
+
+    void format_op_dims(char * str, const struct ggml_tensor * t) {
+        char * p = str;
+
+        // append src0 and src1 (if any)
+        if (t->src[0]) {
+            p += format_tensor_dims(p, t->src[0]);
+
+            for (int i = 1; i < GGML_MAX_SRC && t->src[i]; i++) {
+                p += sprintf(p, " x ");
+                p += format_tensor_dims(p, t->src[i]);
+            }
+
+            p += sprintf(p, " -> ");
+        }
+
+        // format self dims separately for better visual alignment
+        char self[64];
+        format_tensor_dims(self, t);
+
+        p += sprintf(p, "%s", self);
+    }
+
+    int format_tensor_strides(char * str, const struct ggml_tensor * t) {
+        const char * c = ggml_is_contiguous(t) ? "" : "!";
+
+        if (t->ne[2] == 1 && t->ne[3] == 1) {
+            return sprintf(str, "%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], c);
+        } else {
+            return sprintf(str, "%zu:%zu:%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], (size_t) t->nb[2], (size_t) t->nb[3], c);
+        }
+    }
+
+    void format_op_strides(char * str, const struct ggml_tensor * t) {
+        char * p = str;
+
+        // append src0 and src1 (if any)
+        if (t->src[0]) {
+            p += format_tensor_strides(p, t->src[0]);
+
+            for (int i = 1; i < GGML_MAX_SRC && t->src[i]; i++) {
+                p += sprintf(p, " x ");
+                p += format_tensor_strides(p, t->src[i]);
+            }
+
+            p += sprintf(p, " -> ");
+        }
+
+        // format self dims separately for better visual alignment
+        char self[64];
+        format_tensor_strides(self, t);
+
+        p += sprintf(p, "%s", self);
+    }
+
+    void format_op_types(char * str, const struct ggml_tensor * t) {
+        char * p = str;
+
+        // append src0 and src1 (if any)
+        if (t->src[0]) {
+            p += sprintf(p, "%s", ggml_type_name(t->src[0]->type));
+
+            for (int i = 1; i < GGML_MAX_SRC && t->src[i]; i++) {
+                p += sprintf(p, " x ");
+                p += sprintf(p, "%s", ggml_type_name(t->src[i]->type));
+            }
+
+            p += sprintf(p, " -> ");
+        }
+
+        p += sprintf(p, "%s", ggml_type_name(t->type));
+    }
+
+    const char * tensor_buff_name(const struct ggml_tensor * t) {
+        if (t->buffer) {
+            return ggml_backend_buffer_name(t->buffer);
+        }
+        return "NONE";
+    }
+
+    void format_op_buffs(char * str, const struct ggml_tensor * t) {
+        char * p = str;
+
+        // append src0 and src1 (if any)
+        if (t->src[0]) {
+            p += sprintf(p, "%s", tensor_buff_name(t->src[0]));
+
+            for (int i = 1; i < GGML_MAX_SRC && t->src[i]; i++) {
+                p += sprintf(p, " x ");
+                p += sprintf(p, "%s", tensor_buff_name(t->src[i]));
+            }
+
+            p += sprintf(p, " -> ");
+        }
+
+        p += sprintf(p, "%s", tensor_buff_name(t));
+    }
+
+    void format_op_names(char * str, const struct ggml_tensor * t) {
+        char * p = str;
+
+        // append src0 and src1 (if any)
+        if (t->src[0]) {
+            p += sprintf(p, "%s", t->src[0]->name);
+
+            for (int i = 1; i < GGML_MAX_SRC && t->src[i]; i++) {
+                p += sprintf(p, " x ");
+                p += sprintf(p, "%s", t->src[i]->name);
+            }
+
+            p += sprintf(p, " -> ");
+        }
+
+        p += sprintf(p, "%s", t->name);
+    }
+
+    void format(const ggml_tensor * op) {
+        format_op_dims(dims, op);
+        format_op_strides(strides, op);
+        format_op_types(types, op);
+        format_op_buffs(buffs, op);
+        format_op_names(names, op);
+    }
+
+    op_desc() {}
+    op_desc(const ggml_tensor * op) { format(op); }
+};
+
+#endif // OP_DESC_H

ggml/src/ggml-impl.h

@@ -24,10 +24,6 @@
 #include <arm_neon.h>
 #endif
 
-#if defined(__F16C__)
-#include <immintrin.h>
-#endif
-
 #ifdef __cplusplus
 extern "C" {
 #endif

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

@@ -263,6 +263,32 @@ static ggml_cl_compiler_version get_adreno_cl_compiler_version(const char *drive
     return { type, major, minor, patch };
 }
 
+// cl buffer wrapper
+struct ggml_cl_buffer {
+    cl_mem buffer;
+    size_t size;
+
+    ggml_cl_buffer()
+        : buffer(nullptr), size(0) {}
+
+    ~ggml_cl_buffer() {
+        if (buffer) {
+            CL_CHECK(clReleaseMemObject(buffer));
+        }
+    }
+
+    void allocate(cl_context context, size_t new_size) {
+        if (new_size > size) {
+            size = new_size;
+            if (buffer) {
+                CL_CHECK(clReleaseMemObject(buffer));
+            }
+            cl_int err;
+            CL_CHECK((buffer = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &err), err));
+        }
+    }
+};
+
 // Profiling
 struct ProfilingInfo {
     std::string op_name;
@@ -376,6 +402,11 @@ struct ggml_backend_opencl_context {
     cl_context context;
     cl_command_queue queue;
 
+    // prealloc buffers for transposing weights and activations
+    ggml_cl_buffer prealloc_quant_trans;
+    ggml_cl_buffer prealloc_scales_trans;
+    ggml_cl_buffer prealloc_act_trans;
+
     cl_program program_add;
     cl_program program_add_id;
     cl_program program_clamp;
@@ -494,6 +525,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_convert_block_q8_0, kernel_restore_block_q8_0;
     cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
     cl_kernel kernel_convert_block_q4_0_noshuffle;
+    cl_kernel kernel_restore_block_q4_0_noshuffle;
     cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
     cl_kernel kernel_mul_mv_q6_K_f32;
     cl_kernel kernel_mul_mv_mxfp4_f32, kernel_mul_mv_mxfp4_f32_flat;
@@ -634,12 +666,9 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_transpose_32;
     cl_kernel kernel_transpose_32_16;
     cl_kernel kernel_transpose_16;
+    cl_kernel kernel_transpose_16_buf;
     cl_kernel kernel_transpose_16_4x1;
 
-    cl_mem A_s_d_max;            // max scale buffer size for transpose
-    cl_mem A_q_d_max;            // max weight buffer size for transpose
-    cl_mem B_d_max;              // max activation buffer size for transpose
-
     // Gemm and Gemv related programs, kernels, etc
     cl_program program_CL_gemm;
     cl_program program_CL_gemv_general;
@@ -806,6 +835,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
         CL_CHECK((backend_ctx->kernel_convert_block_q4_0_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_0_noshuffle", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q4_0_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_0_noshuffle", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q4_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_0", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q4_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_0", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_mxfp4 = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_mxfp4", &err), err));
@@ -2004,7 +2034,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_transpose_32_16 = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_32_16", &err), err));
         CL_CHECK((backend_ctx->kernel_transpose_32    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_32", &err), err));
         CL_CHECK((backend_ctx->kernel_transpose_16    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16", &err), err));
-        CL_CHECK((backend_ctx->kernel_transpose_16_4x1    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16_4x1", &err), err));
+        CL_CHECK((backend_ctx->kernel_transpose_16_buf = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16_buf", &err), err));
+        CL_CHECK((backend_ctx->kernel_transpose_16_4x1 = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16_4x1", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -2596,9 +2627,9 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
                       required_B_d_bytes, max_B_d_bytes);
     }
 
-    CL_CHECK((backend_ctx->A_q_d_max = clCreateBuffer(context, 0, max_A_q_d_bytes, NULL, &err), err));
-    CL_CHECK((backend_ctx->A_s_d_max = clCreateBuffer(context, 0, max_A_s_d_bytes, NULL, &err), err));
-    CL_CHECK((backend_ctx->B_d_max   = clCreateBuffer(context, 0, max_B_d_bytes,   NULL, &err), err));
+    backend_ctx->prealloc_quant_trans.allocate(context, max_A_q_d_bytes);
+    backend_ctx->prealloc_scales_trans.allocate(context, max_A_s_d_bytes);
+    backend_ctx->prealloc_act_trans.allocate(context, max_B_d_bytes);
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
 
     backend_ctx->disable_fusion = getenv("GGML_OPENCL_DISABLE_FUSION") != nullptr;
@@ -3603,32 +3634,35 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         // use sub_buffer of max buffer size instead
 
         size_t q_size_bytes = K * M / 8 * sizeof(float);
+        backend_ctx->prealloc_quant_trans.allocate(context, q_size_bytes);
+
         cl_buffer_region region;
         region.origin = 0;
         region.size = q_size_bytes;
         cl_mem qT_d = clCreateSubBuffer(
-            backend_ctx->A_q_d_max,
+            backend_ctx->prealloc_quant_trans.buffer,
             0,
             CL_BUFFER_CREATE_TYPE_REGION,
             &region,
             &err);
-        // cl_mem qT_d = clCreateBuffer(context, CL_MEM_READ_WRITE, q_size_bytes, NULL, &err);
         CL_CHECK(err);
 
         bool K_tile_trans = true;
         if ((K / 32) % 4 != 0){
             K_tile_trans =false;
         }
+
         size_t d_size_bytes = M * (K / 32) * 2;
+        backend_ctx->prealloc_scales_trans.allocate(context, d_size_bytes);
+
         region.origin = 0;
         region.size = d_size_bytes;
         cl_mem dT_d = clCreateSubBuffer(
-            backend_ctx->A_s_d_max,
+            backend_ctx->prealloc_scales_trans.buffer,
             0,
             CL_BUFFER_CREATE_TYPE_REGION,
             &region,
             &err);
-        // cl_mem dT_d = clCreateBuffer(context, CL_MEM_READ_WRITE, d_size_bytes, NULL, &err);
         CL_CHECK(err);
 
         // <----------------------------------------------------------------------------------> //
@@ -3933,6 +3967,91 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
     if (tensor->type == GGML_TYPE_Q4_0) {
         ggml_tensor_extra_cl_q4_0 * extra = (ggml_tensor_extra_cl_q4_0 *)tensor->extra;
 
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            cl_int err;
+            cl_kernel kernel;
+
+            cl_int M = tensor->ne[1];   // ne01
+            cl_int K = tensor->ne[0];   // ne00
+
+            GGML_ASSERT(K % 32 == 0);
+            GGML_ASSERT(M % 4 == 0);
+
+            size_t size_q = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*ggml_blck_size(tensor->type)/2;
+            size_t size_d = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*sizeof(ggml_fp16_t);
+            GGML_ASSERT(size_d + size_q == ggml_nbytes(tensor) && "Incorrect tensor size");
+
+            cl_mem buf_trans_q;
+            cl_mem buf_trans_d;
+
+            CL_CHECK((buf_trans_q = clCreateBuffer(context, CL_MEM_READ_WRITE,
+                size_q, NULL, &err), err));
+            CL_CHECK((buf_trans_d = clCreateBuffer(context, CL_MEM_READ_WRITE,
+                size_d, NULL, &err), err));
+
+            kernel = backend_ctx->kernel_transpose_16_buf;
+
+            // transpose q back
+            cl_int stride_k_q = K/4;
+            size_t local_size_q[3] = {64, 1, 1};
+            size_t global_size_q[3] = {(size_t)M, (size_t)stride_k_q, 1};
+
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &buf_trans_q));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_int), &M));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_int), &stride_k_q));
+
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+                global_size_q, local_size_q, 0, NULL, NULL));
+
+            // transpose scales back
+            cl_int stride_k_d = K/32;
+            size_t local_size_d[3] = {64, 1, 1};
+            size_t global_size_d[3] = {(size_t)M, (size_t)stride_k_d, 1};
+
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->d));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &buf_trans_d));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_int), &M));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_int), &stride_k_d));
+
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+                global_size_d, local_size_d, 0, NULL, NULL));
+
+            // unpack
+            cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+                ggml_nbytes(tensor), NULL, &err);
+            CL_CHECK(err);
+
+            cl_uchar mask_0F = 0x0F;
+            cl_uchar mask_F0 = 0xF0;
+
+            size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+            size_t local_work_size[] = {1, 1, 1};
+
+            kernel = backend_ctx->kernel_restore_block_q4_0_noshuffle;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &buf_trans_q));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &buf_trans_d));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &data_device));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_uchar), &mask_0F));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_uchar), &mask_F0));
+
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+                global_work_size, local_work_size, 0, NULL, NULL));
+
+            // read back to host
+            CL_CHECK(clEnqueueReadBuffer(
+                queue, data_device, CL_TRUE, offset,
+                size, data, 0, NULL, NULL));
+
+            CL_CHECK(clReleaseMemObject(data_device));
+            CL_CHECK(clReleaseMemObject(buf_trans_q));
+            CL_CHECK(clReleaseMemObject(buf_trans_d));
+
+            return;
+        }
+#endif
+
         cl_int err;
         cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
             ggml_nbytes(tensor), NULL, &err);
@@ -7306,8 +7425,10 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
             region.origin = 0;
             // Specify the size of the sub-buffer (divide by 2 for FP16)
             region.size = K * (N + padding) * sizeof(float)/2;
+            backend_ctx->prealloc_act_trans.allocate(context, region.size);
+
             B_d = clCreateSubBuffer(
-                backend_ctx->B_d_max,
+                backend_ctx->prealloc_act_trans.buffer,
                 0,
                 CL_BUFFER_CREATE_TYPE_REGION,
                 &region,

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

@@ -117,6 +117,27 @@ kernel void kernel_convert_block_q4_0_noshuffle(
     }
 }
 
+kernel void kernel_restore_block_q4_0_noshuffle(
+    global uchar * src_q,
+    global half  * src_d,
+    global struct block_q4_0 * dst,
+    uchar mask_0F,
+    uchar mask_F0
+) {
+    global struct block_q4_0 * b = (global struct block_q4_0 *) dst + get_global_id(0);
+    global uchar * q = (global uchar *) src_q + QK4_0/2*get_global_id(0);
+    global half  * d = (global half *) src_d + get_global_id(0);
+
+    b->d = *d;
+    for (int i = 0; i < QK4_0/4; ++i) {
+        uchar x0 = q[i + 0      ] ;
+        uchar x1 = q[i + QK4_0/4];
+
+        b->qs[2*i + 0] = convert_uchar((x0 & mask_0F) | ((x1 & mask_0F) << 4));
+        b->qs[2*i + 1] = convert_uchar(((x0 & mask_F0) >> 4) | (x1 & mask_F0));
+    }
+}
+
 //------------------------------------------------------------------------------
 // block_mxfp4
 //------------------------------------------------------------------------------

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

@@ -44,6 +44,19 @@ kernel void kernel_transpose_16_4x1(
     write_imageh(output, i * rows + j, (half4)(temp0, temp1, temp2, temp3));
 }
 
+// Transpose treating each element as 16-bit using buffer
+kernel void kernel_transpose_16_buf(
+    global const ushort * input,
+    global ushort * output,
+    const int ldi,
+    const int ldo
+) {
+    const int x = get_global_id(0);
+    const int y = get_global_id(1);
+
+    output[x*ldo + y] = input[y*ldi + x];
+}
+
 // 32-bit transpose, loading/storing a 4x4 tile of elements
 kernel void kernel_transpose_32(
     __read_only image1d_buffer_t input,

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

@@ -524,6 +524,7 @@ static std::shared_ptr<socket_t> get_socket(const std::string & endpoint) {
     std::string host;
     int port;
     if (!parse_endpoint(endpoint, host, port)) {
+        GGML_LOG_ERROR("Failed to parse endpoint: %s\n", endpoint.c_str());
         return nullptr;
     }
 #ifdef _WIN32
@@ -571,6 +572,10 @@ static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
     return ctx->base_ptr;
 }
 
+static bool ggml_backend_buffer_is_rpc(ggml_backend_buffer_t buffer) {
+    return buffer->iface.free_buffer == ggml_backend_rpc_buffer_free_buffer;
+}
+
 static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
     rpc_tensor result;
     if (!tensor) {
@@ -580,10 +585,10 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
 
     result.id = reinterpret_cast<uint64_t>(tensor);
     result.type = tensor->type;
-    if (tensor->buffer) {
+    if (tensor->buffer && ggml_backend_buffer_is_rpc(tensor->buffer)) {
         ggml_backend_buffer_t buffer = tensor->buffer;
         ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
-        result.buffer = ctx->remote_ptr;
+        result.buffer = ctx != nullptr ? ctx->remote_ptr : 0;
     } else {
         result.buffer = 0;
     }
@@ -664,10 +669,6 @@ static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, con
     RPC_STATUS_ASSERT(status);
 }
 
-static bool ggml_backend_buffer_is_rpc(ggml_backend_buffer_t buffer) {
-    return buffer->iface.free_buffer == ggml_backend_rpc_buffer_free_buffer;
-}
-
 static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
     if (ggml_backend_buffer_is_rpc(src->buffer)) {
         // check if src and dst are on the same server
@@ -2053,6 +2054,10 @@ ggml_backend_reg_t ggml_backend_rpc_reg(void) {
 
 static uint32_t ggml_backend_rpc_get_device_count(const char * endpoint) {
     auto sock = get_socket(endpoint);
+    if (sock == nullptr) {
+        GGML_LOG_ERROR("Failed to connect to %s\n", endpoint);
+        return 0;
+    }
     rpc_msg_device_count_rsp response;
     bool status = send_rpc_cmd(sock, RPC_CMD_DEVICE_COUNT, nullptr, 0, &response, sizeof(response));
     RPC_STATUS_ASSERT(status);

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

@@ -0,0 +1,51 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+
+#include "types.glsl"
+
+layout (push_constant) uniform parameter
+{
+    uint32_t ne00;
+    uint32_t ne01;
+    uint32_t nb00;
+    uint32_t nb01;
+    uint32_t a_offset;
+} p;
+
+#define BLOCK_SIZE 256
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {uint data_a[];};
+layout (binding = 1) writeonly buffer D {uint data_d[];};
+
+shared uint vals[BLOCK_SIZE];
+
+void main() {
+    const uint expert_id = gl_WorkGroupID.x;
+    const uint num_elements = p.ne00 * p.ne01;
+    const uint tid = gl_LocalInvocationID.x;
+
+    uint count = 0;
+    for (uint idx = tid; idx < num_elements; idx += BLOCK_SIZE) {
+        const uint i01 = idx / p.ne00;
+        const uint i00 = idx % p.ne00;
+        const uint a = data_a[p.a_offset + i01 * p.nb01 + i00 * p.nb00];
+
+        count += uint(a == expert_id);
+    }
+
+    vals[tid] = count;
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            vals[tid] += vals[tid + s];
+        }
+        barrier();
+    }
+
+    if (tid == 0) {
+        data_d[expert_id] = vals[0];
+    }
+}

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

@@ -401,13 +401,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
     const uint sl = (data_a[a_offset + ib].scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF;
     const uint sh = (data_a[a_offset + ib].scales_h >> (2 * ib32)) & 3;
     const uint qshift = (iqs & 16) >> 2;
-    u8vec4 qs = u8vec4(
-        data_a[a_offset + ib].qs[iq + 0],
-        data_a[a_offset + ib].qs[iq + 1],
-        data_a[a_offset + ib].qs[iq + 2],
-        data_a[a_offset + ib].qs[iq + 3]
-    );
-    qs = (qs >> qshift) & uint8_t(0xF);
+    const u8vec4 qs = unpack8((data_a_packed32[a_offset + ib].qs[iq/4] >> qshift) & 0x0F0F0F0F);
 
     const float dl = float(int(sl | (sh << 4)) - 32);
     return dl * vec4(

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

@@ -6,4 +6,6 @@ layout (push_constant) uniform parameter
     uint KY;
     float param1;
     float param2;
+    float param3;
+    float param4;
 } p;

ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp

@@ -19,6 +19,7 @@ layout (push_constant) uniform parameter
     int s0; int s1;
     int p0; int p1;
     int d0; int d1;
+    uint batch_IC;
 } p;
 
 layout(constant_id = 0) const uint BLOCK_SIZE = 32;
@@ -34,12 +35,12 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 layout (buffer_reference) buffer D_ptr {D_TYPE d;};
 #endif
 
-void main() {
+void im2col(const uint y, const uint z) {
     const uint gidx = gl_GlobalInvocationID.x;
 
-    const uint oh = gl_GlobalInvocationID.y;
-    const uint batch = gl_GlobalInvocationID.z / p.IC;
-    const uint ic = gl_GlobalInvocationID.z % p.IC;
+    const uint oh = y;
+    const uint batch = z / p.IC;
+    const uint ic = z % p.IC;
 
     const uint src_base = ic * p.offset_delta + batch * p.batch_offset;
     const BDA_OFFSET_T dst_base = ((BDA_OFFSET_T(batch) * p.OH + oh) * p.OW) * p.CHW + BDA_OFFSET_T(ic) * (p.KW * p.KH);
@@ -101,3 +102,15 @@ void main() {
 #endif
     }
 }
+
+void main() {
+    uint y = gl_GlobalInvocationID.y;
+    while (y < p.OH) {
+        uint z = gl_GlobalInvocationID.z;
+        while (z < p.batch_IC) {
+            im2col(y, z);
+            z += gl_NumWorkGroups.z;
+        }
+        y += gl_NumWorkGroups.y;
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp

@@ -11,36 +11,54 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + 16 * itid;
     const uint nibble_shift = 4 * (itid & 1);
     const uint ib32 = itid / 2; // 0..7
-
     uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    // Precompute db multiplication factors
+    float db_vals[NUM_ROWS];
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const float d = float(data_a[ibi].d);
-        const uint scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF;
-        const float db = d * (0.5 + scale) * 0.25;
-
+        const uint scale_raw = data_a[ibi].scales[ib32];
+        const uint scale = (scale_raw >> nibble_shift) & 0xF;
+        // Merge constant calculations d * (0.5 + scale) * 0.25 = d*0.125 + d*scale*0.25
+        db_vals[n] = d * (0.125f + float(scale) * 0.25f);
+        ibi += num_blocks_per_row;
+    }
+    ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        // Preload grid and sign data for all l values
+        vec4 grid0_vals[2], grid1_vals[2];
+        uint sign_vals[2], sign7_vals[2];
         [[unroll]] for (uint l = 0; l < 2; ++l) {
             const uint qs = data_a[ibi].qs[2 * itid + l];
-            const uint sign = qs >> 9;
-            const uint sign7 = bitCount(sign);
-            const vec4 grid0 = vec4(unpack8(iq2xs_grid[qs & 511].x));
-            const vec4 grid1 = vec4(unpack8(iq2xs_grid[qs & 511].y));
-
-            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
-                vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
-                vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
-
-                FLOAT_TYPE sum =
-                      fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
-                      fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
-                      fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
-                      fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w),
-                      fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x),
-                      fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
-                      fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
-                      fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 &  1) != 0 ? -grid1.w : grid1.w),
-                      FLOAT_TYPE(0.0)))))))));
-                temp[j][n] = fma(db, sum, temp[j][n]);
+            sign_vals[l] = qs >> 9;
+            sign7_vals[l] = bitCount(sign_vals[l]);
+            const uvec2 grid_data = iq2xs_grid[qs & 511];
+            grid0_vals[l] = vec4(unpack8(grid_data.x));
+            grid1_vals[l] = vec4(unpack8(grid_data.y));
+        }
+        // Preload B data for all j columns (reduce repeated index calculations)
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+            [[unroll]] for (uint l = 0; l < 2; ++l) {
+                const uint sign = sign_vals[l];
+                const uint sign7 = sign7_vals[l];
+                const vec4 grid0 = grid0_vals[l];
+                const vec4 grid1 = grid1_vals[l];
+                // Precompute indices
+                const uint b_idx = (j * p.batch_stride_b + b_offset + y_idx) / 4 + 2 * l;
+                const vec4 b0 = vec4(data_b_v4[b_idx + 0]);
+                const vec4 b4 = vec4(data_b_v4[b_idx + 1]);
+                sum +=
+                    fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
+                    fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
+                    fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
+                    fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w),
+                    fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x),
+                    fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
+                    fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
+                    fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 &  1) != 0 ? -grid1.w : grid1.w),
+                    FLOAT_TYPE(0.0)))))))));
             }
+            temp[j][n] = fma(FLOAT_TYPE(db_vals[n]), sum, temp[j][n]);
         }
         ibi += num_blocks_per_row;
     }

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

@@ -68,6 +68,7 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 
 #ifdef MUL_MAT_ID
 layout (binding = 3) readonly buffer IDS {int data_ids[];};
+layout (binding = 4) readonly buffer Counts {int data_expert_count[];};
 #endif
 
 layout (push_constant) uniform parameter
@@ -135,13 +136,19 @@ shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
 #include "mul_mm_funcs.glsl"
 
 void main() {
+    const uint ic = gl_WorkGroupID.y;
+
+#ifdef MUL_MAT_ID
+    const uint expert_idx = gl_GlobalInvocationID.z;
+    if (ic * BN >= data_expert_count[expert_idx]) {
+        return;
+    }
+#endif
 #ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
 #endif
 
-#ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.z;
-#else
+#ifndef MUL_MAT_ID
     const uint batch_idx = gl_GlobalInvocationID.z;
 
     const uint i13 = batch_idx / p.ne12;
@@ -156,7 +163,6 @@ void main() {
     const uint blocks_m = (p.M + BM - 1) / BM;
     const uint ir = gl_WorkGroupID.x % blocks_m;
     const uint ik = gl_WorkGroupID.x / blocks_m;
-    const uint ic = gl_WorkGroupID.y;
 
     const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER);
     const uint WSUBM = WM / WMITER;

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

@@ -92,6 +92,7 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 
 #ifdef MUL_MAT_ID
 layout (binding = 3) readonly buffer IDS {int data_ids[];};
+layout (binding = 4) readonly buffer Counts {int data_expert_count[];};
 
 shared u16vec4 row_ids[BN];
 
@@ -107,11 +108,7 @@ B_TYPE decodeFuncB(const in decodeBufB bl, const in uint blockCoords[2], const i
 {
     const uint row_i = blockCoords[0];
 
-    if (row_i >= _ne1) {
-        return B_TYPE(0.0);
-    }
-
-    const u16vec4 row_idx = row_ids[row_i & (BN - 1)];
+    const u16vec4 row_idx = row_ids[row_i];
     B_TYPE ret = data_b[row_idx.y * p.batch_stride_b + row_idx.x * p.stride_b + blockCoords[1]];
 
     return ret;
@@ -138,6 +135,8 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
     uint ids[16];
     uint iter = 0;
 
+    uint expert_count = data_expert_count[expert_idx];
+
     for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
         // prefetch up to 16 elements
         if (iter == 0) {
@@ -185,7 +184,7 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
         }
         _ne1 += total;
         iter &= 15;
-        if (_ne1 >= (ic + 1) * BN) {
+        if (_ne1 >= (ic + 1) * BN || _ne1 == expert_count) {
             break;
         }
     }
@@ -194,15 +193,28 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
 #endif
 
 void main() {
+    const uint tid = gl_LocalInvocationIndex;
+    const uint ic = gl_WorkGroupID.y;
+
+#ifdef MUL_MAT_ID
+    const uint expert_idx = gl_GlobalInvocationID.z;
+    if (ic * BN >= data_expert_count[expert_idx]) {
+        return;
+    }
+    // initialize to row 0 so we don't need to bounds check
+    if (tid < BN) {
+        row_ids[tid] = u16vec4(0);
+    }
+#if !defined(NEEDS_INIT_IQ_SHMEM)
+    barrier();
+#endif
+#endif
+
 #ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
 #endif
 
-    const uint tid = gl_LocalInvocationIndex;
-
-#ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.z;
-#else
+#ifndef MUL_MAT_ID
     const uint batch_idx = gl_GlobalInvocationID.z;
 
     const uint i13 = batch_idx / p.ne12;
@@ -217,7 +229,6 @@ void main() {
     const uint blocks_m = (p.M + BM - 1) / BM;
     const uint ir = gl_WorkGroupID.x % blocks_m;
     const uint ik = gl_WorkGroupID.x / blocks_m;
-    const uint ic = gl_WorkGroupID.y;
 
 #ifdef MUL_MAT_ID
     if (bitCount(p.nei0) == 1) {
@@ -482,7 +493,7 @@ void main() {
                     coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
-                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);
+                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);
 
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 } else {
@@ -490,7 +501,7 @@ void main() {
                     coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
-                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);
+                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);
 
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 }
@@ -526,7 +537,7 @@ void main() {
                     coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
-                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
+                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
 
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 } else {
@@ -534,7 +545,7 @@ void main() {
                     coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
-                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
+                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
 
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 }
@@ -571,7 +582,7 @@ void main() {
 
                 coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
 #ifdef MUL_MAT_ID
-                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
 #else
                 coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
 #endif
@@ -583,7 +594,7 @@ void main() {
 
                 coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
 #ifdef MUL_MAT_ID
-                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
 #else
                 coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
 #endif

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

@@ -159,14 +159,16 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const uint is = iqs / 8;                     // 0..15
             const uint halfsplit = ((iqs % 64) / 16);    // 0,1,2,3
             const uint qsshift = halfsplit * 2;          // 0,2,4,6
-            const uint m = 1 << (4 * n + halfsplit);     // 1,2,4,8,16,32,64,128
 
             const int8_t us = int8_t(((data_a[ib].scales[is % 8] >> (4 * int(is / 8))) & 0xF)
                                   | (((data_a[ib].scales[8 + (is % 4)] >> (2 * int(is / 4))) & 3) << 4));
             const float dl = float(data_a[ib].d) * float(us - 32);
 
-            buf_a[buf_idx] = FLOAT_TYPE_VEC2(dl * float(int8_t((data_a[ib].qs[qsi    ] >> qsshift) & 3) - (((data_a[ib].hmask[hmi    ] & m) != 0) ? 0 : 4)),
-                                             dl * float(int8_t((data_a[ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[ib].hmask[hmi + 1] & m) != 0) ? 0 : 4)));
+            const vec2 qs = vec2(unpack8((uint(data_a_packed16[ib].qs[qsi / 2]) >> qsshift) & 0x0303).xy);
+            const vec2 hm = vec2(unpack8(((uint(data_a_packed16[ib].hmask[hmi / 2]) >> (4 * n + halfsplit)) & 0x0101 ^ 0x0101) << 2).xy);
+
+            buf_a[buf_idx] = FLOAT_TYPE_VEC2(dl * (qs.x - hm.x),
+                                             dl * (qs.y - hm.y));
 #elif defined(DATA_A_Q4_K)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
             const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
@@ -198,8 +200,10 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const float d = loadd.x * sc;
             const float m = -loadd.y * mbyte;
 
-            buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF), m),
-                                             fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF), m));
+            const vec2 q = vec2(unpack8((uint(data_a_packed16[ib].qs[qsi / 2]) >> (b * 4)) & 0x0F0F).xy);
+
+            buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, q.x, m),
+                                             fma(d, q.y, m));
 #elif defined(DATA_A_Q5_K)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
             const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
@@ -213,8 +217,6 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
             const uint qhi = (iqs % 16) * 2;           // 0,2,4..30
 
-            const uint8_t hm = uint8_t(1 << (iqs / 16));
-
             const vec2 loadd = vec2(data_a[ib].dm);
 
             const uint scidx0 = (is < 4) ? is : (is + 4);
@@ -234,8 +236,12 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const float d = loadd.x * sc;
             const float m = -loadd.y * mbyte;
 
-            buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi    ] & hm) != 0 ? 16 : 0), m),
-                                             fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0), m));
+            const uint qs = (uint(data_a_packed16[ib].qs[qsi / 2]) >> (b * 4)) & 0x0F0F;
+            const uint qh = ((uint(data_a_packed16[ib].qh[qhi / 2]) >> (iqs / 16)) & 0x0101) << 4;
+            const vec2 q = vec2(unpack8(qs | qh).xy);
+
+            buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, q.x, m),
+                                             fma(d, q.y, m));
 #elif defined(DATA_A_Q6_K)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
             const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
@@ -394,11 +400,9 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
 
             const float d = float(data_a[ib].d);
             const uint qs = data_a[ib].qs[iqs];
-            const uint signs = pack32(u8vec4(
-                data_a[ib].qs[is+0],
-                data_a[ib].qs[is+1],
-                data_a[ib].qs[is+2],
-                data_a[ib].qs[is+3]
+            const uint signs = pack32(u16vec2(
+                data_a_packed16[ib].qs[is/2],
+                data_a_packed16[ib].qs[is/2+1]
             ));
             const float db = d * 0.5 * (0.5 + (signs >> 28));
             const uint32_t sign7 = bitfieldExtract(signs, 7 * (int(iqs / 2) % 4), 7);
@@ -443,8 +447,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const uint sl = (data_a[ib].scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF;
             const uint sh = ((data_a[ib].scales_h) >> (2 * ib32)) & 3;
             const uint qshift = (idx & 8) >> 1;
-            u8vec2 qs = u8vec2(data_a[ib].qs[iq], data_a[ib].qs[iq + 1]);
-            qs = (qs >> qshift) & uint8_t(0xF);
+            u8vec2 qs = unpack8((uint(data_a_packed16[ib].qs[iq/2]) >> qshift) & 0x0F0F).xy;
 
             const float d = float(data_a[ib].d);
             const vec2 v = d * float(int(sl | (sh << 4)) - 32) * vec2(kvalues_iq4nl[qs.x], kvalues_iq4nl[qs.y]);

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

@@ -13,6 +13,8 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
     uint ids[16];
     uint iter = 0;
 
+    uint expert_count = data_expert_count[expert_idx];
+
     for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
         // prefetch up to 16 elements
         if (iter == 0) {
@@ -60,7 +62,7 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
         }
         _ne1 += total;
         iter &= 15;
-        if (_ne1 >= (ic + 1) * BN) {
+        if (_ne1 >= (ic + 1) * BN || _ne1 == expert_count) {
             break;
         }
     }

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

@@ -35,6 +35,7 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 
 #ifdef MUL_MAT_ID
 layout (binding = 3) readonly buffer IDS {int data_ids[];};
+layout (binding = 4) readonly buffer Counts {int data_expert_count[];};
 #endif
 
 layout (push_constant) uniform parameter
@@ -104,13 +105,19 @@ block_b_cache cache_b;
 #include "mul_mmq_funcs.glsl"
 
 void main() {
+    const uint ic = gl_WorkGroupID.y;
+
+#ifdef MUL_MAT_ID
+    const uint expert_idx = gl_GlobalInvocationID.z;
+    if (ic * BN >= data_expert_count[expert_idx]) {
+        return;
+    }
+#endif
 #ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
 #endif
 
-#ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.z;
-#else
+#ifndef MUL_MAT_ID
     const uint batch_idx = gl_GlobalInvocationID.z;
 
     const uint i13 = batch_idx / p.ne12;
@@ -125,7 +132,6 @@ void main() {
     const uint blocks_m = (p.M + BM - 1) / BM;
     const uint ir = gl_WorkGroupID.x % blocks_m;
     const uint ik = gl_WorkGroupID.x / blocks_m;
-    const uint ic = gl_WorkGroupID.y;
 
     const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER);
     const uint WSUBM = WM / WMITER;

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

@@ -49,8 +49,8 @@ void rope_norm(const uint i0, const uint i1, rope_params p) {
     uint idst = i1*ne0 + i0;
     const uint ix = rope_a_coord(i0, i01, i02, p);
 
-    // Fusion optimization: ROPE + VIEW + SET_ROWS..
-    // The rope output is viewed as a 1D tensor and offset based on a row index in data_i.
+    // Fusion optimization: ROPE + VIEW + SET_ROWS.
+    // The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
     if (p.set_rows_stride != 0) {
         idst = i01*ne0 + i0;
         idst += rope_data_i[i02].x * p.set_rows_stride;
@@ -91,7 +91,7 @@ void rope_neox(const uint i0, const uint i1, rope_params p) {
     uint idst = i1*ne0 + i0/2;
     const uint ix = rope_a_coord(i0/2, i01, i02, p);
 
-    // Fusion optimization: ROPE + VIEW + SET_ROWS..
+    // Fusion optimization: ROPE + VIEW + SET_ROWS.
     // The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
     if (p.set_rows_stride != 0) {
         idst = i01*ne0 + i0/2;
@@ -132,9 +132,16 @@ void rope_multi(const uint i0, const uint i1, rope_params p) {
     const uint i01 = i1 % ne1;
     const uint i02 = i1 / ne1;
 
-    const uint idst = i1*ne0 + i0/2;
+    uint idst = i1*ne0 + i0/2;
     const uint ix = rope_a_coord(i0/2, i01, i02, p);
 
+    // Fusion optimization: ROPE + VIEW + SET_ROWS.
+    // The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
+    if (p.set_rows_stride != 0) {
+        idst = i01*ne0 + i0/2;
+        idst += rope_data_i[i02].x * p.set_rows_stride;
+    }
+
     if (i0 >= p.n_dims) {
         rope_data_d[idst + i0/2 + 0] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 0]);
         rope_data_d[idst + i0/2 + 1] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 1]);

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

@@ -6,6 +6,9 @@
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
     // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x;
+    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (i1 >= pc.nrows) {
+        return;
+    }
     rope_multi(i0, i1, pc);
 }

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

@@ -6,6 +6,9 @@
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
     // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x;
+    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (i1 >= pc.nrows) {
+        return;
+    }
     rope_neox(i0, i1, pc);
 }

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

@@ -6,6 +6,9 @@
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
     // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x;
+    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (i1 >= pc.nrows) {
+        return;
+    }
     rope_norm(i0, i1, pc);
 }

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

@@ -6,6 +6,7 @@
 struct rope_params {
     uint rope_mode;
     uint ncols;
+    uint nrows;
     uint n_dims;
     float freq_scale;
     uint p_delta_rows;

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

@@ -6,6 +6,9 @@
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
     // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x;
+    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (i1 >= pc.nrows) {
+        return;
+    }
     rope_vision(i0, i1, pc);
 }

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

@@ -172,16 +172,12 @@ struct block_q8_0
     float16_t d;
     int8_t qs[32];
 };
+
 struct block_q8_0_packed16
 {
     float16_t d;
     int16_t qs[32/2];
 };
-struct block_q8_0_packed32
-{
-    float16_t d;
-    int32_t qs[32/4];
-};
 
 #if defined(DATA_A_Q8_0)
 #define QUANT_K QUANT_K_Q8_0
@@ -189,7 +185,6 @@ struct block_q8_0_packed32
 #define QUANT_AUXF 1
 #define A_TYPE block_q8_0
 #define A_TYPE_PACKED16 block_q8_0_packed16
-#define A_TYPE_PACKED32 block_q8_0_packed32
 #define DATA_A_QUANT_LEGACY
 #endif
 
@@ -201,11 +196,13 @@ struct block_q8_1
     f16vec2 ds;
     int8_t qs[32];
 };
+
 struct block_q8_1_packed16
 {
     f16vec2 ds;
     int16_t qs[16];
 };
+
 struct block_q8_1_packed32
 {
     f16vec2 ds;
@@ -218,6 +215,7 @@ struct block_q8_1_x4
     f16vec2 ds[4];
     int32_t qs[32];
 };
+
 struct block_q8_1_x4_packed128
 {
     f16vec2 ds[4];
@@ -1346,10 +1344,28 @@ struct block_iq4_xs
     uint8_t qs[QUANT_K_IQ4_XS/2];
 };
 
+struct block_iq4_xs_packed16
+{
+    float16_t d;
+    uint16_t scales_h;
+    uint16_t scales_l[QUANT_K_IQ4_XS/128];
+    uint16_t qs[QUANT_K_IQ4_XS/4];
+};
+
+struct block_iq4_xs_packed32
+{
+    float16_t d;
+    uint16_t scales_h;
+    uint32_t scales_l;
+    uint32_t qs[QUANT_K_IQ4_XS/8];
+};
+
 #if defined(DATA_A_IQ4_XS)
 #define QUANT_K QUANT_K_IQ4_XS
 #define QUANT_R QUANT_R_IQ4_XS
 #define A_TYPE block_iq4_xs
+#define A_TYPE_PACKED16 block_iq4_xs_packed16
+#define A_TYPE_PACKED32 block_iq4_xs_packed32
 #endif
 
 #define QUANT_K_IQ4_NL 32

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

@@ -21,6 +21,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 #define NEAREST  0
 #define BILINEAR 1
 #define BICUBIC  2
+#define BILINEAR_ANTIALIAS 513
 
 layout (constant_id = 0) const uint scale_mode = 0;
 
@@ -62,6 +63,56 @@ float interpolate_bilinear(uint i10, uint i11, uint i12, uint i13) {
     return fetch_bilinear(c0, c1, d, i12, i13);
 }
 
+float triangle_filter(float x) {
+    return max(1.0f - abs(x), 0.0f);
+}
+
+float interpolate_bilinear_antialias(uint i10, uint i11, uint i12, uint i13) {
+    const float support1  = max(1.0f, 1.0f / p.sf1);
+    const float invscale1 = 1.0f / support1;
+    const float support0  = max(1.0f, 1.0f / p.sf0);
+    const float invscale0 = 1.0f / support0;
+
+    const uint i02 = uint(i12 / p.sf2);
+    const uint i03 = uint(i13 / p.sf3);
+
+    const float y = (float(i11) + p.pixel_offset) / p.sf1;
+    const float x = (float(i10) + p.pixel_offset) / p.sf0;
+
+    // the range of source pixels that contribute
+    const int x_min = max(int(x - support0 + p.pixel_offset), 0);
+    const int x_max = min(int(x + support0 + p.pixel_offset), int(p.ne00));
+    const int y_min = max(int(y - support1 + p.pixel_offset), 0);
+    const int y_max = min(int(y + support1 + p.pixel_offset), int(p.ne01));
+
+    // bilinear filter with antialiasing
+    float val = 0.0f;
+    float total_weight = 0.0f;
+
+    for (int sy = y_min; sy < y_max; sy++) {
+        const float weight_y = triangle_filter((sy - y + p.pixel_offset) * invscale1);
+
+        for (int sx = x_min; sx < x_max; sx++) {
+            const float weight_x = triangle_filter((sx - x + p.pixel_offset) * invscale0);
+            const float weight = weight_x * weight_y;
+
+            if (weight <= 0.0f) {
+                continue;
+            }
+
+            const float pixel = data_a[p.a_offset + i03 * p.nb03 + i02 * p.nb02 + sy * p.nb01 + sx * p.nb00];
+            val += pixel * weight;
+            total_weight += weight;
+        }
+    }
+
+    if (total_weight > 0.0f) {
+        val /= total_weight;
+    }
+
+    return val;
+}
+
 // Bicubic interpolation with alpha = -0.75
 // https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm
 const vec4 bcoeffs1 = vec4( 1.25, -2.25,  0.0, 1.0);
@@ -118,6 +169,9 @@ void main() {
         case BICUBIC:
             result = interpolate_bicubic(i10, i11, i12, i13);
             break;
+        case BILINEAR_ANTIALIAS:
+            result = interpolate_bilinear_antialias(i10, i11, i12, i13);
+            break;
     }
 
     data_d[p.d_offset + idx] = D_TYPE(result);

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

@@ -853,6 +853,8 @@ void process_shaders() {
     string_to_spv("hardswish_f32",  "hardswish.comp",   {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("abs_f16",        "abs.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("abs_f32",        "abs.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("xielu_f16",      "xielu.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("xielu_f32",      "xielu.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
 
     string_to_spv("tri_f16",        "tri.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("tri_f32",        "tri.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
@@ -925,6 +927,8 @@ void process_shaders() {
     string_to_spv("rope_multi_f32", "rope_multi.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
     string_to_spv("rope_multi_f16", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
     string_to_spv("rope_multi_f16_rte", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_multi_f32_f16", "rope_multi.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}});
+    string_to_spv("rope_multi_f32_f16_rte", "rope_multi.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
     string_to_spv("rope_vision_f32", "rope_vision.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
     string_to_spv("rope_vision_f16", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
@@ -941,6 +945,8 @@ void process_shaders() {
     string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}}));
     string_to_spv("cumsum_f32", "cumsum.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
+    string_to_spv("count_experts", "count_experts.comp", merge_maps(base_dict, {{"A_TYPE", "uint"}, {"D_TYPE", "uint"}}));
+
     for (std::string dim_str : {"", "_3d"}) {
         for (bool bda : {false, true}) {
             std::string bda_str = bda ? "_bda" : "";

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

@@ -0,0 +1,35 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    float x = float(data_a[i]);
+
+    float alpha_n = p.param1;
+    float alpha_p = p.param2;
+    float beta = p.param3;
+    float eps = p.param4;
+
+    if (x > 0.0f) {
+        x = alpha_p * x * x + beta * x;
+    } else {
+        const float min_x_eps = min(x, eps);
+        x = (exp(min_x_eps) - 1 - x) * alpha_n + beta * x;
+    }
+
+    data_d[i] = D_TYPE(x);
+}

gguf-py/gguf/constants.py

@@ -181,6 +181,7 @@ class Keys:
         DIMENSION_COUNT          = "{arch}.rope.dimension_count"
         DIMENSION_SECTIONS       = "{arch}.rope.dimension_sections"
         FREQ_BASE                = "{arch}.rope.freq_base"
+        FREQ_BASE_SWA            = "{arch}.rope.freq_base_swa"
         SCALING_TYPE             = "{arch}.rope.scaling.type"
         SCALING_FACTOR           = "{arch}.rope.scaling.factor"
         SCALING_ATTN_FACTOR      = "{arch}.rope.scaling.attn_factor"
@@ -354,6 +355,7 @@ class MODEL_ARCH(IntEnum):
     STARCODER        = auto()
     REFACT           = auto()
     BERT             = auto()
+    MODERN_BERT      = auto()
     NOMIC_BERT       = auto()
     NOMIC_BERT_MOE   = auto()
     NEO_BERT         = auto()
@@ -375,6 +377,7 @@ class MODEL_ARCH(IntEnum):
     PHIMOE           = auto()
     PLAMO            = auto()
     PLAMO2           = auto()
+    PLAMO3           = auto()
     CODESHELL        = auto()
     ORION            = auto()
     INTERNLM2        = auto()
@@ -447,6 +450,8 @@ class MODEL_ARCH(IntEnum):
     RND1             = auto()
     PANGU_EMBED      = auto()
     MISTRAL3         = auto()
+    MIMO2            = auto()
+    LLAMA_EMBED      = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -747,6 +752,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.STARCODER:        "starcoder",
     MODEL_ARCH.REFACT:           "refact",
     MODEL_ARCH.BERT:             "bert",
+    MODEL_ARCH.MODERN_BERT:      "modern-bert",
     MODEL_ARCH.NOMIC_BERT:       "nomic-bert",
     MODEL_ARCH.NOMIC_BERT_MOE:   "nomic-bert-moe",
     MODEL_ARCH.NEO_BERT:         "neo-bert",
@@ -768,6 +774,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.PHIMOE:           "phimoe",
     MODEL_ARCH.PLAMO:            "plamo",
     MODEL_ARCH.PLAMO2:           "plamo2",
+    MODEL_ARCH.PLAMO3:           "plamo3",
     MODEL_ARCH.CODESHELL:        "codeshell",
     MODEL_ARCH.ORION:            "orion",
     MODEL_ARCH.INTERNLM2:        "internlm2",
@@ -841,6 +848,8 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.RND1:             "rnd1",
     MODEL_ARCH.PANGU_EMBED:      "pangu-embedded",
     MODEL_ARCH.MISTRAL3:         "mistral3",
+    MODEL_ARCH.MIMO2:            "mimo2",
+    MODEL_ARCH.LLAMA_EMBED:      "llama-embed",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -1367,6 +1376,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.CLS,
         MODEL_TENSOR.CLS_OUT,
     ],
+    MODEL_ARCH.MODERN_BERT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.CLS,
+        MODEL_TENSOR.CLS_OUT,
+    ],
     MODEL_ARCH.NOMIC_BERT: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.TOKEN_EMBD_NORM,
@@ -1743,6 +1765,21 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.SSM_B_NORM,
         MODEL_TENSOR.SSM_C_NORM,
     ],
+    MODEL_ARCH.PLAMO3: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_POST_NORM,
+    ],
     MODEL_ARCH.GPT2: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.POS_EMBD,
@@ -3180,6 +3217,46 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
     ],
+    MODEL_ARCH.MIMO2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_SINKS,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+    ],
+    MODEL_ARCH.LLAMA_EMBED: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+    ],
     # TODO
 }