Merge branch 'master' into compilade/bitnet-ternary

.devops/full-cuda.Dockerfile

@@ -1,18 +1,16 @@
 ARG UBUNTU_VERSION=22.04
-
 # This needs to generally match the container host's environment.
-ARG CUDA_VERSION=11.7.1
+ARG CUDA_VERSION=12.6.0
-
 # Target the CUDA build image
 ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 
 FROM ${BASE_CUDA_DEV_CONTAINER} AS build
 
-# Unless otherwise specified, we make a fat build.
+# CUDA architecture to build for (defaults to all supported archs)
-ARG CUDA_DOCKER_ARCH=all
+ARG CUDA_DOCKER_ARCH=default
 
 RUN apt-get update && \
-    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1
+    apt-get install -y build-essential cmake python3 python3-pip git libcurl4-openssl-dev libgomp1
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -24,13 +22,12 @@ WORKDIR /app
 
 COPY . .
 
-# Set nvcc architecture
+# Use the default CUDA archs if not specified
-ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
-# Enable CUDA
+        export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
-ENV GGML_CUDA=1
+    fi && \
-# Enable cURL
+    cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
-ENV LLAMA_CURL=1
+    cmake --build build --config Release -j$(nproc) && \
-
+    cp build/bin/* .
-RUN make -j$(nproc)
 
 ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/llama-cli-cuda.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG CUDA_VERSION=11.7.1
+ARG CUDA_VERSION=12.6.0
 # Target the CUDA build image
 ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 # Target the CUDA runtime image
@@ -8,28 +8,30 @@ ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_V
 
 FROM ${BASE_CUDA_DEV_CONTAINER} AS build
 
-# Unless otherwise specified, we make a fat build.
+# CUDA architecture to build for (defaults to all supported archs)
-ARG CUDA_DOCKER_ARCH=all
+ARG CUDA_DOCKER_ARCH=default
 
 RUN apt-get update && \
-    apt-get install -y build-essential git
+    apt-get install -y build-essential git cmake
 
 WORKDIR /app
 
 COPY . .
 
-# Set nvcc architecture
+# Use the default CUDA archs if not specified
-ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
-# Enable CUDA
+        export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
-ENV GGML_CUDA=1
+    fi && \
-
+    cmake -B build -DGGML_CUDA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
-RUN make -j$(nproc) llama-cli
+    cmake --build build --config Release --target llama-cli -j$(nproc)
 
 FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
 
 RUN apt-get update && \
     apt-get install -y libgomp1
 
-COPY --from=build /app/llama-cli /llama-cli
+COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
+COPY --from=build /app/build/src/libllama.so /libllama.so
+COPY --from=build /app/build/bin/llama-cli /llama-cli
 
 ENTRYPOINT [ "/llama-cli" ]

.devops/llama-server-cuda.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG CUDA_VERSION=11.7.1
+ARG CUDA_VERSION=12.6.0
 # Target the CUDA build image
 ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 # Target the CUDA runtime image
@@ -8,31 +8,34 @@ ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_V
 
 FROM ${BASE_CUDA_DEV_CONTAINER} AS build
 
-# Unless otherwise specified, we make a fat build.
+# CUDA architecture to build for (defaults to all supported archs)
-ARG CUDA_DOCKER_ARCH=all
+ARG CUDA_DOCKER_ARCH=default
 
 RUN apt-get update && \
-    apt-get install -y build-essential git libcurl4-openssl-dev
+    apt-get install -y build-essential git cmake libcurl4-openssl-dev
 
 WORKDIR /app
 
 COPY . .
 
-# Set nvcc architecture
+# Use the default CUDA archs if not specified
-ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
-# Enable CUDA
+        export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
-ENV GGML_CUDA=1
+    fi && \
-# Enable cURL
+    cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
-ENV LLAMA_CURL=1
+    cmake --build build --config Release --target llama-server -j$(nproc)
-
-RUN make -j$(nproc) llama-server
 
 FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
 
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev libgomp1 curl
 
-COPY --from=build /app/llama-server /llama-server
+COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
+COPY --from=build /app/build/src/libllama.so /libllama.so
+COPY --from=build /app/build/bin/llama-server /llama-server
+
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
 
 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
 

.devops/llama-server-intel.Dockerfile

@@ -26,6 +26,8 @@ RUN apt-get update && \
 COPY --from=build /app/build/bin/llama-server /llama-server
 
 ENV LC_ALL=C.utf8
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
 
 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
 

.devops/llama-server-rocm.Dockerfile

@@ -39,6 +39,8 @@ ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
 ENV GGML_HIPBLAS=1
 ENV CC=/opt/rocm/llvm/bin/clang
 ENV CXX=/opt/rocm/llvm/bin/clang++
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
 
 # Enable cURL
 ENV LLAMA_CURL=1

.devops/llama-server-vulkan.Dockerfile

@@ -23,6 +23,8 @@ RUN cp /app/build/bin/llama-server /llama-server && \
     rm -rf /app
 
 ENV LC_ALL=C.utf8
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
 
 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
 

.devops/llama-server.Dockerfile

@@ -21,6 +21,8 @@ RUN apt-get update && \
 COPY --from=build /app/llama-server /llama-server
 
 ENV LC_ALL=C.utf8
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
 
 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
 

.devops/nix/devshells.nix

@@ -1,13 +1,52 @@
+{ inputs, ... }:
+
 {
   perSystem =
-    { config, lib, ... }:
+    {
+      config,
+      lib,
+      system,
+      ...
+    }:
     {
       devShells =
-        lib.concatMapAttrs
+        let
-          (name: package: {
+          pkgs = import inputs.nixpkgs { inherit system; };
-            ${name} = package.passthru.shell;
+          stdenv = pkgs.stdenv;
-            ${name + "-extra"} = package.passthru.shell-extra;
+          scripts = config.packages.python-scripts;
-          })
+        in
-          config.packages;
+        lib.pipe (config.packages) [
+          (lib.concatMapAttrs (
+            name: package: {
+              ${name} = pkgs.mkShell {
+                name = "${name}";
+                inputsFrom = [ package ];
+                shellHook = ''
+                  echo "Entering ${name} devShell"
+                '';
+              };
+              "${name}-extra" =
+                if (name == "python-scripts") then
+                  null
+                else
+                  pkgs.mkShell {
+                    name = "${name}-extra";
+                    inputsFrom = [
+                      package
+                      scripts
+                    ];
+                    # Extra packages that *may* be used by some scripts
+                    packages = [
+                        pkgs.python3Packages.tiktoken
+                    ];
+                    shellHook = ''
+                      echo "Entering ${name} devShell"
+                      addToSearchPath "LD_LIBRARY_PATH" "${lib.getLib stdenv.cc.cc}/lib"
+                    '';
+                  };
+            }
+          ))
+          (lib.filterAttrs (name: value: value != null))
+        ];
     };
 }

.devops/nix/nixpkgs-instances.nix

@@ -26,16 +26,14 @@
           config.cudaSupport = true;
           config.allowUnfreePredicate =
             p:
-            builtins.all
+            builtins.all (
-              (
               license:
               license.free
               || builtins.elem license.shortName [
                 "CUDA EULA"
                 "cuDNN EULA"
               ]
-              )
+            ) (p.meta.licenses or [ p.meta.license ]);
-              (p.meta.licenses or [ p.meta.license ]);
         };
         # Ensure dependencies use ROCm consistently
         pkgsRocm = import inputs.nixpkgs {

.devops/nix/package-gguf-py.nix

@@ -0,0 +1,36 @@
+{
+  lib,
+  llamaVersion,
+  numpy,
+  tqdm,
+  sentencepiece,
+  pyyaml,
+  poetry-core,
+  buildPythonPackage,
+  pytestCheckHook,
+}:
+
+buildPythonPackage {
+  pname = "gguf";
+  version = llamaVersion;
+  pyproject = true;
+  nativeBuildInputs = [ poetry-core ];
+  propagatedBuildInputs = [
+    numpy
+    tqdm
+    sentencepiece
+    pyyaml
+  ];
+  src = lib.cleanSource ../../gguf-py;
+  pythonImportsCheck = [
+    "numpy"
+    "gguf"
+  ];
+  nativeCheckInputs = [ pytestCheckHook ];
+  doCheck = true;
+  meta = with lib; {
+    description = "Python package for writing binary files in the GGUF format";
+    license = licenses.mit;
+    maintainers = [ maintainers.ditsuke ];
+  };
+}

.devops/nix/package.nix

@@ -3,31 +3,33 @@
   glibc,
   config,
   stdenv,
-  mkShell,
   runCommand,
   cmake,
   ninja,
   pkg-config,
   git,
-  python3,
   mpi,
   blas,
   cudaPackages,
+  autoAddDriverRunpath,
   darwin,
   rocmPackages,
   vulkan-headers,
   vulkan-loader,
   curl,
   shaderc,
-  useBlas ? builtins.all (x: !x) [
+  useBlas ?
+    builtins.all (x: !x) [
       useCuda
       useMetalKit
       useRocm
       useVulkan
-  ] && blas.meta.available,
+    ]
+    && blas.meta.available,
   useCuda ? config.cudaSupport,
   useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin,
-  useMpi ? false, # Increases the runtime closure size by ~700M
+  # Increases the runtime closure size by ~700M
+  useMpi ? false,
   useRocm ? config.rocmSupport,
   enableCurl ? true,
   useVulkan ? false,
@@ -37,8 +39,8 @@
   # otherwise we get libstdc++ errors downstream.
   effectiveStdenv ? if useCuda then cudaPackages.backendStdenv else stdenv,
   enableStatic ? effectiveStdenv.hostPlatform.isStatic,
-  precompileMetalShaders ? false
+  precompileMetalShaders ? false,
-}@inputs:
+}:
 
 let
   inherit (lib)
@@ -46,7 +48,6 @@ let
     cmakeFeature
     optionals
     strings
-    versionOlder
     ;
 
   stdenv = throw "Use effectiveStdenv instead";
@@ -62,52 +63,9 @@ let
   pnameSuffix =
     strings.optionalString (suffices != [ ])
       "-${strings.concatMapStringsSep "-" strings.toLower suffices}";
-  descriptionSuffix =
+  descriptionSuffix = strings.optionalString (
-    strings.optionalString (suffices != [ ])
+    suffices != [ ]
-      ", accelerated with ${strings.concatStringsSep ", " suffices}";
+  ) ", accelerated with ${strings.concatStringsSep ", " suffices}";
-
-  executableSuffix = effectiveStdenv.hostPlatform.extensions.executable;
-
-  # TODO: package the Python in this repository in a Nix-like way.
-  # It'd be nice to migrate to buildPythonPackage, as well as ensure this repo
-  # is PEP 517-compatible, and ensure the correct .dist-info is generated.
-  # https://peps.python.org/pep-0517/
-  #
-  # TODO: Package up each Python script or service appropriately, by making
-  # them into "entrypoints"
-  llama-python = python3.withPackages (
-    ps: [
-      ps.numpy
-      ps.sentencepiece
-    ]
-  );
-
-  # TODO(Green-Sky): find a better way to opt-into the heavy ml python runtime
-  llama-python-extra = python3.withPackages (
-    ps: [
-      ps.numpy
-      ps.sentencepiece
-      ps.tiktoken
-      ps.torchWithoutCuda
-      ps.transformers
-
-      # server bench
-      ps.matplotlib
-
-      # server tests
-      ps.openai
-      ps.behave
-      ps.prometheus-client
-
-      # for examples/pydantic-models-to-grammar-examples.py
-      ps.docstring-parser
-      ps.pydantic
-
-      # for scripts/compare-llama-bench.py
-      ps.gitpython
-      ps.tabulate
-    ]
-  );
 
   xcrunHost = runCommand "xcrunHost" { } ''
     mkdir -p $out/bin
@@ -144,8 +102,7 @@ let
   ];
 in
 
-effectiveStdenv.mkDerivation (
+effectiveStdenv.mkDerivation (finalAttrs: {
-  finalAttrs: {
   pname = "llama-cpp${pnameSuffix}";
   version = llamaVersion;
 
@@ -193,15 +150,10 @@ effectiveStdenv.mkDerivation (
     ++ optionals useCuda [
       cudaPackages.cuda_nvcc
 
-        # TODO: Replace with autoAddDriverRunpath
+      autoAddDriverRunpath
-        # once https://github.com/NixOS/nixpkgs/pull/275241 has been merged
-        cudaPackages.autoAddOpenGLRunpathHook
     ]
-      ++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [
+    ++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [ glibc.static ]
-        glibc.static
+    ++ optionals (effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders) [ xcrunHost ];
-      ] ++ optionals (effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders) [
-        xcrunHost
-      ];
 
   buildInputs =
     optionals effectiveStdenv.isDarwin darwinBuildInputs
@@ -256,35 +208,6 @@ effectiveStdenv.mkDerivation (
     cp $src/include/llama.h $out/include/
   '';
 
-    # Define the shells here, but don't add in the inputsFrom to avoid recursion.
-    passthru = {
-      inherit
-        useBlas
-        useCuda
-        useMetalKit
-        useMpi
-        useRocm
-        useVulkan
-        ;
-
-      shell = mkShell {
-        name = "shell-${finalAttrs.finalPackage.name}";
-        description = "contains numpy and sentencepiece";
-        buildInputs = [ llama-python ];
-        inputsFrom = [ finalAttrs.finalPackage ];
-        shellHook = ''
-          addToSearchPath "LD_LIBRARY_PATH" "${lib.getLib effectiveStdenv.cc.cc}/lib"
-        '';
-      };
-
-      shell-extra = mkShell {
-        name = "shell-extra-${finalAttrs.finalPackage.name}";
-        description = "contains numpy, sentencepiece, torchWithoutCuda, and transformers";
-        buildInputs = [ llama-python-extra ];
-        inputsFrom = [ finalAttrs.finalPackage ];
-      };
-    };
-
   meta = {
     # Configurations we don't want even the CI to evaluate. Results in the
     # "unsupported platform" messages. This is mostly a no-op, because
@@ -320,5 +243,4 @@ effectiveStdenv.mkDerivation (
     # Extend `badPlatforms` instead
     platforms = lib.platforms.all;
   };
-  }
+})
-)

.devops/nix/python-scripts.nix

@@ -0,0 +1,66 @@
+{
+  lib,
+  stdenv,
+  buildPythonPackage,
+  poetry-core,
+  mkShell,
+  python3Packages,
+  gguf-py,
+}@inputs:
+
+let
+  llama-python-deps = with python3Packages; [
+    numpy
+    sentencepiece
+    transformers
+    protobuf
+    torchWithoutCuda
+    gguf-py
+    tqdm
+
+    # for scripts/compare-llama-bench.py
+    gitpython
+    tabulate
+
+    # for examples/pydantic-models-to-grammar-examples.py
+    docstring-parser
+    pydantic
+
+  ];
+
+  llama-python-test-deps = with python3Packages; [
+    # Server bench
+    matplotlib
+
+    # server tests
+    openai
+    behave
+    prometheus-client
+  ];
+in
+
+buildPythonPackage ({
+  pname = "llama-scripts";
+  version = "0.0.0";
+  pyproject = true;
+
+  # NOTE: The files filtered out here are not visible in the build sandbox, neither
+  # do they affect the output hash. They can be modified without triggering a rebuild.
+  src = lib.cleanSourceWith {
+    filter =
+      name: type:
+      let
+        any = builtins.any (x: x);
+        baseName = builtins.baseNameOf name;
+      in
+      any [
+        (lib.hasSuffix ".py" name)
+        (baseName == "README.md")
+        (baseName == "pyproject.toml")
+      ];
+    src = lib.cleanSource ../../.;
+  };
+  nativeBuildInputs = [ poetry-core ];
+  nativeCheckInputs = llama-python-test-deps;
+  dependencies = llama-python-deps;
+})

.devops/nix/scope.nix

@@ -1,19 +1,41 @@
 {
   lib,
   newScope,
+  python3,
   llamaVersion ? "0.0.0",
 }:
 
+let
+  pythonPackages = python3.pkgs;
+  buildPythonPackage = pythonPackages.buildPythonPackage;
+  numpy = pythonPackages.numpy;
+  tqdm = pythonPackages.tqdm;
+  sentencepiece = pythonPackages.sentencepiece;
+  pyyaml = pythonPackages.pyyaml;
+  poetry-core = pythonPackages.poetry-core;
+  pytestCheckHook = pythonPackages.pytestCheckHook;
+in
+
 # We're using `makeScope` instead of just writing out an attrset
 # because it allows users to apply overlays later using `overrideScope'`.
 # Cf. https://noogle.dev/f/lib/makeScope
 
-lib.makeScope newScope (
+lib.makeScope newScope (self: {
-  self: {
   inherit llamaVersion;
+  gguf-py = self.callPackage ./package-gguf-py.nix {
+    inherit
+      buildPythonPackage
+      numpy
+      tqdm
+      sentencepiece
+      poetry-core
+      pyyaml
+      pytestCheckHook
+      ;
+  };
+  python-scripts = self.callPackage ./python-scripts.nix { inherit buildPythonPackage poetry-core; };
   llama-cpp = self.callPackage ./package.nix { };
   docker = self.callPackage ./docker.nix { };
   docker-min = self.callPackage ./docker.nix { interactive = false; };
   sif = self.callPackage ./sif.nix { };
-  }
+})
-)

.ecrc

@@ -1,5 +1,5 @@
 {
-  "Exclude": ["^\\.gitmodules$"],
+  "Exclude": ["^\\.gitmodules$", "stb_image\\.h"],
   "Disable": {
     "IndentSize": true
   }

.github/workflows/docker.yml

@@ -96,21 +96,12 @@ jobs:
         env:
           GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
 
-      - name: Build and push Docker image (versioned)
+      - name: Build and push Docker image (tagged + versioned)
         if: github.event_name == 'push'
-        uses: docker/build-push-action@v4
+        uses: docker/build-push-action@v6
         with:
           context: .
           push: true
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
+          tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
-          file: ${{ matrix.config.dockerfile }}
-
-      - name: Build and push Docker image (tagged)
-        uses: docker/build-push-action@v4
-        with:
-          context: .
-          push: ${{ github.event_name == 'push' }}
-          platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
           file: ${{ matrix.config.dockerfile }}

README.md

@@ -10,32 +10,14 @@
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
-> [!IMPORTANT]
-[2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggerganov/llama.cpp/pull/7809)
-
 ## Recent API changes
 
-- [2024 Jun 26] The source code and CMake build scripts have been restructured https://github.com/ggerganov/llama.cpp/pull/8006
+- [Changelog for `libllama` API](https://github.com/ggerganov/llama.cpp/issues/9289)
-- [2024 Apr 21] `llama_token_to_piece` can now optionally render special tokens https://github.com/ggerganov/llama.cpp/pull/6807
+- [Changelog for `llama-server` REST API](https://github.com/ggerganov/llama.cpp/issues/9291)
-- [2024 Apr 4] State and session file functions reorganized under `llama_state_*` https://github.com/ggerganov/llama.cpp/pull/6341
-- [2024 Mar 26] Logits and embeddings API updated for compactness https://github.com/ggerganov/llama.cpp/pull/6122
-- [2024 Mar 13] Add `llama_synchronize()` + `llama_context_params.n_ubatch` https://github.com/ggerganov/llama.cpp/pull/6017
-- [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_seq_max()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328
-- [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796
-- [2024 Mar 3] `struct llama_context_params` https://github.com/ggerganov/llama.cpp/pull/5849
 
 ## Hot topics
 
-- **`convert.py` has been deprecated and moved to `examples/convert_legacy_llama.py`, please use `convert_hf_to_gguf.py`** https://github.com/ggerganov/llama.cpp/pull/7430
+- *add hot topics here*
-- Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021
-- BPE pre-tokenization support has been added: https://github.com/ggerganov/llama.cpp/pull/6920
-- MoE memory layout has been updated - reconvert models for `mmap` support and regenerate `imatrix` https://github.com/ggerganov/llama.cpp/pull/6387
-- Model sharding instructions using `gguf-split` https://github.com/ggerganov/llama.cpp/discussions/6404
-- Fix major bug in Metal batched inference https://github.com/ggerganov/llama.cpp/pull/6225
-- Multi-GPU pipeline parallelism support https://github.com/ggerganov/llama.cpp/pull/6017
-- Looking for contributions to add Deepseek support: https://github.com/ggerganov/llama.cpp/issues/5981
-- Quantization blind testing: https://github.com/ggerganov/llama.cpp/discussions/5962
-- Initial Mamba support has been added: https://github.com/ggerganov/llama.cpp/pull/5328
 
 ----
 

ci/run.sh

@@ -13,6 +13,9 @@
 # # with SYCL support
 # GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
+# # with VULKAN support
+# GG_BUILD_VULKAN=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
 
 if [ -z "$2" ]; then
     echo "usage: $0 <output-dir> <mnt-dir>"
@@ -40,7 +43,7 @@ if [ ! -z ${GG_BUILD_METAL} ]; then
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=1"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native"
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then
@@ -52,6 +55,10 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then
 
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
 fi
+
+if [ ! -z ${GG_BUILD_VULKAN} ]; then
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1"
+fi
 ## helpers
 
 # download a file if it does not exist or if it is outdated
@@ -107,7 +114,7 @@ function gg_run_ctest_debug {
     gg_check_build_requirements
 
     (time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                          ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j$(nproc)                                  ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
 
@@ -138,7 +145,7 @@ function gg_run_ctest_release {
     gg_check_build_requirements
 
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     if [ -z ${GG_BUILD_LOW_PERF} ]; then
         (time ctest --output-on-failure -L main ) 2>&1 | tee -a $OUT/${ci}-ctest.log
@@ -266,7 +273,6 @@ function gg_sum_ctest_with_model_release {
 }
 
 # open_llama_7b_v2
-# requires: GG_BUILD_CUDA
 
 function gg_run_open_llama_7b_v2 {
     cd ${SRC}
@@ -290,8 +296,8 @@ function gg_run_open_llama_7b_v2 {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DGGML_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
@@ -425,7 +431,7 @@ function gg_run_pythia_1_4b {
     set -e
 
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
@@ -535,7 +541,6 @@ function gg_sum_pythia_1_4b {
 }
 
 # pythia_2_8b
-# requires: GG_BUILD_CUDA
 
 function gg_run_pythia_2_8b {
     cd ${SRC}
@@ -556,8 +561,8 @@ function gg_run_pythia_2_8b {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DGGML_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
@@ -692,7 +697,7 @@ function gg_run_embd_bge_small {
     set -e
 
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
@@ -761,7 +766,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
     fi
 
     if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
-        if [ -z ${GG_BUILD_CUDA} ]; then
+        if [ -z ${GG_BUILD_CUDA} ] && [ -z ${GG_BUILD_VULKAN} ]; then
             test $ret -eq 0 && gg_run pythia_1_4b
         else
             test $ret -eq 0 && gg_run pythia_2_8b

common/common.cpp

@@ -251,6 +251,57 @@ int32_t cpu_get_num_math() {
     return cpu_get_num_physical_cores();
 }
 
+// Helper for setting process priority
+
+#if defined(_WIN32)
+
+bool set_process_priority(enum ggml_sched_priority prio) {
+    if (prio == GGML_SCHED_PRIO_NORMAL) {
+        return true;
+    }
+
+    DWORD p = NORMAL_PRIORITY_CLASS;
+    switch (prio) {
+        case GGML_SCHED_PRIO_NORMAL:   p = NORMAL_PRIORITY_CLASS;       break;
+        case GGML_SCHED_PRIO_MEDIUM:   p = ABOVE_NORMAL_PRIORITY_CLASS; break;
+        case GGML_SCHED_PRIO_HIGH:     p = HIGH_PRIORITY_CLASS;         break;
+        case GGML_SCHED_PRIO_REALTIME: p = REALTIME_PRIORITY_CLASS;     break;
+    }
+
+    if (!SetPriorityClass(GetCurrentProcess(), p)) {
+        fprintf(stderr, "warn: failed to set process priority class %d : (%d)\n", prio, (int) GetLastError());
+        return false;
+    }
+
+    return true;
+}
+
+#else // MacOS and POSIX
+#include <sys/types.h>
+#include <sys/resource.h>
+
+bool set_process_priority(enum ggml_sched_priority prio) {
+    if (prio == GGML_SCHED_PRIO_NORMAL) {
+        return true;
+    }
+
+    int p = 0;
+    switch (prio) {
+        case GGML_SCHED_PRIO_NORMAL:   p =  0;  break;
+        case GGML_SCHED_PRIO_MEDIUM:   p = -5;  break;
+        case GGML_SCHED_PRIO_HIGH:     p = -10; break;
+        case GGML_SCHED_PRIO_REALTIME: p = -20; break;
+    }
+
+    if (!setpriority(PRIO_PROCESS, 0, p)) {
+        fprintf(stderr, "warn: failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno);
+        return false;
+    }
+    return true;
+}
+
+#endif
+
 //
 // CLI argument parsing
 //
@@ -277,6 +328,30 @@ void gpt_params_handle_model_default(gpt_params & params) {
     }
 }
 
+void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model) {
+    int32_t n_set = 0;
+
+    if (cpuparams.n_threads < 0) {
+        // Assuming everything about cpuparams is invalid
+        if (role_model != nullptr) {
+            cpuparams = *role_model;
+        } else {
+            cpuparams.n_threads = cpu_get_num_math();
+        }
+    }
+
+    for (int32_t i = 0; i < GGML_MAX_N_THREADS; i++) {
+        if (cpuparams.cpumask[i]) {
+            n_set++;
+        }
+    }
+
+    if (n_set && n_set < cpuparams.n_threads) {
+        // Not enough set bits, may experience performance issues.
+        fprintf(stderr, "warn: Not enough set bits in CPU mask (%d) to satisfy requested thread count: %d\n", n_set, cpuparams.n_threads);
+    }
+}
+
 bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
     bool invalid_param = false;
     std::string arg;
@@ -296,6 +371,11 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
         }
     }
 
+    postprocess_cpu_params(params.cpuparams, nullptr);
+    postprocess_cpu_params(params.cpuparams_batch, &params.cpuparams);
+    postprocess_cpu_params(params.draft_cpuparams, &params.cpuparams);
+    postprocess_cpu_params(params.draft_cpuparams_batch, &params.cpuparams_batch);
+
     if (params.prompt_cache_all && (params.interactive || params.interactive_first)) {
         throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
     }
@@ -327,7 +407,11 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
 void gpt_params_parse_from_env(gpt_params & params) {
     // we only care about server-related params for now
     get_env("LLAMA_ARG_MODEL",            params.model);
-    get_env("LLAMA_ARG_THREADS",          params.n_threads);
+    get_env("LLAMA_ARG_MODEL_URL",        params.model_url);
+    get_env("LLAMA_ARG_MODEL_ALIAS",      params.model_alias);
+    get_env("LLAMA_ARG_HF_REPO",          params.hf_repo);
+    get_env("LLAMA_ARG_HF_FILE",          params.hf_file);
+    get_env("LLAMA_ARG_THREADS",          params.cpuparams.n_threads);
     get_env("LLAMA_ARG_CTX_SIZE",         params.n_ctx);
     get_env("LLAMA_ARG_N_PARALLEL",       params.n_parallel);
     get_env("LLAMA_ARG_BATCH",            params.n_batch);
@@ -341,6 +425,9 @@ void gpt_params_parse_from_env(gpt_params & params) {
     get_env("LLAMA_ARG_EMBEDDINGS",       params.embedding);
     get_env("LLAMA_ARG_FLASH_ATTN",       params.flash_attn);
     get_env("LLAMA_ARG_DEFRAG_THOLD",     params.defrag_thold);
+    get_env("LLAMA_ARG_CONT_BATCHING",    params.cont_batching);
+    get_env("LLAMA_ARG_HOST",             params.hostname);
+    get_env("LLAMA_ARG_PORT",             params.port);
 }
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
@@ -361,6 +448,79 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
     return true;
 }
 
+bool parse_cpu_range(const std::string & range, bool (&boolmask)[GGML_MAX_N_THREADS]) {
+    size_t dash_loc = range.find('-');
+    if (dash_loc == std::string::npos) {
+        fprintf(stderr, "Format of CPU range is invalid! Expected [<start>]-[<end>].\n");
+        return false;
+    }
+
+    size_t start_i;
+    size_t end_i;
+
+    if (dash_loc == 0) {
+        start_i = 0;
+    } else {
+        start_i = std::stoull(range.substr(0, dash_loc));
+        if (start_i >= GGML_MAX_N_THREADS) {
+            fprintf(stderr, "Start index out of bounds!\n");
+            return false;
+        }
+    }
+
+    if (dash_loc == range.length() - 1) {
+        end_i = GGML_MAX_N_THREADS - 1;
+    } else {
+        end_i = std::stoull(range.substr(dash_loc + 1));
+        if (end_i >= GGML_MAX_N_THREADS) {
+            fprintf(stderr, "End index out of bounds!\n");
+            return false;
+        }
+    }
+
+    for (size_t i = start_i; i <= end_i; i++) {
+        boolmask[i] = true;
+    }
+
+    return true;
+}
+
+bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREADS]) {
+    // Discard potential 0x prefix
+    size_t start_i = 0;
+    if (mask.length() >= 2 && mask.substr(0, 2) == "0x") {
+        start_i = 2;
+    }
+
+    size_t num_digits = mask.length() - start_i;
+    if (num_digits > 128) num_digits = 128;
+
+    size_t end_i = num_digits + start_i;
+
+    for (size_t i = start_i, n = (num_digits*4 - 1); i < end_i; i++, n-=4) {
+        char c = mask.at(i);
+        int8_t id = c;
+
+        if ((c >= '0' && c <= '9')) {
+            id -= '0';
+        } else if (c >= 'a' && c <= 'f') {
+            id -= 'a' - 10;
+        } else if (c >= 'A' && c <= 'F') {
+            id -= 'A' - 10;
+        } else {
+            fprintf(stderr, "Invalid hex character '%c' at position %d\n", c, int32_t(i));
+            return false;
+        }
+
+        boolmask[  n  ] = boolmask[  n  ] || ((id & 8) != 0);
+        boolmask[n - 1] = boolmask[n - 1] || ((id & 4) != 0);
+        boolmask[n - 2] = boolmask[n - 2] || ((id & 2) != 0);
+        boolmask[n - 3] = boolmask[n - 3] || ((id & 1) != 0);
+    }
+
+    return true;
+}
+
 #define CHECK_ARG if (++i >= argc) { invalid_param = true; return true; }
 
 bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param) {
@@ -377,36 +537,142 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
     }
     if (arg == "-t" || arg == "--threads") {
         CHECK_ARG
-        params.n_threads = std::stoi(argv[i]);
+        params.cpuparams.n_threads = std::stoi(argv[i]);
-        if (params.n_threads <= 0) {
+        if (params.cpuparams.n_threads <= 0) {
-            params.n_threads = std::thread::hardware_concurrency();
+            params.cpuparams.n_threads = std::thread::hardware_concurrency();
         }
         return true;
     }
+    if (arg == "-C" || arg == "--cpu-mask") {
+        CHECK_ARG
+        std::string mask = argv[i];
+        params.cpuparams.mask_valid = true;
+        invalid_param = !parse_cpu_mask(mask, params.cpuparams.cpumask);
+        return true;
+    }
+    if (arg == "-Cr" || arg == "--cpu-range") {
+        CHECK_ARG
+        std::string range = argv[i];
+        params.cpuparams.mask_valid = true;
+        invalid_param = !parse_cpu_range(range, params.cpuparams.cpumask);
+        return true;
+    }
+    if (arg == "--prio") {
+        CHECK_ARG
+        params.cpuparams.priority = (enum ggml_sched_priority) std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "--cpu-strict") {
+        CHECK_ARG
+        params.cpuparams.strict_cpu = std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "--poll") {
+        CHECK_ARG
+        params.cpuparams.poll = std::stoul(argv[i]);
+        return true;
+    }
     if (arg == "-tb" || arg == "--threads-batch") {
         CHECK_ARG
-        params.n_threads_batch = std::stoi(argv[i]);
+        params.cpuparams_batch.n_threads = std::stoi(argv[i]);
-        if (params.n_threads_batch <= 0) {
+        if (params.cpuparams_batch.n_threads <= 0) {
-            params.n_threads_batch = std::thread::hardware_concurrency();
+            params.cpuparams_batch.n_threads = std::thread::hardware_concurrency();
         }
         return true;
     }
+    if (arg == "-Cb" || arg == "--cpu-mask-batch") {
+        CHECK_ARG
+        std::string mask = argv[i];
+        params.cpuparams_batch.mask_valid = true;
+        invalid_param = !parse_cpu_mask(mask, params.cpuparams_batch.cpumask);
+        return true;
+    }
+    if (arg == "-Crb" || arg == "--cpu-range_batch") {
+        CHECK_ARG
+        std::string range = argv[i];
+        params.cpuparams_batch.mask_valid = true;
+        invalid_param = !parse_cpu_range(range, params.cpuparams_batch.cpumask);
+        return true;
+    }
+    if (arg == "--prio-batch") {
+        CHECK_ARG
+        params.cpuparams_batch.priority = (enum ggml_sched_priority) std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "--cpu-strict-batch") {
+        params.cpuparams_batch.strict_cpu = true;
+        return true;
+    }
+    if (arg == "--poll-batch") {
+        CHECK_ARG
+        params.cpuparams_batch.poll = std::stoul(argv[i]);
+        return true;
+    }
     if (arg == "-td" || arg == "--threads-draft") {
         CHECK_ARG
-        params.n_threads_draft = std::stoi(argv[i]);
+        params.draft_cpuparams.n_threads = std::stoi(argv[i]);
-        if (params.n_threads_draft <= 0) {
+        if (params.draft_cpuparams.n_threads <= 0) {
-            params.n_threads_draft = std::thread::hardware_concurrency();
+            params.draft_cpuparams.n_threads = std::thread::hardware_concurrency();
         }
         return true;
+    }
+        if (arg == "-Cd" || arg == "--cpu-mask-draft") {
+        CHECK_ARG
+        std::string mask = argv[i];
+        params.draft_cpuparams.mask_valid = true;
+        invalid_param = !parse_cpu_mask(mask, params.draft_cpuparams.cpumask);
+        return true;
+    }
+    if (arg == "-Crd" || arg == "--cpu-range-draft") {
+        CHECK_ARG
+        std::string range = argv[i];
+        params.draft_cpuparams.mask_valid = true;
+        invalid_param = !parse_cpu_range(range, params.draft_cpuparams.cpumask);
+        return true;
+    }
+    if (arg == "--prio-draft") {
+        CHECK_ARG
+        params.draft_cpuparams.priority = (enum ggml_sched_priority) std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "--cpu-strict-draft") {
+        params.draft_cpuparams.strict_cpu = true;
+        return true;
+    }
+    if (arg == "--poll-draft") {
+        CHECK_ARG
+        params.draft_cpuparams.poll = std::stoul(argv[i]);
+        return true;
     }
     if (arg == "-tbd" || arg == "--threads-batch-draft") {
         CHECK_ARG
-        params.n_threads_batch_draft = std::stoi(argv[i]);
+        params.draft_cpuparams_batch.n_threads = std::stoi(argv[i]);
-        if (params.n_threads_batch_draft <= 0) {
+        if (params.draft_cpuparams_batch.n_threads <= 0) {
-            params.n_threads_batch_draft = std::thread::hardware_concurrency();
+            params.draft_cpuparams_batch.n_threads = std::thread::hardware_concurrency();
         }
         return true;
     }
+    if (arg == "-Crbd" || arg == "--cpu-range-batch-draft") {
+        CHECK_ARG
+        std::string range = argv[i];
+        params.draft_cpuparams_batch.mask_valid = true;
+        invalid_param = !parse_cpu_range(range, params.draft_cpuparams_batch.cpumask);
+        return true;
+    }
+    if (arg == "--prio-batch-draft") {
+        CHECK_ARG
+        params.draft_cpuparams_batch.priority = (enum ggml_sched_priority) std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "--cpu-strict-batch-draft") {
+        params.draft_cpuparams_batch.strict_cpu = true;
+        return true;
+    }
+    if (arg == "--poll-batch-draft") {
+        CHECK_ARG
+        params.draft_cpuparams_batch.poll = std::stoul(argv[i]);
+        return true;
+    }
     if (arg == "-p" || arg == "--prompt") {
         CHECK_ARG
         params.prompt = argv[i];
@@ -901,7 +1167,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         }
         return true;
     }
-    if (arg == "-ngld" || arg == "--gpu-layers-draft" || arg == "--gpu-layers-draft") {
+    if (arg == "-ngld" || arg == "--gpu-layers-draft" || arg == "--n-gpu-layers-draft") {
         CHECK_ARG
         params.n_gpu_layers_draft = std::stoi(argv[i]);
         if (!llama_supports_gpu_offload()) {
@@ -968,11 +1234,13 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
 #endif // GGML_USE_CUDA_SYCL_VULKAN
         return true;
     }
+#ifdef GGML_USE_RPC
     if (arg == "--rpc") {
         CHECK_ARG
         params.rpc_servers = argv[i];
         return true;
     }
+#endif
     if (arg == "--no-mmap") {
         params.use_mmap = false;
         return true;
@@ -1491,11 +1759,40 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     options.push_back({ "*",           "       --no-display-prompt",    "don't print prompt at generation (default: %s)", !params.display_prompt ? "true" : "false" });
     options.push_back({ "*",           "-co,   --color",                "colorise output to distinguish prompt and user input from generations (default: %s)", params.use_color ? "true" : "false" });
     options.push_back({ "*",           "-s,    --seed SEED",            "RNG seed (default: %d, use random seed for < 0)", params.seed });
-    options.push_back({ "*",           "-t,    --threads N",            "number of threads to use during generation (default: %d)", params.n_threads });
+    options.push_back({ "*",           "-t,    --threads N",            "number of threads to use during generation (default: %d)", params.cpuparams.n_threads });
     options.push_back({ "*",           "-tb,   --threads-batch N",      "number of threads to use during batch and prompt processing (default: same as --threads)" });
     options.push_back({ "speculative", "-td,   --threads-draft N",      "number of threads to use during generation (default: same as --threads)" });
-    options.push_back({ "speculative", "-tbd,  --threads-batch-draft N",
+    options.push_back({ "speculative", "-tbd,  --threads-batch-draft N","number of threads to use during batch and prompt processing (default: same as --threads-draft)" });
-                                                                        "number of threads to use during batch and prompt processing (default: same as --threads-draft)" });
+
+#ifndef GGML_USE_OPENMP
+    // these options are available only with the internal threadpool
+    options.push_back({ "*",           "-C,    --cpu-mask M",            "CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: \"\")"});
+    options.push_back({ "*",           "-Cr,   --cpu-range lo-hi",       "range of CPUs for affinity. Complements --cpu-mask"});
+    options.push_back({ "*",           "       --cpu-strict <0|1>",      "use strict CPU placement (default: %u)\n", (unsigned) params.cpuparams.strict_cpu});
+    options.push_back({ "*",           "       --priority N",            "set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.cpuparams.priority});
+    options.push_back({ "*",           "       --poll <0...100>",        "use polling level to wait for work (0 - no polling, default: %u)\n", (unsigned) params.cpuparams.poll});
+
+    options.push_back({ "*",           "-Cb,   --cpu-mask-batch M",      "CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask)"});
+    options.push_back({ "*",           "-Crb,  --cpu-range-batch lo-hi", "ranges of CPUs for affinity. Complements --cpu-mask-batch"});
+    options.push_back({ "*",           "       --cpu-strict-batch <0|1>","use strict CPU placement (default: same as --cpu-strict)"});
+    options.push_back({ "*",           "       --priority-batch N",      "set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: --priority)"});
+    options.push_back({ "*",           "       --poll-batch <0|1>",      "use polling to wait for work (default: same as --poll"});
+
+    options.push_back({ "speculative", "-Cd,   --cpu-mask-draft M",      "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)"});
+    options.push_back({ "speculative", "-Crd,  --cpu-range-draft lo-hi", "Ranges of CPUs for affinity. Complements --cpu-mask-draft"});
+    options.push_back({ "speculative", "       --cpu-strict-draft <0|1>","Use strict CPU placement for draft model (default: same as --cpu-strict)"});
+    options.push_back({ "speculative", "       --priority-draft N",      "Set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: same as --priority)"});
+    options.push_back({ "speculative", "       --poll-draft <0|1>",      "Use polling to wait for draft model work (default: same as --poll])"});
+
+    options.push_back({ "speculative", "-Cbd,  --cpu-mask-batch-draft M","Draft model CPU affinity mask. Complements cpu-range-draft-batch (default: same as --cpu-mask-draft)"});
+    options.push_back({ "speculative", "-Crbd, --cpu-range-batch-draft lo-hi",
+                                                                         "Ranges of CPUs for affinity. Complements --cpu-mask-draft-batch)"});
+    options.push_back({ "speculative", "       --cpu-strict-batch-draft <0|1>",
+                                                                         "Use strict CPU placement for draft model (default: --cpu-strict-draft)"});
+    options.push_back({ "speculative", "       --priority-batch-draft N","Set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: --priority-draft)"});
+    options.push_back({ "speculative", "       --poll-batch-draft <0|1>","Use polling to wait for draft model work (default: --poll-draft)"});
+#endif // GGML_USE_OPENMP
+
     options.push_back({ "speculative", "       --draft N",              "number of tokens to draft for speculative decoding (default: %d)", params.n_draft });
     options.push_back({ "speculative", "-ps,   --p-split N",            "speculative decoding split probability (default: %.1f)", (double)params.p_split });
     options.push_back({ "*",           "-lcs,  --lookup-cache-static FNAME",
@@ -1634,7 +1931,9 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     options.push_back({ "*",           "       --image FILE",           "path to an image file. use with multimodal models. Specify multiple times for batching" });
 
     options.push_back({ "backend" });
+#ifdef GGML_USE_RPC
     options.push_back({ "*",           "       --rpc SERVERS",          "comma separated list of RPC servers" });
+#endif
 
     if (llama_supports_mlock()) {
         options.push_back({ "*",           "       --mlock",                "force system to keep model in RAM rather than swapping or compressing" });
@@ -1767,7 +2066,6 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     options.push_back({ "export-lora", "-m,    --model",                "model path from which to load base model (default '%s')", params.model.c_str() });
     options.push_back({ "export-lora", "       --lora FNAME",           "path to LoRA adapter  (can be repeated to use multiple adapters)" });
     options.push_back({ "export-lora", "       --lora-scaled FNAME S",  "path to LoRA adapter with user defined scaling S  (can be repeated to use multiple adapters)" });
-    options.push_back({ "*",           "-t,    --threads N",            "number of threads to use during computation (default: %d)", params.n_threads });
     options.push_back({ "export-lora", "-o,    --output FNAME",         "output file (default: '%s')", params.lora_outfile.c_str() });
 
     printf("usage: %s [options]\n", argv[0]);
@@ -1799,9 +2097,9 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
 std::string gpt_params_get_system_info(const gpt_params & params) {
     std::ostringstream os;
 
-    os << "system_info: n_threads = " << params.n_threads;
+    os << "system_info: n_threads = " << params.cpuparams.n_threads;
-    if (params.n_threads_batch != -1) {
+    if (params.cpuparams_batch.n_threads != -1) {
-        os << " (n_threads_batch = " << params.n_threads_batch << ")";
+        os << " (n_threads_batch = " << params.cpuparams_batch.n_threads << ")";
     }
 #if defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later
     // TODO: windows + arm64 + mingw64
@@ -1861,13 +2159,19 @@ std::string string_get_sortable_timestamp() {
 
 void string_replace_all(std::string & s, const std::string & search, const std::string & replace) {
     if (search.empty()) {
-        return; // Avoid infinite loop if 'search' is an empty string
+        return;
     }
+    std::string builder;
+    builder.reserve(s.length());
     size_t pos = 0;
-    while ((pos = s.find(search, pos)) != std::string::npos) {
+    size_t last_pos = 0;
-        s.replace(pos, search.length(), replace);
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
-        pos += replace.length();
+        builder.append(s, last_pos, pos - last_pos);
+        builder.append(replace);
+        last_pos = pos + search.length();
     }
+    builder.append(s, last_pos, std::string::npos);
+    s = std::move(builder);
 }
 
 void string_process_escapes(std::string & input) {
@@ -2319,8 +2623,9 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.n_seq_max         = params.n_parallel;
     cparams.n_batch           = params.n_batch;
     cparams.n_ubatch          = params.n_ubatch;
-    cparams.n_threads         = params.n_threads;
+    cparams.n_threads         = params.cpuparams.n_threads;
-    cparams.n_threads_batch   = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    cparams.n_threads_batch   = params.cpuparams_batch.n_threads == -1 ?
+                                    params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
     cparams.seed              = params.seed;
     cparams.logits_all        = params.logits_all;
     cparams.embeddings        = params.embedding;
@@ -2346,6 +2651,22 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     return cparams;
 }
 
+struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params) {
+    struct ggml_threadpool_params tpp;
+
+    ggml_threadpool_params_init(&tpp, params.n_threads); // setup the defaults
+
+    if (params.mask_valid) {
+        std::memcpy(&tpp.cpumask, &params.cpumask, GGML_MAX_N_THREADS);
+    }
+
+    tpp.prio       = params.priority;
+    tpp.poll       = params.poll;
+    tpp.strict_cpu = params.strict_cpu;
+
+    return tpp;
+}
+
 #ifdef LLAMA_USE_CURL
 
 static bool starts_with(const std::string & str, const std::string & prefix) {
@@ -3335,7 +3656,7 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
     yaml_dump_vector_float(stream, "tensor_split", tensor_split_vector);
 
     fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z);
-    fprintf(stream, "threads: %d # default: %u\n", params.n_threads, std::thread::hardware_concurrency());
+    fprintf(stream, "threads: %d # default: %u\n", params.cpuparams.n_threads, std::thread::hardware_concurrency());
     fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
     fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
     fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);

common/common.h

@@ -67,13 +67,18 @@ enum dimre_method {
     DIMRE_METHOD_MEAN,
 };
 
+struct cpu_params {
+    int      n_threads                   = -1;
+    bool     cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask.
+    bool     mask_valid                  = false;   // Default: any CPU
+    enum ggml_sched_priority  priority   = GGML_SCHED_PRIO_NORMAL;  // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime)
+    bool     strict_cpu                  = false;   // Use strict CPU placement
+    uint32_t poll                        = 50;      // Polling (busywait) level (0 - no polling, 100 - mostly polling)
+};
+
 struct gpt_params {
     uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
 
-    int32_t n_threads             = cpu_get_num_math();
-    int32_t n_threads_draft       =    -1;
-    int32_t n_threads_batch       =    -1; // number of threads to use for batch processing (-1 = use n_threads)
-    int32_t n_threads_batch_draft =    -1;
     int32_t n_predict             =    -1; // new tokens to predict
     int32_t n_ctx                 =     0; // context size
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
@@ -100,6 +105,11 @@ struct gpt_params {
     int32_t yarn_orig_ctx         =     0; // YaRN original context length
     float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
 
+    struct cpu_params cpuparams;
+    struct cpu_params cpuparams_batch;
+    struct cpu_params draft_cpuparams;
+    struct cpu_params draft_cpuparams_batch;
+
     ggml_backend_sched_eval_callback cb_eval = nullptr;
     void * cb_eval_user_data                 = nullptr;
 
@@ -204,7 +214,7 @@ struct gpt_params {
     int32_t port           = 8080;         // server listens on this network port
     int32_t timeout_read   = 600;          // http read timeout in seconds
     int32_t timeout_write  = timeout_read; // http write timeout in seconds
-    int32_t n_threads_http = -1;           // number of threads to process HTTP requests
+    int     n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";
@@ -277,6 +287,11 @@ void gpt_params_print_usage(int argc, char ** argv, const gpt_params & params);
 
 std::string gpt_params_get_system_info(const gpt_params & params);
 
+bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]);
+bool parse_cpu_mask(const std::string& mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
+void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model = nullptr);
+bool set_process_priority(enum ggml_sched_priority prio);
+
 //
 // String utils
 //
@@ -329,6 +344,7 @@ struct llama_init_result    llama_init_from_gpt_params(gpt_params & params);
 
 struct llama_model_params     llama_model_params_from_gpt_params    (const gpt_params & params);
 struct llama_context_params   llama_context_params_from_gpt_params  (const gpt_params & params);
+struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
 
 struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
 struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);

convert_hf_to_gguf.py

@@ -3,6 +3,7 @@
 
 from __future__ import annotations
 
+import ast
 import logging
 import argparse
 import contextlib
@@ -63,6 +64,7 @@ class Model:
     model_name: str | None
     metadata_override: Path | None
     dir_model_card: Path
+    is_lora: bool
 
     # subclasses should define this!
     model_arch: gguf.MODEL_ARCH
@@ -70,7 +72,7 @@ class Model:
     def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False,
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
-                 split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False):
+                 split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False, is_lora: bool = False):
         if type(self) is Model:
             raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
 
@@ -92,6 +94,7 @@ class Model:
         self.metadata_override = metadata_override
         self.model_name = model_name
         self.dir_model_card = dir_model  # overridden in convert_lora_to_gguf.py
+        self.is_lora = is_lora  # true if model is used inside convert_lora_to_gguf.py
 
         # Apply heuristics to figure out typical tensor encoding based on first layer tensor encoding type
         if self.ftype == gguf.LlamaFileType.GUESSED:
@@ -296,9 +299,12 @@ class Model:
                             gguf.MODEL_TENSOR.POS_EMBD,
                             gguf.MODEL_TENSOR.TOKEN_TYPES,
                             gguf.MODEL_TENSOR.SSM_CONV1D,
+                            gguf.MODEL_TENSOR.TIME_MIX_FIRST,
+                            gguf.MODEL_TENSOR.TIME_MIX_W1,
+                            gguf.MODEL_TENSOR.TIME_MIX_W2,
                         )
                     )
-                    or not name.endswith(".weight")
+                    or not new_name.endswith(".weight")
                 ):
                     data_qtype = gguf.GGMLQuantizationType.F32
 
@@ -1588,7 +1594,7 @@ class LlamaModel(Model):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
                 freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
 
                 factor = rope_scaling.get("factor", 8.0)
@@ -1611,6 +1617,7 @@ class LlamaModel(Model):
                         smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
                         rope_factors.append(1 / ((1 - smooth) / factor + smooth))
 
+                if not self.is_lora:
                     self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
 
         super().prepare_tensors()
@@ -2157,6 +2164,7 @@ class Phi3MiniModel(Model):
         if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
             raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
 
+        if not self.is_lora:
             self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG]  + ".weight", np.array(long_factors, dtype=np.float32))
             self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
 
@@ -2729,6 +2737,84 @@ class StarCoder2Model(Model):
     model_arch = gguf.MODEL_ARCH.STARCODER2
 
 
+@Model.register("Rwkv6ForCausalLM")
+class Rwkv6Model(Model):
+    model_arch = gguf.MODEL_ARCH.RWKV6
+
+    def set_vocab(self):
+        assert (self.dir_model / "rwkv_vocab_v20230424.txt").is_file()
+        vocab_size = self.hparams.get("vocab_size", 65536)
+
+        tokens: list[bytes] = ['<s>'.encode("utf-8")]
+        toktypes: list[int] = [gguf.TokenType.CONTROL]
+
+        with open(self.dir_model / "rwkv_vocab_v20230424.txt", "r", encoding="utf-8") as f:
+            lines = f.readlines()
+            for line in lines:
+                parts = line.split(' ')
+                assert len(parts) >= 3
+                token, token_len = ast.literal_eval(' '.join(parts[1:-1])), int(parts[-1])
+                token = token.encode("utf-8") if isinstance(token, str) else token
+                assert isinstance(token, bytes)
+                assert len(token) == token_len
+                token_text: str = repr(token)[2:-1]  # "b'\xff'" -> "\xff"
+                tokens.append(token_text.encode("utf-8"))
+                toktypes.append(gguf.TokenType.NORMAL)
+        remainder = vocab_size - len(tokens)
+        assert remainder >= 0
+        for i in range(len(tokens), vocab_size):
+            tokens.append(f"[PAD{i}]".encode("utf-8"))
+            toktypes.append(gguf.TokenType.UNUSED)
+
+        self.gguf_writer.add_tokenizer_model("rwkv")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        head_size = self.hparams["head_size"]
+        hidden_size = self.hparams["hidden_size"]
+        layer_norm_eps = self.hparams["layer_norm_epsilon"]
+        rescale_every_n_layers = self.hparams["rescale_every"]
+        intermediate_size = self.hparams["intermediate_size"] if self.hparams["intermediate_size"] is not None else int((hidden_size * 3.5) // 32 * 32)
+        time_mix_extra_dim = 64 if hidden_size == 4096 else 32
+        time_decay_extra_dim = 128 if hidden_size == 4096 else 64
+
+        # RWKV isn't context limited
+        self.gguf_writer.add_context_length(1048576)
+        self.gguf_writer.add_embedding_length(hidden_size)
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_layer_norm_eps(layer_norm_eps)
+        self.gguf_writer.add_rescale_every_n_layers(rescale_every_n_layers)
+        self.gguf_writer.add_wkv_head_size(head_size)
+        self.gguf_writer.add_time_mix_extra_dim(time_mix_extra_dim)
+        self.gguf_writer.add_time_decay_extra_dim(time_decay_extra_dim)
+        self.gguf_writer.add_feed_forward_length(intermediate_size)
+        self.gguf_writer.add_file_type(self.ftype)
+
+        # required by llama.cpp, unused
+        self.gguf_writer.add_head_count(0)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        new_name = self.map_tensor_name(name)
+
+        if not (new_name.endswith(".weight") or new_name.endswith(".bias")):
+            new_name += ".weight"
+
+        if new_name.endswith("time_mix_w1.weight") or new_name.endswith("time_mix_decay_w1.weight") or new_name.endswith("time_mix_decay_w2.weight"):
+            data_torch = data_torch.transpose(0, 1)
+
+        if new_name.endswith("time_mix_w2.weight"):
+            data_torch = data_torch.permute(0, 2, 1)
+
+        rescale_every_n_layers = self.hparams["rescale_every"]
+        if rescale_every_n_layers > 0:
+            if new_name.endswith("time_mix_output.weight") or new_name.endswith("channel_mix_value.weight"):
+                data_torch = data_torch.div_(2 ** int(bid // rescale_every_n_layers))
+
+        yield (new_name, data_torch)
+
+
 @Model.register("MambaForCausalLM", "MambaLMHeadModel", "FalconMambaForCausalLM")
 class MambaModel(Model):
     model_arch = gguf.MODEL_ARCH.MAMBA
@@ -3833,7 +3919,7 @@ class ExaoneModel(Model):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
                 freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
 
                 factor = rope_scaling.get("factor", 8.0)
@@ -3856,6 +3942,7 @@ class ExaoneModel(Model):
                         smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
                         rope_factors.append(1 / ((1 - smooth) / factor + smooth))
 
+                if not self.is_lora:
                     self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
 
         super().prepare_tensors()

convert_lora_to_gguf.py

@@ -386,6 +386,7 @@ if __name__ == '__main__':
             dry_run=args.dry_run,
             dir_lora_model=dir_lora,
             lora_alpha=alpha,
+            is_lora=True,
         )
 
         logger.info("Exporting model...")

docs/backend/SYCL.md

@@ -336,12 +336,12 @@ Choose one of following methods to run.
 - Use device 0:
 
 ```sh
-./examples/sycl/run_llama2.sh 0
+./examples/sycl/run-llama2.sh 0
 ```
 - Use multiple devices:
 
 ```sh
-./examples/sycl/run_llama2.sh
+./examples/sycl/run-llama2.sh
 ```
 
 2. Command line

docs/docker.md

@@ -20,7 +20,7 @@ Additionally, there the following images, similar to the above:
 - `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 - `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 
-The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
+The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
 
 ## Usage
 
@@ -66,8 +66,8 @@ You may want to pass in some different `ARGS`, depending on the CUDA environment
 
 The defaults are:
 
-- `CUDA_VERSION` set to `11.7.1`
+- `CUDA_VERSION` set to `12.6.0`
-- `CUDA_DOCKER_ARCH` set to `all`
+- `CUDA_DOCKER_ARCH` set to the cmake build default, which includes all the supported architectures
 
 The resulting images, are essentially the same as the non-CUDA images:
 

examples/baby-llama/baby-llama.cpp

@@ -18,7 +18,7 @@ constexpr float rms_norm_eps = 5e-6f;
 #endif
 
 static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
-    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr);
 
     if (plan.work_size > 0) {
         buf.resize(plan.work_size);

examples/benchmark/benchmark-matmult.cpp

@@ -21,7 +21,7 @@
 #endif
 
 static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
-    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr);
 
     if (plan.work_size > 0) {
         buf.resize(plan.work_size);
@@ -54,7 +54,7 @@ static void tensor_dump(const ggml_tensor * tensor, const char * name) {
 #define TENSOR_DUMP(tensor) tensor_dump(tensor, #tensor)
 
 struct benchmark_params_struct {
-    int32_t n_threads     = 1;
+    int     n_threads     = 1;
     int32_t n_iterations  = 10;
 };
 

examples/cvector-generator/cvector-generator.cpp

@@ -486,7 +486,7 @@ int main(int argc, char ** argv) {
     if (use_pca) {
         // run PCA
         PCA::pca_params pca_params;
-        pca_params.n_threads = params.n_threads;
+        pca_params.n_threads    = params.cpuparams.n_threads;
         pca_params.n_batch      = params.n_pca_batch;
         pca_params.n_iterations = params.n_pca_iterations;
         PCA::run_pca(pca_params, ctx_train.v_diff, ctx_train.v_final);

examples/export-lora/export-lora.cpp

@@ -410,7 +410,7 @@ int main(int argc, char ** argv) {
 
     g_verbose = (params.verbosity == 1);
     try {
-        lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.n_threads);
+        lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.cpuparams.n_threads);
         ctx.run_merge();
     } catch (const std::exception & err) {
         fprintf(stderr, "%s\n", err.what());

examples/llama-bench/README.md

@@ -14,7 +14,8 @@ Performance testing tool for llama.cpp.
     1. [Markdown](#markdown)
     2. [CSV](#csv)
     3. [JSON](#json)
-    4. [SQL](#sql)
+    4. [JSONL](#jsonl)
+    5. [SQL](#sql)
 
 ## Syntax
 
@@ -26,13 +27,17 @@ options:
   -m, --model <filename>                    (default: models/7B/ggml-model-q4_0.gguf)
   -p, --n-prompt <n>                        (default: 512)
   -n, --n-gen <n>                           (default: 128)
-  -pg <pp,tg>                         (default: 512,128)
+  -pg <pp,tg>                               (default: )
   -b, --batch-size <n>                      (default: 2048)
   -ub, --ubatch-size <n>                    (default: 512)
   -ctk, --cache-type-k <t>                  (default: f16)
   -ctv, --cache-type-v <t>                  (default: f16)
-  -t, --threads <n>                   (default: 16)
+  -t, --threads <n>                         (default: 8)
+  -C, --cpu-mask <hex,hex>                  (default: 0x0)
+  --cpu-strict <0|1>                        (default: 0)
+  --poll <0...100>                          (default: 50)
   -ngl, --n-gpu-layers <n>                  (default: 99)
+  -rpc, --rpc <rpc_servers>                 (default: )
   -sm, --split-mode <none|layer|row>        (default: layer)
   -mg, --main-gpu <i>                       (default: 0)
   -nkvo, --no-kv-offload <0|1>              (default: 0)
@@ -42,7 +47,10 @@ options:
   -embd, --embeddings <0|1>                 (default: 0)
   -ts, --tensor-split <ts0/ts1/..>          (default: 0)
   -r, --repetitions <n>                     (default: 5)
-  -o, --output <csv|json|md|sql>      (default: md)
+  --prio <0|1|2|3>                          (default: 0)
+  --delay <0...N> (seconds)                 (default: 0)
+  -o, --output <csv|json|jsonl|md|sql>      (default: md)
+  -oe, --output-err <csv|json|jsonl|md|sql> (default: none)
   -v, --verbose                             (default: 0)
 
 Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.
@@ -238,6 +246,19 @@ $ ./llama-bench -o json
 ]
 ```
 
+
+### JSONL
+
+```sh
+$ ./llama-bench -o jsonl
+```
+
+```json lines
+{"build_commit":"3469684","build_number":1275,"cuda":true,"metal":false,"gpu_blas":true,"blas":true,"cpu_info":"13th Gen Intel(R) Core(TM) i9-13900K","gpu_info":"NVIDIA GeForce RTX 3090 Ti","model_filename":"models/7B/ggml-model-q4_0.gguf","model_type":"llama 7B mostly Q4_0","model_size":3825065984,"model_n_params":6738415616,"n_batch":512,"n_threads":16,"f16_kv":true,"n_gpu_layers":99,"main_gpu":0,"mul_mat_q":true,"tensor_split":"0.00","n_prompt":512,"n_gen":0,"test_time":"2023-09-23T12:09:57Z","avg_ns":212365953,"stddev_ns":985423,"avg_ts":2410.974041,"stddev_ts":11.163766,"samples_ns":[213837238,211635853,212328053,211329715,212698907],"samples_ts":[2394.34,2419.25,2411.36,2422.75,2407.16]}
+{"build_commit":"3469684","build_number":1275,"cuda":true,"metal":false,"gpu_blas":true,"blas":true,"cpu_info":"13th Gen Intel(R) Core(TM) i9-13900K","gpu_info":"NVIDIA GeForce RTX 3090 Ti","model_filename":"models/7B/ggml-model-q4_0.gguf","model_type":"llama 7B mostly Q4_0","model_size":3825065984,"model_n_params":6738415616,"n_batch":512,"n_threads":16,"f16_kv":true,"n_gpu_layers":99,"main_gpu":0,"mul_mat_q":true,"tensor_split":"0.00","n_prompt":0,"n_gen":128,"test_time":"2023-09-23T12:09:59Z","avg_ns":977425219,"stddev_ns":9268593,"avg_ts":130.965708,"stddev_ts":1.238924,"samples_ns":[984472709,974901233,989474741,970729355,967548060],"samples_ts":[130.019,131.295,129.362,131.86,132.293]}
+```
+
+
 ### SQL
 
 SQL output is suitable for importing into a SQLite database. The output can be piped into the `sqlite3` command line tool to add the results to a database.

examples/llama-bench/llama-bench.cpp

@@ -16,6 +16,7 @@
 #include <sstream>
 #include <string>
 #include <vector>
+#include <thread>
 
 #include "ggml.h"
 #include "llama.h"
@@ -170,13 +171,14 @@ static std::string get_gpu_info() {
 }
 
 // command line params
-enum output_formats {NONE, CSV, JSON, MARKDOWN, SQL};
+enum output_formats {NONE, CSV, JSON, JSONL, MARKDOWN, SQL};
 
 static const char * output_format_str(output_formats format) {
     switch (format) {
         case NONE:     return "none";
         case CSV:      return "csv";
         case JSON:     return "json";
+        case JSONL:    return "jsonl";
         case MARKDOWN: return "md";
         case SQL:      return "sql";
         default: GGML_ABORT("invalid output format");
@@ -190,6 +192,8 @@ static bool output_format_from_str(const std::string & s, output_formats & forma
         format = CSV;
     } else if (s == "json") {
         format = JSON;
+    } else if (s == "jsonl") {
+        format = JSONL;
     } else if (s == "md") {
         format = MARKDOWN;
     } else if (s == "sql") {
@@ -225,6 +229,9 @@ struct cmd_params {
     std::vector<ggml_type> type_k;
     std::vector<ggml_type> type_v;
     std::vector<int> n_threads;
+    std::vector<std::string> cpu_mask;
+    std::vector<bool> cpu_strict;
+    std::vector<int> poll;
     std::vector<int> n_gpu_layers;
     std::vector<std::string> rpc_servers;
     std::vector<llama_split_mode> split_mode;
@@ -236,6 +243,8 @@ struct cmd_params {
     std::vector<bool> embeddings;
     ggml_numa_strategy numa;
     int reps;
+    ggml_sched_priority prio;
+    int delay;
     bool verbose;
     output_formats output_format;
     output_formats output_format_stderr;
@@ -251,6 +260,9 @@ static const cmd_params cmd_params_defaults = {
     /* type_k               */ {GGML_TYPE_F16},
     /* type_v               */ {GGML_TYPE_F16},
     /* n_threads            */ {cpu_get_num_math()},
+    /* cpu_mask             */ {"0x0"},
+    /* cpu_strict           */ {false},
+    /* poll                 */ {50},
     /* n_gpu_layers         */ {99},
     /* rpc_servers          */ {""},
     /* split_mode           */ {LLAMA_SPLIT_MODE_LAYER},
@@ -262,6 +274,8 @@ static const cmd_params cmd_params_defaults = {
     /* embeddings           */ {false},
     /* numa                 */ GGML_NUMA_STRATEGY_DISABLED,
     /* reps                 */ 5,
+    /* prio                 */ GGML_SCHED_PRIO_NORMAL,
+    /* delay                */ 0,
     /* verbose              */ false,
     /* output_format        */ MARKDOWN,
     /* output_format_stderr */ NONE,
@@ -281,8 +295,13 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -ctk, --cache-type-k <t>                  (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
     printf("  -ctv, --cache-type-v <t>                  (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
     printf("  -t, --threads <n>                         (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
+    printf("  -C, --cpu-mask <hex,hex>                  (default: %s)\n", join(cmd_params_defaults.cpu_mask, ",").c_str());
+    printf("  --cpu-strict <0|1>                        (default: %s)\n", join(cmd_params_defaults.cpu_strict, ",").c_str());
+    printf("  --poll <0...100>                          (default: %s)\n", join(cmd_params_defaults.poll, ",").c_str());
     printf("  -ngl, --n-gpu-layers <n>                  (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
+#ifdef GGML_USE_RPC
     printf("  -rpc, --rpc <rpc_servers>                 (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str());
+#endif
     printf("  -sm, --split-mode <none|layer|row>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
     printf("  -mg, --main-gpu <i>                       (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
     printf("  -nkvo, --no-kv-offload <0|1>              (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
@@ -292,8 +311,10 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -embd, --embeddings <0|1>                 (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
     printf("  -ts, --tensor-split <ts0/ts1/..>          (default: 0)\n");
     printf("  -r, --repetitions <n>                     (default: %d)\n", cmd_params_defaults.reps);
-    printf("  -o, --output <csv|json|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
+    printf("  --prio <0|1|2|3>                          (default: %d)\n", cmd_params_defaults.prio);
-    printf("  -oe, --output-err <csv|json|md|sql> (default: %s)\n", output_format_str(cmd_params_defaults.output_format_stderr));
+    printf("  --delay <0...N> (seconds)                 (default: %d)\n", cmd_params_defaults.delay);
+    printf("  -o, --output <csv|json|jsonl|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
+    printf("  -oe, --output-err <csv|json|jsonl|md|sql> (default: %s)\n", output_format_str(cmd_params_defaults.output_format_stderr));
     printf("  -v, --verbose                             (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
     printf("\n");
     printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
@@ -338,6 +359,8 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     params.output_format_stderr = cmd_params_defaults.output_format_stderr;
     params.reps = cmd_params_defaults.reps;
     params.numa = cmd_params_defaults.numa;
+    params.prio = cmd_params_defaults.prio;
+    params.delay = cmd_params_defaults.delay;
 
     for (int i = 1; i < argc; i++) {
         arg = argv[i];
@@ -433,6 +456,27 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = string_split<int>(argv[i], split_delim);
             params.n_threads.insert(params.n_threads.end(), p.begin(), p.end());
+        } else if (arg == "-C" || arg == "--cpu-mask") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<std::string>(argv[i], split_delim);
+            params.cpu_mask.insert(params.cpu_mask.end(), p.begin(), p.end());
+        } else if (arg == "--cpu-strict") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<bool>(argv[i], split_delim);
+            params.cpu_strict.insert(params.cpu_strict.end(), p.begin(), p.end());
+        } else if (arg == "--poll") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<int>(argv[i], split_delim);
+            params.poll.insert(params.poll.end(), p.begin(), p.end());
         } else if (arg == "-ngl" || arg == "--n-gpu-layers") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -440,12 +484,14 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = string_split<int>(argv[i], split_delim);
             params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
+#ifdef GGML_USE_RPC
         } else if (arg == "-rpc" || arg == "--rpc") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params.rpc_servers.push_back(argv[i]);
+#endif
         } else if (arg == "-sm" || arg == "--split-mode") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -541,6 +587,18 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 break;
             }
             params.reps = std::stoi(argv[i]);
+        } else if (arg == "--prio") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.prio = (enum ggml_sched_priority) std::stoi(argv[i]);
+        } else if (arg == "--delay") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.delay = std::stoi(argv[i]);
         } else if (arg == "-o" || arg == "--output") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -585,6 +643,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.use_mmap.empty())     { params.use_mmap = cmd_params_defaults.use_mmap; }
     if (params.embeddings.empty())   { params.embeddings = cmd_params_defaults.embeddings; }
     if (params.n_threads.empty())    { params.n_threads = cmd_params_defaults.n_threads; }
+    if (params.cpu_mask.empty())     { params.cpu_mask  = cmd_params_defaults.cpu_mask;  }
+    if (params.cpu_strict.empty())   { params.cpu_strict = cmd_params_defaults.cpu_strict; }
+    if (params.poll.empty())         { params.poll = cmd_params_defaults.poll; }
 
     return params;
 }
@@ -598,6 +659,9 @@ struct cmd_params_instance {
     ggml_type type_k;
     ggml_type type_v;
     int n_threads;
+    std::string cpu_mask;
+    bool cpu_strict;
+    int poll;
     int n_gpu_layers;
     std::string rpc_servers;
     llama_split_mode split_mode;
@@ -667,7 +731,10 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & tv : params.type_v)
     for (const auto & nkvo : params.no_kv_offload)
     for (const auto & fa : params.flash_attn)
-    for (const auto & nt : params.n_threads) {
+    for (const auto & nt : params.n_threads)
+    for (const auto & cm : params.cpu_mask)
+    for (const auto & cs : params.cpu_strict)
+    for (const auto & pl : params.poll) {
         for (const auto & n_prompt : params.n_prompt) {
             if (n_prompt == 0) {
                 continue;
@@ -681,6 +748,9 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
+                /* .cpu_mask     = */ cm,
+                /* .cpu_strict   = */ cs,
+                /* .poll         = */ pl,
                 /* .n_gpu_layers = */ nl,
                 /* .rpc_servers  = */ rpc,
                 /* .split_mode   = */ sm,
@@ -707,6 +777,9 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
+                /* .cpu_mask     = */ cm,
+                /* .cpu_strict   = */ cs,
+                /* .poll         = */ pl,
                 /* .n_gpu_layers = */ nl,
                 /* .rpc_servers  = */ rpc,
                 /* .split_mode   = */ sm,
@@ -733,6 +806,9 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
+                /* .cpu_mask     = */ cm,
+                /* .cpu_strict   = */ cs,
+                /* .poll         = */ pl,
                 /* .n_gpu_layers = */ nl,
                 /* .rpc_servers  = */ rpc,
                 /* .split_mode   = */ sm,
@@ -769,6 +845,9 @@ struct test {
     int n_batch;
     int n_ubatch;
     int n_threads;
+    std::string cpu_mask;
+    bool cpu_strict;
+    int poll;
     bool has_rpc;
     ggml_type type_k;
     ggml_type type_v;
@@ -795,6 +874,9 @@ struct test {
         n_batch = inst.n_batch;
         n_ubatch = inst.n_ubatch;
         n_threads = inst.n_threads;
+        cpu_mask = inst.cpu_mask;
+        cpu_strict = inst.cpu_strict;
+        poll = inst.poll;
         has_rpc = !inst.rpc_servers.empty();
         type_k = inst.type_k;
         type_v = inst.type_v;
@@ -872,13 +954,14 @@ struct test {
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_ubatch",
-            "n_threads", "type_k", "type_v",
+            "n_threads", "cpu_mask", "cpu_strict", "poll",
+            "type_k", "type_v",
             "n_gpu_layers", "split_mode",
             "main_gpu", "no_kv_offload", "flash_attn",
             "tensor_split", "use_mmap", "embeddings",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
-            "avg_ts", "stddev_ts"
+            "avg_ts", "stddev_ts",
         };
         return fields;
     }
@@ -887,7 +970,7 @@ struct test {
 
     static field_type get_field_type(const std::string & field) {
         if (field == "build_number" || field == "n_batch" || field == "n_ubatch" ||
-            field == "n_threads" ||
+            field == "n_threads" || field == "poll" ||
             field == "model_size" || field == "model_n_params" ||
             field == "n_gpu_layers" || field == "main_gpu" ||
             field == "n_prompt" || field == "n_gen" ||
@@ -896,6 +979,7 @@ struct test {
         }
         if (field == "cuda" || field == "vulkan" || field == "kompute" || field == "metal" ||
             field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" ||
+            field == "cpu_strict" ||
             field == "flash_attn" || field == "use_mmap" || field == "embeddings") {
             return BOOL;
         }
@@ -928,7 +1012,8 @@ struct test {
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
             std::to_string(n_batch), std::to_string(n_ubatch),
-            std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
+            std::to_string(n_threads), cpu_mask, std::to_string(cpu_strict), std::to_string(poll),
+            ggml_type_name(type_k), ggml_type_name(type_v),
             std::to_string(n_gpu_layers), split_mode_str(split_mode),
             std::to_string(main_gpu), std::to_string(no_kv_offload), std::to_string(flash_attn),
             tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings),
@@ -996,8 +1081,6 @@ struct csv_printer : public printer {
     }
 };
 
-struct json_printer : public printer {
-    bool first = true;
 
 static std::string escape_json(const std::string & value) {
     std::string escaped;
@@ -1017,7 +1100,7 @@ struct json_printer : public printer {
     return escaped;
 }
 
-    static std::string format_value(const std::string & field, const std::string & value) {
+static std::string format_json_value(const std::string & field, const std::string & value) {
     switch (test::get_field_type(field)) {
         case test::STRING:
             return "\"" + escape_json(value) + "\"";
@@ -1028,6 +1111,9 @@ struct json_printer : public printer {
     }
 }
 
+struct json_printer : public printer {
+    bool first = true;
+
     void print_header(const cmd_params & params) override {
         fprintf(fout, "[\n");
         (void) params;
@@ -1036,7 +1122,7 @@ struct json_printer : public printer {
     void print_fields(const std::vector<std::string> & fields, const std::vector<std::string> & values) {
         assert(fields.size() == values.size());
         for (size_t i = 0; i < fields.size(); i++) {
-            fprintf(fout, "    \"%s\": %s,\n", fields.at(i).c_str(), format_value(fields.at(i), values.at(i)).c_str());
+            fprintf(fout, "    \"%s\": %s,\n", fields.at(i).c_str(), format_json_value(fields.at(i), values.at(i)).c_str());
         }
     }
 
@@ -1059,6 +1145,25 @@ struct json_printer : public printer {
     }
 };
 
+
+struct jsonl_printer : public printer {
+    void print_fields(const std::vector<std::string> & fields, const std::vector<std::string> & values) {
+        assert(fields.size() == values.size());
+        for (size_t i = 0; i < fields.size(); i++) {
+            fprintf(fout, "\"%s\": %s, ", fields.at(i).c_str(), format_json_value(fields.at(i), values.at(i)).c_str());
+        }
+    }
+
+    void print_test(const test & t) override {
+        fprintf(fout, "{");
+        print_fields(test::get_fields(), t.get_values());
+        fprintf(fout, "\"samples_ns\": [ %s ],", join(t.samples_ns, ", ").c_str());
+        fprintf(fout, "\"samples_ts\": [ %s ]", join(t.get_ts(), ", ").c_str());
+        fprintf(fout, "}\n");
+        fflush(fout);
+    }
+};
+
 struct markdown_printer : public printer {
     std::vector<std::string> fields;
 
@@ -1067,7 +1172,7 @@ struct markdown_printer : public printer {
             return -30;
         }
         if (field == "t/s") {
-            return 16;
+            return 20;
         }
         if (field == "size" || field == "params") {
             return 10;
@@ -1149,6 +1254,15 @@ struct markdown_printer : public printer {
         if (params.n_threads.size() > 1 || params.n_threads != cmd_params_defaults.n_threads || is_cpu_backend) {
             fields.emplace_back("n_threads");
         }
+        if (params.cpu_mask.size() > 1 || params.cpu_mask != cmd_params_defaults.cpu_mask) {
+            fields.emplace_back("cpu_mask");
+        }
+        if (params.cpu_strict.size() > 1 || params.cpu_strict != cmd_params_defaults.cpu_strict) {
+            fields.emplace_back("cpu_strict");
+        }
+        if (params.poll.size() > 1 || params.poll != cmd_params_defaults.poll) {
+            fields.emplace_back("poll");
+        }
         if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
             fields.emplace_back("n_batch");
         }
@@ -1350,6 +1464,8 @@ static std::unique_ptr<printer> create_printer(output_formats format) {
             return std::unique_ptr<printer>(new csv_printer());
         case JSON:
             return std::unique_ptr<printer>(new json_printer());
+        case JSONL:
+            return std::unique_ptr<printer>(new jsonl_printer());
         case MARKDOWN:
             return std::unique_ptr<printer>(new markdown_printer());
         case SQL:
@@ -1383,6 +1499,8 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
+    set_process_priority(params.prio);
+
     // initialize printer
     std::unique_ptr<printer> p = create_printer(params.output_format);
     std::unique_ptr<printer> p_err = create_printer(params.output_format_stderr);
@@ -1428,6 +1546,28 @@ int main(int argc, char ** argv) {
 
         llama_kv_cache_clear(ctx);
 
+        // cool off before the test
+        if (params.delay) {
+            std::this_thread::sleep_for(std::chrono::seconds(params.delay));
+        }
+
+        struct ggml_threadpool_params tpp = ggml_threadpool_params_default(t.n_threads);
+        if (!parse_cpu_mask(t.cpu_mask, tpp.cpumask)) {
+            LOG_TEE("%s: failed to parse cpu-mask: %s\n", __func__, t.cpu_mask.c_str());
+            exit(1);
+        }
+        tpp.strict_cpu = t.cpu_strict;
+        tpp.poll       = t.poll;
+        tpp.prio       = params.prio;
+
+        struct ggml_threadpool* threadpool = ggml_threadpool_new(&tpp);
+        if (!threadpool) {
+            LOG_TEE("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
+            exit(1);
+        }
+
+        llama_attach_threadpool(ctx, threadpool, NULL);
+
         // warmup run
         if (t.n_prompt > 0) {
             //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
@@ -1466,6 +1606,8 @@ int main(int argc, char ** argv) {
         llama_print_timings(ctx);
 
         llama_free(ctx);
+
+        ggml_threadpool_free(threadpool);
     }
 
     llama_free_model(lmodel);

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -71,8 +71,8 @@ actor LlamaContext {
         var ctx_params = llama_context_default_params()
         ctx_params.seed  = 1234
         ctx_params.n_ctx = 2048
-        ctx_params.n_threads       = UInt32(n_threads)
+        ctx_params.n_threads       = Int32(n_threads)
-        ctx_params.n_threads_batch = UInt32(n_threads)
+        ctx_params.n_threads_batch = Int32(n_threads)
 
         let context = llama_new_context_with_model(model, ctx_params)
         guard let context else {

examples/llava/README-minicpmv2.5.md

@@ -15,8 +15,8 @@ cd llama.cpp
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us)
 
 ```bash
-python ./examples/minicpmv/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
+python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
-python ./examples/minicpmv/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2
+python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2
 python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model
 
 # quantize int4 version

examples/llava/clip.cpp

@@ -216,13 +216,19 @@ static std::string gguf_data_to_str(enum gguf_type type, const void * data, int
 
 static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
     if (search.empty()) {
-        return; // Avoid infinite loop if 'search' is an empty string
+        return;
     }
+    std::string builder;
+    builder.reserve(s.length());
     size_t pos = 0;
-    while ((pos = s.find(search, pos)) != std::string::npos) {
+    size_t last_pos = 0;
-        s.replace(pos, search.length(), replace);
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
-        pos += replace.length();
+        builder.append(s, last_pos, pos - last_pos);
+        builder.append(replace);
+        last_pos = pos + search.length();
     }
+    builder.append(s, last_pos, std::string::npos);
+    s = std::move(builder);
 }
 
 static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
@@ -1617,7 +1623,7 @@ static void normalize_image_u8_to_f32(const clip_image_u8* src, clip_image_f32*
     }
 }
 
-inline float clip(float x, float lower, float upper) {
+inline int clip(int x, int lower, int upper) {
     return std::max(lower, std::min(x, upper));
 }
 
@@ -1821,10 +1827,6 @@ static std::pair<int, int> uhd_get_refine_size(std::pair<int, int> original_size
     return refine_size;
 }
 
-inline int clip(int x, int lower, int upper) {
-    return std::max(lower, std::min(x, upper));
-}
-
 static std::pair<int, int> uhd_best_grid(const int max_slice_nums, const int multiple, const float log_ratio) {
     std::vector<int> candidate_split_grids_nums;
     for (int i : {multiple - 1, multiple, multiple + 1}) {

examples/llava/llava-cli.cpp

@@ -129,14 +129,14 @@ static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_para
         if (!params->image.empty()) {
             LOG_TEE("using base64 encoded image instead of command line image path\n");
         }
-        embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->n_threads, prompt);
+        embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt);
         if (!embed) {
             LOG_TEE("%s: can't load image from prompt\n", __func__);
             return NULL;
         }
         params->prompt = remove_image_from_prompt(prompt);
     } else {
-        embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->n_threads, fname.c_str());
+        embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->cpuparams.n_threads, fname.c_str());
         if (!embed) {
             fprintf(stderr, "%s: is %s really an image file?\n", __func__, fname.c_str());
             return NULL;

examples/llava/minicpmv-cli.cpp

@@ -180,7 +180,7 @@ static const char * sample(struct llama_sampling_context * ctx_sampling,
 
 static struct llava_context * minicpmv_init(gpt_params * params, const std::string & fname, int &n_past){
     auto ctx_clip = clip_init_context(params);
-    auto embeds = llava_image_embed_make_with_filename(ctx_clip, params->n_threads, fname.c_str());
+    auto embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str());
     if (!embeds) {
         std::cerr << "error: failed to load image " << fname << ". Terminating\n\n";
         return NULL;

examples/main/main.cpp

@@ -221,6 +221,40 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    LOG("%s: llama threadpool init = n_threads = %d\n",
+        __func__,
+        (int) params.cpuparams.n_threads
+    );
+    struct ggml_threadpool_params tpp_batch =
+            ggml_threadpool_params_from_cpu_params(params.cpuparams_batch);
+    struct ggml_threadpool_params tpp =
+            ggml_threadpool_params_from_cpu_params(params.cpuparams);
+
+    set_process_priority(params.cpuparams.priority);
+
+    struct ggml_threadpool * threadpool_batch = NULL;
+    if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
+        threadpool_batch = ggml_threadpool_new(&tpp_batch);
+        if (!threadpool_batch) {
+            LOG_TEE("%s: batch threadpool create failed : n_threads %d\n", __func__, tpp_batch.n_threads);
+            exit(1);
+        }
+
+        // Start the non-batch threadpool in the paused state
+        tpp.paused = true;
+    }
+
+    struct ggml_threadpool * threadpool = ggml_threadpool_new(&tpp);
+    if (!threadpool) {
+        LOG_TEE("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
+        exit(1);
+    }
+
+    llama_attach_threadpool(ctx, threadpool, threadpool_batch);
+    if (ctx_guidance) {
+        llama_attach_threadpool(ctx_guidance, threadpool, threadpool_batch);
+    }
+
     const int n_ctx_train = llama_n_ctx_train(model);
     const int n_ctx = llama_n_ctx(ctx);
     LOG("n_ctx: %d\n", n_ctx);
@@ -352,8 +386,8 @@ int main(int argc, char ** argv) {
     }
 
     LOGLN(
-            "recalculate the cached logits (check): embd_inp.empty() %s, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu, embd_inp.size() %zu",
+            "recalculate the cached logits (check): embd_inp.empty() %s, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu",
-            log_tostr(embd_inp.empty()), n_matching_session_tokens, embd_inp.size(), session_tokens.size(), embd_inp.size());
+            log_tostr(embd_inp.empty()), n_matching_session_tokens, embd_inp.size(), session_tokens.size());
 
     // if we will use the cache for the full prompt without reaching the end of the cache, force
     // reevaluation of the last token to recalculate the cached logits
@@ -989,6 +1023,9 @@ int main(int argc, char ** argv) {
     llama_sampling_free(ctx_sampling);
     llama_backend_free();
 
+    ggml_threadpool_free(threadpool);
+    ggml_threadpool_free(threadpool_batch);
+
 #ifndef LOG_DISABLE_LOGS
     LOG_TEE("Log end\n");
 #endif // LOG_DISABLE_LOGS

examples/quantize/quantize.cpp

@@ -106,7 +106,7 @@ static void usage(const char * executable) {
     printf("  --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n");
     printf("  --output-tensor-type ggml_type: use this ggml_type for the output.weight tensor\n");
     printf("  --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n");
-    printf("  --keep-split: will generate quatized model in the same shards as input");
+    printf("  --keep-split: will generate quantized model in the same shards as input\n");
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("      Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n");
     printf("Note: --include-weights and --exclude-weights cannot be used together\n");

examples/server/README.md

@@ -249,23 +249,49 @@ logging:
 
 Available environment variables (if specified, these variables will override parameters specified in arguments):
 
-- `LLAMA_CACHE` (cache directory, used by `--hf-repo`)
+- `LLAMA_CACHE`: cache directory, used by `--hf-repo`
-- `HF_TOKEN` (Hugging Face access token, used when accessing a gated model with `--hf-repo`)
+- `HF_TOKEN`: Hugging Face access token, used when accessing a gated model with `--hf-repo`
-- `LLAMA_ARG_MODEL`
+- `LLAMA_ARG_MODEL`: equivalent to `-m`
-- `LLAMA_ARG_THREADS`
+- `LLAMA_ARG_MODEL_URL`: equivalent to `-mu`
-- `LLAMA_ARG_CTX_SIZE`
+- `LLAMA_ARG_MODEL_ALIAS`: equivalent to `-a`
-- `LLAMA_ARG_N_PARALLEL`
+- `LLAMA_ARG_HF_REPO`: equivalent to `--hf-repo`
-- `LLAMA_ARG_BATCH`
+- `LLAMA_ARG_HF_FILE`: equivalent to `--hf-file`
-- `LLAMA_ARG_UBATCH`
+- `LLAMA_ARG_THREADS`: equivalent to `-t`
-- `LLAMA_ARG_N_GPU_LAYERS`
+- `LLAMA_ARG_CTX_SIZE`: equivalent to `-c`
-- `LLAMA_ARG_THREADS_HTTP`
+- `LLAMA_ARG_N_PARALLEL`: equivalent to `-np`
-- `LLAMA_ARG_CHAT_TEMPLATE`
+- `LLAMA_ARG_BATCH`: equivalent to `-b`
-- `LLAMA_ARG_N_PREDICT`
+- `LLAMA_ARG_UBATCH`: equivalent to `-ub`
-- `LLAMA_ARG_ENDPOINT_METRICS`
+- `LLAMA_ARG_N_GPU_LAYERS`: equivalent to `-ngl`
-- `LLAMA_ARG_ENDPOINT_SLOTS`
+- `LLAMA_ARG_THREADS_HTTP`: equivalent to `--threads-http`
-- `LLAMA_ARG_EMBEDDINGS`
+- `LLAMA_ARG_CHAT_TEMPLATE`: equivalent to `--chat-template`
-- `LLAMA_ARG_FLASH_ATTN`
+- `LLAMA_ARG_N_PREDICT`: equivalent to `-n`
-- `LLAMA_ARG_DEFRAG_THOLD`
+- `LLAMA_ARG_ENDPOINT_METRICS`: if set to `1`, it will enable metrics endpoint (equivalent to `--metrics`)
+- `LLAMA_ARG_ENDPOINT_SLOTS`: if set to `0`, it will **disable** slots endpoint (equivalent to `--no-slots`). This feature is enabled by default.
+- `LLAMA_ARG_EMBEDDINGS`: if set to `1`, it will enable embeddings endpoint (equivalent to `--embeddings`)
+- `LLAMA_ARG_FLASH_ATTN`: if set to `1`, it will enable flash attention (equivalent to `-fa`)
+- `LLAMA_ARG_CONT_BATCHING`: if set to `0`, it will **disable** continuous batching (equivalent to `--no-cont-batching`). This feature is enabled by default.
+- `LLAMA_ARG_DEFRAG_THOLD`: equivalent to `-dt`
+- `LLAMA_ARG_HOST`: equivalent to `--host`
+- `LLAMA_ARG_PORT`: equivalent to `--port`
+
+Example usage of docker compose with environment variables:
+
+```yml
+services:
+  llamacpp-server:
+    image: ghcr.io/ggerganov/llama.cpp:server
+    ports:
+      - 8080:8080
+    volumes:
+      - ./models:/models
+    environment:
+      # alternatively, you can use "LLAMA_ARG_MODEL_URL" to download the model
+      LLAMA_ARG_MODEL: /models/my_model.gguf
+      LLAMA_ARG_CTX_SIZE: 4096
+      LLAMA_ARG_N_PARALLEL: 2
+      LLAMA_ARG_ENDPOINT_METRICS: 1  # to disable, either remove or set to 0
+      LLAMA_ARG_PORT: 8080
+```
 
 ## Build
 

examples/server/tests/features/steps/steps.py

@@ -23,6 +23,8 @@ from prometheus_client import parser
 
 # pyright: reportRedeclaration=false
 
+DEFAULT_TIMEOUT_SECONDS = aiohttp.ClientTimeout(total=600)
+
 @step("a server listening on {server_fqdn}:{server_port}")
 def step_server_config(context, server_fqdn: str, server_port: str):
     context.server_fqdn = server_fqdn
@@ -689,7 +691,7 @@ def step_tokenize_set_add_special(context):
 @async_run_until_complete
 async def step_tokenize(context):
     context.tokenized_text = context_text(context)
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         tokenize_args = {
             "content": context.tokenized_text,
         }
@@ -706,7 +708,7 @@ async def step_tokenize(context):
 @async_run_until_complete
 async def step_detokenize(context):
     assert len(context.tokens) > 0
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with session.post(f'{context.base_url}/detokenize',
                                 json={
                                     "tokens": context.tokens,
@@ -735,7 +737,7 @@ def step_strings_for_tokenization(context):
 @step('an OPTIONS request is sent from {origin}')
 @async_run_until_complete
 async def step_options_request(context, origin):
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         headers = {'Authorization': f'Bearer {context.user_api_key}', 'Origin': origin}
         async with session.options(f'{context.base_url}/v1/chat/completions',
                                     headers=headers) as response:
@@ -751,7 +753,7 @@ def step_check_options_header_value(context, cors_header, cors_header_value):
 @step('prometheus metrics are exposed')
 @async_run_until_complete
 async def step_prometheus_metrics_exported(context):
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with await session.get(f'{context.base_url}/metrics') as metrics_response:
             assert metrics_response.status == 200
             assert metrics_response.headers['Content-Type'] == "text/plain; version=0.0.4"
@@ -818,13 +820,13 @@ async def concurrent_requests(context, f_completion, *args, **kwargs):
     for prompt_no in range(context.n_prompts):
         shifted_args = [context.prompts.pop(), seeds[prompt_no], *args]
         context.concurrent_tasks.append(asyncio.create_task(f_completion(*shifted_args, **kwargs)))
-    await asyncio.sleep(0.1)
+    await asyncio.sleep(0.01)
 
 
 @step('the slot {slot_id:d} is saved with filename "{filename}"')
 @async_run_until_complete
 async def step_save_slot(context, slot_id, filename):
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with session.post(f'{context.base_url}/slots/{slot_id}?action=save',
                                 json={"filename": filename},
                                 headers={"Content-Type": "application/json"}) as response:
@@ -834,7 +836,7 @@ async def step_save_slot(context, slot_id, filename):
 @step('the slot {slot_id:d} is restored with filename "{filename}"')
 @async_run_until_complete
 async def step_restore_slot(context, slot_id, filename):
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with session.post(f'{context.base_url}/slots/{slot_id}?action=restore',
                                 json={"filename": filename},
                                 headers={"Content-Type": "application/json"}) as response:
@@ -844,7 +846,7 @@ async def step_restore_slot(context, slot_id, filename):
 @step('the slot {slot_id:d} is erased')
 @async_run_until_complete
 async def step_erase_slot(context, slot_id):
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with session.post(f'{context.base_url}/slots/{slot_id}?action=erase',
                                 headers={"Content-Type": "application/json"}) as response:
             context.response = response
@@ -853,7 +855,7 @@ async def step_erase_slot(context, slot_id):
 @step('switch {on_or_off} lora adapter {lora_id:d}')
 @async_run_until_complete
 async def toggle_lora_adapter(context, on_or_off: str, lora_id: int):
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with session.post(f'{context.base_url}/lora-adapters',
                                 json=[{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}],
                                 headers={"Content-Type": "application/json"}) as response:
@@ -889,7 +891,7 @@ async def request_completion(prompt,
             print(f"Set user_api_key: {user_api_key}")
         headers['Authorization'] = f'Bearer {user_api_key}'
 
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with session.post(f'{base_url}/completion',
                                 json={
                                     "input_prefix": prompt_prefix,
@@ -902,8 +904,7 @@ async def request_completion(prompt,
                                     "temperature": temperature if temperature is not None else 0.8,
                                     "n_probs": 2,
                                 },
-                                headers=headers,
+                                headers=headers) as response:
-                                timeout=3600) as response:
             if expect_api_error is None or not expect_api_error:
                 assert response.status == 200
                 assert response.headers['Access-Control-Allow-Origin'] == origin
@@ -961,7 +962,7 @@ async def oai_chat_completions(user_prompt,
     if async_client:
         origin = 'llama.cpp'
         headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin}
-        async with aiohttp.ClientSession() as session:
+        async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
             async with session.post(f'{base_url}{base_path}',
                                     json=payload,
                                     headers=headers) as response:
@@ -1048,7 +1049,7 @@ async def oai_chat_completions(user_prompt,
 
 
 async def request_embedding(content, seed, base_url=None) -> list[list[float]]:
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with session.post(f'{base_url}/embedding',
                                 json={
                                     "content": content,
@@ -1068,14 +1069,13 @@ async def request_oai_embeddings(input, seed,
         headers=[]
         if user_api_key is not None:
             headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin}
-        async with aiohttp.ClientSession() as session:
+        async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
             async with session.post(f'{base_url}/v1/embeddings',
                                     json={
                                         "input": input,
                                         "model": model,
                                     },
-                                    headers=headers,
+                                    headers=headers) as response:
-                                    timeout=3600) as response:
                 assert response.status == 200, f"received status code not expected: {response.status}"
                 assert response.headers['Access-Control-Allow-Origin'] == origin
                 assert response.headers['Content-Type'] == "application/json; charset=utf-8"
@@ -1194,7 +1194,7 @@ async def wait_for_slots_status(context,
     if 'GITHUB_ACTIONS' in os.environ:
         timeout *= 2
 
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         while True:
             async with await session.get(f'{base_url}/slots', params=params) as slots_response:
                 status_code = slots_response.status
@@ -1237,7 +1237,7 @@ def assert_embeddings(embeddings):
 
 
 async def request_slots_status(context, expected_slots):
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with await session.get(f'{context.base_url}/slots') as slots_response:
             assert slots_response.status == 200
             slots = await slots_response.json()

examples/server/tests/features/wrong_usages.feature

@@ -8,9 +8,12 @@ Feature: Wrong usage of llama.cpp server
   Scenario: Infinite loop
     Given a server listening on localhost:8080
     And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
+    And   42 as server seed
+    And   2048 KV cache size
     # Uncomment below to fix the issue
     #And   64 server max tokens to predict
     Then  the server is starting
+    Then  the server is healthy
     Given a prompt:
       """
       Go to: infinite loop

examples/server/utils.hpp

@@ -3,6 +3,14 @@
 #include "llama.h"
 #include "common.h"
 
+#ifndef NDEBUG
+// crash the server in debug mode, otherwise send an http 500 error
+#define CPPHTTPLIB_NO_EXCEPTIONS 1
+#endif
+// increase max payload length to allow use of larger context size
+#define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576
+#include "httplib.h"
+
 // Change JSON_ASSERT from assert() to GGML_ASSERT:
 #define JSON_ASSERT GGML_ASSERT
 #include "json.hpp"
@@ -279,6 +287,18 @@ static size_t find_partial_stop_string(const std::string &stop, const std::strin
     return std::string::npos;
 }
 
+static bool json_is_array_of_numbers(json data) {
+    if (data.is_array()) {
+        for (const auto & e : data) {
+            if (!e.is_number()) {
+                return false;
+            }
+        }
+        return true;
+    }
+    return false;
+}
+
 // TODO: reuse llama_detokenize
 template <class Iter>
 static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
@@ -343,6 +363,19 @@ static json probs_vector_to_json(const llama_context * ctx, const std::vector<co
     return out;
 }
 
+static bool server_sent_event(httplib::DataSink & sink, const char * event, json & data) {
+    const std::string str =
+        std::string(event) + ": " +
+        data.dump(-1, ' ', false, json::error_handler_t::replace) +
+        "\n\n";
+
+    LOG_VERBOSE("data stream", {
+        { "to_send", str }
+    });
+
+    return sink.write(str.c_str(), str.size());
+}
+
 //
 // OAI utils
 //

examples/speculative/speculative.cpp

@@ -73,10 +73,11 @@ int main(int argc, char ** argv) {
     // load the draft model
     params.model = params.model_draft;
     params.n_gpu_layers = params.n_gpu_layers_draft;
-    if (params.n_threads_draft > 0) {
+    if (params.draft_cpuparams.n_threads > 0) {
-        params.n_threads = params.n_threads_draft;
+        params.cpuparams.n_threads = params.draft_cpuparams.n_threads;
     }
-    params.n_threads_batch = params.n_threads_batch_draft;
+
+    params.cpuparams_batch.n_threads = params.draft_cpuparams_batch.n_threads;
     llama_init_result llama_init_dft = llama_init_from_gpt_params(params);
     model_dft = llama_init_dft.model;
     ctx_dft = llama_init_dft.context;

flake.lock

@@ -5,11 +5,11 @@
         "nixpkgs-lib": "nixpkgs-lib"
       },
       "locked": {
-        "lastModified": 1722555600,
+        "lastModified": 1725024810,
-        "narHash": "sha256-XOQkdLafnb/p9ij77byFQjDf5m5QYl9b2REiVClC+x4=",
+        "narHash": "sha256-ODYRm8zHfLTH3soTFWE452ydPYz2iTvr9T8ftDMUQ3E=",
         "owner": "hercules-ci",
         "repo": "flake-parts",
-        "rev": "8471fe90ad337a8074e957b69ca4d0089218391d",
+        "rev": "af510d4a62d071ea13925ce41c95e3dec816c01d",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1723637854,
+        "lastModified": 1724819573,
-        "narHash": "sha256-med8+5DSWa2UnOqtdICndjDAEjxr5D7zaIiK4pn0Q7c=",
+        "narHash": "sha256-GnR7/ibgIH1vhoy8cYdmXE6iyZqKqFxQSVkFgosBh6w=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "c3aa7b8938b17aebd2deecf7be0636000d62a2b9",
+        "rev": "71e91c409d1e654808b2621f28a327acfdad8dc2",
         "type": "github"
       },
       "original": {

flake.nix

@@ -145,7 +145,9 @@
             # the same path you would with an overlay.
             legacyPackages = {
               llamaPackages = pkgs.callPackage .devops/nix/scope.nix { inherit llamaVersion; };
-              llamaPackagesWindows = pkgs.pkgsCross.mingwW64.callPackage .devops/nix/scope.nix { inherit llamaVersion; };
+              llamaPackagesWindows = pkgs.pkgsCross.mingwW64.callPackage .devops/nix/scope.nix {
+                inherit llamaVersion;
+              };
               llamaPackagesCuda = pkgsCuda.callPackage .devops/nix/scope.nix { inherit llamaVersion; };
               llamaPackagesRocm = pkgsRocm.callPackage .devops/nix/scope.nix { inherit llamaVersion; };
             };
@@ -157,6 +159,7 @@
                 default = config.legacyPackages.llamaPackages.llama-cpp;
                 vulkan = config.packages.default.override { useVulkan = true; };
                 windows = config.legacyPackages.llamaPackagesWindows.llama-cpp;
+                python-scripts = config.legacyPackages.llamaPackages.python-scripts;
               }
               // lib.optionalAttrs pkgs.stdenv.isLinux {
                 cuda = config.legacyPackages.llamaPackagesCuda.llama-cpp;

ggml/include/ggml-backend.h

@@ -63,6 +63,7 @@ extern "C" {
     GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
 
+    // "offset" refers to the offset of the tensor data for setting/getting data
     GGML_API GGML_CALL void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
 
@@ -102,6 +103,7 @@ extern "C" {
 
     GGML_API GGML_CALL bool ggml_backend_is_cpu                (ggml_backend_t backend);
     GGML_API           void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
+    GGML_API           void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
     GGML_API           void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
 
     // Create a backend buffer from an existing pointer

ggml/include/ggml.h

@@ -220,7 +220,7 @@
 #include <stdio.h>
 
 #define GGML_FILE_MAGIC   0x67676d6c // "ggml"
-#define GGML_FILE_VERSION 1
+#define GGML_FILE_VERSION 2
 
 #define GGML_QNT_VERSION        2    // bump this on quantization format changes
 #define GGML_QNT_VERSION_FACTOR 1000 // do not change this
@@ -231,6 +231,8 @@
 #define GGML_MAX_SRC            10
 #ifndef GGML_MAX_NAME
 #define GGML_MAX_NAME           64
+#define GGML_MAX_N_THREADS      512
+
 #endif
 #define GGML_MAX_OP_PARAMS      64
 #define GGML_DEFAULT_N_THREADS  4
@@ -455,6 +457,8 @@ extern "C" {
         GGML_OP_SQR,
         GGML_OP_SQRT,
         GGML_OP_LOG,
+        GGML_OP_SIN,
+        GGML_OP_COS,
         GGML_OP_SUM,
         GGML_OP_SUM_ROWS,
         GGML_OP_MEAN,
@@ -492,9 +496,11 @@ extern "C" {
         GGML_OP_CLAMP,
         GGML_OP_CONV_TRANSPOSE_1D,
         GGML_OP_IM2COL,
+        GGML_OP_IM2COL_BACK,
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
+        GGML_OP_POOL_2D_BACK,
         GGML_OP_UPSCALE, // nearest interpolate
         GGML_OP_PAD,
         GGML_OP_ARANGE,
@@ -510,6 +516,7 @@ extern "C" {
         GGML_OP_WIN_UNPART,
         GGML_OP_GET_REL_POS,
         GGML_OP_ADD_REL_POS,
+        GGML_OP_RWKV_WKV,
 
         GGML_OP_UNARY,
 
@@ -544,6 +551,7 @@ extern "C" {
         GGML_UNARY_OP_SILU,
         GGML_UNARY_OP_HARDSWISH,
         GGML_UNARY_OP_HARDSIGMOID,
+        GGML_UNARY_OP_EXP,
 
         GGML_UNARY_OP_COUNT,
     };
@@ -626,6 +634,29 @@ extern "C" {
     // If it returns true, the computation is aborted
     typedef bool (*ggml_abort_callback)(void * data);
 
+    // Scheduling priorities
+    enum ggml_sched_priority {
+        GGML_SCHED_PRIO_NORMAL,
+        GGML_SCHED_PRIO_MEDIUM,
+        GGML_SCHED_PRIO_HIGH,
+        GGML_SCHED_PRIO_REALTIME
+    };
+
+    // Threadpool params
+    // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
+    struct ggml_threadpool_params {
+        bool                cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
+        int                 n_threads;                   // number of threads
+        enum ggml_sched_priority prio;                   // thread priority
+        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
+        bool                strict_cpu;                  // strict cpu placement
+        bool                paused;                      // start in paused state
+    };
+
+    struct ggml_threadpool;     // forward declaration, see ggml.c
+
+    typedef struct  ggml_threadpool * ggml_threadpool_t;
+
     // the compute plan that needs to be prepared for ggml_graph_compute()
     // since https://github.com/ggerganov/ggml/issues/287
     struct ggml_cplan {
@@ -633,6 +664,7 @@ extern "C" {
         uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
 
         int n_threads;
+        struct ggml_threadpool * threadpool;
 
         // abort ggml_graph_compute when true
         ggml_abort_callback abort_callback;
@@ -971,6 +1003,22 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_sin(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_sin_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_cos(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_cos_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // return scalar
     GGML_API struct ggml_tensor * ggml_sum(
             struct ggml_context * ctx,
@@ -1121,6 +1169,14 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_exp(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_exp_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // normalize along rows
     GGML_API struct ggml_tensor * ggml_norm(
             struct ggml_context * ctx,
@@ -1568,34 +1624,49 @@ extern "C" {
             float                 min,
             float                 max);
 
+    // im2col
+    // converts data into a format that effectively results in a convolution when combined with matrix multiplication
     GGML_API struct ggml_tensor * ggml_im2col(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
+            struct ggml_tensor  * a,  // convolution kernel
-            struct ggml_tensor  * b,
+            struct ggml_tensor  * b,  // data
-            int                  s0,
+            int                   s0, // stride dimension 0
-            int                  s1,
+            int                   s1, // stride dimension 1
-            int                  p0,
+            int                   p0, // padding dimension 0
-            int                  p1,
+            int                   p1, // padding dimension 1
-            int                  d0,
+            int                   d0, // dilation dimension 0
-            int                  d1,
+            int                   d1, // dilation dimension 1
             bool                  is_2D,
             enum ggml_type        dst_type);
 
+    GGML_API struct ggml_tensor * ggml_im2col_back(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,  // convolution kernel
+        struct ggml_tensor  * b,  // gradient of im2col output
+        int64_t             * ne, // shape of im2col input
+        int                   s0, // stride dimension 0
+        int                   s1, // stride dimension 1
+        int                   p0, // padding dimension 0
+        int                   p1, // padding dimension 1
+        int                   d0, // dilation dimension 0
+        int                   d1, // dilation dimension 1
+        bool                  is_2D);
+
     GGML_API struct ggml_tensor * ggml_conv_depthwise_2d(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
+            struct ggml_tensor  * a,  // convolution kernel
-            struct ggml_tensor  * b,
+            struct ggml_tensor  * b,  // data
-            int                  s0,
+            int                  s0,  // stride dimension 0
-            int                  s1,
+            int                  s1,  // stride dimension 1
-            int                  p0,
+            int                  p0,  // padding dimension 0
-            int                  p1,
+            int                  p1,  // padding dimension 1
-            int                  d0,
+            int                  d0,  // dilation dimension 0
-            int                  d1);
+            int                  d1); // dilation dimension 1
 
     GGML_API struct ggml_tensor * ggml_conv_1d(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
+            struct ggml_tensor  * a,   // convolution kernel
-            struct ggml_tensor  * b,
+            struct ggml_tensor  * b,   // data
             int                   s0,  // stride
             int                   p0,  // padding
             int                   d0); // dilation
@@ -1604,29 +1675,29 @@ extern "C" {
     // alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d)
     GGML_API struct ggml_tensor* ggml_conv_1d_ph(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
+            struct ggml_tensor  * a,  // convolution kernel
-            struct ggml_tensor  * b,
+            struct ggml_tensor  * b,  // data
-            int                   s,
+            int                   s,  // stride
-            int                   d);
+            int                   d); // dilation
 
     GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
+            struct ggml_tensor  * a,   // convolution kernel
-            struct ggml_tensor  * b,
+            struct ggml_tensor  * b,   // data
-            int                   s0,
+            int                   s0,  // stride
-            int                   p0,
+            int                   p0,  // padding
-            int                   d0);
+            int                   d0); // dilation
 
     GGML_API struct ggml_tensor * ggml_conv_2d(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
+            struct ggml_tensor  * a,   // convolution kernel
-            struct ggml_tensor  * b,
+            struct ggml_tensor  * b,   // data
-            int                   s0,
+            int                   s0,  // stride dimension 0
-            int                   s1,
+            int                   s1,  // stride dimension 1
-            int                   p0,
+            int                   p0,  // padding dimension 0
-            int                   p1,
+            int                   p1,  // padding dimension 1
-            int                   d0,
+            int                   d0,  // dilation dimension 0
-            int                   d1);
+            int                   d1); // dilation dimension 1
 
 
     // kernel size is a->ne[0] x a->ne[1]
@@ -1688,6 +1759,18 @@ extern "C" {
             float                 p0,
             float                 p1);
 
+    GGML_API struct ggml_tensor * ggml_pool_2d_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * af, // "a"/input used in forward pass
+            enum ggml_op_pool     op,
+            int                   k0,
+            int                   k1,
+            int                   s0,
+            int                   s1,
+            float                 p0,
+            float                 p1);
+
     // nearest interpolate
     // multiplies ne0 and ne1 by scale factor
     // used in stable-diffusion
@@ -1762,7 +1845,8 @@ extern "C" {
             struct ggml_tensor  * v,
             struct ggml_tensor  * mask,
             float                 scale,
-            float                 max_bias);
+            float                 max_bias,
+            float                 logit_softcap);
 
     GGML_API void ggml_flash_attn_ext_set_prec(
             struct ggml_tensor * a,
@@ -1841,6 +1925,15 @@ extern "C" {
             struct ggml_tensor  * pw,
             struct ggml_tensor  * ph);
 
+    GGML_API struct ggml_tensor * ggml_rwkv_wkv(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * k,
+            struct ggml_tensor  * v,
+            struct ggml_tensor  * r,
+            struct ggml_tensor  * tf,
+            struct ggml_tensor  * td,
+            struct ggml_tensor  * state);
+
     // custom operators
 
     typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
@@ -2011,10 +2104,23 @@ extern "C" {
     GGML_API size_t ggml_graph_overhead(void);
     GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
 
+    GGML_API struct ggml_threadpool_params   ggml_threadpool_params_default(int n_threads);
+    GGML_API void                            ggml_threadpool_params_init  (struct ggml_threadpool_params *p, int n_threads);
+    GGML_API bool                            ggml_threadpool_params_match (const struct ggml_threadpool_params *p0, const struct ggml_threadpool_params *p1);
+    GGML_API struct ggml_threadpool*         ggml_threadpool_new          (struct ggml_threadpool_params  * params);
+    GGML_API void                            ggml_threadpool_free         (struct ggml_threadpool * threadpool);
+    GGML_API int                             ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
+    GGML_API void                            ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
+    GGML_API void                            ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
+
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
-    GGML_API struct ggml_cplan ggml_graph_plan   (const struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
+    GGML_API struct ggml_cplan ggml_graph_plan(
+                  const struct ggml_cgraph * cgraph,
+                                       int   n_threads, /* = GGML_DEFAULT_N_THREADS */
+                    struct ggml_threadpool * threadpool /* = NULL */ );
     GGML_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+
     // same as ggml_graph_compute() but the work data is allocated as a part of the context
     // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
     GGML_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);

ggml/src/CMakeLists.txt

@@ -1247,7 +1247,7 @@ endif()
 
 # Data types, macros and functions related to controlling CPU affinity and
 # some memory allocation are available on Linux through GNU extensions in libc
-if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+if (CMAKE_SYSTEM_NAME MATCHES "Linux" OR CMAKE_SYSTEM_NAME MATCHES "Android")
     add_compile_definitions(_GNU_SOURCE)
 endif()
 

ggml/src/ggml-aarch64.c

@@ -36,6 +36,84 @@
 //                         from bias offset form to pure sign form (this saves subtract
 //                         operations durin unpacking)
 //
+#if defined(__AVX__)
+#if defined(__F16C__)
+// the  _mm256_cvt intrinsics require F16C
+#define GGML_F32Cx8_LOAD(x)     _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(x)))
+#define GGML_F32Cx8_REPEAT_LOAD(x, loadMask)     _mm256_cvtph_ps(_mm_shuffle_epi32(_mm_maskload_epi32((int const*)(x), loadMask), 68))
+#define GGML_F32Cx8_REARRANGE_LOAD(x, arrangeMask)     _mm256_cvtph_ps(_mm_shuffle_epi8(_mm_loadu_si128((const __m128i *) x), arrangeMask))
+#else
+static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
+    float tmp[8];
+
+    for (int i = 0; i < 8; i++) {
+        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+    }
+
+    return _mm256_loadu_ps(tmp);
+}
+static inline __m256 __avx_repeat_f32cx8_load(ggml_fp16_t *x) {
+    float tmp[8];
+
+    for (int i = 0; i < 4; i++) {
+        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i + 4] = GGML_FP16_TO_FP32(x[i]);
+    }
+
+    return _mm256_loadu_ps(tmp);
+}
+static inline __m256 __avx_rearranged_f32cx8_load(ggml_fp16_t *x, __m128i arrangeMask) {
+    uint16_t tmphalf[8];
+    float tmp[8];
+
+    _mm_storeu_si128((__m128i*)tmphalf, _mm_shuffle_epi8(_mm_loadu_si128((const __m128i *) x), arrangeMask));
+    for (int i = 0; i < 8; i++) {
+        tmp[i] = GGML_FP16_TO_FP32(tmphalf[i]);
+    }
+
+    return _mm256_loadu_ps(tmp);
+}
+
+#define GGML_F32Cx8_LOAD(x)     __avx_f32cx8_load(x)
+#define GGML_F32Cx8_REPEAT_LOAD(x, loadMask)     __avx_repeat_f32cx8_load(x)
+#define GGML_F32Cx8_REARRANGE_LOAD(x, arrangeMask)     __avx_rearranged_f32cx8_load(x, arrangeMask)
+#endif
+#endif
+
+
+#if defined(__AVX2__) || defined(__AVX512F__)
+static inline __m256i sum_i16_pairs_int(const __m256i x) {
+    const __m256i ones = _mm256_set1_epi16(1);
+    return _mm256_madd_epi16(ones, x);
+}
+
+static inline __m256i mul_sum_us8_pairs_int(const __m256i ax, const __m256i sy) {
+#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+    const __m256i zero = _mm256_setzero_si256();
+    return _mm256_dpbusd_epi32(zero, ax, sy);
+#else
+    // Perform multiplication and create 16-bit values
+    const __m256i dot = _mm256_maddubs_epi16(ax, sy);
+    return sum_i16_pairs_int(dot);
+#endif
+}
+
+// Integer variant of the function defined in ggml-quants.c
+// multiply int8_t, add results pairwise twice and return as float vector
+static inline __m256i mul_sum_i8_pairs_int(const __m256i x, const __m256i y) {
+#if __AVXVNNIINT8__
+    const __m256i zero = _mm256_setzero_si256();
+    return _mm256_dpbssd_epi32(zero, x, y);
+#else
+    // Get absolute values of x vectors
+    const __m256i ax = _mm256_sign_epi8(x, x);
+    // Sign the values of the y vectors
+    const __m256i sy = _mm256_sign_epi8(y, x);
+    return mul_sum_us8_pairs_int(ax, sy);
+#endif
+}
+#endif
+
 static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
     block_q4_0x4 out;
 
@@ -255,6 +333,103 @@ void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int64_t k)
             y[i].qs[32 * j + 31] = vgetq_lane_s32(vi, 3);
         }
     }
+#elif defined(__AVX2__) || defined(__AVX__)
+    float id[4];
+    __m256 srcv[4][4];
+    __m256 idvec[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            // Load elements into 4 AVX vectors
+            __m256 v0 = _mm256_loadu_ps( x + row_iter * k + i * 32 );
+            __m256 v1 = _mm256_loadu_ps( x + row_iter * k + i * 32 + 8 );
+            __m256 v2 = _mm256_loadu_ps( x + row_iter * k + i * 32 + 16 );
+            __m256 v3 = _mm256_loadu_ps( x + row_iter * k + i * 32 + 24 );
+
+            // Compute max(abs(e)) for the block
+            const __m256 signBit = _mm256_set1_ps( -0.0f );
+            __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
+            maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
+            maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
+            maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
+
+            __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
+            max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
+            max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
+            const float maxScalar = _mm_cvtss_f32( max4 );
+
+            // Divided by 127.f to mirror results in quantize_row_q8_0
+            const float d = maxScalar  / 127.f;
+            id[row_iter] = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; //d ? 1.0f / d : 0.0f;
+
+            // Store the scale for the individual block
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+
+            // Store the values in blocks of eight values - Aim is to use these later for block interleaving
+            srcv[row_iter][0] = v0;
+            srcv[row_iter][1] = v1;
+            srcv[row_iter][2] = v2;
+            srcv[row_iter][3] = v3;
+            idvec[row_iter] = _mm256_set1_ps(id[row_iter]);
+        }
+
+        // The loop iterates four times - The aim is to get 4 corresponding chunks of eight bytes from the original weight blocks that are interleaved
+        for (int j = 0; j < 4; j++) {
+            // Apply the multiplier
+            __m256 v0 = _mm256_mul_ps(srcv[0][j], idvec[0]);
+            __m256 v1 = _mm256_mul_ps(srcv[1][j], idvec[1]);
+            __m256 v2 = _mm256_mul_ps(srcv[2][j], idvec[2]);
+            __m256 v3 = _mm256_mul_ps(srcv[3][j], idvec[3]);
+
+            // Round to nearest integer
+            v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
+            v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
+            v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
+            v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
+
+            // Convert floats to integers
+            __m256i i0 = _mm256_cvtps_epi32( v0 );
+            __m256i i1 = _mm256_cvtps_epi32( v1 );
+            __m256i i2 = _mm256_cvtps_epi32( v2 );
+            __m256i i3 = _mm256_cvtps_epi32( v3 );
+
+#if defined(__AVX2__)
+            // Convert int32 to int16
+            i0 = _mm256_packs_epi32( i0, i1 );
+            i2 = _mm256_packs_epi32( i2, i3 );
+            // Convert int16 to int8
+            i0 = _mm256_packs_epi16( i0, i2 );
+
+            //  Permute and store the quantized weights in the required order after the pack instruction
+            const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
+            i0 = _mm256_permutevar8x32_epi32( i0, perm );
+
+            _mm256_storeu_si256((__m256i *)(y[i].qs + 32 * j), i0);
+#else
+            // Since we don't have in AVX some necessary functions,
+            // we split the registers in half and call AVX2 analogs from SSE
+            __m128i ni0 = _mm256_castsi256_si128( i0 );
+            __m128i ni1 = _mm256_extractf128_si256( i0, 1);
+            __m128i ni2 = _mm256_castsi256_si128( i1 );
+            __m128i ni3 = _mm256_extractf128_si256( i1, 1);
+            __m128i ni4 = _mm256_castsi256_si128( i2 );
+            __m128i ni5 = _mm256_extractf128_si256( i2, 1);
+            __m128i ni6 = _mm256_castsi256_si128( i3 );
+            __m128i ni7 = _mm256_extractf128_si256( i3, 1);
+
+            // Convert int32 to int16
+            ni0 = _mm_packs_epi32( ni0, ni1 );
+            ni2 = _mm_packs_epi32( ni2, ni3 );
+            ni4 = _mm_packs_epi32( ni4, ni5 );
+            ni6 = _mm_packs_epi32( ni6, ni7 );
+            // Convert int16 to int8
+            ni0 = _mm_packs_epi16( ni0, ni2 );
+            ni4 = _mm_packs_epi16( ni4, ni6 );
+            _mm_storeu_si128((__m128i *)(y[i].qs + 32 * j), ni0);
+            _mm_storeu_si128((__m128i *)(y[i].qs + 32 * j + 16), ni4);
+#endif
+        }
+    }
 #else
     // scalar
     const int blck_size_interleave = 8;
@@ -337,34 +512,19 @@ static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict ds
 }
 
 size_t quantize_q4_0_4x4(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    if (!quant_weights) {
+    UNUSED(quant_weights);
     return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4);
 }
-    else {
-        assert(false);
-        return 0;
-    }
-}
 
 size_t quantize_q4_0_4x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    if (!quant_weights) {
+    UNUSED(quant_weights);
     return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8);
 }
-    else {
-        assert(false);
-        return 0;
-    }
-}
 
 size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    if (!quant_weights) {
+    UNUSED(quant_weights);
     return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
 }
-    else {
-        assert(false);
-        return 0;
-    }
-}
 
 void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
     const int qk = QK8_0;
@@ -699,6 +859,96 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
     GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) &&
                 "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal "
                 "performance");
+#elif defined(__AVX2__)
+    // Lookup table to convert signed nibbles to signed bytes
+    __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
+    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+    __m128i changemask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
+    __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
+
+    // Permute mask used for easier vector processing at later stages
+    const __m256i m4b = _mm256_set1_epi8(0x0F);
+
+    int64_t b_nb = n / QK4_0;
+
+    const block_q4_0x8 * b_ptr_start = (const block_q4_0x8 *)vx;
+    const block_q8_0 * a_ptr_start = (const block_q8_0 *)vy;
+
+    // Process Q8_0 blocks one by one
+    for (int64_t y = 0; y < nr; y++) {
+
+        // Pointers to LHS blocks of block_q8_0 format
+        const block_q8_0 * a_ptr = a_ptr_start + (y * nb);
+
+        // Take group of eight block_q4_0x8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = 0; x < nc / 8; x++) {
+
+            // Pointers to RHS blocks
+            const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb);
+
+            // Master FP accumulator
+            __m256 acc_row = _mm256_setzero_ps();
+
+            for (int64_t b = 0; b < nb; b++) {
+                // Load 8 blocks of Q4_0 interleaved as 8 bytes (B0 - B7)
+                const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
+                const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 1);
+                const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 2);
+                const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 3);
+
+                // 4-bit -> 8-bit - Sign is maintained
+                const __m256i rhs_vec_0123_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_vec_0123_0, m4b)); // B0(0-7) B1(0-7) B2(0-7) B3(0-7)
+                const __m256i rhs_vec_4567_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_vec_4567_0, m4b)); // B4(0-7) B5(0-7) B6(0-7) B7(0-7)
+                const __m256i rhs_vec_0123_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_vec_0123_1, m4b)); // B0(8-15) B1(8-15) B2(8-15) B3(8-15)
+                const __m256i rhs_vec_4567_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_vec_4567_1, m4b)); // B0(8-15) B1(8-15) B2(8-15) B3(8-15)
+
+                const __m256i rhs_vec_0123_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m4b)); // B0(16-23) B1(16-23) B2(16-23) B3(16-23)
+                const __m256i rhs_vec_4567_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m4b)); // B4(16-23) B5(16-23) B6(16-23) B7(16-23)
+                const __m256i rhs_vec_0123_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b)); // B0(24-31) B1(24-31) B2(24-31) B3(24-31)
+                const __m256i rhs_vec_4567_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b)); // B4(24-31) B5(24-31) B6(24-31) B7(24-31)
+
+                // Load the scale values for the 8 blocks interleaved in block_q4_0x8
+                const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);
+
+                // Load and convert to FP32 scale from block_q8_0
+                const __m256 row_scale_f32 = _mm256_set1_ps(GGML_FP16_TO_FP32(a_ptr[b].d));
+
+                // Load the block values in block_q8_0 in batches of 16 bytes and replicate the same across 256 bit vector
+                __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)a_ptr[b].qs));
+                __m256i lhs_vec_1 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16)));
+
+                lhs_vec_0 = _mm256_permute2f128_si256(lhs_vec_0, lhs_vec_0, 0); // A0 (0-15) A0(0-15)
+                lhs_vec_1 = _mm256_permute2f128_si256(lhs_vec_1, lhs_vec_1, 0); // A0 (16-31) A0(16-31))
+
+                __m256i iacc = _mm256_setzero_si256();
+
+                // Dot product done within 32 bit lanes and accumulated in the same vector
+                // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
+                // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
+                // ...........................................................................
+                // B0(28-31) B4(28-31) B1(28-31) B5(28-31) B2(28-31) B6(28-31) B3(28-31) B7(28-31) with A0(28-31)
+
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int(_mm256_blend_epi32(rhs_vec_0123_0 ,_mm256_shuffle_epi32(rhs_vec_4567_0, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 0)));
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_0, 177) ,rhs_vec_4567_0, 170), _mm256_shuffle_epi32(lhs_vec_0, 85)));
+
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int(_mm256_blend_epi32(rhs_vec_0123_1 ,_mm256_shuffle_epi32(rhs_vec_4567_1, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 170)));
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_1, 177) ,rhs_vec_4567_1, 170), _mm256_shuffle_epi32(lhs_vec_0, 255)));
+
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int(_mm256_blend_epi32(rhs_vec_0123_2 ,_mm256_shuffle_epi32(rhs_vec_4567_2, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 0)));
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_2, 177) ,rhs_vec_4567_2, 170), _mm256_shuffle_epi32(lhs_vec_1, 85)));
+
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int(_mm256_blend_epi32(rhs_vec_0123_3 ,_mm256_shuffle_epi32(rhs_vec_4567_3, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 170)));
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_3, 177) ,rhs_vec_4567_3, 170), _mm256_shuffle_epi32(lhs_vec_1, 255)));
+
+                // Accumulated values multipled with appropriate scales
+                acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
+            }
+
+            // Accumulated output values permuted so as to be stored in appropriate order post accumulation
+            acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
+            _mm256_storeu_ps(s + (y * nr + x * 8), acc_row);
+        }
+    }
 #else
     float sumf[8];
     int sumi;
@@ -2158,6 +2408,353 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
     GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) &&
                 "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal "
                 "performance");
+#elif defined(__AVX2__) || defined(__AVX512F__)
+    const block_q4_0x8 * b_ptr_start = (const block_q4_0x8 *)vx;
+    const block_q8_0x4 * a_ptr_start = (const block_q8_0x4 *)vy;
+    int64_t b_nb = n / QK4_0;
+    int64_t y = 0;
+    // Mask to mask out nibbles from packed bytes
+    const __m256i m4b = _mm256_set1_epi8(0x0F);
+    const __m128i loadMask = _mm_blend_epi32(_mm_setzero_si128(), _mm_set1_epi32(0xFFFFFFFF), 3);
+    // Lookup table to convert signed nibbles to signed bytes
+    __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
+    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+    // Permute mask used for easier vector processing at later stages
+    __m256i requiredOrder = _mm256_set_epi32(3 ,2 ,1 ,0, 7 ,6, 5, 4);
+
+    // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
+    int anr = nr - nr %16; // Used to align nr with boundary of 16
+
+    for (; y < anr / 4; y += 4) {
+        const block_q8_0x4 * a_ptrs[4];
+
+        a_ptrs[0] = a_ptr_start + (y * nb);
+        for (int i = 0; i < 3; ++i) {
+            a_ptrs[i + 1] = a_ptrs[i] + nb;
+        }
+
+        // Take group of eight block_q4_0x8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = 0; x < nc / 8; x++) {
+
+            const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb);
+
+            // Master FP accumulators
+            __m256 acc_rows[16];
+            for (int i = 0; i < 16; i++) {
+                acc_rows[i] = _mm256_setzero_ps();
+            }
+
+            for (int64_t b = 0; b < nb; b++) {
+                // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+                const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
+                const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
+                const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
+                const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+
+                // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
+                const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                // 4-bit -> 8-bit - Sign is maintained
+                const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
+                const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+
+                const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
+                const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+
+                const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
+                const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+
+                const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
+                const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+
+                // Shuffle pattern one - right side input
+                const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136);  //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
+                const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136);  //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+
+                const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136);  //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
+                const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136);  //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+
+                const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136);  //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
+                const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136);  //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+
+                const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136);  //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
+                const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136);  //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+
+                // Shuffle pattern two - right side input
+
+                const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221);  //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
+                const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221);  //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+
+                const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221);  //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
+                const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221);  //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+
+                const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221);  //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
+                const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221);  //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+
+                const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221);  //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
+                const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221);  //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+
+                // Scale values - Load the wight scale values of block_q4_0x8
+                const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+
+                // Process LHS in groups of four
+                for (int rp = 0; rp < 4; rp++) {
+                    // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+                    // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
+                    __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
+                    __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
+                    __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
+                    __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
+                    __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
+                    __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
+                    __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
+                    __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
+                    __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
+                    __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
+                    __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
+                    __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+
+                    // Shuffle pattern one - left side input
+                    const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                    const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+
+                    const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                    const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+
+                    const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                    const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+
+                    const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                    const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+
+                    // Shuffle pattern two - left side input
+                    const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                    const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+
+                    const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                    const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+
+                    const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                    const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+
+                    const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                    const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+
+                    // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                    // Resembles MMLAs into 2x2 matrices in ARM Version
+                    __m256i iacc_mat_00_sp1 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), mul_sum_i8_pairs_int(lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1)), mul_sum_i8_pairs_int(lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1)), mul_sum_i8_pairs_int(lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1));
+                    __m256i iacc_mat_01_sp1 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), mul_sum_i8_pairs_int(lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1)), mul_sum_i8_pairs_int(lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1)), mul_sum_i8_pairs_int(lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1));
+                    __m256i iacc_mat_10_sp1 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), mul_sum_i8_pairs_int(lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1)), mul_sum_i8_pairs_int(lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1)), mul_sum_i8_pairs_int(lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1));
+                    __m256i iacc_mat_11_sp1 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), mul_sum_i8_pairs_int(lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1)), mul_sum_i8_pairs_int(lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1)), mul_sum_i8_pairs_int(lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1));
+                    __m256i iacc_mat_00_sp2 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), mul_sum_i8_pairs_int(lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2)), mul_sum_i8_pairs_int(lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2)), mul_sum_i8_pairs_int(lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2));
+                    __m256i iacc_mat_01_sp2 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), mul_sum_i8_pairs_int(lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2)), mul_sum_i8_pairs_int(lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2)), mul_sum_i8_pairs_int(lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2));
+                    __m256i iacc_mat_10_sp2 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), mul_sum_i8_pairs_int(lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2)), mul_sum_i8_pairs_int(lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2)), mul_sum_i8_pairs_int(lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2));
+                    __m256i iacc_mat_11_sp2 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), mul_sum_i8_pairs_int(lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2)), mul_sum_i8_pairs_int(lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2)), mul_sum_i8_pairs_int(lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2));
+
+                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                    __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+                    __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+                    __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+                    __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+
+                    // Straighten out to make 4 row vectors
+                    __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
+                    __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
+                    __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
+                    __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+
+                    // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+                    const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
+
+                    // Multiply with appropiate scales and accumulate
+                    acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
+                    acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
+                    acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
+                    acc_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32,  255)), acc_rows[rp * 4 + 3]);
+                }
+            }
+
+            // Store the accumulated values
+            for (int i = 0; i < 16; i++) {
+                _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+            }
+        }
+    }
+
+    // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
+    for (; y < nr / 4; y ++) {
+
+        const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
+
+        // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+        for (int64_t x = 0; x < nc / 8; x++) {
+
+            const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb);
+
+            // Master FP accumulators
+            __m256 acc_rows[4];
+            for (int i = 0; i < 4; i++) {
+                acc_rows[i] = _mm256_setzero_ps();
+            }
+
+            for (int64_t b = 0; b < nb; b++) {
+                // Load the eight block_q8_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+                const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
+                const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
+                const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
+                const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+
+                // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
+                const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                // 4-bit -> 8-bit - Sign is maintained
+                const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b));  //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
+                const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b));  //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+
+                const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b));  //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
+                const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b));  //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+
+                const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b));  //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
+                const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b));  //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+
+                const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b));  //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
+                const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b));  //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+
+                // Shuffle pattern one - right side input
+                const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136);  //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
+                const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136);  //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+
+                const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136);  //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
+                const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136);  //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+
+                const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136);  //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
+                const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136);  //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+
+                const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136);  //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
+                const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136);  //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+
+                // Shuffle pattern two - right side input
+
+                const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221);  //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
+                const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221);  //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+
+                const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221);  //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
+                const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221);  //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+
+                const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221);  //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
+                const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221);  //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+
+                const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221);  //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
+                const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221);  //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+
+                // Scale values - Load the wight scale values of block_q4_0x8
+                const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+
+                // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+                // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
+                __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
+                __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
+                __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
+                __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
+                __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
+                __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
+                __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
+                __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
+                __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
+                __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
+                __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
+                __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+
+                // Shuffle pattern one - left side input
+
+                const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+
+                const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+
+                const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+
+                const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+
+                // Shuffle pattern two - left side input
+
+                const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+
+                const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+
+                const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+
+                const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+
+                // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                // Resembles MMLAs into 2x2 matrices in ARM Version
+                __m256i iacc_mat_00_sp1 =
+                    _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), mul_sum_i8_pairs_int(lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1)), mul_sum_i8_pairs_int(lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1)), mul_sum_i8_pairs_int(lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1));
+                __m256i iacc_mat_01_sp1 =
+                    _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), mul_sum_i8_pairs_int(lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1)), mul_sum_i8_pairs_int(lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1)), mul_sum_i8_pairs_int(lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1));
+                __m256i iacc_mat_10_sp1 =
+                    _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), mul_sum_i8_pairs_int(lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1)), mul_sum_i8_pairs_int(lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1)), mul_sum_i8_pairs_int(lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1));
+                __m256i iacc_mat_11_sp1 =
+                    _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), mul_sum_i8_pairs_int(lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1)), mul_sum_i8_pairs_int(lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1)), mul_sum_i8_pairs_int(lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1));
+                __m256i iacc_mat_00_sp2 =
+                    _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), mul_sum_i8_pairs_int(lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2)), mul_sum_i8_pairs_int(lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2)), mul_sum_i8_pairs_int(lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2));
+                __m256i iacc_mat_01_sp2 =
+                    _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), mul_sum_i8_pairs_int(lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2)), mul_sum_i8_pairs_int(lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2)), mul_sum_i8_pairs_int(lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2));
+                __m256i iacc_mat_10_sp2 =
+                    _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), mul_sum_i8_pairs_int(lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2)), mul_sum_i8_pairs_int(lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2)), mul_sum_i8_pairs_int(lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2));
+                __m256i iacc_mat_11_sp2 =
+                    _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int(lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), mul_sum_i8_pairs_int(lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2)), mul_sum_i8_pairs_int(lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2)), mul_sum_i8_pairs_int(lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2));
+
+                // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+                __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+                __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+                __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+
+
+                // Straighten out to make 4 row vectors
+                __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
+                __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
+                __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
+                __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+
+                // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+                const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask);
+
+                // Multiply with appropiate scales and accumulate
+                acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
+                acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
+                acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
+                acc_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
+            }
+
+            // Store the accumulated values
+            for (int i = 0; i < 4; i++) {
+                _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+            }
+        }
+    }
 #else
     float sumf[4][8];
     int sumi;

ggml/src/ggml-backend.c

@@ -723,6 +723,8 @@ ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
 
 struct ggml_backend_cpu_context {
     int                 n_threads;
+    ggml_threadpool_t   threadpool;
+
     void *              work_data;
     size_t              work_size;
 
@@ -759,7 +761,7 @@ GGML_CALL static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(gg
 
     struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu));
 
-    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
+    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
     cpu_plan->cgraph = *cgraph; // FIXME: deep copy
 
     if (cpu_plan->cplan.work_size > 0) {
@@ -796,7 +798,7 @@ GGML_CALL static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backe
 GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
 
-    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
+    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
 
     if (cpu_ctx->work_size < cplan.work_size) {
         free(cpu_ctx->work_data);
@@ -873,6 +875,7 @@ ggml_backend_t ggml_backend_cpu_init(void) {
     }
 
     ctx->n_threads           = GGML_DEFAULT_N_THREADS;
+    ctx->threadpool          = NULL;
     ctx->work_data           = NULL;
     ctx->work_size           = 0;
     ctx->abort_callback      = NULL;
@@ -903,6 +906,18 @@ void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
     ctx->n_threads = n_threads;
 }
 
+void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+
+    if (ctx->threadpool && ctx->threadpool != threadpool) {
+        // already had a different threadpool, pause/suspend it before switching
+        ggml_threadpool_pause(ctx->threadpool);
+    }
+    ctx->threadpool = threadpool;
+}
+
 void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
     GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
 

ggml/src/ggml-cuda.cu

@@ -9,8 +9,10 @@
 #include "ggml-cuda/binbcast.cuh"
 #include "ggml-cuda/clamp.cuh"
 #include "ggml-cuda/concat.cuh"
+#include "ggml-cuda/conv-transpose-1d.cuh"
 #include "ggml-cuda/convert.cuh"
 #include "ggml-cuda/cpy.cuh"
+#include "ggml-cuda/cross-entropy-loss.cuh"
 #include "ggml-cuda/diagmask.cuh"
 #include "ggml-cuda/dmmv.cuh"
 #include "ggml-cuda/fattn.cuh"
@@ -29,7 +31,6 @@
 #include "ggml-cuda/tsembd.cuh"
 #include "ggml-cuda/unary.cuh"
 #include "ggml-cuda/upscale.cuh"
-#include "ggml-cuda/conv-transpose-1d.cuh"
 
 #include <algorithm>
 #include <array>
@@ -2181,6 +2182,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_ADD:
             ggml_cuda_op_add(ctx, dst);
             break;
+        case GGML_OP_SUB:
+            ggml_cuda_op_sub(ctx, dst);
+            break;
         case GGML_OP_ACC:
             ggml_cuda_op_acc(ctx, dst);
             break;
@@ -2267,6 +2271,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SQRT:
             ggml_cuda_op_sqrt(ctx, dst);
             break;
+        case GGML_OP_SIN:
+            ggml_cuda_op_sin(ctx, dst);
+            break;
+        case GGML_OP_COS:
+            ggml_cuda_op_cos(ctx, dst);
+            break;
         case GGML_OP_CLAMP:
             ggml_cuda_op_clamp(ctx, dst);
             break;
@@ -2303,6 +2313,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_FLASH_ATTN_EXT:
             ggml_cuda_flash_attn_ext(ctx, dst);
             break;
+        case GGML_OP_CROSS_ENTROPY_LOSS:
+            ggml_cuda_cross_entropy_loss(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -2610,6 +2623,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
                 assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
                 for (int j = 0; j < GGML_MAX_SRC; j++) {
                     if (node->src[j] != nullptr) {
+                        assert(node->src[j]->buffer);
                         assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) || ggml_backend_buffer_is_cuda_split(node->src[j]->buffer));
                     }
                 }
@@ -2853,12 +2867,15 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_TRANSPOSE:
         case GGML_OP_NORM:
         case GGML_OP_ADD:
+        case GGML_OP_SUB:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_RMS_NORM:
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
+        case GGML_OP_SIN:
+        case GGML_OP_COS:
         case GGML_OP_CLAMP:
         case GGML_OP_CONT:
         case GGML_OP_DIAG_MASK_INF:
@@ -2890,6 +2907,8 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
             }
             return ggml_cuda_info().devices[cuda_ctx->device].cc >= CC_VOLTA &&
                 op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
+        case GGML_OP_CROSS_ENTROPY_LOSS:
+            return true;
 #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
         default:
             return false;

ggml/src/ggml-cuda/binbcast.cu

@@ -9,6 +9,10 @@ static __device__ __forceinline__ float op_add(const float a, const float b) {
     return a + b;
 }
 
+static __device__ __forceinline__ float op_sub(const float a, const float b) {
+    return a - b;
+}
+
 static __device__ __forceinline__ float op_mul(const float a, const float b) {
     return a * b;
 }
@@ -271,6 +275,10 @@ void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_add>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
 }
 
+void ggml_cuda_op_sub(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_sub>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
+}
+
 void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_mul>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
 }

ggml/src/ggml-cuda/binbcast.cuh

@@ -2,5 +2,6 @@
 
 void ggml_cuda_op_repeat(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_sub(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/cross-entropy-loss.cu

@@ -0,0 +1,106 @@
+#include "common.cuh"
+#include "cross-entropy-loss.cuh"
+#include "sumrows.cuh"
+
+#include <cmath>
+#include <cstdint>
+
+static __global__ void cross_entropy_loss_f32(const float * logits, const float * labels, float * dst, const int nclasses, const int k) {
+    const int warp_id = threadIdx.x / WARP_SIZE;
+    const int lane_id = threadIdx.x % WARP_SIZE;
+    const int i0 = blockDim.x*blockIdx.x + warp_id*WARP_SIZE;
+
+    const int ne_tmp = WARP_SIZE*nclasses;
+
+    extern __shared__ float tmp_all[];
+    float * tmp_logits = tmp_all + (2*warp_id + 0)*ne_tmp;
+    float * tmp_labels = tmp_all + (2*warp_id + 1)*ne_tmp;
+
+    // Each warp first loads ne_tmp logits/labels into shared memory:
+    for (int i = lane_id; i < ne_tmp; i += WARP_SIZE) {
+        const int ig = i0*nclasses + i; // ig == i global
+
+        tmp_logits[i] = ig < k*nclasses ? logits[ig] : 0.0f;
+        tmp_labels[i] = ig < k*nclasses ? labels[ig] : 0.0f;
+    }
+
+    // Each thread in the warp then calculates the cross entropy loss for a single row.
+    // TODO: pad in order to avoid shared memory bank conflicts.
+
+    // Find maximum for softmax:
+    float max = -INFINITY;
+    for (int i = 0; i < nclasses; ++i) {
+        max = fmaxf(max, tmp_logits[lane_id*nclasses + i]);
+    }
+
+    // Calculate log(softmax(logits)) which is just logits - max:
+    float sum = 0.0f;
+    for (int i = 0; i < nclasses; ++i) {
+        float val = tmp_logits[lane_id*nclasses + i] - max;
+        sum += expf(val);
+        tmp_logits[lane_id*nclasses + i] = val;
+    }
+    sum = logf(sum);
+
+    // log(exp(logits - max) / sum) = (logits - max) - log(sum)
+    float loss = 0.0f;
+    for (int i = 0; i < nclasses; ++i) {
+        loss += (tmp_logits[lane_id*nclasses + i] - sum) * tmp_labels[lane_id*nclasses + i];
+    }
+    loss = -warp_reduce_sum(loss) / (float)k;
+
+    __syncthreads();
+
+    if (lane_id == 0) {
+        tmp_all[warp_id] = loss;
+    }
+
+    __syncthreads();
+
+    if (warp_id != 0) {
+        return;
+    }
+
+    loss = lane_id < CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE/WARP_SIZE ? tmp_all[lane_id] : 0.0f;
+    loss = warp_reduce_sum(loss);
+
+    if (lane_id != 0) {
+        return;
+    }
+
+    dst[blockIdx.x] = loss;
+}
+
+void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    const int64_t ne00  = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    const float * src0_d = (const float *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       * dst_d  = (float       *) dst->data;
+
+    ggml_cuda_pool & pool = ctx.pool();
+    cudaStream_t stream = ctx.stream();
+
+    const dim3 blocks_dim(CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE, 1, 1);
+    const dim3 blocks_num((nrows + CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE - 1) / CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE, 1, 1);
+    const int shmem = 2*CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE*ne00*sizeof(float);
+
+    ggml_cuda_pool_alloc<float> dst_tmp(pool, blocks_num.x);
+
+    cross_entropy_loss_f32<<<blocks_num, blocks_dim, shmem, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
+
+    // Combine results from individual blocks:
+    sum_rows_f32_cuda(dst_tmp.ptr, dst_d, blocks_num.x, 1, stream);
+}

ggml/src/ggml-cuda/cross-entropy-loss.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE 256
+
+void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -22,6 +22,7 @@ typedef void (* fattn_kernel_t)(
         const float m0,
         const float m1,
         const uint32_t n_head_log2,
+        const float logit_softcap,
         const int ne00,
         const int ne01,
         const int ne02,
@@ -659,9 +660,15 @@ void launch_fattn(
 
     float scale         = 1.0f;
     float max_bias      = 0.0f;
+    float logit_softcap = 0.0f;
 
     memcpy(&scale,         (float *) KQV->op_params + 0, sizeof(float));
     memcpy(&max_bias,      (float *) KQV->op_params + 1, sizeof(float));
+    memcpy(&logit_softcap, (float *) KQV->op_params + 2, sizeof(float));
+
+    if (logit_softcap != 0.0f) {
+        scale /= logit_softcap;
+    }
 
     const uint32_t n_head      = Q->ne[2];
     const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
@@ -675,7 +682,7 @@ void launch_fattn(
         V_data,
         mask ? ((const char *) mask->data) : nullptr,
         (parallel_blocks) == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
-        scale, max_bias, m0, m1, n_head_log2,
+        scale, max_bias, m0, m1, n_head_log2, logit_softcap,
         Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
         K->ne[0], K->ne[1], K->ne[2], K->ne[3],
         mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,

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

@@ -4,7 +4,7 @@
 
 #define FATTN_KQ_STRIDE_TILE_F16 64
 
-template<int D, int ncols, int nwarps, int parallel_blocks> // D == head size
+template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
 #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
@@ -20,6 +20,7 @@ static __global__ void flash_attn_tile_ext_f16(
         const float m0,
         const float m1,
         const uint32_t n_head_log2,
+        const float logit_softcap,
         const int ne00,
         const int ne01,
         const int ne02,
@@ -44,6 +45,12 @@ static __global__ void flash_attn_tile_ext_f16(
         const int ne2,
         const int ne3) {
 #ifdef FP16_AVAILABLE
+    // Skip unused kernel variants for faster compilation:
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
     //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
     const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
@@ -154,7 +161,13 @@ static __global__ void flash_attn_tile_ext_f16(
             for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
                 const int j_KQ = j_KQ_0 + threadIdx.y;
 
-                half sum = __low2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]) + __high2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
+                half sum;
+                if (use_logit_softcap) {
+                    const float2 tmp = __half22float2(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
+                    sum = logit_softcap * tanhf(tmp.x + tmp.y);
+                } else {
+                    sum = __low2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]) + __high2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
+                }
                 sum += mask ? slopeh*maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
 
                 kqmax_new[j_KQ_0/nwarps] = ggml_cuda_hmax(kqmax_new[j_KQ_0/nwarps], sum);
@@ -270,20 +283,20 @@ static __global__ void flash_attn_tile_ext_f16(
 #endif // FP16_AVAILABLE
 }
 
-template <int cols_per_block, int parallel_blocks>
+template <int cols_per_block, int parallel_blocks, bool use_logit_softcap>
 void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * Q = dst->src[0];
     switch (Q->ne[0]) {
         case  64: {
             constexpr int      D = 64;
             constexpr int nwarps = 8;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks>;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
             launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
         } break;
         case 128: {
             constexpr int      D = 128;
             constexpr int nwarps = 8;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks>;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
             launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
         } break;
         default: {
@@ -296,24 +309,45 @@ void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_ten
     const ggml_tensor * KQV = dst;
     const ggml_tensor * Q   = dst->src[0];
 
-    const int32_t precision = KQV->op_params[2];
+    const int32_t precision = KQV->op_params[3];
     GGML_ASSERT(precision == GGML_PREC_DEFAULT);
 
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
     if (Q->ne[1] <= 16) {
         constexpr int cols_per_block = 16;
         constexpr int parallel_blocks = 4;
-        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        }
         return;
     }
 
     if (Q->ne[1] <= 32) {
         constexpr int cols_per_block = 32;
         constexpr int parallel_blocks = 4;
-        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        }
         return;
     }
 
     constexpr int cols_per_block = 32;
     constexpr int parallel_blocks = 1;
-    launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+    } else {
+        constexpr bool use_logit_softcap = true;
+        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+    }
 }

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

@@ -4,7 +4,7 @@
 
 #define FATTN_KQ_STRIDE_TILE_F32 32
 
-template<int D, int ncols, int nwarps, int parallel_blocks> // D == head size
+template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
 #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
@@ -20,6 +20,7 @@ static __global__ void flash_attn_tile_ext_f32(
         const float m0,
         const float m1,
         const uint32_t n_head_log2,
+        const float logit_softcap,
         const int ne00,
         const int ne01,
         const int ne02,
@@ -43,6 +44,12 @@ static __global__ void flash_attn_tile_ext_f32(
         const int ne1,
         const int ne2,
         const int ne3) {
+    // Skip unused kernel variants for faster compilation:
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
     //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
     const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
@@ -151,6 +158,10 @@ static __global__ void flash_attn_tile_ext_f32(
             for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
                 const int j_KQ = j_KQ_0 + threadIdx.y;
 
+                if (use_logit_softcap) {
+                    sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] = logit_softcap * tanhf(sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
+                }
+
                 sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += mask ? slope*__half2float(maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
 
                 kqmax_new[j_KQ_0/nwarps] = fmaxf(kqmax_new[j_KQ_0/nwarps], sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
@@ -267,20 +278,20 @@ static __global__ void flash_attn_tile_ext_f32(
     }
 }
 
-template <int cols_per_block, int parallel_blocks>
+template <int cols_per_block, int parallel_blocks, bool use_logit_softcap>
 void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * Q = dst->src[0];
     switch (Q->ne[0]) {
         case  64: {
             constexpr int      D = 64;
             constexpr int nwarps = 8;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks>;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
             launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
         } break;
         case 128: {
             constexpr int      D = 128;
             constexpr int nwarps = 8;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks>;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
             launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
         } break;
         default: {
@@ -290,23 +301,45 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
 }
 
 void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
     const ggml_tensor * Q = dst->src[0];
 
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
     if (Q->ne[1] <= 16) {
         constexpr int cols_per_block = 16;
         constexpr int parallel_blocks = 4;
-        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        }
         return;
     }
 
     if (Q->ne[1] <= 32) {
         constexpr int cols_per_block = 32;
         constexpr int parallel_blocks = 4;
-        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        }
         return;
     }
 
     constexpr int cols_per_block = 32;
     constexpr int parallel_blocks = 1;
-    launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+    } else {
+        constexpr bool use_logit_softcap = true;
+        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+    }
 }

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

@@ -1,7 +1,7 @@
 #include "common.cuh"
 #include "fattn-common.cuh"
 
-template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V> // D == head size
+template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
 #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
@@ -17,6 +17,7 @@ static __global__ void flash_attn_vec_ext_f16(
         const float m0,
         const float m1,
         const uint32_t n_head_log2,
+        const float logit_softcap,
         const int ne00,
         const int ne01,
         const int ne02,
@@ -41,6 +42,12 @@ static __global__ void flash_attn_vec_ext_f16(
         const int ne2,
         const int ne3) {
 #ifdef FP16_AVAILABLE
+    // Skip unused kernel variants for faster compilation:
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
     //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
     constexpr vec_dot_KQ_f16_t vec_dot_KQ = get_vec_dot_KQ_f16<D>(type_K);
@@ -190,6 +197,11 @@ static __global__ void flash_attn_vec_ext_f16(
             for (int j = 0; j < ncols; ++j) {
                 half sum = vec_dot_KQ(K + (k_VKQ_0 + i_KQ)*nb11, Q_h2[j], Q_i32[j], Q_ds[j]);
                 sum = warp_reduce_sum(sum);
+
+                if (use_logit_softcap) {
+                    sum = logit_softcap*tanhf(sum);
+                }
+
                 sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
 
                 if (ncols == 1) {
@@ -286,10 +298,10 @@ static __global__ void flash_attn_vec_ext_f16(
 #endif // FP16_AVAILABLE
 }
 
-template <int D, int cols_per_block, int parallel_blocks, ggml_type type_K, ggml_type type_V>
+template <int D, int cols_per_block, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
 void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     constexpr int nwarps = D/WARP_SIZE;
-    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks, type_K, type_V>;
+    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>;
     constexpr bool need_f16_K = D != 128;
     constexpr bool need_f16_V = D != 128 && D != 64;
     launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, need_f16_K, need_f16_V);
@@ -297,48 +309,81 @@ void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx,
 
 template <int D, ggml_type type_K, ggml_type type_V>
 void ggml_cuda_flash_attn_ext_vec_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * KQV = dst;
+    const ggml_tensor * KQV = dst;
-    ggml_tensor * Q   = dst->src[0];
+    const ggml_tensor * Q   = dst->src[0];
-    ggml_tensor * K   = dst->src[1];
+    const ggml_tensor * K   = dst->src[1];
-    ggml_tensor * V   = dst->src[2];
+    const ggml_tensor * V   = dst->src[2];
 
-    const int32_t precision = KQV->op_params[2];
+    const int32_t precision = KQV->op_params[3];
     GGML_ASSERT(precision == GGML_PREC_DEFAULT);
 
     GGML_ASSERT(K->type == type_K);
     GGML_ASSERT(V->type == type_V);
 
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
     if (Q->ne[1] == 1) {
         constexpr int cols_per_block  = 1;
         constexpr int parallel_blocks = 4;
-        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V>(ctx, dst);
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
         return;
     }
 
     if (Q->ne[1] == 2) {
         constexpr int cols_per_block  = 2;
         constexpr int parallel_blocks = 4;
-        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V>(ctx, dst);
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
         return;
     }
 
     if (Q->ne[1] <= 4) {
         constexpr int cols_per_block  = 4;
         constexpr int parallel_blocks = 4;
-        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V>(ctx, dst);
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
         return;
     }
 
     if (Q->ne[1] <= 8) {
         constexpr int cols_per_block  = 8;
         constexpr int parallel_blocks = 4;
-        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V>(ctx, dst);
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
         return;
     }
 
     constexpr int cols_per_block  = 8;
     constexpr int parallel_blocks = 1;
-    ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V>(ctx, dst);
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+    } else {
+        constexpr bool use_logit_softcap = true;
+        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+    }
 }
 
 #define DECL_FATTN_VEC_F16_CASE(D, type_K, type_V)                          \

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

@@ -1,7 +1,7 @@
 #include "common.cuh"
 #include "fattn-common.cuh"
 
-template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V> // D == head size
+template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
 #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
@@ -17,6 +17,7 @@ static __global__ void flash_attn_vec_ext_f32(
         const float m0,
         const float m1,
         const uint32_t n_head_log2,
+        const float logit_softcap,
         const int ne00,
         const int ne01,
         const int ne02,
@@ -40,6 +41,12 @@ static __global__ void flash_attn_vec_ext_f32(
         const int ne1,
         const int ne2,
         const int ne3) {
+    // Skip unused kernel variants for faster compilation:
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
     //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
     constexpr vec_dot_KQ_f32_t vec_dot_KQ = get_vec_dot_KQ_f32<D>(type_K);
@@ -180,6 +187,11 @@ static __global__ void flash_attn_vec_ext_f32(
             for (int j = 0; j < ncols; ++j) {
                 float sum = vec_dot_KQ(K + (k_VKQ_0 + i_KQ)*nb11, Q_f2[j], Q_i32[j], Q_ds[j]);
                 sum = warp_reduce_sum(sum);
+
+                if (use_logit_softcap) {
+                    sum = logit_softcap*tanhf(sum);
+                }
+
                 sum += mask ? slope*__half2float(maskh[j*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
 
                 kqmax_new_arr[j] = fmaxf(kqmax_new_arr[j], sum);
@@ -267,10 +279,10 @@ static __global__ void flash_attn_vec_ext_f32(
     }
 }
 
-template <int D, int cols_per_block, int parallel_blocks, ggml_type type_K, ggml_type type_V>
+template <int D, int cols_per_block, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
 void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     constexpr int nwarps = D/WARP_SIZE;
-    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks, type_K, type_V>;
+    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>;
     constexpr bool need_f16_K = D != 128;
     constexpr bool need_f16_V = D != 128 && D != 64;
     launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, need_f16_K, need_f16_V);
@@ -278,44 +290,78 @@ void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx,
 
 template <int D, ggml_type type_K, ggml_type type_V>
 void ggml_cuda_flash_attn_ext_vec_f32_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * Q   = dst->src[0];
+    const ggml_tensor * KQV = dst;
-    ggml_tensor * K   = dst->src[1];
+    const ggml_tensor * Q   = dst->src[0];
-    ggml_tensor * V   = dst->src[2];
+    const ggml_tensor * K   = dst->src[1];
+    const ggml_tensor * V   = dst->src[2];
 
     GGML_ASSERT(K->type == type_K);
     GGML_ASSERT(V->type == type_V);
 
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
     if (Q->ne[1] == 1) {
         constexpr int cols_per_block  = 1;
         constexpr int parallel_blocks = 4;
-        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V>(ctx, dst);
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
         return;
     }
 
     if (Q->ne[1] == 2) {
         constexpr int cols_per_block  = 2;
         constexpr int parallel_blocks = 4;
-        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V>(ctx, dst);
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
         return;
     }
 
     if (Q->ne[1] <= 4) {
         constexpr int cols_per_block  = 4;
         constexpr int parallel_blocks = 4;
-        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V>(ctx, dst);
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
         return;
     }
 
     if (Q->ne[1] <= 8) {
         constexpr int cols_per_block  = 8;
         constexpr int parallel_blocks = 4;
-        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V>(ctx, dst);
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
         return;
     }
 
     constexpr int cols_per_block  = 8;
     constexpr int parallel_blocks = 1;
-    ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V>(ctx, dst);
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+    } else {
+        constexpr bool use_logit_softcap = true;
+        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+    }
 }
 
 #define DECL_FATTN_VEC_F32_CASE(D, type_K, type_V)                          \

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

@@ -6,7 +6,7 @@
 #endif // FP16_MMA_AVAILABLE
 
 // D == head size, VKQ_stride == num VKQ rows calculated in parallel:
-template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t>
+template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t, bool use_logit_softcap>
 #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
 #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
@@ -22,6 +22,7 @@ static __global__ void flash_attn_ext_f16(
         const float m0,
         const float m1,
         const uint32_t n_head_log2,
+        const float logit_softcap,
         const int ne00,
         const int ne01,
         const int ne02,
@@ -46,6 +47,12 @@ static __global__ void flash_attn_ext_f16(
         const int ne2,
         const int ne3) {
 #ifdef FP16_MMA_AVAILABLE
+    // Skip unused kernel variants for faster compilation:
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
     //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
     const int ic0 = ncols*(blockIdx.x / parallel_blocks); // Index of the first Q/QKV column to work on.
@@ -85,6 +92,8 @@ static __global__ void flash_attn_ext_f16(
     const half  slopeh = __float2half(slopef);
     const half2 slope2 = make_half2(slopef, slopef);
 
+    const half2 logit_softcap_2 = make_half2(logit_softcap, logit_softcap);
+
     frag_b Q_b[D/16][ncols/frag_n];
 
     // A single buffer for temporarily holding tiles of KQ and VKQ parts:
@@ -194,6 +203,10 @@ static __global__ void flash_attn_ext_f16(
                     const int k = k0 + threadIdx.x;
 
                     KQ_f_tmp[k0/WARP_SIZE] = KQ_f[j*kqs_padded + k];
+
+                    if (use_logit_softcap) {
+                        KQ_f_tmp[k0/WARP_SIZE] = logit_softcap*tanhf(KQ_f_tmp[k0/WARP_SIZE]);
+                    }
                 }
 
                 float KQ_max_new = KQ_max_f[j0/nwarps];
@@ -237,6 +250,15 @@ static __global__ void flash_attn_ext_f16(
                     const int k = k0 + threadIdx.x;
 
                     KQ2_tmp[k0/WARP_SIZE] = KQ2[j*(kqs_padded/2) + k];
+
+                    if (use_logit_softcap) {
+                        // There is no dedicated tangens hyperbolicus function for half2.
+                        KQ2_tmp[k0/WARP_SIZE] = h2exp(KQ2_tmp[k0/WARP_SIZE]*make_half2(2.0f, 2.0f));
+                        KQ2_tmp[k0/WARP_SIZE] = (KQ2_tmp[k0/WARP_SIZE] - make_half2(1.0f, 1.0f))
+                                               /(KQ2_tmp[k0/WARP_SIZE] + make_half2(1.0f, 1.0f));
+
+                        KQ2_tmp[k0/WARP_SIZE] *= logit_softcap_2;
+                    }
                 }
 
                 half2 KQ_max_new = KQ_max_h2[j0/nwarps];
@@ -427,6 +449,7 @@ static_assert(get_VKQ_stride( 80, 4, 16) ==  16, "Test failed.");
 
 template <int D, int cols_per_block, typename KQ_acc_t>
 void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
     const ggml_tensor * Q   = dst->src[0];
 
     constexpr int nwarps = 4;
@@ -435,20 +458,50 @@ void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggm
     const int blocks_num_pb1 = ((Q->ne[1] + cols_per_block - 1) / cols_per_block)*Q->ne[2]*Q->ne[3];
     const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm;
 
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
     if (4*blocks_num_pb1 < 2*nsm) {
         constexpr int parallel_blocks = 4;
-        fattn_kernel_t fattn_kernel = flash_attn_ext_f16<D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t>;
+        fattn_kernel_t fattn_kernel;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            fattn_kernel = flash_attn_ext_f16<
+                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+        } else {
+            constexpr bool use_logit_softcap = true;
+            fattn_kernel = flash_attn_ext_f16<
+                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+        }
         launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
         return;
     }
     if (2*blocks_num_pb1 < 2*nsm) {
         constexpr int parallel_blocks = 2;
-        fattn_kernel_t fattn_kernel = flash_attn_ext_f16<D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t>;
+        fattn_kernel_t fattn_kernel;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            fattn_kernel = flash_attn_ext_f16<
+                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+        } else {
+            constexpr bool use_logit_softcap = true;
+            fattn_kernel = flash_attn_ext_f16<
+                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+        }
         launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
         return;
     }
     constexpr int parallel_blocks = 1;
-    fattn_kernel_t fattn_kernel = flash_attn_ext_f16<D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t>;
+    fattn_kernel_t fattn_kernel;
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        fattn_kernel = flash_attn_ext_f16<
+            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+    } else {
+        constexpr bool use_logit_softcap = true;
+        fattn_kernel = flash_attn_ext_f16<
+            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+    }
     launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
 }
 

ggml/src/ggml-cuda/fattn.cu

@@ -13,7 +13,7 @@ static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, g
     const ggml_tensor * KQV = dst;
     const ggml_tensor * Q   = dst->src[0];
 
-    const int32_t precision = KQV->op_params[2];
+    const int32_t precision = KQV->op_params[3];
 
     if (precision != GGML_PREC_DEFAULT) {
         if (Q->ne[1] <= 32 || Q->ne[0] > 128) {
@@ -301,7 +301,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
 
     ggml_cuda_set_device(ctx.device);
     const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-    const int32_t precision = KQV->op_params[2];
+    const int32_t precision = KQV->op_params[3];
 
     // On AMD the tile kernels perform poorly, use the vec kernel instead:
     if (cc >= CC_OFFSET_AMD) {

ggml/src/ggml-cuda/sumrows.cu

@@ -16,7 +16,7 @@ static __global__ void k_sum_rows_f32(const float * x, float * dst, const int nc
     }
 }
 
-static void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     const dim3 block_dims(WARP_SIZE, 1, 1);
     const dim3 block_nums(nrows, 1, 1);
     k_sum_rows_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
@@ -32,7 +32,6 @@ void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
     GGML_ASSERT(ggml_is_contiguous(src0));
 
-
     const int64_t ncols = src0->ne[0];
     const int64_t nrows = ggml_nrows(src0);
 

ggml/src/ggml-cuda/sumrows.cuh

@@ -1,3 +1,5 @@
 #include "common.cuh"
 
+void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream);
+
 void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/unary.cu

@@ -101,6 +101,24 @@ static __global__ void sqrt_f32(const float * x, float * dst, const int k) {
     dst[i] = sqrtf(x[i]);
 }
 
+static __global__ void sin_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = sinf(x[i]);
+}
+
+static __global__ void cos_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = cosf(x[i]);
+}
+
 static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
     gelu_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
@@ -156,6 +174,16 @@ static void sqrt_f32_cuda(const float * x, float * dst, const int k, cudaStream_
     sqrt_f32<<<num_blocks, CUDA_SQRT_BLOCK_SIZE, 0, stream>>>(x, dst, k);
 }
 
+static void sin_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SIN_BLOCK_SIZE - 1) / CUDA_SIN_BLOCK_SIZE;
+    sin_f32<<<num_blocks, CUDA_SIN_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void cos_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_COS_BLOCK_SIZE - 1) / CUDA_COS_BLOCK_SIZE;
+    cos_f32<<<num_blocks, CUDA_COS_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
 void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const float * src0_d = (const float *)src0->data;
@@ -312,3 +340,31 @@ void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     sqrt_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
 }
+
+void ggml_cuda_op_sin(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    sin_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    cos_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}

ggml/src/ggml-cuda/unary.cuh

@@ -9,6 +9,8 @@
 #define CUDA_HARDSWISH_BLOCK_SIZE 256
 #define CUDA_SQR_BLOCK_SIZE 256
 #define CUDA_SQRT_BLOCK_SIZE 256
+#define CUDA_SIN_BLOCK_SIZE 256
+#define CUDA_COS_BLOCK_SIZE 256
 
 void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
@@ -31,3 +33,7 @@ void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
 void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_sin(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-metal.m

@@ -31,6 +31,8 @@ struct ggml_metal_kernel {
 enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_ADD,
     GGML_METAL_KERNEL_TYPE_ADD_ROW,
+    GGML_METAL_KERNEL_TYPE_SUB,
+    GGML_METAL_KERNEL_TYPE_SUB_ROW,
     GGML_METAL_KERNEL_TYPE_MUL,
     GGML_METAL_KERNEL_TYPE_MUL_ROW,
     GGML_METAL_KERNEL_TYPE_DIV,
@@ -82,6 +84,8 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_RMS_NORM,
     GGML_METAL_KERNEL_TYPE_GROUP_NORM,
     GGML_METAL_KERNEL_TYPE_NORM,
+    GGML_METAL_KERNEL_TYPE_SSM_CONV_F32,
+    GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,
@@ -205,6 +209,9 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL,
     GGML_METAL_KERNEL_TYPE_CONCAT,
     GGML_METAL_KERNEL_TYPE_SQR,
+    GGML_METAL_KERNEL_TYPE_SQRT,
+    GGML_METAL_KERNEL_TYPE_SIN,
+    GGML_METAL_KERNEL_TYPE_COS,
     GGML_METAL_KERNEL_TYPE_SUM_ROWS,
 
     GGML_METAL_KERNEL_TYPE_COUNT
@@ -491,6 +498,8 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) {
 
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD,                           add,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW,                       add_row,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB,                           sub,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB_ROW,                       sub_row,                        true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL,                           mul,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW,                       mul_row,                        true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV,                           div,                            true);
@@ -542,6 +551,8 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM,                      rms_norm,                       ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM,                    group_norm,                     ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM,                          norm,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_CONV_F32,                  ssm_conv_f32,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32,                  ssm_scan_f32,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,                mul_mv_f32_f32,                 ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,                mul_mv_f16_f16,                 ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,                mul_mv_f16_f32,                 ctx->support_simdgroup_reduction);
@@ -665,6 +676,9 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL,                cpy_f32_iq4_nl,                 true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT,                        concat,                         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR,                           sqr,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQRT,                          sqrt,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIN,                           sin,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS,                           cos,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                      sum_rows,                       true);
     }
 
@@ -765,15 +779,20 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx
         case GGML_OP_PERMUTE:
         case GGML_OP_CONCAT:
         case GGML_OP_ADD:
+        case GGML_OP_SUB:
         case GGML_OP_ACC:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_REPEAT:
         case GGML_OP_SCALE:
         case GGML_OP_CLAMP:
-        case GGML_OP_SQR:
-        case GGML_OP_SUM_ROWS:
             return true;
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+        case GGML_OP_SIN:
+        case GGML_OP_COS:
+            return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_SUM_ROWS:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_RMS_NORM:
         case GGML_OP_GROUP_NORM:
@@ -803,6 +822,9 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx
                 return false;
             }
             return ctx->support_simdgroup_mm; // TODO: over-restricted for vec-kernels
+        case GGML_OP_SSM_CONV:
+        case GGML_OP_SSM_SCAN:
+            return true;
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             return ctx->support_simdgroup_reduction &&
@@ -1050,6 +1072,7 @@ static enum ggml_status ggml_metal_graph_compute(
                         [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                     } break;
                 case GGML_OP_ADD:
+                case GGML_OP_SUB:
                 case GGML_OP_MUL:
                 case GGML_OP_DIV:
                     {
@@ -1073,6 +1096,7 @@ static enum ggml_status ggml_metal_graph_compute(
                             nb = ne00 / 4;
                             switch (dst->op) {
                                 case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW].pipeline; break;
+                                case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB_ROW].pipeline; break;
                                 case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_ROW].pipeline; break;
                                 case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV_ROW].pipeline; break;
                                 default: GGML_ABORT("fatal error");
@@ -1082,6 +1106,7 @@ static enum ggml_status ggml_metal_graph_compute(
                         } else {
                             switch (dst->op) {
                                 case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline; break;
+                                case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB].pipeline; break;
                                 case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL].pipeline; break;
                                 case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV].pipeline; break;
                                 default: GGML_ABORT("fatal error");
@@ -1409,6 +1434,48 @@ static enum ggml_status ggml_metal_graph_compute(
 
                         const int64_t n = ggml_nelements(dst);
 
+                        [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                    } break;
+                case GGML_OP_SQRT:
+                    {
+                        GGML_ASSERT(ggml_is_contiguous(src0));
+
+                        id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SQRT].pipeline;
+
+                        [encoder setComputePipelineState:pipeline];
+                        [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                        [encoder setBuffer:id_dst  offset:offs_dst atIndex:1];
+
+                        const int64_t n = ggml_nelements(dst);
+
+                        [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                    } break;
+                case GGML_OP_SIN:
+                    {
+                        GGML_ASSERT(ggml_is_contiguous(src0));
+
+                        id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SIN].pipeline;
+
+                        [encoder setComputePipelineState:pipeline];
+                        [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                        [encoder setBuffer:id_dst  offset:offs_dst atIndex:1];
+
+                        const int64_t n = ggml_nelements(dst);
+
+                        [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                    } break;
+                case GGML_OP_COS:
+                    {
+                        GGML_ASSERT(ggml_is_contiguous(src0));
+
+                        id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_COS].pipeline;
+
+                        [encoder setComputePipelineState:pipeline];
+                        [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                        [encoder setBuffer:id_dst  offset:offs_dst atIndex:1];
+
+                        const int64_t n = ggml_nelements(dst);
+
                         [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                     } break;
                 case GGML_OP_SUM_ROWS:
@@ -1538,6 +1605,121 @@ static enum ggml_status ggml_metal_graph_compute(
                             [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         }
                     } break;
+                case GGML_OP_SSM_CONV:
+                    {
+                        GGML_ASSERT(src0t == GGML_TYPE_F32);
+                        GGML_ASSERT(src1t == GGML_TYPE_F32);
+
+                        GGML_ASSERT(ggml_is_contiguous(src0));
+                        GGML_ASSERT(ggml_is_contiguous(src1));
+
+                        id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_CONV_F32].pipeline;
+
+                        [encoder setComputePipelineState:pipeline];
+                        [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
+                        [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
+                        [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
+                        [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:3];
+                        [encoder setBytes:&ne01    length:sizeof(ne01) atIndex:4];
+                        [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:5];
+                        [encoder setBytes:&nb00    length:sizeof(nb00) atIndex:6];
+                        [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:7];
+                        [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:8];
+                        [encoder setBytes:&ne10    length:sizeof(ne10) atIndex:9];
+                        [encoder setBytes:&ne11    length:sizeof(ne11) atIndex:10];
+                        [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:11];
+                        [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:12];
+                        [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:13];
+                        [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:14];
+                        [encoder setBytes:&ne2     length:sizeof(ne2)  atIndex:15];
+                        [encoder setBytes:&nb0     length:sizeof(nb0)  atIndex:16];
+                        [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:17];
+                        [encoder setBytes:&nb2     length:sizeof(nb2)  atIndex:18];
+
+                        [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne1, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                    } break;
+                case GGML_OP_SSM_SCAN:
+                    {
+                        struct ggml_tensor * src3 = gf->nodes[i]->src[3];
+                        struct ggml_tensor * src4 = gf->nodes[i]->src[4];
+                        struct ggml_tensor * src5 = gf->nodes[i]->src[5];
+
+                        GGML_ASSERT(src3);
+                        GGML_ASSERT(src4);
+                        GGML_ASSERT(src5);
+
+                        size_t offs_src3 = 0;
+                        size_t offs_src4 = 0;
+                        size_t offs_src5 = 0;
+
+                        id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
+                        id<MTLBuffer> id_src4 = src4 ? ggml_metal_get_buffer(src4, &offs_src4) : nil;
+                        id<MTLBuffer> id_src5 = src5 ? ggml_metal_get_buffer(src5, &offs_src5) : nil;
+
+                        const int64_t  ne30 = src3->ne[0]; GGML_UNUSED(ne30);
+                        const int64_t  ne31 = src3->ne[1]; GGML_UNUSED(ne31);
+
+                        const uint64_t nb30 = src3->nb[0];
+                        const uint64_t nb31 = src3->nb[1];
+
+                        const int64_t  ne40 = src4->ne[0]; GGML_UNUSED(ne40);
+                        const int64_t  ne41 = src4->ne[1]; GGML_UNUSED(ne41);
+                        const int64_t  ne42 = src4->ne[2]; GGML_UNUSED(ne42);
+
+                        const uint64_t nb40 = src4->nb[0];
+                        const uint64_t nb41 = src4->nb[1];
+                        const uint64_t nb42 = src4->nb[2];
+
+                        const int64_t  ne50 = src5->ne[0]; GGML_UNUSED(ne50);
+                        const int64_t  ne51 = src5->ne[1]; GGML_UNUSED(ne51);
+                        const int64_t  ne52 = src5->ne[2]; GGML_UNUSED(ne52);
+
+                        const uint64_t nb50 = src5->nb[0];
+                        const uint64_t nb51 = src5->nb[1];
+                        const uint64_t nb52 = src5->nb[2];
+
+                        const int64_t d_state      = ne00;
+                        const int64_t d_inner      = ne01;
+                        const int64_t n_seq_tokens = ne11;
+                        const int64_t n_seqs       = ne02;
+
+                        id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline;
+
+                        [encoder setComputePipelineState:pipeline];
+                        [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                        [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                        [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2];
+                        [encoder setBuffer:id_src3 offset:offs_src3 atIndex:3];
+                        [encoder setBuffer:id_src4 offset:offs_src4 atIndex:4];
+                        [encoder setBuffer:id_src5 offset:offs_src5 atIndex:5];
+                        [encoder setBuffer:id_dst  offset:offs_dst  atIndex:6];
+
+                        [encoder setBytes:&d_state      length:sizeof(d_state)      atIndex:7];
+                        [encoder setBytes:&d_inner      length:sizeof(d_inner)      atIndex:8];
+                        [encoder setBytes:&n_seq_tokens length:sizeof(n_seq_tokens) atIndex:9];
+                        [encoder setBytes:&n_seqs       length:sizeof(n_seqs)       atIndex:10];
+
+                        [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:11];
+                        [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:12];
+                        [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:13];
+                        [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
+                        [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
+                        [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
+                        [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17];
+                        [encoder setBytes:&nb20 length:sizeof(nb20) atIndex:18];
+                        [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:19];
+                        [encoder setBytes:&nb22 length:sizeof(nb22) atIndex:20];
+                        [encoder setBytes:&nb30 length:sizeof(nb30) atIndex:21];
+                        [encoder setBytes:&nb31 length:sizeof(nb31) atIndex:22];
+                        [encoder setBytes:&nb40 length:sizeof(nb40) atIndex:23];
+                        [encoder setBytes:&nb41 length:sizeof(nb41) atIndex:24];
+                        [encoder setBytes:&nb42 length:sizeof(nb42) atIndex:25];
+                        [encoder setBytes:&nb50 length:sizeof(nb50) atIndex:26];
+                        [encoder setBytes:&nb51 length:sizeof(nb51) atIndex:27];
+                        [encoder setBytes:&nb52 length:sizeof(nb52) atIndex:28];
+
+                        [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                    } break;
                 case GGML_OP_MUL_MAT:
                     {
                         GGML_ASSERT(ne00 == ne10);
@@ -2624,9 +2806,14 @@ static enum ggml_status ggml_metal_graph_compute(
 
                         float scale;
                         float max_bias;
-
+                        float logit_softcap;
                         memcpy(&scale,         ((int32_t *) dst->op_params) + 0, sizeof(scale));
                         memcpy(&max_bias,      ((int32_t *) dst->op_params) + 1, sizeof(max_bias));
+                        memcpy(&logit_softcap, ((int32_t *) dst->op_params) + 2, sizeof(logit_softcap));
+
+                        if (logit_softcap != 0.0f) {
+                            scale /= logit_softcap;
+                        }
 
                         const uint32_t n_head      = src0->ne[2];
                         const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
@@ -2701,6 +2888,7 @@ static enum ggml_status ggml_metal_graph_compute(
                         [encoder setBytes:&m0            length:sizeof(m0)            atIndex:25];
                         [encoder setBytes:&m1            length:sizeof(m1)            atIndex:26];
                         [encoder setBytes:&n_head_log2   length:sizeof(n_head_log2)   atIndex:27];
+                        [encoder setBytes:&logit_softcap length:sizeof(logit_softcap) atIndex:28];
 
                         if (!use_vec_kernel) {
                             // half8x8 kernel

ggml/src/ggml-metal.metal

@@ -17,7 +17,7 @@ enum ggml_sort_order {
     GGML_SORT_ORDER_DESC,
 };
 
-// general-purpose kernel for addition, multiplication and division of two tensors
+// general-purpose kernel for addition, subtraction, multiplication and division of two tensors
 // pros: works for non-contiguous tensors, supports broadcast across all dims
 // cons: not very efficient
 kernel void kernel_add(
@@ -70,6 +70,56 @@ kernel void kernel_add(
     }
 }
 
+kernel void kernel_sub(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant  int64_t & ne00,
+        constant  int64_t & ne01,
+        constant  int64_t & ne02,
+        constant  int64_t & ne03,
+        constant uint64_t & nb00,
+        constant uint64_t & nb01,
+        constant uint64_t & nb02,
+        constant uint64_t & nb03,
+        constant  int64_t & ne10,
+        constant  int64_t & ne11,
+        constant  int64_t & ne12,
+        constant  int64_t & ne13,
+        constant uint64_t & nb10,
+        constant uint64_t & nb11,
+        constant uint64_t & nb12,
+        constant uint64_t & nb13,
+        constant  int64_t & ne0,
+        constant  int64_t & ne1,
+        constant  int64_t & ne2,
+        constant  int64_t & ne3,
+        constant uint64_t & nb0,
+        constant uint64_t & nb1,
+        constant uint64_t & nb2,
+        constant uint64_t & nb3,
+        constant  int64_t & offs,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig.z;
+    const int64_t i02 = tgpig.y;
+    const int64_t i01 = tgpig.x;
+
+    const int64_t i13 = i03 % ne13;
+    const int64_t i12 = i02 % ne12;
+    const int64_t i11 = i01 % ne11;
+
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + offs;
+    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + offs;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        const int i10 = i0 % ne10;
+        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) - *((device float *)(src1_ptr + i10*nb10));
+    }
+}
+
 kernel void kernel_mul(
         device const char * src0,
         device const char * src1,
@@ -226,6 +276,15 @@ kernel void kernel_add_row(
     dst[tpig] = src0[tpig] + src1[tpig % nb];
 }
 
+kernel void kernel_sub_row(
+        device const float4 * src0,
+        device const float4 * src1,
+        device       float4 * dst,
+        constant   uint64_t & nb [[buffer(28)]],
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] - src1[tpig % nb];
+}
+
 kernel void kernel_mul_row(
         device const float4 * src0,
         device const float4 * src1,
@@ -358,6 +417,27 @@ kernel void kernel_sqr(
     dst[tpig] = src0[tpig] * src0[tpig];
 }
 
+kernel void kernel_sqrt(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = sqrt(src0[tpig]);
+}
+
+kernel void kernel_sin(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = sin(src0[tpig]);
+}
+
+kernel void kernel_cos(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = cos(src0[tpig]);
+}
+
 kernel void kernel_sum_rows(
         device const float * src0,
         device       float * dst,
@@ -667,6 +747,127 @@ kernel void kernel_diag_mask_inf_8(
     }
 }
 
+// ref: ggml.c:ggml_compute_forward_ssm_conv_f32
+// TODO: optimize
+kernel void kernel_ssm_conv_f32(
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i2 = tgpig.y;
+    const int64_t i3 = tgpig.z;
+
+    const int64_t nc  = ne10;
+    const int64_t ncs = ne00;
+    const int64_t nr  = ne01;
+    const int64_t n_t = ne1;
+    const int64_t n_s = ne2;
+
+    device const float * s = (device const float *) ((device const char *) src0 + ir*nb01 + i2*nb00 + i3*nb02);
+    device const float * c = (device const float *) ((device const char *) src1 + ir*nb11);
+    device       float * x = (device       float *) ((device       char *) dst  + ir*nb0  + i2*nb1  + i3*nb2);
+
+    float sumf = 0.0f;
+
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+// ref: ggml.c:ggml_compute_forward_ssm_scan_f32
+// TODO: optimize
+kernel void kernel_ssm_scan_f32(
+        device const void * src0,
+        device const void * src1,
+        device const void * src2,
+        device const void * src3,
+        device const void * src4,
+        device const void * src5,
+        device      float * dst,
+        constant  int64_t & d_state,
+        constant  int64_t & d_inner,
+        constant  int64_t & n_seq_tokens,
+        constant  int64_t & n_seqs,
+        constant uint64_t & nb00,
+        constant uint64_t & nb01,
+        constant uint64_t & nb02,
+        constant uint64_t & nb10,
+        constant uint64_t & nb11,
+        constant uint64_t & nb12,
+        constant uint64_t & nb13,
+        constant uint64_t & nb20,
+        constant uint64_t & nb21,
+        constant uint64_t & nb22,
+        constant uint64_t & nb30,
+        constant uint64_t & nb31,
+        constant uint64_t & nb40,
+        constant uint64_t & nb41,
+        constant uint64_t & nb42,
+        constant uint64_t & nb50,
+        constant uint64_t & nb51,
+        constant uint64_t & nb52,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i3 = tgpig.y;
+
+    const int64_t nc  = d_state;
+    const int64_t nr  = d_inner;
+    const int64_t n_t = n_seq_tokens;
+    const int64_t n_s = n_seqs;
+
+    for (int64_t i2 = 0; i2 < n_t; ++i2) {
+        device const float * s0 = (device const float *) ((device const char *) src0 + ir*nb01 + i3*nb02);
+        device const float * x  = (device const float *) ((device const char *) src1 + ir*nb10 + i2*nb11 + i3*nb12);
+        device const float * dt = (device const float *) ((device const char *) src2 + ir*nb20 + i2*nb21 + i3*nb22);
+        device const float * A  = (device const float *) ((device const char *) src3 + ir*nb31);
+        device const float * B  = (device const float *) ((device const char *) src4 + i2*nb41 + i3*nb42);
+        device const float * C  = (device const float *) ((device const char *) src5 + i2*nb51 + i3*nb52);
+        device       float * y  = (device       float *) ((device       char *) dst  + ir*nb10 + i2*nb11 + i3*nb12); // TODO: do not use src1 strides
+        device       float * s  = (device       float *) ((device       char *) dst  + ir*nb01 + i3*nb02 +    nb13);
+
+        if (i2 > 0) {
+            s0 = s;
+        }
+
+        // i1 == 0
+        float dt_soft_plus = dt[0] <= 20.0f ? log(1.0f + exp(dt[0])) : dt[0];
+        float x_dt = x[0] * dt_soft_plus;
+        float sumf = 0.0f;
+
+        for (int64_t i0 = 0; i0 < nc; ++i0) {
+            int64_t i = i0;
+            float state = (s0[i] * exp(dt_soft_plus * A[i])) + (B[i0] * x_dt);
+            sumf += state * C[i0];
+            s[i] = state;
+        }
+
+        y[0] = sumf;
+    }
+}
+
 kernel void kernel_norm(
         device const  void * src0,
         device       float * dst,
@@ -1976,6 +2177,7 @@ typedef void (flash_attn_ext_f16_t)(
         constant     float & m0,
         constant     float & m1,
         constant  uint32_t & n_head_log2,
+        constant     float & logit_softcap,
         threadgroup   half * shared,
         uint3  tgpig[[threadgroup_position_in_grid]],
         uint3  tpitg[[thread_position_in_threadgroup]],
@@ -2014,6 +2216,7 @@ kernel void kernel_flash_attn_ext_f16(
         constant     float & m0,
         constant     float & m1,
         constant  uint32_t & n_head_log2,
+        constant     float & logit_softcap,
         threadgroup   half * shared [[threadgroup(0)]],
         uint3  tgpig[[threadgroup_position_in_grid]],
         uint3  tpitg[[thread_position_in_threadgroup]],
@@ -2138,19 +2341,6 @@ kernel void kernel_flash_attn_ext_f16(
                     }
 
                     simdgroup_store(mqk, ss + 8*cc, TF, 0, false);
-
-                    const short tx = tiisg%4;
-                    const short ty = tiisg/4;
-
-                    if (mask != q) {
-                        // mqk = mqk*scale + mask*slope
-                        ss[8*cc + ty*TF + 2*tx + 0] = scale*ss[8*cc + ty*TF + 2*tx + 0] + slope*mp[ic + 8*cc + ty*nb31/sizeof(half) + 2*tx + 0];
-                        ss[8*cc + ty*TF + 2*tx + 1] = scale*ss[8*cc + ty*TF + 2*tx + 1] + slope*mp[ic + 8*cc + ty*nb31/sizeof(half) + 2*tx + 1];
-                    } else {
-                        // mqk = mqk*scale
-                        ss[8*cc + ty*TF + 2*tx + 0] *= scale;
-                        ss[8*cc + ty*TF + 2*tx + 1] *= scale;
-                    }
                 }
             }
 
@@ -2162,10 +2352,19 @@ kernel void kernel_flash_attn_ext_f16(
                 float ms[Q];
 
                 for (short j = 0; j < Q; ++j) {
-                    const short p = tiisg;
-
                     const float m = M[j];
-                    const float s = ss[j*TF + p];
+
+                    // scale and apply the logitcap / mask
+                    float s = ss[j*TF + tiisg]*scale;
+
+                    if (logit_softcap != 0.0f) {
+                        s = logit_softcap*precise::tanh(s);
+                    }
+
+                    if (mask != q) {
+                        // mqk = mqk + mask*slope
+                        s += slope*mp[ic + j*nb31/sizeof(half) + tiisg];
+                    }
 
                     smax = simd_max(max(smax, s));
                     M[j] = simd_max(max(M[j], s));
@@ -2176,7 +2375,7 @@ kernel void kernel_flash_attn_ext_f16(
                     S[j] = S[j]*ms[j] + simd_sum(vs);
 
                     // the P matrix from the paper (Q rows, C columns)
-                    ss[j*TF + p] = vs;
+                    ss[j*TF + tiisg] = vs;
                 }
 
                 // create a QxQ diagonal matrix for rescaling the output
@@ -2345,6 +2544,7 @@ kernel void kernel_flash_attn_ext_vec_f16(
         constant     float & m0,
         constant     float & m1,
         constant  uint32_t & n_head_log2,
+        constant     float & logit_softcap,
         threadgroup   half * shared [[threadgroup(0)]],
         uint3  tgpig[[threadgroup_position_in_grid]],
         uint3  tpitg[[thread_position_in_threadgroup]],
@@ -2479,7 +2679,13 @@ kernel void kernel_flash_attn_ext_vec_f16(
 
                     // mqk = mqk*scale + mask*slope
                     if (tiisg == 0) {
-                        mqk = mqk*scale + ((mask != q) ? ((float4) mp4[ic/4 + cc])*slope : (float4) 0.0f);
+                        mqk *= scale;
+
+                        if (logit_softcap != 0.0f) {
+                            mqk = logit_softcap*precise::tanh(mqk);
+                        }
+
+                        mqk += (mask != q) ? ((float4) mp4[ic/4 + cc])*slope : (float4) 0.0f;
 
                         ss4[cc] = mqk;
                     }

ggml/src/ggml-quants.c

@@ -3829,7 +3829,7 @@ void quantize_row_q8_K(const float * restrict x, void * restrict y, int64_t k) {
     quantize_row_q8_K_ref(x, y, k);
 }
 
-//===================================== Dot ptoducts =================================
+//===================================== Dot products =================================
 
 //
 // Helper functions

ggml/src/ggml-sycl/dmmv.cpp

@@ -76,8 +76,8 @@ static void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat *
     }
 
     // sum up partial sums and write back result
-#pragma unroll
+    const int mask_start = ncols > GGML_SYCL_DMMV_X ? WARP_SIZE >> 1 : WARP_SIZE >> 2;
-    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+    for (int mask = mask_start; mask > 0; mask >>= 1) {
         tmp +=
             dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
     }

ggml/src/ggml-vulkan.cpp

@@ -188,6 +188,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_upscale_f32;
     vk_pipeline pipeline_scale_f32;
     vk_pipeline pipeline_sqr_f32;
+    vk_pipeline pipeline_sin_f32;
+    vk_pipeline pipeline_cos_f32;
     vk_pipeline pipeline_clamp_f32;
     vk_pipeline pipeline_pad_f32;
     vk_pipeline pipeline_repeat_f32;
@@ -1702,6 +1704,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_sin_f32, "sin_f32", sin_f32_len, sin_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cos_f32, "cos_f32", cos_f32_len, cos_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_clamp_f32, "clamp_f32", clamp_f32_len, clamp_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
@@ -4023,6 +4027,16 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_sqr_f32;
         }
         return nullptr;
+    case GGML_OP_SIN:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_sin_f32;
+        }
+        return nullptr;
+    case GGML_OP_COS:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_cos_f32;
+        }
+        return nullptr;
     case GGML_OP_CLAMP:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_clamp_f32;
@@ -4171,6 +4185,8 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_UPSCALE:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
+    case GGML_OP_SIN:
+    case GGML_OP_COS:
     case GGML_OP_CLAMP:
     case GGML_OP_PAD:
     case GGML_OP_REPEAT:
@@ -4381,6 +4397,8 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_MUL:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
+    case GGML_OP_SIN:
+    case GGML_OP_COS:
     case GGML_OP_CLAMP:
     case GGML_OP_PAD:
     case GGML_OP_REPEAT:
@@ -4598,6 +4616,32 @@ static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const
     }, dryrun);
 }
 
+static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SIN, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f,
+    });
+}
+
+static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_COS, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f,
+    });
+}
+
 static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     float * op_params = (float *)dst->op_params;
     const uint32_t src0_type_size = ggml_type_size(src0->type);
@@ -5658,6 +5702,8 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_UPSCALE:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
+    case GGML_OP_SIN:
+    case GGML_OP_COS:
     case GGML_OP_CLAMP:
     case GGML_OP_PAD:
     case GGML_OP_CPY:
@@ -5735,6 +5781,14 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_SQR:
         ggml_vk_sqr(ctx, compute_ctx, src0, node, dryrun);
 
+        break;
+    case GGML_OP_SIN:
+        ggml_vk_sin(ctx, compute_ctx, src0, node);
+
+        break;
+    case GGML_OP_COS:
+        ggml_vk_cos(ctx, compute_ctx, src0, node);
+
         break;
     case GGML_OP_CLAMP:
         ggml_vk_clamp(ctx, compute_ctx, src0, node, dryrun);
@@ -5851,6 +5905,8 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_UPSCALE:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
+    case GGML_OP_SIN:
+    case GGML_OP_COS:
     case GGML_OP_CLAMP:
     case GGML_OP_PAD:
     case GGML_OP_CPY:
@@ -6582,6 +6638,8 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const
         case GGML_OP_UPSCALE:
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
+        case GGML_OP_SIN:
+        case GGML_OP_COS:
         case GGML_OP_CLAMP:
         case GGML_OP_PAD:
         case GGML_OP_CONT:
@@ -7024,6 +7082,10 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         tensor_clone = ggml_scale(ggml_ctx, src0_clone, ((float *)tensor->op_params)[0]);
     } else if (tensor->op == GGML_OP_SQR) {
         tensor_clone = ggml_sqr(ggml_ctx, src0_clone);
+    } else if (tensor->op == GGML_OP_SIN) {
+        tensor_clone = ggml_sin(ggml_ctx, src0_clone);
+    } else if (tensor->op == GGML_OP_COS) {
+        tensor_clone = ggml_cos(ggml_ctx, src0_clone);
     } else if (tensor->op == GGML_OP_CLAMP) {
         tensor_clone = ggml_clamp(ggml_ctx, src0_clone, ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]);
     } else if (tensor->op == GGML_OP_PAD) {

ggml/src/llamafile/sgemm.cpp

@@ -606,17 +606,29 @@ class tinyBLAS_Q0_AVX {
         case 0x44:
             mc = 4;
             nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<4>(m0, m, n0, n);
+#else
             gemm<4, 4>(m0, m, n0, n);
+#endif
             break;
         case 0x43:
             mc = 4;
             nc = 3;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<3>(m0, m, n0, n);
+#else
             gemm<4, 3>(m0, m, n0, n);
+#endif
             break;
         case 0x34:
             mc = 3;
             nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<3>(m0, m, n0, n);
+#else
             gemm<3, 4>(m0, m, n0, n);
+#endif
             break;
         case 0x33:
             mc = 3;
@@ -626,12 +638,20 @@ class tinyBLAS_Q0_AVX {
         case 0x42:
             mc = 4;
             nc = 2;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<2>(m0, m, n0, n);
+#else
             gemm<4, 2>(m0, m, n0, n);
+#endif
             break;
         case 0x24:
             mc = 2;
             nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<2>(m0, m, n0, n);
+#else
             gemm<2, 4>(m0, m, n0, n);
+#endif
             break;
 #else
         case 0x44:
@@ -639,13 +659,21 @@ class tinyBLAS_Q0_AVX {
         case 0x42:
             mc = 4;
             nc = 2;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<2>(m0, m, n0, n);
+#else
             gemm<4, 2>(m0, m, n0, n);
+#endif
             break;
         case 0x34:
         case 0x24:
             mc = 2;
             nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<2>(m0, m, n0, n);
+#else
             gemm<2, 4>(m0, m, n0, n);
+#endif
             break;
         case 0x33:
 #endif
@@ -662,7 +690,11 @@ class tinyBLAS_Q0_AVX {
         case 0x41:
             mc = 4;
             nc = 1;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<1>(m0, m, n0, n);
+#else
             gemm<4, 1>(m0, m, n0, n);
+#endif
             break;
         case 0x22:
             mc = 2;
@@ -672,7 +704,11 @@ class tinyBLAS_Q0_AVX {
         case 0x14:
             mc = 1;
             nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<1>(m0, m, n0, n);
+#else
             gemm<1, 4>(m0, m, n0, n);
+#endif
             break;
         case 0x31:
             mc = 3;
@@ -708,6 +744,119 @@ class tinyBLAS_Q0_AVX {
         mnpack(m0, m, np, n);
     }
 
+#if defined(__AVX2__) && defined(__F16C__)
+// Templated functions for gemm of dimensions 4xN
+    template <int RN>
+    NOINLINE void gemm4xN(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / 4;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * 4;
+            int64_t jj = n0 + job % xtiles * RN;
+            __m256 Cv[RN][4] = {};
+            for (int64_t l = 0; l < k; ++l) {
+                uint64_t a_delta = ((uint64_t)A[lda * (ii + 3) + l].d << 48) | ((uint64_t)A[lda * (ii + 2) + l].d << 32) | ((uint64_t)A[lda * (ii + 1) + l].d << 16) | (A[lda * (ii + 0) + l].d);
+                // Convert delta values for four blocks to float values
+                __m128 da = _mm_cvtph_ps(_mm_set_epi64x(0, a_delta));
+                __m256i avec0 = load(A + lda * (ii + 0) + l);
+                __m256i avec1 = load(A + lda * (ii + 1) + l);
+                __m256i avec2 = load(A + lda * (ii + 2) + l);
+                __m256i avec3 = load(A + lda * (ii + 3) + l);
+                for (int64_t j = 0; j < RN; ++j) {
+                        __m128 db = _mm_set1_ps(unhalf(B[ldb * (jj + j) + l].d));
+                        // Computation of product of delta values for four blocks and replicate it across 256 bit lane
+                        __m256 dvec =  _mm256_castps128_ps256(_mm_mul_ps(da, db));
+                        dvec = _mm256_permute2f128_ps(dvec ,dvec, 0);
+                        // Computation of dot product and multiplication with appropriate delta value products
+                        Cv[j][0] = madd(_mm256_shuffle_ps(dvec, dvec, 0),
+                                    updot(_mm256_sign_epi8(avec0, avec0),
+                                          _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec0)),
+                                    Cv[j][0]);
+                        Cv[j][1] = madd(_mm256_shuffle_ps(dvec, dvec, 85),
+                                    updot(_mm256_sign_epi8(avec1, avec1),
+                                            _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec1)),
+                                    Cv[j][1]);
+                        Cv[j][2] = madd(_mm256_shuffle_ps(dvec, dvec, 170),
+                                    updot(_mm256_sign_epi8(avec2, avec2),
+                                            _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec2)),
+                                    Cv[j][2]);
+                        Cv[j][3] = madd(_mm256_shuffle_ps(dvec, dvec, 255),
+                                    updot(_mm256_sign_epi8(avec3, avec3),
+                                            _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec3)),
+                                    Cv[j][3]);
+                }
+            }
+
+            for (int64_t j = 0; j < RN; ++j)
+                for (int64_t i = 0; i < 4; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+
+    // Templated functions for gemm of dimensions Mx4
+    template <int RM>
+    NOINLINE void gemmMx4(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / 4;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * 4;
+            __m256 Cv[4][RM] = {};
+            for (int64_t l = 0; l < k; ++l) {
+                uint64_t b_delta = ((uint64_t)B[ldb * (jj + 3) + l].d << 48) | ((uint64_t)B[ldb * (jj + 2) + l].d << 32) | ((uint64_t)B[ldb * (jj + 1) + l].d << 16) | (B[ldb * (jj + 0) + l].d);
+                // Convert delta values for four blocks to float values
+                __m128 db = _mm_cvtph_ps(_mm_set_epi64x(0, b_delta));
+                __m256i bvec0 = load(B + ldb * (jj + 0) + l);
+                __m256i bvec1 = load(B + ldb * (jj + 1) + l);
+                __m256i bvec2 = load(B + ldb * (jj + 2) + l);
+                __m256i bvec3 = load(B + ldb * (jj + 3) + l);
+                for (int64_t i = 0; i < RM; ++i) {
+                    __m128 da = _mm_set1_ps(unhalf((A[lda * (ii + i) + l].d)));
+                    // Computation of product of delta values for four blocks and replicate it across 256 bit lane
+                    __m256 dvec =  _mm256_castps128_ps256(_mm_mul_ps(da, db));
+                    dvec = _mm256_permute2f128_ps(dvec ,dvec, 0);
+                    // Computation of dot product and multiplication with appropriate delta value products
+                    Cv[0][i] = madd(_mm256_shuffle_ps(dvec, dvec, 0),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec0, load(A + lda * (ii + i) + l))),
+                                    Cv[0][i]);
+                    Cv[1][i] = madd(_mm256_shuffle_ps(dvec, dvec, 85),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec1, load(A + lda * (ii + i) + l))),
+                                    Cv[1][i]);
+                    Cv[2][i] = madd(_mm256_shuffle_ps(dvec, dvec, 170),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec2, load(A + lda * (ii + i) + l))),
+                                    Cv[2][i]);
+                    Cv[3][i] = madd(_mm256_shuffle_ps(dvec, dvec, 255),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec3, load(A + lda * (ii + i) + l))),
+                                    Cv[3][i]);
+                }
+            }
+            for (int64_t j = 0; j < 4; ++j)
+                for (int64_t i = 0; i < RM; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+#endif
+
     template <int RM, int RN>
     NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
         int64_t ytiles = (m - m0) / RM;

ggml/src/vulkan-shaders/cos.comp

@@ -0,0 +1,15 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(cos(val));
+}

ggml/src/vulkan-shaders/sin.comp

@@ -0,0 +1,15 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(sin(val));
+}

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

@@ -396,6 +396,14 @@ void process_shaders(std::vector<std::future<void>>& tasks) {
         string_to_spv("sqr_f32", "square.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }));
 
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("sin_f32", "sin.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("cos_f32", "cos.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    }));
+
     tasks.push_back(std::async(std::launch::async, [] {
         string_to_spv("clamp_f32", "clamp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }));

gguf-py/gguf/constants.py

@@ -94,6 +94,9 @@ class Keys:
         DECODER_START_TOKEN_ID            = "{arch}.decoder_start_token_id"
         ATTN_LOGIT_SOFTCAPPING            = "{arch}.attn_logit_softcapping"
         FINAL_LOGIT_SOFTCAPPING           = "{arch}.final_logit_softcapping"
+        RESCALE_EVERY_N_LAYERS            = "{arch}.rescale_every_n_layers"
+        TIME_MIX_EXTRA_DIM                = "{arch}.time_mix_extra_dim"
+        TIME_DECAY_EXTRA_DIM              = "{arch}.time_decay_extra_dim"
 
     class Attention:
         HEAD_COUNT        = "{arch}.attention.head_count"
@@ -132,6 +135,9 @@ class Keys:
         TIME_STEP_RANK = "{arch}.ssm.time_step_rank"
         DT_B_C_RMS     = "{arch}.ssm.dt_b_c_rms"
 
+    class WKV:
+        HEAD_SIZE = "{arch}.wkv.head_size"
+
     class Tokenizer:
         MODEL                = "tokenizer.ggml.model"
         PRE                  = "tokenizer.ggml.pre"
@@ -207,6 +213,7 @@ class MODEL_ARCH(IntEnum):
     GEMMA        = auto()
     GEMMA2       = auto()
     STARCODER2   = auto()
+    RWKV6        = auto()
     MAMBA        = auto()
     XVERSE       = auto()
     COMMAND_R    = auto()
@@ -270,6 +277,29 @@ class MODEL_TENSOR(IntEnum):
     SSM_A                = auto()
     SSM_D                = auto()
     SSM_OUT              = auto()
+    TIME_MIX_W1          = auto()
+    TIME_MIX_W2          = auto()
+    TIME_MIX_LERP_X      = auto()
+    TIME_MIX_LERP_K      = auto()
+    TIME_MIX_LERP_V      = auto()
+    TIME_MIX_LERP_R      = auto()
+    TIME_MIX_LERP_G      = auto()
+    TIME_MIX_LERP_W      = auto()
+    TIME_MIX_FIRST       = auto()
+    TIME_MIX_DECAY       = auto()
+    TIME_MIX_DECAY_W1    = auto()
+    TIME_MIX_DECAY_W2    = auto()
+    TIME_MIX_KEY         = auto()
+    TIME_MIX_VALUE       = auto()
+    TIME_MIX_RECEPTANCE  = auto()
+    TIME_MIX_GATE        = auto()
+    TIME_MIX_LN          = auto()
+    TIME_MIX_OUTPUT      = auto()
+    CHANNEL_MIX_LERP_K   = auto()
+    CHANNEL_MIX_LERP_R   = auto()
+    CHANNEL_MIX_KEY      = auto()
+    CHANNEL_MIX_RECEPTANCE = auto()
+    CHANNEL_MIX_VALUE    = auto()
     ATTN_Q_A             = auto()
     ATTN_Q_B             = auto()
     ATTN_KV_A_MQA        = auto()
@@ -337,6 +367,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.GEMMA:          "gemma",
     MODEL_ARCH.GEMMA2:         "gemma2",
     MODEL_ARCH.STARCODER2:     "starcoder2",
+    MODEL_ARCH.RWKV6:          "rwkv6",
     MODEL_ARCH.MAMBA:          "mamba",
     MODEL_ARCH.XVERSE:         "xverse",
     MODEL_ARCH.COMMAND_R:      "command-r",
@@ -400,6 +431,29 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.SSM_A:                     "blk.{bid}.ssm_a",
     MODEL_TENSOR.SSM_D:                     "blk.{bid}.ssm_d",
     MODEL_TENSOR.SSM_OUT:                   "blk.{bid}.ssm_out",
+    MODEL_TENSOR.TIME_MIX_W1:               "blk.{bid}.time_mix_w1",
+    MODEL_TENSOR.TIME_MIX_W2:               "blk.{bid}.time_mix_w2",
+    MODEL_TENSOR.TIME_MIX_LERP_X:           "blk.{bid}.time_mix_lerp_x",
+    MODEL_TENSOR.TIME_MIX_LERP_K:           "blk.{bid}.time_mix_lerp_k",
+    MODEL_TENSOR.TIME_MIX_LERP_V:           "blk.{bid}.time_mix_lerp_v",
+    MODEL_TENSOR.TIME_MIX_LERP_R:           "blk.{bid}.time_mix_lerp_r",
+    MODEL_TENSOR.TIME_MIX_LERP_G:           "blk.{bid}.time_mix_lerp_g",
+    MODEL_TENSOR.TIME_MIX_LERP_W:           "blk.{bid}.time_mix_lerp_w",
+    MODEL_TENSOR.TIME_MIX_FIRST:            "blk.{bid}.time_mix_first",
+    MODEL_TENSOR.TIME_MIX_DECAY:            "blk.{bid}.time_mix_decay",
+    MODEL_TENSOR.TIME_MIX_DECAY_W1:         "blk.{bid}.time_mix_decay_w1",
+    MODEL_TENSOR.TIME_MIX_DECAY_W2:         "blk.{bid}.time_mix_decay_w2",
+    MODEL_TENSOR.TIME_MIX_KEY:              "blk.{bid}.time_mix_key",
+    MODEL_TENSOR.TIME_MIX_VALUE:            "blk.{bid}.time_mix_value",
+    MODEL_TENSOR.TIME_MIX_RECEPTANCE:       "blk.{bid}.time_mix_receptance",
+    MODEL_TENSOR.TIME_MIX_GATE:             "blk.{bid}.time_mix_gate",
+    MODEL_TENSOR.TIME_MIX_LN:               "blk.{bid}.time_mix_ln",
+    MODEL_TENSOR.TIME_MIX_OUTPUT:           "blk.{bid}.time_mix_output",
+    MODEL_TENSOR.CHANNEL_MIX_LERP_K:        "blk.{bid}.channel_mix_lerp_k",
+    MODEL_TENSOR.CHANNEL_MIX_LERP_R:        "blk.{bid}.channel_mix_lerp_r",
+    MODEL_TENSOR.CHANNEL_MIX_KEY:           "blk.{bid}.channel_mix_key",
+    MODEL_TENSOR.CHANNEL_MIX_RECEPTANCE:    "blk.{bid}.channel_mix_receptance",
+    MODEL_TENSOR.CHANNEL_MIX_VALUE:         "blk.{bid}.channel_mix_value",
     MODEL_TENSOR.ATTN_Q_A:                  "blk.{bid}.attn_q_a",
     MODEL_TENSOR.ATTN_Q_B:                  "blk.{bid}.attn_q_b",
     MODEL_TENSOR.ATTN_KV_A_MQA:             "blk.{bid}.attn_kv_a_mqa",
@@ -856,6 +910,37 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.RWKV6: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_NORM_2,
+        MODEL_TENSOR.TIME_MIX_W1,
+        MODEL_TENSOR.TIME_MIX_W2,
+        MODEL_TENSOR.TIME_MIX_LERP_X,
+        MODEL_TENSOR.TIME_MIX_LERP_K,
+        MODEL_TENSOR.TIME_MIX_LERP_V,
+        MODEL_TENSOR.TIME_MIX_LERP_R,
+        MODEL_TENSOR.TIME_MIX_LERP_G,
+        MODEL_TENSOR.TIME_MIX_LERP_W,
+        MODEL_TENSOR.TIME_MIX_FIRST,
+        MODEL_TENSOR.TIME_MIX_DECAY,
+        MODEL_TENSOR.TIME_MIX_DECAY_W1,
+        MODEL_TENSOR.TIME_MIX_DECAY_W2,
+        MODEL_TENSOR.TIME_MIX_KEY,
+        MODEL_TENSOR.TIME_MIX_VALUE,
+        MODEL_TENSOR.TIME_MIX_RECEPTANCE,
+        MODEL_TENSOR.TIME_MIX_GATE,
+        MODEL_TENSOR.TIME_MIX_LN,
+        MODEL_TENSOR.TIME_MIX_OUTPUT,
+        MODEL_TENSOR.CHANNEL_MIX_LERP_K,
+        MODEL_TENSOR.CHANNEL_MIX_LERP_R,
+        MODEL_TENSOR.CHANNEL_MIX_KEY,
+        MODEL_TENSOR.CHANNEL_MIX_RECEPTANCE,
+        MODEL_TENSOR.CHANNEL_MIX_VALUE,
+    ],
     MODEL_ARCH.MAMBA: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,

gguf-py/gguf/gguf_writer.py

@@ -670,6 +670,18 @@ class GGUFWriter:
     def add_expert_weights_scale(self, value: float) -> None:
         self.add_float32(Keys.LLM.EXPERT_WEIGHTS_SCALE.format(arch=self.arch), value)
 
+    def add_rescale_every_n_layers(self, count: int) -> None:
+        self.add_uint32(Keys.LLM.RESCALE_EVERY_N_LAYERS.format(arch=self.arch), count)
+
+    def add_time_mix_extra_dim(self, dim: int) -> None:
+        self.add_uint32(Keys.LLM.TIME_MIX_EXTRA_DIM.format(arch=self.arch), dim)
+
+    def add_time_decay_extra_dim(self, dim: int) -> None:
+        self.add_uint32(Keys.LLM.TIME_DECAY_EXTRA_DIM.format(arch=self.arch), dim)
+
+    def add_wkv_head_size(self, size: int) -> None:
+        self.add_uint32(Keys.WKV.HEAD_SIZE.format(arch=self.arch), size)
+
     def add_layer_norm_eps(self, value: float) -> None:
         self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value)
 

gguf-py/gguf/tensor_mapping.py

@@ -27,6 +27,7 @@ class TensorNameMap:
             "embedding.word_embeddings",                 # chatglm
             "transformer.token_embeddings",              # openelm
             "shared",                                    # t5
+            "rwkv.embeddings",                           # rwkv
         ),
 
         # Token type embeddings
@@ -40,6 +41,7 @@ class TensorNameMap:
             "embeddings.LayerNorm",       # bert
             "emb_ln",                     # nomic-bert
             "transformer.norm",           # openelm
+            "rwkv.blocks.0.pre_ln",       # rwkv
         ),
 
         # Position embeddings
@@ -57,6 +59,7 @@ class TensorNameMap:
             "word_embeddings_for_head",  # persimmon
             "lm_head.linear",            # phi2
             "output_layer",              # chatglm
+            "head",                      # rwkv
         ),
 
         # Output norm
@@ -76,6 +79,7 @@ class TensorNameMap:
             "encoder.final_layernorm",                 # chatglm
             "transformer.norm",                        # openelm
             "model.norm",                              # nemotron
+            "rwkv.ln_out",                             # rwkv
         ),
 
         # Rope frequencies
@@ -108,12 +112,14 @@ class TensorNameMap:
             "transformer.blocks.{bid}.norm_attn_norm.norm_1",       # dbrx
             "encoder.layers.{bid}.input_layernorm",                 # chatglm
             "transformer.layers.{bid}.attn_norm",                   # openelm
+            "rwkv.blocks.{bid}.ln1",                                # rwkv
         ),
 
         # Attention norm 2
         MODEL_TENSOR.ATTN_NORM_2: (
             "transformer.h.{bid}.ln_attn",                  # falcon40b
             "encoder.layer.{bid}.layer_norm_1",             # jina-v2-code
+            "rwkv.blocks.{bid}.ln2",                        # rwkv
         ),
 
         # Attention query-key-value
@@ -434,6 +440,98 @@ class TensorNameMap:
             "backbone.layers.{bid}.mixer.out_proj",
         ),
 
+        MODEL_TENSOR.TIME_MIX_W1: (
+            "rwkv.blocks.{bid}.attention.time_maa_w1",  # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_W2: (
+            "rwkv.blocks.{bid}.attention.time_maa_w2",  # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_LERP_X: (
+            "rwkv.blocks.{bid}.attention.time_maa_x",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_LERP_K: (
+            "rwkv.blocks.{bid}.attention.time_maa_k",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_LERP_V: (
+            "rwkv.blocks.{bid}.attention.time_maa_v",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_LERP_R: (
+            "rwkv.blocks.{bid}.attention.time_maa_r",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_LERP_G: (
+            "rwkv.blocks.{bid}.attention.time_maa_g",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_LERP_W: (
+            "rwkv.blocks.{bid}.attention.time_maa_w",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_FIRST: (
+            "rwkv.blocks.{bid}.attention.time_faaaa",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_DECAY: (
+            "rwkv.blocks.{bid}.attention.time_decay",   # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_DECAY_W1: (
+            "rwkv.blocks.{bid}.attention.time_decay_w1",  # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_DECAY_W2: (
+            "rwkv.blocks.{bid}.attention.time_decay_w2",  # rwkv v6
+        ),
+
+        MODEL_TENSOR.TIME_MIX_KEY: (
+            "rwkv.blocks.{bid}.attention.key", # rwkv
+        ),
+
+        MODEL_TENSOR.TIME_MIX_VALUE: (
+            "rwkv.blocks.{bid}.attention.value", # rwkv
+        ),
+
+        MODEL_TENSOR.TIME_MIX_RECEPTANCE: (
+            "rwkv.blocks.{bid}.attention.receptance", # rwkv
+        ),
+
+        MODEL_TENSOR.TIME_MIX_GATE: (
+            "rwkv.blocks.{bid}.attention.gate", # rwkv
+        ),
+
+        MODEL_TENSOR.TIME_MIX_LN: (
+            "rwkv.blocks.{bid}.attention.ln_x", # rwkv
+        ),
+
+        MODEL_TENSOR.TIME_MIX_OUTPUT: (
+            "rwkv.blocks.{bid}.attention.output", # rwkv
+        ),
+
+        MODEL_TENSOR.CHANNEL_MIX_LERP_K: (
+            "rwkv.blocks.{bid}.feed_forward.time_maa_k", # rwkv v6
+        ),
+
+        MODEL_TENSOR.CHANNEL_MIX_LERP_R: (
+            "rwkv.blocks.{bid}.feed_forward.time_maa_r", # rwkv v6
+        ),
+
+        MODEL_TENSOR.CHANNEL_MIX_KEY: (
+            "rwkv.blocks.{bid}.feed_forward.key", # rwkv
+        ),
+
+        MODEL_TENSOR.CHANNEL_MIX_RECEPTANCE: (
+            "rwkv.blocks.{bid}.feed_forward.receptance", # rwkv
+        ),
+
+        MODEL_TENSOR.CHANNEL_MIX_VALUE: (
+            "rwkv.blocks.{bid}.feed_forward.value", # rwkv
+        ),
+
         MODEL_TENSOR.ATTN_Q_A: (
             "model.layers.{bid}.self_attn.q_a_proj", # deepseek2
         ),

gguf-py/pyproject.toml

@@ -23,6 +23,7 @@ python = ">=3.8"
 numpy = ">=1.17"
 tqdm = ">=4.27"
 pyyaml = ">=5.1"
+sentencepiece = ">=0.1.98,<=0.2.0"
 
 [tool.poetry.dev-dependencies]
 pytest = "^5.2"

grammars/README.md

@@ -120,7 +120,7 @@ You can use GBNF grammars:
 
 - In [llama-server](../examples/server):
     - For any completion endpoints, passed as the `json_schema` body field
-    - For the `/chat/completions` endpoint, passed inside the `result_format` body field (e.g. `{"type", "json_object", "schema": {"items": {}}}`)
+    - For the `/chat/completions` endpoint, passed inside the `response_format` body field (e.g. `{"type", "json_object", "schema": {"items": {}}}`)
 - In [llama-cli](../examples/main), passed as the `--json` / `-j` flag
 - To convert to a grammar ahead of time:
     - in CLI, with [examples/json_schema_to_grammar.py](../examples/json_schema_to_grammar.py)

include/llama.h

@@ -66,6 +66,7 @@ extern "C" {
         LLAMA_VOCAB_TYPE_BPE  = 2, // GPT-2 tokenizer based on byte-level BPE
         LLAMA_VOCAB_TYPE_WPM  = 3, // BERT tokenizer based on WordPiece
         LLAMA_VOCAB_TYPE_UGM  = 4, // T5 tokenizer based on Unigram
+        LLAMA_VOCAB_TYPE_RWKV = 5, // RWKV tokenizer based on greedy tokenization
     };
 
     // pre-tokenization types
@@ -269,9 +270,9 @@ extern "C" {
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
 
         // main_gpu interpretation depends on split_mode:
-        // LLAMA_SPLIT_NONE: the GPU that is used for the entire model
+        // LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model
-        // LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results
+        // LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results
-        // LLAMA_SPLIT_LAYER: ignored
+        // LLAMA_SPLIT_MODE_LAYER: ignored
         int32_t main_gpu;
 
         // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
@@ -306,8 +307,8 @@ extern "C" {
         uint32_t n_batch;           // logical maximum batch size that can be submitted to llama_decode
         uint32_t n_ubatch;          // physical maximum batch size
         uint32_t n_seq_max;         // max number of sequences (i.e. distinct states for recurrent models)
-        uint32_t n_threads;         // number of threads to use for generation
+        int32_t  n_threads;         // number of threads to use for generation
-        uint32_t n_threads_batch;   // number of threads to use for batch processing
+        int32_t  n_threads_batch;   // number of threads to use for batch processing
 
         enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
         enum llama_pooling_type      pooling_type;      // whether to pool (sum) embedding results by sequence id
@@ -430,6 +431,13 @@ extern "C" {
     //optional:
     LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa);
 
+    // Optional: an auto threadpool gets created in ggml if not passed explicitly
+    LLAMA_API void llama_attach_threadpool(
+               struct   llama_context * ctx,
+            ggml_threadpool_t   threadpool,
+            ggml_threadpool_t   threadpool_batch);
+    LLAMA_API void llama_detach_threadpool(struct llama_context * ctx);
+
     // Call once at the end of the program - currently only used for MPI
     LLAMA_API void llama_backend_free(void);
 
@@ -839,13 +847,13 @@ extern "C" {
     // Set the number of threads used for decoding
     // n_threads is the number of threads used for generation (single token)
     // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens)
-    LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch);
+    LLAMA_API void llama_set_n_threads(struct llama_context * ctx, int32_t n_threads, int32_t n_threads_batch);
 
     // Get the number of threads used for generation of a single token.
-    LLAMA_API uint32_t llama_n_threads(struct llama_context * ctx);
+    LLAMA_API int32_t llama_n_threads(struct llama_context * ctx);
 
     // Get the number of threads used for prompt and batch processing (multiple token).
-    LLAMA_API uint32_t llama_n_threads_batch(struct llama_context * ctx);
+    LLAMA_API int32_t llama_n_threads_batch(struct llama_context * ctx);
 
     // Set whether the model is in embeddings mode or not
     // If true, embeddings will be returned but logits will not

pyproject.toml

@@ -17,7 +17,7 @@ classifiers = [
 [tool.poetry.dependencies]
 python = ">=3.9"
 numpy = "^1.25.0"
-sentencepiece = ">=0.1.98,<0.2.0"
+sentencepiece = ">=0.1.98,<=0.2.0"
 transformers = ">=4.35.2,<5.0.0"
 protobuf = ">=4.21.0,<5.0.0"
 gguf = { path = "./gguf-py" }

scripts/sync-ggml.last

@@ -1 +1 @@
-797faa25af14126eb30134d4033139ae3c5428ed
+28b7633d733bbeef0026570fbc61c79c5e9aa5ae

src/llama-impl.h

@@ -31,11 +31,17 @@ void llama_log_callback_default(ggml_log_level level, const char * text, void *
 
 static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
     if (search.empty()) {
-        return; // Avoid infinite loop if 'search' is an empty string
+        return;
     }
+    std::string builder;
+    builder.reserve(s.length());
     size_t pos = 0;
-    while ((pos = s.find(search, pos)) != std::string::npos) {
+    size_t last_pos = 0;
-        s.replace(pos, search.length(), replace);
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
-        pos += replace.length();
+        builder.append(s, last_pos, pos - last_pos);
+        builder.append(replace);
+        last_pos = pos + search.length();
     }
+    builder.append(s, last_pos, std::string::npos);
+    s = std::move(builder);
 }

src/llama-vocab.cpp

@@ -58,17 +58,17 @@ struct naive_trie {
         auto res = children.find(c);
         if (res != children.end()) {
             return res->second.get_longest_prefix(key, len, offset + 1);
-        } else {
+        }
+
         return std::make_pair(key, offset);
     }
-    }
+    const struct naive_trie * traverse(const char c) const {
-    struct naive_trie * traverse(const char c) {
         auto res = children.find(c);
         if (res != children.end()) {
             return &res->second;
-        } else {
-            return NULL;
         }
+
+        return NULL;
     }
     std::map<char, struct naive_trie> children;
     bool has_value;
@@ -843,7 +843,7 @@ struct llm_tokenizer_ugm {
             // traverse the token matcher trie to find a matching token
             bool single_codepoint_token_found = false;
             const struct best_tokenization & current_best = tokenization_results[input_offset];
-            struct naive_trie * node  = token_matcher.traverse(normalized[prefix_offset++]);
+            const struct naive_trie * node = token_matcher.traverse(normalized[prefix_offset++]);
 
             while (prefix_offset <= input_len && node != NULL) {
                 // check if we found valid token in prefix
@@ -963,7 +963,7 @@ private:
     /*
      * This structure is a view wrapper for XOR-compressed double array (XCDA)
      * See Shunsuke Kanda (2018). Space- and Time-Efficient String Dictionaries.
-     * Eeach bit-packed entry contains:
+     * Each bit-packed entry contains:
      * - BASE array value in bits 10-30
      * - LCHECK array value in bits 0-7
      * - LEAF array value in bit 9
@@ -1097,6 +1097,111 @@ private:
     struct naive_trie token_matcher;
 };
 
+//
+// RWKV tokenizer
+//
+
+static std::vector<uint8_t> llama_unescape_rwkv_token(const std::string & escaped) {
+    std::vector<uint8_t> output;
+    output.reserve(escaped.size());
+
+    // Parser state
+    bool escaping = false;
+    uint8_t hex_remaining = 0;
+    uint8_t hex_acc = 0;
+
+    // Step through characters, performing parsing
+    for (const char & c : escaped) {
+        // If we're parsing a hex code, interpret the next character
+        if (hex_remaining != 0) {
+            uint8_t value = (c >= 'a') ? (c - 'a' + 10) : (c - '0');
+            hex_acc = (hex_acc << 4) + value;
+
+            hex_remaining -= 1;
+            if (hex_remaining == 0) {
+                output.push_back(hex_acc);
+                hex_acc = 0;
+            }
+
+            continue;
+        }
+
+        // If we got an escape character, interpret it
+        if (escaping) {
+            if (c == 't') {
+                output.push_back('\t');
+            } else if (c == 'n') {
+                output.push_back('\n');
+            } else if (c == 'r') {
+                output.push_back('\r');
+            } else if (c == 'x') {
+                hex_remaining = 2;
+            } else {
+                output.push_back(c);
+            }
+
+            escaping = false;
+            continue;
+        }
+
+        if (c == '\\') {
+            escaping = true;
+            continue;
+        }
+
+        output.push_back(c);
+    }
+
+    return output;
+}
+
+struct llm_tokenizer_rwkv {
+    llm_tokenizer_rwkv(const llama_vocab & vocab): vocab(vocab) {
+        // RWKV supports arbitrary byte tokens, but the vocab struct only supports string tokens.
+        // For now, we decode the vocab here into the lookup we'll use for tokenization.
+
+        // build trie
+        for (unsigned int id = 0; id < vocab.id_to_token.size(); ++id) {
+            const auto & token = vocab.id_to_token[id];
+            const auto data = llama_unescape_rwkv_token(token.text);
+            token_matcher.insert((const char *) data.data(), data.size(), id);
+        }
+    }
+
+    void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
+        uint32_t position = 0;
+
+        while (position < text.size()) {
+            const struct naive_trie * node = token_matcher.traverse(text[position]);
+            if (node == NULL) {
+                // no matching token found, add unknown token
+                output.push_back(vocab.special_unk_id);
+                position += 1;
+                continue;
+            }
+
+            // traverse the trie to find the longest matching token
+            uint32_t token_id = 0;
+            uint32_t token_length = 0;
+            while (node != NULL) {
+                if (node->has_value) {
+                    token_id = node->value;
+                    token_length = position + 1;
+                }
+                node = node->traverse(text[++position]);
+            }
+
+            // add the longest matching token
+            output.push_back(token_id);
+            position = token_length;
+        }
+    }
+
+    const llama_vocab & vocab;
+
+    struct naive_trie token_matcher;
+};
+
 //
 // (de-) tokenize
 //
@@ -1401,6 +1506,23 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
                     output.push_back(vocab.special_eos_id);
                 }
             } break;
+        case LLAMA_VOCAB_TYPE_RWKV:
+            {
+                for (const auto & fragment : fragment_buffer) {
+                    if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
+                        auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
+
+#ifdef PRETOKENIZERDEBUG
+                        LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
+#endif
+
+                        llm_tokenizer_rwkv tokenizer(vocab);
+                        tokenizer.tokenize(raw_text, output);
+                    } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
+                        output.push_back(fragment.token);
+                    }
+                }
+            } break;
         case LLAMA_VOCAB_TYPE_NONE:
             GGML_ABORT("fatal error");
     }
@@ -1616,6 +1738,17 @@ int32_t llama_token_to_piece_impl(const struct llama_vocab & vocab, llama_token
                 }
                 break;
             }
+            case LLAMA_VOCAB_TYPE_RWKV: {
+                std::vector<uint8_t> result = llama_unescape_rwkv_token(token_text);
+
+                // If we don't have enough space, return an error
+                if (result.size() > (size_t)length) {
+                    return -(int)result.size();
+                }
+
+                memcpy(buf, result.data(), result.size());
+                return (int)result.size();
+            }
             default:
                 GGML_ABORT("fatal error");
         }

tests/test-backend-ops.cpp

@@ -949,6 +949,58 @@ struct test_rms_norm : public test_case {
     }
 };
 
+// GGML_OP_SSM_CONV
+struct test_ssm_conv : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne_a;
+    const std::array<int64_t, 4> ne_b;
+
+    std::string vars() override {
+        return VARS_TO_STR3(type, ne_a, ne_b);
+    }
+
+    test_ssm_conv(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne_a = {10, 10, 10, 1},
+            std::array<int64_t, 4> ne_b = {3, 3, 1, 1})
+        : type(type), ne_a(ne_a), ne_b(ne_b) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a   = ggml_new_tensor(ctx, type, 4, ne_a.data());
+        ggml_tensor * b   = ggml_new_tensor(ctx, type, 4, ne_b.data());
+        ggml_tensor * out = ggml_ssm_conv(ctx, a, b);
+        return out;
+    }
+};
+
+// GGML_OP_SSM_SCAN
+struct test_ssm_scan : public test_case {
+    const ggml_type type;
+
+    const int64_t d_state;
+    const int64_t d_inner;
+    const int64_t n_seq_tokens;
+    const int64_t n_seqs;
+
+    std::string vars() override {
+        return VARS_TO_STR5(type, d_state, d_inner, n_seq_tokens, n_seqs);
+    }
+
+    test_ssm_scan(ggml_type type = GGML_TYPE_F32,
+            int64_t d_state = 32, int64_t d_inner = 32, int64_t n_seq_tokens = 32, int64_t n_seqs = 32)
+        : type(type), d_state(d_state), d_inner(d_inner), n_seq_tokens(n_seq_tokens), n_seqs(n_seqs) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * s   = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_state, d_inner,      n_seqs, 1 }.data());
+        ggml_tensor * x   = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_inner, n_seq_tokens, n_seqs, 1 }.data());
+        ggml_tensor * dt  = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_inner, n_seq_tokens, n_seqs, 1 }.data());
+        ggml_tensor * A   = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_state, d_inner,      1     , 1 }.data());
+        ggml_tensor * B   = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_state, n_seq_tokens, n_seqs, 1 }.data());
+        ggml_tensor * C   = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_state, n_seq_tokens, n_seqs, 1 }.data());
+        ggml_tensor * out = ggml_ssm_scan(ctx, s, x, dt, A, B, C);
+        return out;
+    }
+};
+
 // GGML_OP_MUL_MAT
 struct test_mul_mat : public test_case {
     const ggml_type type_a;
@@ -1108,6 +1160,58 @@ struct test_sqrt : public test_case {
     }
 };
 
+// GGML_OP_SIN
+struct test_sin : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_sin(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 10})
+        : type(type), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_sin(ctx, a);
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            init_tensor_uniform(t, -100.0f, 100.0f);
+        }
+    }
+};
+
+// GGML_OP_COS
+struct test_cos : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_cos(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 10})
+        : type(type), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_cos(ctx, a);
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            init_tensor_uniform(t, -100.0f, 100.0f);
+        }
+    }
+};
+
 // GGML_OP_CLAMP
 struct test_clamp : public test_case {
     const ggml_type type;
@@ -1652,19 +1756,20 @@ struct test_flash_attn_ext : public test_case {
     const bool mask; // use mask
 
     const float max_bias; // ALiBi
+    const float logit_softcap; // Gemma 2
 
     const ggml_type type_KV;
 
     std::string vars() override {
-        return VARS_TO_STR7(hs, nh, kv, nb, mask, max_bias, type_KV);
+        return VARS_TO_STR8(hs, nh, kv, nb, mask, max_bias, logit_softcap, type_KV);
     }
 
     double max_nmse_err() override {
         return 5e-4;
     }
 
-    test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8, bool mask = true, float max_bias = 0.0f, ggml_type type_KV = GGML_TYPE_F16)
+    test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8, bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_type type_KV = GGML_TYPE_F16)
-        : hs(hs), nh(nh), kv(kv), nb(nb), mask(mask), max_bias(max_bias), type_KV(type_KV) {}
+        : hs(hs), nh(nh), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), type_KV(type_KV) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         const int64_t hs_padded = GGML_PAD(hs, ggml_blck_size(type_KV));
@@ -1673,7 +1778,28 @@ struct test_flash_attn_ext : public test_case {
         ggml_tensor * k = ggml_new_tensor_4d(ctx, type_KV,       hs_padded, kv, nh, 1);
         ggml_tensor * v = ggml_new_tensor_4d(ctx, type_KV,       hs_padded, kv, nh, 1);
         ggml_tensor * m = mask ? ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), 1, 1) : nullptr;
-        ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hs), max_bias);
+        ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hs), max_bias, logit_softcap);
+        return out;
+    }
+};
+
+// GGML_OP_CROSS_ENTROPY_LOSS
+struct test_cross_entropy_loss : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_cross_entropy_loss(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 10})
+        : type(type), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * logits = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * labels = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_cross_entropy_loss(ctx, logits, labels);
         return out;
     }
 };
@@ -2239,6 +2365,12 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         test_cases.emplace_back(new test_rms_norm(GGML_TYPE_F32, {64, 10, 10, 10}, eps));
     }
 
+    test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, 1536, 1, 1}, {4, 1536, 1, 1}));
+    test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {8, 1536, 1, 1}, {4, 1536, 1, 1}));
+    test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, 1536, 4, 1}, {4, 1536, 1, 1}));
+
+    test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 16, 1024, 32, 4));
+
 #if 1
     for (ggml_type type_a : base_types) {
         for (ggml_type type_b : {GGML_TYPE_F32, GGML_TYPE_F16}) {
@@ -2334,6 +2466,8 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
 
     test_cases.emplace_back(new test_sqr());
     test_cases.emplace_back(new test_sqrt());
+    test_cases.emplace_back(new test_sin());
+    test_cases.emplace_back(new test_cos());
     test_cases.emplace_back(new test_clamp());
 
     test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10,  1,  1}, 5));
@@ -2437,11 +2571,13 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         for (bool mask : { true, false } ) {
             for (float max_bias : { 0.0f, 8.0f }) {
                 if (!mask && max_bias > 0.0f) continue;
+                for (float logit_softcap : {0.0f, 10.0f}) {
+                    if (hs != 128 && logit_softcap != 0.0f) continue;
                     for (int nh : { 32, }) {
                         for (int kv : { 512, 1024, }) {
                             for (int nb : { 1, 2, 4, 8, }) {
                                 for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
-                                test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb, mask, max_bias, type_KV));
+                                    test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb, mask, max_bias, logit_softcap, type_KV));
                                 }
                             }
                         }
@@ -2449,6 +2585,9 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
                 }
             }
         }
+    }
+
+    test_cases.emplace_back(new test_cross_entropy_loss());
 
     // these tests are disabled to save execution time, but they can be handy for debugging
 #if 0
@@ -2483,7 +2622,6 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     }
 
     GGML_ABORT("fatal error");
-    return false;
 }
 
 static void usage(char ** argv) {

tests/test-grad0.cpp

@@ -1,10 +1,14 @@
 #define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnings on Windows
 #include "ggml.h"
 
+#include <cfloat>
 #include <cmath>
+#include <cstdint>
 #include <cstdio>
 #include <cstdlib>
 #include <cassert>
+#include <initializer_list>
+#include <vector>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -217,7 +221,8 @@ static bool check_gradient(
         int nargs,
         float eps,
         float max_error_abs,
-        float max_error_rel) {
+        float max_error_rel,
+        std::vector<double> expected_vals) {
 
     static int n_threads = -1;
     if (n_threads < 0) {
@@ -248,9 +253,10 @@ static bool check_gradient(
     // ggml_graph_dump_dot(gb, gf,  "test-grad0-backward.dot");
 
     for (int i = 0; i < nargs; ++i) {
+        bool all_g0_bad = true;
         const int nelements = ggml_nelements(x[i]);
         for (int k = 0; k < nelements; ++k) {
-            // compute gradient using finite differences
+            // Calculate gradient numerically:
             const float x0 = ggml_get_f32_1d(x[i], k);
             const float xm = x0 - eps;
             const float xp = x0 + eps;
@@ -267,6 +273,28 @@ static bool check_gradient(
             const double f1 = ggml_get_f32_1d(f, 0);
             const double g0 = (f0 - f1)/(2.0*(double) eps);
 
+            // The numerical calculation of the gradient fails around noncontinuities (e.g. 0 for ReLU).
+            // In such cases, provide a vector of expected values and skip the comparison for failed calculations.
+            if (!expected_vals.empty()) {
+                bool matches_any = false;
+                for (const double & ev : expected_vals) {
+                    const double error_abs = std::fabs(g0 - ev);
+                    if (error_abs > max_error_abs) {
+                        continue;
+                    }
+                    const double error_rel = g0 != 0.0 ? fabs(g0 - ev)/fabs(g0) : 0.0;
+                    if (error_rel > max_error_rel) {
+                        continue;
+                    }
+                    matches_any = true;
+                    break;
+                }
+                if (!matches_any) {
+                    continue;
+                }
+            }
+            all_g0_bad = false;
+
             ggml_set_f32_1d(x[i], k, x0);
 
             // compute gradient using backward graph
@@ -278,7 +306,7 @@ static bool check_gradient(
             const double g1 = ggml_get_f32_1d(x[i]->grad, k);
 
             const double error_abs = fabs(g0 - g1);
-            const double error_rel = g0 != 0 ? fabs(g0 - g1)/fabs(g0) : 0;
+            const double error_rel = g0 != 0.0 ? fabs(g0 - g1)/fabs(g0) : 0.0;
 
             if (error_abs > max_error_abs || error_rel > max_error_rel) {
                 printf("%s: ndims=%d, i=%d, k=%d, x0=%f, xm=%f, xp=%f, f0=%f, f1=%f, g0=%f, g1=%f, eps=%f, error_abs=%f, error_rel=%f\n",
@@ -287,6 +315,10 @@ static bool check_gradient(
                 return false;
             }
         }
+        if (all_g0_bad) {
+            printf("%s: numerical calculation of the gradient failed for all values\n", op_name);
+            return false;
+        }
     }
 
     return true;
@@ -404,7 +436,7 @@ int main(int argc, const char ** argv) {
         seed_iter = rand();
         unsigned seed = rand();
 
-        printf("test-grad0: iter:%d/%d\n", iter, niter);
+        printf("test-grad0: iter:%d/%d\n", (iter+1), niter);
         struct ggml_context * ctx0 = ggml_init(params);
 
         get_random_dims(ne, 4);
@@ -424,7 +456,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_add(ctx0, x[0], x[1]));
 
-                check_gradient("add f32", ctx0, x, f, ndims, nargs, 1e-3f, 2e-3f, 2e-3f);
+                check_gradient("add f32", ctx0, x, f, ndims, nargs, 1e-3f, 2e-3f, 2e-3f, {});
             }
         }
 
@@ -441,7 +473,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_add(ctx0, x[0], x[1]));
 
-                check_gradient("add f16", ctx0, x, f, ndims, nargs, 1e-1f, 2e-1f, 2e-1f);
+                check_gradient("add f16", ctx0, x, f, ndims, nargs, 1e-1f, 2e-1f, 2e-1f, {});
             }
         }
 
@@ -458,7 +490,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_sub(ctx0, x[0], x[1]));
 
-                check_gradient("sub", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+                check_gradient("sub", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
             }
         }
 
@@ -475,7 +507,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_mul(ctx0, x[0], x[1]));
 
-                check_gradient("mul", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("mul", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -492,7 +524,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_div(ctx0, x[0], x[1]));
 
-                check_gradient("div", ctx0, x, f, ndims, nargs, 1e-3f, 1e-1f, 1e-1f);
+                check_gradient("div", ctx0, x, f, ndims, nargs, 1e-3f, 1e-1f, 1e-1f, {});
             }
         }
 
@@ -509,7 +541,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqr(ctx0, x[0]));
 
-                check_gradient("sqr", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("sqr", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -526,7 +558,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqrt(ctx0, x[0]));
 
-                check_gradient("sqrt", ctx0, x, f, ndims, nargs, 1e-3f, 2e-2f, 1e-1f);
+                check_gradient("sqrt", ctx0, x, f, ndims, nargs, 1e-3f, 2e-2f, 1e-1f, {});
             }
         }
 
@@ -543,7 +575,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_log(ctx0, x[0]));
 
-                check_gradient("log", ctx0, x, f, ndims, nargs, 1e-3f, INFINITY, 1e-1f);
+                check_gradient("log", ctx0, x, f, ndims, nargs, 1e-3f, INFINITY, 1e-1f, {});
             }
         }
 
@@ -560,7 +592,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, x[0]);
 
-                check_gradient("sum", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+                check_gradient("sum", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
             }
         }
 
@@ -578,7 +610,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqr(ctx0, ggml_sum_rows(ctx0, x[0])));
 
-                check_gradient("sum_rows", ctx0, x, f, ndims, nargs, 1e-3f, 1e-2f, INFINITY);
+                check_gradient("sum_rows", ctx0, x, f, ndims, nargs, 1e-3f, 1e-2f, INFINITY, {});
             }
         }
 
@@ -596,7 +628,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_mean(ctx0, x[0]));
 
-                check_gradient("mean", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+                check_gradient("mean", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
             }
         }
 
@@ -614,7 +646,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_argmax(ctx0, x[0]));
 
-                check_gradient("argmax", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+                check_gradient("argmax", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
             }
         }
 
@@ -637,7 +669,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x[1], ggml_repeat(ctx0, x[0], x[1]))));
 
-                check_gradient("repeat", ctx0, x, f, ndims, nargs, 1e-3f, 1e-2f, INFINITY);
+                check_gradient("repeat", ctx0, x, f, ndims, nargs, 1e-3f, 1e-2f, INFINITY, {});
             }
         }
 
@@ -660,25 +692,25 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x[0], ggml_repeat_back(ctx0, x[1], x[0]))));
 
-                check_gradient("repeat back", ctx0, x, f, ndims, nargs, 1e-3f, 1e-2f, INFINITY);
+                check_gradient("repeat back", ctx0, x, f, ndims, nargs, 1e-3f, 1e-2f, INFINITY, {});
             }
         }
 
-        // abs (finite differences do not work)
+        // abs
-        //{
+        {
-        //    const int nargs = 1;
+           const int nargs = 1;
 
-        //    for (int ndims = 1; ndims <= 2; ++ndims) {
+           for (int ndims = 1; ndims <= 4; ++ndims) {
-        //        for (int i = 0; i < nargs; ++i) {
+               for (int i = 0; i < nargs; ++i) {
-        //            x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+                   x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-        //            ggml_set_param(ctx0, x[i]);
+                   ggml_set_param(ctx0, x[i]);
-        //        }
+               }
 
-        //        struct ggml_tensor * f = ggml_sum(ctx0, ggml_abs(ctx0, x[0]));
+               struct ggml_tensor * f = ggml_sum(ctx0, ggml_abs(ctx0, x[0]));
 
-        //        check_gradient("abs", ctx0, x, f, ndims, nargs, 1e-3f, INFINITY, 1e-3f);
+               check_gradient("abs", ctx0, x, f, ndims, nargs, 1e-3f, INFINITY, 1e-3f, {-1.0, 1.0});
-        //    }
+           }
-        //}
+        }
 
         // sgn
         {
@@ -693,7 +725,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor* f = ggml_sum(ctx0, ggml_sgn(ctx0, x[0]));
 
-                check_gradient("sgn", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+                check_gradient("sgn", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {0.0});
             }
         }
 
@@ -710,7 +742,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor* f = ggml_sum(ctx0, ggml_neg(ctx0, x[0]));
 
-                check_gradient("neg", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+                check_gradient("neg", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
             }
         }
 
@@ -727,7 +759,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor* f = ggml_sum(ctx0, ggml_step(ctx0, x[0]));
 
-                check_gradient("step", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+                check_gradient("step", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {0.0});
             }
         }
 
@@ -745,7 +777,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor* f = ggml_sum(ctx0, ggml_tanh(ctx0, x[0]));
 
-                check_gradient("tanh", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+                check_gradient("tanh", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
             }
         }
 
@@ -776,7 +808,7 @@ int main(int argc, const char ** argv) {
 
                         GGML_PRINT_DEBUG("testing: mul_mat, [%lld, %lld] (%d) * [%lld, %lld] (%d)\n", x[1]->ne[0], x[1]->ne[1], x[1]->n_dims, x[0]->ne[0], x[0]->ne[1], x[0]->n_dims);
 
-                        check_gradient("mul_mat", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                        check_gradient("mul_mat", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
                         if (ndims == 2) {
                             // check_mat_mul does not support ndims > 2
                             check_mat_mul(m, x[1], x[0]);
@@ -800,7 +832,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor* f = ggml_sum(ctx0, ggml_elu(ctx0, x[0]));
 
-                check_gradient("elu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+                check_gradient("elu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
             }
         }
 
@@ -817,7 +849,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor* f = ggml_sum(ctx0, ggml_relu(ctx0, x[0]));
 
-                check_gradient("relu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("relu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {0.0, 1.0});
             }
         }
 
@@ -835,7 +867,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor* f = ggml_sum(ctx0, ggml_gelu(ctx0, x[0]));
 
-                check_gradient("gelu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f);
+                check_gradient("gelu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
             }
         }
 
@@ -854,9 +886,9 @@ int main(int argc, const char ** argv) {
 
 #ifdef GGML_SILU_FP16
                 // due to GGML_SILU_FP16 the finite difference method will be slightly wrong -> increase error bounds.
-                check_gradient("silu", ctx0, x, f, ndims, nargs, 1e-3f, 0.5, INFINITY);
+                check_gradient("silu", ctx0, x, f, ndims, nargs, 1e-3f, 0.5, INFINITY, {});
 #else
-                check_gradient("silu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("silu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
 #endif
             }
         }
@@ -874,7 +906,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_rms_norm(ctx0, x[0], 1e-6f));
 
-                check_gradient("rms_norm", ctx0, x, f, ndims, nargs, 1e-4f, 1.0f, INFINITY);
+                check_gradient("rms_norm", ctx0, x, f, ndims, nargs, 1e-4f, 1.0f, INFINITY, {});
             }
         }
 
@@ -892,7 +924,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_scale(ctx0, x[0], s));
 
-                check_gradient("scale", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("scale", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -910,7 +942,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_cpy(ctx0, x[0], x[1]));
 
-                check_gradient("cpy f32", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("cpy f32", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -928,7 +960,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_cpy(ctx0, x[0], x[1]));
 
-                check_gradient("cpy f16", ctx0, x, f, ndims, nargs, 1e-1f, 1e-1f, INFINITY);
+                check_gradient("cpy f16", ctx0, x, f, ndims, nargs, 1e-1f, 1e-1f, INFINITY, {});
             }
         }
 
@@ -952,7 +984,7 @@ int main(int argc, const char ** argv) {
 
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_reshape(ctx0, x[0], x[1]));
-                check_gradient("reshape", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("reshape", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -976,7 +1008,7 @@ int main(int argc, const char ** argv) {
 
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_reshape(ctx0, x[0], x[1]));
-                check_gradient("reshape", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("reshape", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -1004,7 +1036,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_acc(ctx0, x[0], x[1], x[0]->nb[1], x[0]->nb[2], x[0]->nb[3], offset));
 
-                check_gradient("acc 1d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("acc 1d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -1037,7 +1069,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_acc(ctx0, x[0], x[1], x[0]->nb[1], x[0]->nb[2], x[0]->nb[3], offset));
 
-                check_gradient("acc 2d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("acc 2d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -1072,7 +1104,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_acc(ctx0, x[0], x[1], x[0]->nb[1], x[0]->nb[2], x[0]->nb[3], offset));
 
-                check_gradient("acc 3d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("acc 3d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -1109,7 +1141,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_acc(ctx0, x[0], x[1], x[0]->nb[1], x[0]->nb[2], x[0]->nb[3], offset));
 
-                check_gradient("acc 4d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("acc 4d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -1137,7 +1169,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_set_1d(ctx0, x[0], x[1], offset));
 
-                check_gradient("set_1d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("set_1d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -1170,7 +1202,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_set_2d(ctx0, x[0], x[1], x[1]->nb[1], offset));
 
-                check_gradient("set_2d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("set_2d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -1194,7 +1226,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_view_1d(ctx0, x[0], nelem, offset));
 
-                check_gradient("view_1d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("view_1d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -1225,7 +1257,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_view_2d(ctx0, x[0], ne2[0], ne2[1], nb2[1], offset));
 
-                check_gradient("view_2d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("view_2d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -1257,7 +1289,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_view_3d(ctx0, x[0], ne2[0], ne2[1], ne2[2], nb2[1], nb2[2], offset));
 
-                check_gradient("view_3d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("view_3d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -1291,7 +1323,7 @@ int main(int argc, const char ** argv) {
                 // sum requires contiguous tensor rows
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_cont(ctx0, ggml_permute(ctx0, x[0], ax0, ax1, ax2, ax3)));
 
-                check_gradient("permute", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("permute", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -1319,7 +1351,7 @@ int main(int argc, const char ** argv) {
                 // sum requires contiguous tensor rows
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, x[0])));
 
-                check_gradient("transpose", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("transpose", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -1337,7 +1369,7 @@ int main(int argc, const char ** argv) {
 
             struct ggml_tensor * f = ggml_sum(ctx0, ggml_get_rows(ctx0, x[0], x[1]));
 
-            check_gradient("get_rows", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+            check_gradient("get_rows", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
         }
 
         // diag_mask_inf
@@ -1353,7 +1385,7 @@ int main(int argc, const char ** argv) {
 
             struct ggml_tensor * f = ggml_sum(ctx0, ggml_diag_mask_inf(ctx0, x[0], n_past));
 
-            check_gradient("diag_mask_inf", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+            check_gradient("diag_mask_inf", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
         }
 
         // diag_mask_zero
@@ -1369,7 +1401,7 @@ int main(int argc, const char ** argv) {
 
             struct ggml_tensor * f = ggml_sum(ctx0, ggml_diag_mask_zero(ctx0, x[0], n_past));
 
-            check_gradient("diag_mask_zero", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+            check_gradient("diag_mask_zero", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
         }
 
         // softmax
@@ -1395,7 +1427,7 @@ int main(int argc, const char ** argv) {
                                                         1.0f - eps),
                                                     ggml_new_f32(ctx0, eps))));
 
-                check_gradient("softmax", ctx0, x, f, ndims, nargs, 1e-3f, 2e-1f, INFINITY);
+                check_gradient("softmax", ctx0, x, f, ndims, nargs, 1e-3f, 2e-1f, INFINITY, {});
                 // NOTE: softmax forward is computed using f16 table lookup instead of using actual expf, but backward assumes actual expf.
                 // this may result in different gradients too finite differences.
                 // when this test reports errors, first try to replace the table lookup with actual expf and test again to see if just that was the cause.
@@ -1412,7 +1444,7 @@ int main(int argc, const char ** argv) {
             get_random_dims(ne2, 4);
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -0.1f, 0.1f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
                 x[1] = get_random_tensor_f32(ctx0, ndims, ne2, 0.0f, 1.0f);
                 // the second argument to cross_entropy_loss must sum up to 1 for each row
                 int nr = ggml_nrows(x[1]);
@@ -1430,7 +1462,7 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_cross_entropy_loss(ctx0, x[0], x[1]);
 
-                check_gradient("cross_entropy_loss", ctx0, x, f, ndims, nargs, 1e-4f, 1e-3f, INFINITY);
+                check_gradient("cross_entropy_loss", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
             }
         }
 
@@ -1468,7 +1500,7 @@ int main(int argc, const char ** argv) {
                         struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], p, n_rot, mode));
 
                         GGML_PRINT_DEBUG("rope f32: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
-                        check_gradient("rope f32", ctx0, x, f, ndims, nargs, 1e-2f, 1e-3f, INFINITY);
+                        check_gradient("rope f32", ctx0, x, f, ndims, nargs, 1e-2f, 1e-3f, INFINITY, {});
                     }
                 }
             }
@@ -1508,12 +1540,93 @@ int main(int argc, const char ** argv) {
                         struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], p, n_rot, mode));
 
                         GGML_PRINT_DEBUG("rope f16: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
-                        check_gradient("rope f16", ctx0, x, f, ndims, nargs, 1e-1f, 1e-1f, INFINITY);
+                        check_gradient("rope f16", ctx0, x, f, ndims, nargs, 1e-1f, 1e-1f, INFINITY, {});
                     }
                 }
             }
         }
 
+        // im2col f32
+        {
+            srand(seed);
+            const int nargs = 1;
+            const int ndims = 4;
+
+            for (const bool is_2D : {false, true}) {
+                int64_t ne0[ndims];
+                int64_t ne1[ndims];
+                get_random_dims(ne0, ndims);
+                get_random_dims(ne1, ndims);
+
+                // // Ensure that the output is not zero-sized:
+                ne1[0] += 8;
+                ne1[1] += 8;
+
+                if (is_2D) {
+                    ne1[2] = ne0[2];
+                } else {
+                    ne1[1] = ne0[1];
+                    ne0[3] = 1;
+                    ne1[3] = 1;
+                }
+
+                // The order of arguments is swapped because the first tensor is only used for its shape.
+                x[1] = get_random_tensor_f16(ctx0, ndims, ne0, -1.0f, 1.0f);
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne1, -1.0f, 1.0f);
+
+                ggml_set_param(ctx0, x[0]);
+
+                const int s0 =         1 + irand(2);
+                const int s1 = is_2D ? 1 + irand(2) : 0;
+                const int p0 =         0 + irand(2);
+                const int p1 = is_2D ? 0 + irand(2) : 0;
+                const int d0 =         1 + irand(2);
+                const int d1 = is_2D ? 1 + irand(2) : 0;
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_im2col(ctx0, x[1], x[0], s0, s1, p0, p1, d0, d1, is_2D, GGML_TYPE_F32));
+
+                GGML_PRINT_DEBUG("im2col f32: is_2D=%s, s0=%d, s1=%d, p0=%d, p1=%d, d0=%d, d1=%d\n", is_2D ? "yes" : "no", s0, s1, p0, p1, d0, d1);
+                check_gradient("im2col f32", ctx0, x, f, ndims, nargs, 1e-2f, 1e-3f, INFINITY, {});
+            }
+        }
+
+        // pool_2d f32
+        {
+            srand(seed);
+            const int nargs = 1;
+            const int ndims = 4;
+
+            for (const enum ggml_op_pool op : {GGML_OP_POOL_AVG, GGML_OP_POOL_MAX}) {
+                int64_t ne0[ndims];
+                get_random_dims(ne0, ndims);
+
+                ne0[0] += 8;
+                ne0[1] += 8;
+
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne0, -1.0f, 1.0f);
+
+                ggml_set_param(ctx0, x[0]);
+
+                const int k0 = 2 + irand(2);
+                const int k1 = 2 + irand(2);
+                const int s0 = 2 + irand(2);
+                const int s1 = 2 + irand(2);
+                const int p0 = 0 + irand(2);
+                const int p1 = 0 + irand(2);
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_pool_2d(ctx0, x[0], op, k0, k1, s0, s1, p0, p1));
+
+                GGML_PRINT_DEBUG("ggml_pool_2d f32: op=%s k0=%d, k1=%d, s0=%d, s1=%d, p0=%d, p1=%d\n",
+                                 op == GGML_OP_POOL_MAX ? "max" : "avg", k0, k1, s0, s1, p0, p1);
+                std::vector<double> expected_vals;
+                if (op == GGML_OP_POOL_MAX) {
+                    expected_vals.push_back(0.0);
+                    expected_vals.push_back(1.0);
+                }
+                check_gradient("ggml_pool_2d f32", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, expected_vals);
+            }
+        }
+
         // flash_attn f32
         // TODO: adapt to ggml_flash_attn_ext() changes
         //{
@@ -1553,7 +1666,7 @@ int main(int argc, const char ** argv) {
 
         //                struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
 
-        //                check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
+        //                check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY, {});
         //            }
         //        }
         //    }

tests/test-lora-conversion-inference.sh

@@ -14,7 +14,7 @@ MODELS_REPO_URL=https://huggingface.co/ggml-org/$MODELS_REPO
 # Clone the Hugging Face repository if the directory does not exist
 if [ ! -d "$MODELS_REPO" ]; then
     echo "Cloning the Hugging Face repository..."
-    git clone $MODELS_REPO_URL
+    git clone $MODELS_REPO_URL --depth 1
 else
     echo "Repository already exists. Skipping clone."
 fi

tests/test-rope.cpp

@@ -113,7 +113,7 @@ static struct ggml_tensor * get_random_tensor_f32(
 }
 
 static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
-    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr);
 
     if (plan.work_size > 0) {
         buf.resize(plan.work_size);

tests/test-sampling.cpp

@@ -166,12 +166,12 @@ static void test_sampler_queue(
     for (auto s : samplers_sequence) {
         switch (s){
             case 'k': llama_sample_top_k    (nullptr, &candidates_p, top_k, 1); break;
-            case 'f': GGML_ABORT("tail_free test not implemented");   break;
+            case 'f': GGML_ABORT("tail_free test not implemented");
-            case 'y': GGML_ABORT("typical test not implemented");     break;
+            case 'y': GGML_ABORT("typical test not implemented");
             case 'p': llama_sample_top_p    (nullptr, &candidates_p, top_p, 1); break;
             case 'm': llama_sample_min_p    (nullptr, &candidates_p, min_p, 1); break;
-            case 't': GGML_ABORT("temperature test not implemented"); break;
+            case 't': GGML_ABORT("temperature test not implemented");
-            default : GGML_ABORT("Unknown sampler");                  break;
+            default : GGML_ABORT("Unknown sampler");
         }
 
         llama_sample_softmax(nullptr, &candidates_p); // make sure tokens are sorted for tests