llama : support RWKV v6 models (#8980)

* convert_hf_to_gguf: Add support for RWKV v6

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Add RWKV tokenization

* Fix build

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Do not use special tokens when matching in RWKV tokenizer

* Fix model loading

* Add (broken) placeholder graph builder for RWKV

* Add workaround for kv cache

* Add logits conversion to rwkv5

* Add rwkv5 layer norms

* Add time mix KVRG & correct merge mistake

* Add remaining time mix parameters

* Add time mix output loading

* Add placeholder llm_build_time_mix

* Fix build

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Load more tensors for rwkv v6

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Fix rwkv tokenizer

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* ggml: Add unary operator Exp

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* RWKV v6 graph building

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Add ``rescale_every_n_layers`` parameter

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Add ``wkv.head_size`` key for RWKV

so it doesn't reuse Mamba ssm parameters

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Fix offloading layers to CUDA

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Fix parallel inferencing for RWKV

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Remove trailing whitespaces

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* build_rwkv: Avoid using inplace operations

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* convert_hf_to_gguf: rwkv: Avoid using ``eval``

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* convert_hf_to_gguf: rwkv tokenizer: Don't escape sequences manually

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Update convert_hf_to_gguf.py

Co-authored-by: compilade <git@compilade.net>

* ggml: Add backward computation for unary op ``exp``

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Update convert_hf_to_gguf.py

Co-authored-by: compilade <git@compilade.net>

* Update convert_hf_to_gguf.py

Co-authored-by: compilade <git@compilade.net>

* Use MODEL_ARCH.RWKV6 instead of MODEL_ARCH.RWKV

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* build_rwkv6: Simplify graph

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: rwkv6: Detect model.type

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: rwkv6: Fix tensor loading for 7B/14B models

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: rwkv6: Fix group_norm assertion failure with Metal

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: rwkv6: Clean up

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: rwkv6: Add quantization tensor exclusion

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: rwkv6: Use the new advanced batch splits

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Update src/llama.cpp

Co-authored-by: compilade <git@compilade.net>

* llama: rwkv6: Use ``ggml_norm`` instead of ``ggml_group_norm``

Co-authored-by: compilade <git@compilade.net>

* llama: rwkv6: Apply code style and misc changes

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* converter: Use class name ``Rwkv6Model``

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: rwkv6: Make use of key ``feed_forward_length``

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: rwkv6: Add kv ``time_mix_extra_dim`` and ``time_decay_extra_dim``

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* converter: Match ``new_name`` instead of ``name`` for float32 explicit tensors

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: rwkv6: Keep ``time_mix_w1/w2`` as F32

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: rwkv6: Remove unused nodes

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: rwkv6: Apply code format changes

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: rwkv6: Add lora for some supported tensors

Currently att.key/receptance/value/gate/output, ffn.receptance/key/value, as well as head.weight

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* rwkv : speed-up tokenization using trie

* minor : style + indentation

* llama: rwkv6: Avoid division by zero

Co-authored-by: compilade <git@compilade.net>

* ggml: rwkv_wkv: Avoid copying the state

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

---------

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>
Co-authored-by: Layl Bongers <3094382+LaylBongers@users.noreply.github.com>
Co-authored-by: compilade <git@compilade.net>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

.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.
-ARG CUDA_DOCKER_ARCH=all
+# CUDA architecture to build for (defaults to all supported archs)
+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
-ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable CUDA
-ENV GGML_CUDA=1
-# Enable cURL
-ENV LLAMA_CURL=1
-
-RUN make -j$(nproc)
+# Use the default CUDA archs if not specified
+RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
+        export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
+    fi && \
+    cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-cli -j$(nproc) && \
+    cp build/bin/* .
 
 ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/llama-cli-cann.Dockerfile

@@ -0,0 +1,44 @@
+ARG ASCEND_VERSION=8.0.rc2.alpha003-910b-openeuler22.03-py3.8
+
+FROM cosdt/cann:$ASCEND_VERSION AS build
+
+WORKDIR /app
+
+COPY . .
+
+RUN yum install -y gcc g++ cmake make
+ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
+ENV LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:$LIBRARY_PATH
+ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/lib64/plugin/opskernel:${ASCEND_TOOLKIT_HOME}/lib64/plugin/nnengine:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe/op_tiling:${LD_LIBRARY_PATH}
+ENV PYTHONPATH=${ASCEND_TOOLKIT_HOME}/python/site-packages:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe:${PYTHONPATH}
+ENV PATH=${ASCEND_TOOLKIT_HOME}/bin:${ASCEND_TOOLKIT_HOME}/compiler/ccec_compiler/bin:${PATH}
+ENV ASCEND_AICPU_PATH=${ASCEND_TOOLKIT_HOME}
+ENV ASCEND_OPP_PATH=${ASCEND_TOOLKIT_HOME}/opp
+ENV TOOLCHAIN_HOME=${ASCEND_TOOLKIT_HOME}/toolkit
+ENV ASCEND_HOME_PATH=${ASCEND_TOOLKIT_HOME}
+
+# find libascend_hal.so, because the drive hasn`t been mounted.
+ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH
+
+RUN echo "Building with static libs" && \
+    source /usr/local/Ascend/ascend-toolkit/set_env.sh --force && \
+    cmake -B build -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF  && \
+    cmake --build build --config Release --target llama-cli
+
+# TODO: use image with NNRT
+FROM cosdt/cann:$ASCEND_VERSION AS runtime
+COPY --from=build /app/build/bin/llama-cli /llama-cli
+
+ENV LC_ALL=C.utf8
+
+ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
+ENV LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:$LIBRARY_PATH
+ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/lib64/plugin/opskernel:${ASCEND_TOOLKIT_HOME}/lib64/plugin/nnengine:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe/op_tiling:${LD_LIBRARY_PATH}
+ENV PYTHONPATH=${ASCEND_TOOLKIT_HOME}/python/site-packages:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe:${PYTHONPATH}
+ENV PATH=${ASCEND_TOOLKIT_HOME}/bin:${ASCEND_TOOLKIT_HOME}/compiler/ccec_compiler/bin:${PATH}
+ENV ASCEND_AICPU_PATH=${ASCEND_TOOLKIT_HOME}
+ENV ASCEND_OPP_PATH=${ASCEND_TOOLKIT_HOME}/opp
+ENV TOOLCHAIN_HOME=${ASCEND_TOOLKIT_HOME}/toolkit
+ENV ASCEND_HOME_PATH=${ASCEND_TOOLKIT_HOME}
+
+ENTRYPOINT ["/llama-cli" ]

.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.
-ARG CUDA_DOCKER_ARCH=all
+# CUDA architecture to build for (defaults to all supported archs)
+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
-ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable CUDA
-ENV GGML_CUDA=1
-
-RUN make -j$(nproc) llama-cli
+# Use the default CUDA archs if not specified
+RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
+        export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
+    fi && \
+    cmake -B build -DGGML_CUDA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-cli -j$(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.
-ARG CUDA_DOCKER_ARCH=all
+# CUDA architecture to build for (defaults to all supported archs)
+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
-ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable CUDA
-ENV GGML_CUDA=1
-# Enable cURL
-ENV LLAMA_CURL=1
-
-RUN make -j$(nproc) llama-server
+# Use the default CUDA archs if not specified
+RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
+        export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
+    fi && \
+    cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-server -j$(nproc)
 
 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

@@ -3,7 +3,7 @@ ARG UBUNTU_VERSION=22.04
 FROM ubuntu:$UBUNTU_VERSION AS build
 
 RUN apt-get update && \
-    apt-get install -y build-essential git libcurl4-openssl-dev curl
+    apt-get install -y build-essential git libcurl4-openssl-dev
 
 WORKDIR /app
 
@@ -16,11 +16,13 @@ RUN make -j$(nproc) llama-server
 FROM ubuntu:$UBUNTU_VERSION AS runtime
 
 RUN apt-get update && \
-    apt-get install -y libcurl4-openssl-dev libgomp1
+    apt-get install -y libcurl4-openssl-dev libgomp1 curl
 
 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/package.nix

@@ -13,6 +13,7 @@
   mpi,
   blas,
   cudaPackages,
+  autoAddDriverRunpath,
   darwin,
   rocmPackages,
   vulkan-headers,
@@ -126,16 +127,9 @@ let
     ++ optionals useMetalKit [ MetalKit ];
 
   cudaBuildInputs = with cudaPackages; [
-    cuda_cccl.dev # <nv/target>
-
-    # A temporary hack for reducing the closure size, remove once cudaPackages
-    # have stopped using lndir: https://github.com/NixOS/nixpkgs/issues/271792
-    cuda_cudart.dev
-    cuda_cudart.lib
-    cuda_cudart.static
-    libcublas.dev
-    libcublas.lib
-    libcublas.static
+    cuda_cudart
+    cuda_cccl # <nv/target>
+    libcublas
   ];
 
   rocmBuildInputs = with rocmPackages; [
@@ -199,10 +193,7 @@ effectiveStdenv.mkDerivation (
       ]
       ++ optionals useCuda [
         cudaPackages.cuda_nvcc
-
-        # TODO: Replace with autoAddDriverRunpath
-        # once https://github.com/NixOS/nixpkgs/pull/275241 has been merged
-        cudaPackages.autoAddOpenGLRunpathHook
+        autoAddDriverRunpath
       ]
       ++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [
         glibc.static

.ecrc

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

.github/workflows/bench.yml.disabled

@@ -1,3 +1,6 @@
+# TODO: there have been some issues with the workflow, so disabling for now
+#       https://github.com/ggerganov/llama.cpp/issues/7893
+#
 # Benchmark
 name: Benchmark
 
@@ -129,6 +132,8 @@ jobs:
 
       - name: Server bench
         id: server_bench
+        env:
+            HEAD_REF: ${{ github.head_ref || github.ref_name }}
         run: |
           set -eux
 
@@ -137,7 +142,7 @@ jobs:
           python bench.py \
               --runner-label ${{ env.RUNNER_LABEL }} \
               --name ${{ github.job }} \
-              --branch ${{ github.head_ref || github.ref_name }} \
+              --branch $HEAD_REF \
               --commit ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha }} \
               --scenario script.js \
               --duration ${{ github.event.inputs.duration || env.DURATION }} \

.github/workflows/build.yml

@@ -47,7 +47,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF ..
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -105,7 +105,7 @@ jobs:
           sysctl -a
           # Metal is disabled due to intermittent failures with Github runners not having a GPU:
           # https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
-          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
+          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -222,7 +222,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -696,22 +696,20 @@ jobs:
     strategy:
       matrix:
         include:
-          - build: 'rpc-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'noavx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx2-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'avx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx512-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'openblas-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'kompute-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'llvm-arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'msvc-arm64'
@@ -860,7 +858,8 @@ jobs:
           mkdir build
           cd build
           cmake .. -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON
-          cmake --build . --config Release -j $((${env:NUMBER_OF_PROCESSORS} - 1))
+          cmake --build . --config Release -j $((${env:NUMBER_OF_PROCESSORS} - 1)) -t ggml
+          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: Determine tag name
         id: tag

.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 }}"
-          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 }}"
+          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 }}

.github/workflows/python-check-requirements.yml

@@ -6,15 +6,13 @@ on:
       - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
-      - 'requirements.txt'
-      - 'requirements/*.txt'
+      - '**/requirements*.txt'
   pull_request:
     paths:
       - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
-      - 'requirements.txt'
-      - 'requirements/*.txt'
+      - '**/requirements*.txt'
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}

.gitignore

@@ -50,6 +50,7 @@ build*
 !docs/build.md
 /libllama.so
 /llama-*
+/vulkan-shaders-gen
 android-ndk-*
 arm_neon.h
 cmake-build-*
@@ -78,7 +79,6 @@ models-mnt
 !models/ggml-vocab-*.gguf*
 
 # Zig
-
 zig-out/
 zig-cache/
 
@@ -129,3 +129,6 @@ poetry.toml
 
 # Scripts
 !/scripts/install-oneapi.bat
+
+# Test models for lora adapters
+/lora-tests

CMakeLists.txt

@@ -139,7 +139,8 @@ set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location o
 # determining _precisely_ which defines are necessary for the llama-config
 # package.
 #
-get_directory_property(GGML_DIR_DEFINES DIRECTORY ggml/src COMPILE_DEFINITIONS)
+get_target_property(GGML_DIRECTORY ggml SOURCE_DIR)
+get_directory_property(GGML_DIR_DEFINES DIRECTORY ${GGML_DIRECTORY} COMPILE_DEFINITIONS)
 get_target_property(GGML_TARGET_DEFINES ggml COMPILE_DEFINITIONS)
 set(GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES} ${GGML_DIR_DEFINES})
 get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES)

CMakePresets.json

@@ -28,6 +28,7 @@
     { "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
     { "name": "reldbg",  "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
     { "name": "static",  "hidden": true, "cacheVariables": { "GGML_STATIC": "ON" } },
+    { "name": "sycl_f16",  "hidden": true, "cacheVariables": { "GGML_SYCL_F16": "ON" } },
 
     {
         "name": "arm64-windows-msvc", "hidden": true,
@@ -60,6 +61,8 @@
     { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] },
 
     { "name": "x64-windows-sycl-debug"  , "inherits": [ "sycl-base", "debug"   ] },
-    { "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] }
+    { "name": "x64-windows-sycl-debug-f16", "inherits": [ "sycl-base", "debug", "sycl_f16" ] },
+    { "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] },
+    { "name": "x64-windows-sycl-release-f16", "inherits": [ "sycl-base", "release", "sycl_f16" ] }
   ]
 }

CONTRIBUTING.md

@@ -5,6 +5,7 @@
   - Execute [the full CI locally on your machine](ci/README.md) before publishing
 - Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs.
   - The PR template has a series of review complexity checkboxes `[ ]` that [you can mark as](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) `[X]` for your convenience
+- Consider allowing write access to your branch for faster review
 - If your PR becomes stale, don't hesitate to ping the maintainers in the comments
 
 # Pull requests (for collaborators)

Makefile

@@ -19,6 +19,7 @@ BUILD_TARGETS = \
 	llama-imatrix \
 	llama-infill \
 	llama-llava-cli \
+	llama-minicpmv-cli\
 	llama-lookahead \
 	llama-lookup \
 	llama-lookup-create \
@@ -325,9 +326,9 @@ ifdef LLAMA_DEBUG
 	endif
 else
 	MK_CPPFLAGS   += -DNDEBUG
-	MK_CFLAGS     += -O3
-	MK_CXXFLAGS   += -O3
-	MK_NVCCFLAGS  += -O3
+	MK_CFLAGS     += -O3 -g
+	MK_CXXFLAGS   += -O3 -g
+	MK_NVCCFLAGS  += -O3 -g
 endif
 
 ifdef LLAMA_SANITIZE_THREAD
@@ -528,10 +529,21 @@ ifndef GGML_NO_ACCELERATE
 	endif
 endif # GGML_NO_ACCELERATE
 
+ifdef GGML_MUSA
+	CC := clang
+	CXX := clang++
+	GGML_CUDA := 1
+	MK_CPPFLAGS += -DGGML_USE_MUSA
+endif
+
 ifndef GGML_NO_OPENMP
 	MK_CPPFLAGS += -DGGML_USE_OPENMP
 	MK_CFLAGS   += -fopenmp
 	MK_CXXFLAGS += -fopenmp
+	ifdef GGML_MUSA
+		MK_CPPFLAGS += -I/usr/lib/llvm-10/include/openmp
+		MK_LDFLAGS  += -L/usr/lib/llvm-10/lib
+	endif # GGML_MUSA
 endif # GGML_NO_OPENMP
 
 ifdef GGML_OPENBLAS
@@ -582,15 +594,27 @@ else
 endif # GGML_CUDA_FA_ALL_QUANTS
 
 ifdef GGML_CUDA
-	ifneq ('', '$(wildcard /opt/cuda)')
-		CUDA_PATH ?= /opt/cuda
-	else
-		CUDA_PATH ?= /usr/local/cuda
-	endif
+	ifdef GGML_MUSA
+		ifneq ('', '$(wildcard /opt/musa)')
+			CUDA_PATH ?= /opt/musa
+		else
+			CUDA_PATH ?= /usr/local/musa
+		endif
 
-	MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
-	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
-	MK_NVCCFLAGS += -use_fast_math
+		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include
+		MK_LDFLAGS   += -lmusa -lmublas -lmusart -lpthread -ldl -lrt -L$(CUDA_PATH)/lib -L/usr/lib64
+		MK_NVCCFLAGS += -x musa -mtgpu --cuda-gpu-arch=mp_22
+	else
+		ifneq ('', '$(wildcard /opt/cuda)')
+			CUDA_PATH ?= /opt/cuda
+		else
+			CUDA_PATH ?= /usr/local/cuda
+		endif
+
+		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
+		MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
+		MK_NVCCFLAGS += -use_fast_math
+	endif # GGML_MUSA
 
 	OBJ_GGML += ggml/src/ggml-cuda.o
 	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
@@ -600,9 +624,11 @@ ifdef LLAMA_FATAL_WARNINGS
 	MK_NVCCFLAGS += -Werror all-warnings
 endif # LLAMA_FATAL_WARNINGS
 
+ifndef GGML_MUSA
 ifndef JETSON_EOL_MODULE_DETECT
 	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
 endif # JETSON_EOL_MODULE_DETECT
+endif # GGML_MUSA
 
 ifdef LLAMA_DEBUG
 	MK_NVCCFLAGS += -lineinfo
@@ -615,8 +641,12 @@ endif # GGML_CUDA_DEBUG
 ifdef GGML_CUDA_NVCC
 	NVCC = $(CCACHE) $(GGML_CUDA_NVCC)
 else
-	NVCC = $(CCACHE) nvcc
-endif #GGML_CUDA_NVCC
+	ifdef GGML_MUSA
+		NVCC = $(CCACHE) mcc
+	else
+		NVCC = $(CCACHE) nvcc
+	endif # GGML_MUSA
+endif # GGML_CUDA_NVCC
 
 ifdef CUDA_DOCKER_ARCH
 	MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
@@ -687,9 +717,15 @@ define NVCC_COMPILE
 	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endef # NVCC_COMPILE
 else
+	ifdef GGML_MUSA
+define NVCC_COMPILE
+	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -c $< -o $@
+endef # NVCC_COMPILE
+	else
 define NVCC_COMPILE
 	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endef # NVCC_COMPILE
+	endif # GGML_MUSA
 endif # JETSON_EOL_MODULE_DETECT
 
 ggml/src/ggml-cuda/%.o: \
@@ -727,6 +763,10 @@ ifdef GGML_VULKAN_MEMORY_DEBUG
 	MK_CPPFLAGS  += -DGGML_VULKAN_MEMORY_DEBUG
 endif
 
+ifdef GGML_VULKAN_PERF
+	MK_CPPFLAGS  += -DGGML_VULKAN_PERF
+endif
+
 ifdef GGML_VULKAN_VALIDATE
 	MK_CPPFLAGS  += -DGGML_VULKAN_VALIDATE
 endif
@@ -853,15 +893,16 @@ ggml/src/ggml-metal-embed.o: \
 	ggml/src/ggml-common.h
 	@echo "Embedding Metal library"
 	@sed -e '/#include "ggml-common.h"/r ggml/src/ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml/src/ggml-metal.metal > ggml/src/ggml-metal-embed.metal
-	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)
-	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f ${TEMP_ASSEMBLY}
+	$(eval TEMP_ASSEMBLY=$(shell mktemp -d))
+	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	$(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@
+	@rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s
+	@rmdir ${TEMP_ASSEMBLY}
 endif
 endif # GGML_METAL
 
@@ -944,6 +985,7 @@ $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
 ifdef GGML_CUDA
 $(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
 CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+ifndef GGML_MUSA
 ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
 
 ifndef CUDA_DOCKER_ARCH
@@ -953,6 +995,7 @@ endif # CUDA_POWER_ARCH
 endif # CUDA_DOCKER_ARCH
 
 endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
+endif # GGML_MUSA
 endif # GGML_CUDA
 $(info )
 
@@ -1168,6 +1211,7 @@ clean:
 	rm -rvf ggml/*.dll
 	rm -rvf ggml/*.so
 	rm -vrf ggml/src/*.o
+	rm -rvf ggml/src/llamafile/*.o
 	rm -rvf common/build-info.cpp
 	rm -vrf ggml/src/ggml-metal-embed.metal
 	rm -vrf ggml/src/ggml-cuda/*.o
@@ -1414,15 +1458,20 @@ libllava.a: examples/llava/llava.cpp \
 	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual
 
 llama-llava-cli: examples/llava/llava-cli.cpp \
-	examples/llava/clip.h \
-	examples/llava/clip.cpp \
-	examples/llava/llava.h \
 	examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
 	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp  -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
-	$(CXX) $(CXXFLAGS) -c examples/llava/llava.cpp -o $(call GET_OBJ_FILE, examples/llava/llava.cpp)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $< examples/llava/clip.cpp examples/llava/llava.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) $(call GET_OBJ_FILE, examples/llava/llava.cpp) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
+
+llama-minicpmv-cli: examples/llava/minicpmv-cli.cpp \
+	examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
 
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift
@@ -1568,42 +1617,41 @@ llama-q8dot: pocs/vdot/q8dot.cpp ggml/src/ggml.o \
 # Mark legacy binary targets as .PHONY so that they are always checked.
 .PHONY: main quantize perplexity embedding server
 
+# Define the object file target
+examples/deprecation-warning/deprecation-warning.o: examples/deprecation-warning/deprecation-warning.cpp
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 # NOTE: We currently will always build the deprecation-warning `main` and `server` binaries to help users migrate.
 #  Eventually we will want to remove these target from building all the time.
-main: examples/deprecation-warning/deprecation-warning.cpp
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+main: examples/deprecation-warning/deprecation-warning.o
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
 	@echo "NOTICE: The 'main' binary is deprecated. Please use 'llama-cli' instead."
 
-server: examples/deprecation-warning/deprecation-warning.cpp
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+server: examples/deprecation-warning/deprecation-warning.o
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
 	@echo "NOTICE: The 'server' binary is deprecated. Please use 'llama-server' instead."
 
-quantize: examples/deprecation-warning/deprecation-warning.cpp
+quantize: examples/deprecation-warning/deprecation-warning.o
 ifneq (,$(wildcard quantize))
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
 	@echo "#########"
 	@echo "WARNING: The 'quantize' binary is deprecated. Please use 'llama-quantize' instead."
 	@echo "  Remove the 'quantize' binary to remove this warning."
 	@echo "#########"
 endif
 
-perplexity: examples/deprecation-warning/deprecation-warning.cpp
+perplexity: examples/deprecation-warning/deprecation-warning.o
 ifneq (,$(wildcard perplexity))
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
 	@echo "#########"
 	@echo "WARNING: The 'perplexity' binary is deprecated. Please use 'llama-perplexity' instead."
 	@echo "  Remove the 'perplexity' binary to remove this warning."
 	@echo "#########"
 endif
 
-embedding: examples/deprecation-warning/deprecation-warning.cpp
+embedding: examples/deprecation-warning/deprecation-warning.o
 ifneq (,$(wildcard embedding))
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
 	@echo "#########"
 	@echo "WARNING: The 'embedding' binary is deprecated. Please use 'llama-embedding' instead."
 	@echo "  Remove the 'embedding' binary to remove this warning."

README.md

@@ -95,8 +95,18 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
+- [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330)
 - [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
+- [x] [Snowflake-Arctic MoE](https://huggingface.co/collections/Snowflake/arctic-66290090abe542894a5ac520)
+- [x] [Smaug](https://huggingface.co/models?search=Smaug)
+- [x] [Poro 34B](https://huggingface.co/LumiOpen/Poro-34B)
+- [x] [Bitnet b1.58 models](https://huggingface.co/1bitLLM)
+- [x] [Flan T5](https://huggingface.co/models?search=flan-t5)
+- [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca)
 - [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b)
+- [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
+- [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
+- [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
 
@@ -145,6 +155,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [Faraday](https://faraday.dev/) (proprietary)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
 - [Layla](https://play.google.com/store/apps/details?id=com.laylalite) (proprietary)
+- [ramalama](https://github.com/containers/ramalama) (MIT)
 - [LocalAI](https://github.com/mudler/LocalAI) (MIT)
 - [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
 - [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)
@@ -177,10 +188,12 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 
 - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
 - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
+- [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
 
 **Infrastructure:**
 
 - [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
+- [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
 
 **Games:**
 - [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you.
@@ -409,9 +422,11 @@ Please refer to [Build llama.cpp locally](./docs/build.md)
 | [BLAS](./docs/build.md#blas-build) | All |
 | [BLIS](./docs/backend/BLIS.md) | All |
 | [SYCL](./docs/backend/SYCL.md) | Intel and Nvidia GPU |
+| [MUSA](./docs/build.md#musa) | Moore Threads GPU |
 | [CUDA](./docs/build.md#cuda) | Nvidia GPU |
 | [hipBLAS](./docs/build.md#hipblas) | AMD GPU |
 | [Vulkan](./docs/build.md#vulkan) | GPU |
+| [CANN](./docs/build.md#cann) | Ascend NPU |
 
 ## Tools
 

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 make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.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 make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.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

@@ -77,6 +77,41 @@
 
 using json = nlohmann::ordered_json;
 
+//
+// Environment variable utils
+//
+
+template<typename T>
+static typename std::enable_if<std::is_same<T, std::string>::value, void>::type
+get_env(std::string name, T & target) {
+    char * value = std::getenv(name.c_str());
+    target = value ? std::string(value) : target;
+}
+
+template<typename T>
+static typename std::enable_if<!std::is_same<T, bool>::value && std::is_integral<T>::value, void>::type
+get_env(std::string name, T & target) {
+    char * value = std::getenv(name.c_str());
+    target = value ? std::stoi(value) : target;
+}
+
+template<typename T>
+static typename std::enable_if<std::is_floating_point<T>::value, void>::type
+get_env(std::string name, T & target) {
+    char * value = std::getenv(name.c_str());
+    target = value ? std::stof(value) : target;
+}
+
+template<typename T>
+static typename std::enable_if<std::is_same<T, bool>::value, void>::type
+get_env(std::string name, T & target) {
+    char * value = std::getenv(name.c_str());
+    if (value) {
+        std::string val(value);
+        target = val == "1" || val == "true";
+    }
+}
+
 //
 // CPU utils
 //
@@ -110,8 +145,34 @@ int32_t cpu_get_num_physical_cores() {
     if (result == 0) {
         return num_physical_cores;
     }
-#elif defined(_WIN32)
-    //TODO: Implement
+#elif defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later
+    // TODO: windows + arm64 + mingw64
+    unsigned int n_threads_win = std::thread::hardware_concurrency();
+    unsigned int default_threads = n_threads_win > 0 ? (n_threads_win <= 4 ? n_threads_win : n_threads_win / 2) : 4;
+
+    DWORD buffer_size = 0;
+    if (!GetLogicalProcessorInformationEx(RelationProcessorCore, nullptr, &buffer_size)) {
+        if (GetLastError() != ERROR_INSUFFICIENT_BUFFER) {
+            return default_threads;
+        }
+    }
+
+    std::vector<char> buffer(buffer_size);
+    if (!GetLogicalProcessorInformationEx(RelationProcessorCore, reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(buffer.data()), &buffer_size)) {
+        return default_threads;
+    }
+
+    int32_t num_physical_cores = 0;
+    PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX info = reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(buffer.data());
+    while (buffer_size > 0) {
+        if (info->Relationship == RelationProcessorCore) {
+            num_physical_cores += info->Processor.GroupCount;
+        }
+        buffer_size -= info->Size;
+        info = reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(reinterpret_cast<char*>(info) + info->Size);
+    }
+
+    return num_physical_cores > 0 ? num_physical_cores : default_threads;
 #endif
     unsigned int n_threads = std::thread::hardware_concurrency();
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
@@ -190,16 +251,61 @@ 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
 //
 
-void gpt_params_handle_hf_token(gpt_params & params) {
-    if (params.hf_token.empty() && std::getenv("HF_TOKEN")) {
-        params.hf_token = std::getenv("HF_TOKEN");
-    }
-}
-
 void gpt_params_handle_model_default(gpt_params & params) {
     if (!params.hf_repo.empty()) {
         // short-hand to avoid specifying --hf-file -> default it to --model
@@ -222,6 +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;
@@ -241,13 +371,20 @@ 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");
     }
 
     gpt_params_handle_model_default(params);
 
-    gpt_params_handle_hf_token(params);
+    if (params.hf_token.empty()) {
+        get_env("HF_TOKEN", params.hf_token);
+    }
 
     if (params.escape) {
         string_process_escapes(params.prompt);
@@ -267,6 +404,32 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
     return true;
 }
 
+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_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);
+    get_env("LLAMA_ARG_UBATCH",           params.n_ubatch);
+    get_env("LLAMA_ARG_N_GPU_LAYERS",     params.n_gpu_layers);
+    get_env("LLAMA_ARG_THREADS_HTTP",     params.n_threads_http);
+    get_env("LLAMA_ARG_CHAT_TEMPLATE",    params.chat_template);
+    get_env("LLAMA_ARG_N_PREDICT",        params.n_predict);
+    get_env("LLAMA_ARG_ENDPOINT_METRICS", params.endpoint_metrics);
+    get_env("LLAMA_ARG_ENDPOINT_SLOTS",   params.endpoint_slots);
+    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) {
     const auto params_org = params; // the example can modify the default params
 
@@ -285,6 +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) {
@@ -301,36 +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]);
-        if (params.n_threads <= 0) {
-            params.n_threads = std::thread::hardware_concurrency();
+        params.cpuparams.n_threads = std::stoi(argv[i]);
+        if (params.cpuparams.n_threads <= 0) {
+            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]);
-        if (params.n_threads_batch <= 0) {
-            params.n_threads_batch = std::thread::hardware_concurrency();
+        params.cpuparams_batch.n_threads = std::stoi(argv[i]);
+        if (params.cpuparams_batch.n_threads <= 0) {
+            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]);
-        if (params.n_threads_draft <= 0) {
-            params.n_threads_draft = std::thread::hardware_concurrency();
+        params.draft_cpuparams.n_threads = std::stoi(argv[i]);
+        if (params.draft_cpuparams.n_threads <= 0) {
+            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]);
-        if (params.n_threads_batch_draft <= 0) {
-            params.n_threads_batch_draft = std::thread::hardware_concurrency();
+        params.draft_cpuparams_batch.n_threads = std::stoi(argv[i]);
+        if (params.draft_cpuparams_batch.n_threads <= 0) {
+            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];
@@ -684,14 +1026,24 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
     }
     if (arg == "--lora") {
         CHECK_ARG
-        params.lora_adapter.emplace_back(argv[i], 1.0f);
+        params.lora_adapters.push_back({
+            std::string(argv[i]),
+            1.0,
+        });
         return true;
     }
     if (arg == "--lora-scaled") {
         CHECK_ARG
-        const char* lora_adapter = argv[i];
+        std::string lora_adapter = argv[i];
         CHECK_ARG
-        params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i]));
+        params.lora_adapters.push_back({
+            lora_adapter,
+            std::stof(argv[i]),
+        });
+        return true;
+    }
+    if (arg == "--lora-init-without-apply") {
+        params.lora_init_without_apply = true;
         return true;
     }
     if (arg == "--control-vector") {
@@ -815,7 +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()) {
@@ -1324,6 +1676,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         else { invalid_param = true; }
         return true;
     }
+    if (arg == "--no-warmup") {
+        params.warmup = false;
+        return true;
+    }
 #ifndef LOG_DISABLE_LOGS
     // Parse args for logging parameters
     if (log_param_single_parse(argv[i])) {
@@ -1401,11 +1757,40 @@ 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",
-                                                                        "number of threads to use during batch and prompt processing (default: same as --threads-draft)" });
+    options.push_back({ "speculative", "-tbd,  --threads-batch-draft N","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",
@@ -1446,6 +1831,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     options.push_back({ "main infill", "       --in-prefix-bos",        "prefix BOS to user inputs, preceding the `--in-prefix` string" });
     options.push_back({ "main infill", "       --in-prefix STRING",     "string to prefix user inputs with (default: empty)" });
     options.push_back({ "main infill", "       --in-suffix STRING",     "string to suffix after user inputs with (default: empty)" });
+    options.push_back({ "main",        "       --no-warmup",            "skip warming up the model with an empty run" });
     options.push_back({ "server infill",
                                        "       --spm-infill",           "use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: %s)", params.spm_infill ? "enabled" : "disabled" });
 
@@ -1629,7 +2015,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     options.push_back({ "server",      "       --host HOST",            "ip address to listen (default: %s)", params.hostname.c_str() });
     options.push_back({ "server",      "       --port PORT",            "port to listen (default: %d)", params.port });
     options.push_back({ "server",      "       --path PATH",            "path to serve static files from (default: %s)", params.public_path.c_str() });
-    options.push_back({ "server",      "       --embedding(s)",         "enable embedding endpoint (default: %s)", params.embedding ? "enabled" : "disabled" });
+    options.push_back({ "server",      "       --embedding(s)",         "restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled" });
     options.push_back({ "server",      "       --api-key KEY",          "API key to use for authentication (default: none)" });
     options.push_back({ "server",      "       --api-key-file FNAME",   "path to file containing API keys (default: none)" });
     options.push_back({ "server",      "       --ssl-key-file FNAME",   "path to file a PEM-encoded SSL private key" });
@@ -1649,6 +2035,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
                                                                         "https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template" });
     options.push_back({ "server",      "-sps,  --slot-prompt-similarity SIMILARITY",
                                                                         "how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity });
+    options.push_back({ "server",      "       --lora-init-without-apply",     "load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: %s)", params.lora_init_without_apply ? "enabled" : "disabled"});
 
 #ifndef LOG_DISABLE_LOGS
     options.push_back({ "logging" });
@@ -1675,7 +2062,6 @@ 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]);
@@ -1707,11 +2093,17 @@ 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;
-    if (params.n_threads_batch != -1) {
-        os << " (n_threads_batch = " << params.n_threads_batch << ")";
+    os << "system_info: n_threads = " << params.cpuparams.n_threads;
+    if (params.cpuparams_batch.n_threads != -1) {
+        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
+    DWORD logicalProcessorCount = GetActiveProcessorCount(ALL_PROCESSOR_GROUPS);
+    os << " / " << logicalProcessorCount << " | " << llama_print_system_info();
+#else
     os << " / " << std::thread::hardware_concurrency() << " | " << llama_print_system_info();
+#endif
 
     return os.str();
 }
@@ -1761,6 +2153,23 @@ std::string string_get_sortable_timestamp() {
     return std::string(timestamp_no_ns) + "." + std::string(timestamp_ns);
 }
 
+void string_replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    if (search.empty()) {
+        return;
+    }
+    std::string builder;
+    builder.reserve(s.length());
+    size_t pos = 0;
+    size_t last_pos = 0;
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
+        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) {
     std::size_t input_len = input.length();
     std::size_t output_idx = 0;
@@ -2034,8 +2443,8 @@ std::string fs_get_cache_file(const std::string & filename) {
 //
 // Model utils
 //
-
-std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params) {
+struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
+    llama_init_result iparams;
     auto mparams = llama_model_params_from_gpt_params(params);
 
     llama_model * model = nullptr;
@@ -2050,7 +2459,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 
     if (model == NULL) {
         fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
-        return std::make_tuple(nullptr, nullptr);
+        return iparams;
     }
 
     auto cparams = llama_context_params_from_gpt_params(params);
@@ -2059,7 +2468,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
     if (lctx == NULL) {
         fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
         llama_free_model(model);
-        return std::make_tuple(nullptr, nullptr);
+        return iparams;
     }
 
     if (!params.control_vectors.empty()) {
@@ -2070,7 +2479,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         if (cvec.n_embd == -1) {
             llama_free(lctx);
             llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
+            return iparams;
         }
 
         int err = llama_control_vector_apply(lctx,
@@ -2082,21 +2491,26 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         if (err) {
             llama_free(lctx);
             llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
+            return iparams;
         }
     }
 
-    for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
-        const std::string & lora_adapter = std::get<0>(params.lora_adapter[i]);
-        float lora_scale = std::get<1>(params.lora_adapter[i]);
-        auto adapter = llama_lora_adapter_init(model, lora_adapter.c_str());
-        if (adapter == nullptr) {
-            fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
+    // load and optionally apply lora adapters
+    for (auto & la : params.lora_adapters) {
+        llama_lora_adapter_container loaded_la;
+        loaded_la.path = la.path;
+        loaded_la.scale = la.scale;
+        loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
+        if (loaded_la.adapter == nullptr) {
+            fprintf(stderr, "%s: error: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
             llama_free(lctx);
             llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
+            return iparams;
         }
-        llama_lora_adapter_set(lctx, adapter, lora_scale);
+        iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
+    }
+    if (!params.lora_init_without_apply) {
+        llama_lora_adapters_apply(lctx, iparams.lora_adapters);
     }
 
     if (params.ignore_eos) {
@@ -2124,13 +2538,26 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
             tmp.clear();
             tmp.push_back(decoder_start_token_id);
         }
-        llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
+        if (llama_model_has_decoder(model)) {
+            llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
+        }
         llama_kv_cache_clear(lctx);
         llama_synchronize(lctx);
         llama_reset_timings(lctx);
     }
 
-    return std::make_tuple(model, lctx);
+    iparams.model   = model;
+    iparams.context = lctx;
+    return iparams;
+}
+
+void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters) {
+    llama_lora_adapter_clear(ctx);
+    for (auto & la : lora_adapters) {
+        if (la.scale != 0.0f) {
+            llama_lora_adapter_set(ctx, la.adapter, la.scale);
+        }
+    }
 }
 
 struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params) {
@@ -2192,8 +2619,9 @@ 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_batch   = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    cparams.n_threads         = params.cpuparams.n_threads;
+    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;
@@ -2219,6 +2647,22 @@ 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) {
@@ -2657,12 +3101,6 @@ std::string llama_detokenize(llama_context * ctx, const std::vector<llama_token>
     return text;
 }
 
-bool llama_should_add_bos_token(const llama_model * model) {
-    const int add_bos = llama_add_bos_token(model);
-
-    return add_bos != -1 ? bool(add_bos) : (llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM);
-}
-
 //
 // Chat template utils
 //
@@ -3155,19 +3593,18 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
     }
 
     fprintf(stream, "lora:\n");
-    for (std::tuple<std::string, float> la : params.lora_adapter) {
-        if (std::get<1>(la) != 1.0f) {
-            continue;
+    for (auto & la : params.lora_adapters) {
+        if (la.scale == 1.0f) {
+            fprintf(stream, "  - %s\n", la.path.c_str());
         }
-        fprintf(stream, "  - %s\n", std::get<0>(la).c_str());
     }
     fprintf(stream, "lora_scaled:\n");
-    for (std::tuple<std::string, float> la : params.lora_adapter) {
-        if (std::get<1>(la) == 1.0f) {
-            continue;
+    for (auto & la : params.lora_adapters) {
+        if (la.scale != 1.0f) {
+            fprintf(stream, "  - %s: %f\n", la.path.c_str(), la.scale);
         }
-        fprintf(stream, "  - %s: %f\n", std::get<0>(la).c_str(), std::get<1>(la));
     }
+    fprintf(stream, "lora_init_without_apply: %s # default: false\n", params.lora_init_without_apply ? "true" : "false");
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
     fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep);
     fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
@@ -3215,7 +3652,7 @@ 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

@@ -33,6 +33,15 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
+struct llama_lora_adapter_info {
+    std::string path;
+    float scale;
+};
+
+struct llama_lora_adapter_container : llama_lora_adapter_info {
+    struct llama_lora_adapter * adapter;
+};
+
 // build info
 extern int LLAMA_BUILD_NUMBER;
 extern char const * LLAMA_COMMIT;
@@ -58,13 +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)
@@ -91,6 +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;
 
@@ -126,8 +145,8 @@ struct gpt_params {
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
 
-    // TODO: avoid tuple, use struct
-    std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
+    bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
+    std::vector<llama_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
 
     std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
 
@@ -195,7 +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   = "";
@@ -258,7 +277,7 @@ struct gpt_params {
     std::string lora_outfile = "ggml-lora-merged-f16.gguf";
 };
 
-void gpt_params_handle_hf_token(gpt_params & params);
+void gpt_params_parse_from_env(gpt_params & params);
 void gpt_params_handle_model_default(gpt_params & params);
 
 bool gpt_params_parse_ex   (int argc, char ** argv, gpt_params & params);
@@ -268,6 +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
 //
@@ -277,6 +301,8 @@ std::vector<std::string> string_split(std::string input, char separator);
 std::string string_strip(const std::string & str);
 std::string string_get_sortable_timestamp();
 
+void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
+
 template<class T>
 static std::vector<T> string_split(const std::string & str, char delim) {
     std::vector<T> values;
@@ -308,15 +334,24 @@ std::string fs_get_cache_file(const std::string & filename);
 // Model utils
 //
 
-// TODO: avoid tuplue, use struct
-std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params);
+struct llama_init_result {
+    struct llama_model   * model   = nullptr;
+    struct llama_context * context = nullptr;
+    std::vector<llama_lora_adapter_container> lora_adapters;
+};
 
-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 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);
 
+// clear LoRA adapters from context, then apply new list of adapters
+void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters);
+
 // Batch utils
 
 void llama_batch_clear(struct llama_batch & batch);
@@ -361,10 +396,6 @@ std::string llama_detokenize(
         const std::vector<llama_token> & tokens,
                                   bool   special = true);
 
-// Uses the value from the model metadata if possible, otherwise
-// defaults to true when model type is SPM, otherwise false.
-bool llama_should_add_bos_token(const llama_model * model);
-
 //
 // Chat template utils
 //

common/grammar-parser.cpp

@@ -369,6 +369,9 @@ namespace grammar_parser {
             }
             // Validate the state to ensure that all rules are defined
             for (const auto & rule : state.rules) {
+                if (rule.empty()) {
+                    throw std::runtime_error("Undefined rule");
+                }
                 for (const auto & elem : rule) {
                     if (elem.type == LLAMA_GRETYPE_RULE_REF) {
                         // Ensure that the rule at that location exists

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:
@@ -251,12 +254,7 @@ class Model:
 
         return [(self.map_tensor_name(name), data_torch)]
 
-    def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
-        del name, new_name, bid, n_dims  # unused
-
-        return False
-
-    def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
+    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
         del name, new_name, bid, n_dims  # unused
 
         return False
@@ -285,55 +283,51 @@ class Model:
             for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)):
                 data: np.ndarray  # type hint
                 n_dims = len(data.shape)
-                data_dtype = data.dtype
-                data_qtype: gguf.GGMLQuantizationType | None = None
-
-                # when both are True, f32 should win
-                extra_f32 = self.extra_f32_tensors(name, new_name, bid, n_dims)
-                extra_f16 = self.extra_f16_tensors(name, new_name, bid, n_dims)
+                data_qtype: gguf.GGMLQuantizationType | bool = self.tensor_force_quant(name, new_name, bid, n_dims)
 
                 # Most of the codebase that takes in 1D tensors or norms only handles F32 tensors
-                # Conditions should closely match those in llama_model_quantize_internal in llama.cpp
-                extra_f32 = any(cond for cond in (
-                    extra_f32,
-                    n_dims == 1,
-                    new_name.endswith("_norm.weight"),
-                ))
-
-                # Some tensor types are always in float32
-                extra_f32 = extra_f32 or any(self.match_model_tensor_name(new_name, key, bid) for key in (
-                    gguf.MODEL_TENSOR.FFN_GATE_INP,
-                    gguf.MODEL_TENSOR.POS_EMBD,
-                    gguf.MODEL_TENSOR.TOKEN_TYPES,
-                ))
-
-                # if f16 desired, convert any float32 2-dim weight tensors to float16
-                extra_f16 = any(cond for cond in (
-                    extra_f16,
-                    (name.endswith(".weight") and n_dims >= 2),
-                ))
-
-                if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
-                    if self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
-                        data = gguf.quantize_bf16(data)
-                        assert data.dtype == np.int16
-                        data_qtype = gguf.GGMLQuantizationType.BF16
-
-                    elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0 and gguf.can_quantize_to_q8_0(data):
-                        data = gguf.quantize_q8_0(data)
-                        assert data.dtype == np.uint8
-                        data_qtype = gguf.GGMLQuantizationType.Q8_0
-
-                    else:  # default to float16 for quantized tensors
-                        if data_dtype != np.float16:
-                            data = data.astype(np.float16)
-                        data_qtype = gguf.GGMLQuantizationType.F16
-
-                if data_qtype is None:  # by default, convert to float32
-                    if data_dtype != np.float32:
-                        data = data.astype(np.float32)
+                if n_dims <= 1 or new_name.endswith("_norm.weight"):
                     data_qtype = gguf.GGMLQuantizationType.F32
 
+                # Conditions should closely match those in llama_model_quantize_internal in llama.cpp
+                # Some tensor types are always in float32
+                if data_qtype is False and (
+                    any(
+                        self.match_model_tensor_name(new_name, key, bid)
+                        for key in (
+                            gguf.MODEL_TENSOR.FFN_GATE_INP,
+                            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 new_name.endswith(".weight")
+                ):
+                    data_qtype = gguf.GGMLQuantizationType.F32
+
+                # No override (data_qtype is False), or wants to be quantized (data_qtype is True)
+                if isinstance(data_qtype, bool):
+                    if self.ftype == gguf.LlamaFileType.ALL_F32:
+                        data_qtype = gguf.GGMLQuantizationType.F32
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_F16:
+                        data_qtype = gguf.GGMLQuantizationType.F16
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
+                        data_qtype = gguf.GGMLQuantizationType.BF16
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0:
+                        data_qtype = gguf.GGMLQuantizationType.Q8_0
+                    else:
+                        raise ValueError(f"Unknown file type: {self.ftype.name}")
+
+                try:
+                    data = gguf.quants.quantize(data, data_qtype)
+                except gguf.QuantError as e:
+                    logger.warning("%s, %s", e, "falling back to F16")
+                    data_qtype = gguf.GGMLQuantizationType.F16
+                    data = gguf.quants.quantize(data, data_qtype)
+
                 shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape
 
                 # reverse shape to make it similar to the internal ggml dimension order
@@ -603,6 +597,15 @@ class Model:
         if chkhsh == "855059429035d75a914d1eda9f10a876752e281a054a7a3d421ef0533e5b6249":
             # ref: https://huggingface.co/HuggingFaceTB/SmolLM-135M
             res = "smollm"
+        if chkhsh == "3c30d3ad1d6b64202cd222813e7736c2db6e1bd6d67197090fc1211fbc612ae7":
+            # ref: https://huggingface.co/bigscience/bloom
+            res = "bloom"
+        if chkhsh == "bc01ce58980e1db43859146dc51b1758b3b88729b217a74792e9f8d43e479d21":
+            # ref: https://huggingface.co/TurkuNLP/gpt3-finnish-small
+            res = "gpt3-finnish"
+        if chkhsh == "4e2b24cc4770243d65a2c9ec19770a72f08cffc161adbb73fcbb6b7dd45a0aae":
+            # ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct
+            res = "exaone"
 
         if res is None:
             logger.warning("\n")
@@ -906,7 +909,7 @@ class GPTNeoXModel(Model):
         return tensors
 
 
-@Model.register("BloomForCausalLM")
+@Model.register("BloomForCausalLM", "BloomModel")
 class BloomModel(Model):
     model_arch = gguf.MODEL_ARCH.BLOOM
 
@@ -1570,6 +1573,35 @@ class LlamaModel(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
     def prepare_tensors(self):
+        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.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)
+                low_freq_factor = rope_scaling.get("low_freq_factor", 1.0)
+                high_freq_factor = rope_scaling.get("high_freq_factor", 4.0)
+                old_context_len = self.hparams.get("original_max_position_embeddings", 8192)
+
+                low_freq_wavelen = old_context_len / low_freq_factor
+                high_freq_wavelen = old_context_len / high_freq_factor
+                assert low_freq_wavelen != high_freq_wavelen
+
+                rope_factors = []
+                for freq in freqs:
+                    wavelen = 2 * math.pi / freq
+                    if wavelen < high_freq_wavelen:
+                        rope_factors.append(1)
+                    elif wavelen > low_freq_wavelen:
+                        rope_factors.append(factor)
+                    else:
+                        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()
 
         if self._experts is not None:
@@ -1737,7 +1769,7 @@ class DbrxModel(Model):
 
         return [(new_name, data_torch)]
 
-    def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
+    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
         del name, new_name, bid  # unused
 
         return n_dims > 1
@@ -2115,8 +2147,9 @@ 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}')
 
-        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))
+        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))
 
 
 @Model.register("PlamoForCausalLM")
@@ -2478,6 +2511,112 @@ class NomicBertModel(BertModel):
         self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
 
 
+@Model.register("XLMRobertaModel")
+class XLMRobertaModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # we need the pad_token_id to know how to chop down position_embd matrix
+        if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
+            self._position_offset = 1 + pad_token_id
+            if "max_position_embeddings" in self.hparams:
+                self.hparams["max_position_embeddings"] -= self._position_offset
+        else:
+            self._position_offset = None
+
+    def set_vocab(self):
+        # to avoid TypeError: Descriptors cannot be created directly
+        # exception when importing sentencepiece_model_pb2
+        os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+        from sentencepiece import SentencePieceProcessor
+        from sentencepiece import sentencepiece_model_pb2 as model
+
+        tokenizer_path = self.dir_model / 'sentencepiece.bpe.model'
+        if not tokenizer_path.is_file():
+            raise FileNotFoundError(f"File not found: {tokenizer_path}")
+
+        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
+        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
+        assert sentencepiece_model.trainer_spec.model_type == 1  # UNIGRAM
+
+        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
+        remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
+        precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap
+
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(SentencePieceTokenTypes.UNUSED)
+
+        # realign tokens (see HF tokenizer code)
+        tokens = [b'<s>', b'<pad>', b'</s>', b'<unk>'] + tokens[3:-1]
+        scores = [0.0, 0.0, 0.0, 0.0] + scores[3:-1]
+        toktypes = [
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.UNKNOWN,
+        ] + toktypes[3:-1]
+
+        self.gguf_writer.add_tokenizer_model("t5")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+        self.gguf_writer.add_add_space_prefix(add_prefix)
+        self.gguf_writer.add_token_type_count(1)
+        self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
+        if precompiled_charsmap:
+            self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+        self.gguf_writer.add_add_bos_token(True)
+        self.gguf_writer.add_add_eos_token(True)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
+        if name == "embeddings.position_embeddings.weight":
+            if self._position_offset is not None:
+                data_torch = data_torch[self._position_offset:,:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @Model.register("GemmaForCausalLM")
 class GemmaModel(Model):
     model_arch = gguf.MODEL_ARCH.GEMMA
@@ -2581,7 +2720,85 @@ class StarCoder2Model(Model):
     model_arch = gguf.MODEL_ARCH.STARCODER2
 
 
-@Model.register("MambaForCausalLM", "MambaLMHeadModel")
+@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
 
@@ -2612,7 +2829,10 @@ class MambaModel(Model):
         # ref: https://github.com/state-spaces/mamba/blob/ce59daea3a090d011d6476c6e5b97f6d58ddad8b/mamba_ssm/modules/mamba_simple.py#L58
         dt_rank      = self.find_hparam(["time_step_rank",     "dt_rank"],      optional=True) or -(d_model // -16)
         rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
-
+        use_dt_b_c_norm = False
+        # For falconmamba we do apply RMS norm on B / DT and C layers
+        if self.find_hparam(["model_type"], optional=True) in ("falcon_mamba",):
+            use_dt_b_c_norm = True
         # Fail early for models which don't have a block expansion factor of 2
         assert d_inner == 2 * d_model
 
@@ -2620,12 +2840,13 @@ class MambaModel(Model):
         self.gguf_writer.add_embedding_length(d_model)
         self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading
         self.gguf_writer.add_head_count(0) # unused, but seemingly required when loading
-        self.gguf_writer.add_block_count(self.hparams["n_layer"])
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_ssm_conv_kernel(d_conv)
         self.gguf_writer.add_ssm_inner_size(d_inner)
         self.gguf_writer.add_ssm_state_size(d_state)
         self.gguf_writer.add_ssm_time_step_rank(dt_rank)
         self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
+        self.gguf_writer.add_ssm_dt_b_c_rms(use_dt_b_c_norm) # For classic Mamba we don't apply rms norm on B / DT layers
         self.gguf_writer.add_file_type(self.ftype)
 
     _tok_embd = None
@@ -2652,19 +2873,6 @@ class MambaModel(Model):
 
         return [(new_name, data_torch)]
 
-    def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
-        del n_dims  # unused
-
-        return bid is not None and new_name in (
-            self.format_tensor_name(n, bid, ".weight" if name.endswith(".weight") else "") for n in [
-                gguf.MODEL_TENSOR.SSM_CONV1D,
-                gguf.MODEL_TENSOR.SSM_X,
-                gguf.MODEL_TENSOR.SSM_DT,
-                gguf.MODEL_TENSOR.SSM_A,
-                gguf.MODEL_TENSOR.SSM_D,
-            ]
-        )
-
 
 @Model.register("CohereForCausalLM")
 class CommandR2Model(Model):
@@ -3199,6 +3407,145 @@ class T5Model(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@Model.register("T5EncoderModel")
+class T5EncoderModel(Model):
+    model_arch = gguf.MODEL_ARCH.T5ENCODER
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.shared_token_embeddings_found = False
+
+    def set_vocab(self):
+        # to avoid TypeError: Descriptors cannot be created directly
+        # exception when importing sentencepiece_model_pb2
+        os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+        from sentencepiece import SentencePieceProcessor
+        from sentencepiece import sentencepiece_model_pb2 as model
+
+        tokenizer_path = self.dir_model / 'tokenizer.model'
+
+        # many older models use spiece.model tokenizer model filename
+        if not tokenizer_path.is_file():
+            tokenizer_path = self.dir_model / 'spiece.model'
+
+        if not tokenizer_path.is_file():
+            raise FileNotFoundError(f"File not found: {tokenizer_path}")
+
+        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
+        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
+
+        # some models like Pile-T5 family use BPE tokenizer instead of Unigram
+        if sentencepiece_model.trainer_spec.model_type == 2:  # BPE
+            # assure the tokenizer model file name is correct
+            assert tokenizer_path.name == 'tokenizer.model'
+            return self._set_vocab_sentencepiece()
+        else:
+            assert sentencepiece_model.trainer_spec.model_type == 1  # UNIGRAM
+
+        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
+        remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
+        precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap
+
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        added_tokens_file = self.dir_model / 'added_tokens.json'
+        if added_tokens_file.is_file():
+            with open(added_tokens_file, "r", encoding="utf-8") as f:
+                added_tokens_json = json.load(f)
+                for key in added_tokens_json:
+                    token_id = added_tokens_json[key]
+                    if token_id >= vocab_size:
+                        logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                        continue
+
+                    tokens[token_id] = key.encode("utf-8")
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(SentencePieceTokenTypes.UNUSED)
+
+        self.gguf_writer.add_tokenizer_model("t5")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+        self.gguf_writer.add_add_space_prefix(add_prefix)
+        self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
+        if precompiled_charsmap:
+            self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+        self.gguf_writer.add_add_bos_token(False)
+        self.gguf_writer.add_add_eos_token(True)
+
+    def set_gguf_parameters(self):
+        if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
+            logger.warning("Couldn't find context length in config.json, assuming default value of 512")
+            n_ctx = 512
+        self.gguf_writer.add_context_length(n_ctx)
+        self.gguf_writer.add_embedding_length(self.hparams["d_model"])
+        self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"])
+        self.gguf_writer.add_block_count(self.hparams["num_layers"])
+        self.gguf_writer.add_head_count(self.hparams["num_heads"])
+        self.gguf_writer.add_key_length(self.hparams["d_kv"])
+        self.gguf_writer.add_value_length(self.hparams["d_kv"])
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_relative_attn_buckets_count(self.hparams["relative_attention_num_buckets"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # T5 based models contain shared token embeddings tensors saved randomly as either "encoder.embed_tokens.weight",
+        # "decoder.embed_tokens.weight" or "shared.weight" tensor. In some models there are even multiple of them stored
+        # in the safetensors files. We use the first tensor from these three as the token embeddings for both encoder
+        # and decoder and ignore the remaining ones.
+        if name in ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight", "shared.weight"]:
+            if not self.shared_token_embeddings_found:
+                name = "shared.weight"
+                self.shared_token_embeddings_found = True
+            else:
+                logger.debug(f"Skipping shared tensor {name!r} in safetensors so that convert can end normally.")
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 @Model.register("JAISLMHeadModel")
 class JaisModel(Model):
     model_arch = gguf.MODEL_ARCH.JAIS
@@ -3470,8 +3817,121 @@ class ChatGLMModel(Model):
         name = name.removeprefix("transformer.")
         return [(self.map_tensor_name(name), data_torch)]
 
-###### CONVERSION LOGIC ######
 
+@Model.register("NemotronForCausalLM")
+class NemotronModel(Model):
+    model_arch = gguf.MODEL_ARCH.NEMOTRON
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+        self.gguf_writer.add_pad_token_id(0)
+        self.gguf_writer.add_unk_token_id(1)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+
+        f_norm_eps = self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon", "norm_eps"])
+        self.gguf_writer.add_layer_norm_eps(f_norm_eps)
+
+        # * Partial RoPE
+        rot_pct = self.find_hparam(["partial_rotary_factor", "rope_pct", "rope_percent"])
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
+        self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
+
+        # * RopeScaling for Nemotron
+        if "rope_scaling" not in self.hparams or self.hparams["rope_scaling"] is None:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+        else:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(self.hparams["factor"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # * Adding +1 to LayerNorm's weights here to implement layernorm1p w/o changing anything on the GGML engine side
+        #   model.layers.{l}.input_layernorm.weight
+        #   model.layers.{l}.post_attention_layernorm.weight
+        #   model.norm.weight
+        if name.endswith("norm.weight"):
+            data_torch = data_torch + 1
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
+@Model.register("ExaoneForCausalLM")
+class ExaoneModel(Model):
+    model_arch = gguf.MODEL_ARCH.EXAONE
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+
+        assert (hparams["activation_function"] == "silu")
+
+        max_position_embeddings = hparams["max_position_embeddings"]
+        embed_dim = hparams["hidden_size"]
+        num_heads = hparams["num_attention_heads"]
+        num_kv_heads = hparams.get("num_key_value_heads", num_heads)
+        layer_norm_eps = hparams["layer_norm_epsilon"]
+        intermediate_size = hparams["intermediate_size"] if "intermediate_size" in hparams else 4 * embed_dim
+        num_layers = hparams["num_layers"]
+        # ignore for now as EXAONE-3.0-7.8B-Instruct attentino_dropout is 0.0
+        # attention_dropout_rate = hparams["attention_dropout"]
+        # ignore for now as EXAONE-3.0-7.8B-Instruct embed_dropout is 0.0
+        # embed_dropout_rate = hparams["embed_dropout"]
+        self.gguf_writer.add_embedding_length(embed_dim)
+        self.gguf_writer.add_head_count(num_heads)
+        self.gguf_writer.add_head_count_kv(num_kv_heads)
+        self.gguf_writer.add_context_length(max_position_embeddings)
+        self.gguf_writer.add_layer_norm_rms_eps(layer_norm_eps)
+        self.gguf_writer.add_feed_forward_length(intermediate_size)
+        self.gguf_writer.add_block_count(num_layers)
+        self.gguf_writer.add_file_type(self.ftype)
+
+        if (rope_theta := self.hparams.get("rope_theta")) is not None:
+            self.gguf_writer.add_rope_freq_base(rope_theta)
+        rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"], optional=True)
+        rotary_factor = rotary_factor if rotary_factor is not None else 1.0
+        self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
+        if hparams.get("rope_scaling") is not None and "factor" in hparams["rope_scaling"]:
+            if hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])
+
+    def prepare_tensors(self):
+        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.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)
+                low_freq_factor = rope_scaling.get("low_freq_factor", 1.0)
+                high_freq_factor = rope_scaling.get("high_freq_factor", 4.0)
+                old_context_len = self.hparams.get("original_max_position_embeddings", 8192)
+
+                low_freq_wavelen = old_context_len / low_freq_factor
+                high_freq_wavelen = old_context_len / high_freq_factor
+                assert low_freq_wavelen != high_freq_wavelen
+
+                rope_factors = []
+                for freq in freqs:
+                    wavelen = 2 * math.pi / freq
+                    if wavelen < high_freq_wavelen:
+                        rope_factors.append(1)
+                    elif wavelen > low_freq_wavelen:
+                        rope_factors.append(factor)
+                    else:
+                        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()
+
+
+###### CONVERSION LOGIC ######
 
 # tree of lazy tensors
 class LazyTorchTensor(gguf.LazyBase):

convert_hf_to_gguf_update.py

@@ -94,6 +94,9 @@ models = [
     {"name": "codeshell",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", },
     {"name": "tekken",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },
     {"name": "smollm",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", },
+    {'name': "bloom",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigscience/bloom", },
+    {'name': "gpt3-finnish",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/TurkuNLP/gpt3-finnish-small", },
+    {"name": "exaone",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct", },
 ]
 
 

convert_llama_ggml_to_gguf.py

@@ -116,7 +116,7 @@ class Tensor:
         assert quant is not None, 'Unknown tensor type'
         (blksize, tysize) = quant
         offset += 12
-        self.dtype= dtype
+        self.dtype= gguf.GGMLQuantizationType(dtype)
         self.dims = struct.unpack(f'<{n_dims}I', data[offset:offset + (4 * n_dims)])
         offset += 4 * n_dims
         self.name = bytes(data[offset:offset + name_len])

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/CANN.md

@@ -0,0 +1,259 @@
+# llama.cpp for CANN
+
+ - [Background](#background)
+ - [News](#news)
+ - [OS](#os)
+ - [Hardware](#hardware)
+ - [Model Supports](#model-supports)
+ - [DataType Supports](#datatype-supports)
+ - [Docker](#docker)
+ - [Linux](#linux)
+ - [TODO](#todo)
+
+
+## Background
+
+**Ascend NPU** is a range of AI processors using Neural Processing Unit. It will efficiently handle matrix-matrix multiplication, dot-product and scalars.
+
+**CANN** (Compute Architecture for Neural Networks) is a heterogeneous computing architecture for AI scenarios, providing support for multiple AI frameworks on the top and serving AI processors and programming at the bottom. It plays a crucial role in bridging the gap between upper and lower layers, and is a key platform for improving the computing efficiency of Ascend AI processors. Meanwhile, it offers a highly efficient and easy-to-use programming interface for diverse application scenarios, allowing users to rapidly build AI applications and services based on the Ascend platform.
+
+**Llama.cpp + CANN**
+
+The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the ability of AscendC and ACLNN which are intergrated to CANN Toolkit and kernels to using Ascend NPU directly.
+
+## News
+
+- 2024.8
+  - Support `Q4_0` and `Q8_0` data type for Ascend NPU.
+- 2024.7
+  - Create CANN backend for Ascend NPU.
+
+## OS
+
+| OS      | Status  | Verified                                       |
+|:-------:|:-------:|:----------------------------------------------:|
+| Linux   | Support | Ubuntu 22.04, OpenEuler22.03                   |
+
+
+## Hardware
+
+### Ascend NPU
+
+**Verified devices**
+| Ascend NPU                    | Status  |
+|:-----------------------------:|:-------:|
+| Atlas 300T A2                 | Support |
+
+*Notes:*
+
+- If you have trouble with Ascend NPU device, please create a issue with **[CANN]** prefix/tag.
+- If you run successfully with your Ascend NPU device, please help update the upper table.
+
+
+## Model Supports
+
+| Model Name                  | FP16  | Q8_0 | Q4_0 |
+|:----------------------------|:-----:|:----:|:----:|
+| AquilaChat2-7B              |   √   |   √  |   √  |
+| Baichuan-7b                 |   √   |   √  |   √  |
+| Baichuan2-7B-Chat           |   √   |   √  |   √  |
+| bitnet_b1_58-large          |   √   |   √  |   √  |
+| bloom-560m                  |   √   |   x  |   √  |
+| bloomz-alpaca-560m          |   √   |   x  |   √  |
+| c4ai-command-r-35B-v01      |   x   |   x  |   x  |
+| chatglm3-6B                 |   x   |   x  |   x  |
+| chinese-alpaca-2-1.3b       |   √   |   √  |   √  |
+| CodeShell-7B                |   √   |   √  |   √  |
+| deepseek-ai_deepseek-coder-1.3B-base | x |   x  |   x  |
+| deepseek-ai_DeepSeek-V2-Lite | x   |   x  |   x   |
+| deepseek-coder-6.7B-instruct | x   |   x  |   x   |
+| DeepSeek-V2-Lite-64x1.5B    |   x   |   x  |   x  |
+| falcon-7b-instruct          |   √   |   √  |   √  |
+| flan-t5-large               |   √   |   √  |   √  |
+| gemma-2-9b-it               |   √   |   √  |   √  |
+| glm-4-9B                    |   x   |   x  |   x  |
+| gpt2                        |   √   |   √  |   √  |
+| Gpt2-163M                   |   √   |   √  |   √  |
+| granite-3B-code-instruct    |   √   |   √  |   √  |
+| GritLM-7B                   |   √   |   √  |   √  |
+| internlm2_5-7b-chat         |   √   |   √  |   √  |
+| koala-7B-HF                 |   √   |   √  |   √  |
+| Llama-2-7b-chat-hf          |   √   |   √  |   √  |
+| Llama-3-Smaug-8B            |   √   |   √  |   √  |
+| Llama2-Chinese-7b-Chat      |   √   |   √  |   √  |
+| Llama3-8B                   |   √   |   √  |   √  |
+| Llama3-8b-chinese           |   √   |   √  |   √  |
+| mamba-130m-hf               |   √   |   √  |   √  |
+| Mistral-7B-Instruct-v0.2    |   √   |   √  |   √  |
+| Mixtral-8x7B-Instruct-v0.1  |   x   |   √  |   √  |
+| mpt-7B                      |   √   |   √  |   √  |
+| OLMo-1B-hf                  |   √   |   √  |   √  |
+| OpenELM-3B-Instruct         |   √   |   √  |   √  |
+| Orion-14b-base              |   √   |   √  |   √  |
+| phi1                        |   x   |   x  |   x  |
+| phi2                        |   x   |   x  |   x  |
+| Phi-3-mini-4k-instruct      |   √   |   √  |   √  |
+| plamo-13b                   |   √   |   √  |   √  |
+| pythia-70M                  |   x   |   x  |   x  |
+| Qwen-7B                     |   √   |   √  |   √  |
+| Qwen2-1.5B-Instruct         |   √   |   x  |   √  |
+| Refact-1_6B-fim             |   √   |   √  |   √  |
+| SmolLM-135M                 |   √   |   √  |   √  |
+| stablelm-zephyr             |   x   |   x  |   x  |
+| stablelm-2-zephyr-1_6b      |   x   |   x  |   x  |
+| starcoderbase-1b            |   √   |   √  |   √  |
+| starcoder2-3b               |   √   |   √  |   √  |
+| vigogne-7b-chat             |   √   |   √  |   √  |
+| xverse-7b-chat              |   √   |   √  |   √  |
+| Yi-6b-Chat                  |   √   |   √  |   √  |
+
+
+
+## DataType Supports
+
+| DataType               | Status  |
+|:----------------------:|:-------:|
+| FP16                   | Support |
+| Q8_0                   | Support |
+| Q4_0                   | Support |
+
+## Docker
+
+### Build Images
+You can get a image with llama.cpp in one command.
+```sh
+docker build -t llama-cpp-cann -f .devops/llama-cli-cann.Dockerfile .
+```
+
+### Run container
+
+```sh
+# Find all cards.
+npu-smi info
+
+# Select the cards that you want to use, make sure these cards are not used by someone.
+# Following using cards of device0.
+docker run --name llamacpp --device /dev/davinci0  --device /dev/davinci_manager --device /dev/devmm_svm --device /dev/hisi_hdc -v /usr/local/dcmi:/usr/local/dcmi -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info -v /PATH_TO_YOUR_MODELS/:/app/models -it llama-cpp-cann -m /app/models/MODEL_PATH -ngl 32 -p "Building a website can be done in 10 simple steps:"
+```
+
+*Notes:*
+
+- You may need to install Ascend Driver and firmware on the **host** machine *(Please refer to the [Linux configuration](#linux) for details)*.
+
+## Linux
+
+### I. Setup Environment
+
+1. **Install Ascend Driver and firmware**
+
+    ```sh
+    # create driver running user.
+    sudo groupadd -g HwHiAiUser
+    sudo useradd -g HwHiAiUser -d /home/HwHiAiUser -m HwHiAiUser -s /bin/bash
+    sudo usermod -aG HwHiAiUser $USER
+
+    # download driver from https://www.hiascend.com/hardware/firmware-drivers/community according to your system
+    # and install driver.
+    sudo sh Ascend-hdk-910b-npu-driver_x.x.x_linux-{arch}.run --full --install-for-all
+    ```
+
+    Once installed, run `npu-smi info` to check whether driver is installed successfully.
+    ```sh
+    +-------------------------------------------------------------------------------------------+
+    | npu-smi 24.1.rc2               Version: 24.1.rc2                                          |
+    +----------------------+---------------+----------------------------------------------------+
+    | NPU   Name           | Health        | Power(W)    Temp(C)           Hugepages-Usage(page)|
+    | Chip                 | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |
+    +======================+===============+====================================================+
+    | 2     xxx            | OK            | 64.4        51                15   / 15            |
+    | 0                    | 0000:01:00.0  | 0           1873 / 15077      0    / 32768         |
+    +======================+===============+====================================================+
+    | 5     xxx            | OK            | 64.0        52                15   / 15            |
+    | 0                    | 0000:81:00.0  | 0           1874 / 15077      0    / 32768         |
+    +======================+===============+====================================================+
+    | No running processes found in NPU 2                                                       |
+    +======================+===============+====================================================+
+    | No running processes found in NPU 5                                                       |
+    +======================+===============+====================================================+
+    ```
+
+2. **Install Ascend Firmware**
+    ```sh
+    # download driver from https://www.hiascend.com/hardware/firmware-drivers/community according to your system
+    # and install driver.
+    sudo sh Ascend-hdk-910b-npu-firmware_x.x.x.x.X.run --full
+    ```
+    If the following messaage appers, firmware is installed successfully.
+    ```sh
+    Firmware package installed successfully!
+    ```
+
+
+3. **Install CANN toolkit and kernels**
+
+    CANN toolkit and kernels can be obtained from the official [CANN Toolkit](https://www.hiascend.com/zh/developer/download/community/result?module=cann) page.
+
+    Please download the corresponding version that satified your system. The minimum version required is 8.0.RC2.alpha002 and here is the install command.
+    ```sh
+    pip3 install attrs numpy decorator sympy cffi pyyaml pathlib2 psutil protobuf scipy requests absl-py wheel typing_extensions
+    sh Ascend-cann-toolkit_8.0.RC2.alpha002_linux-aarch64.run --install
+    sh Ascend-cann-kernels-910b_8.0.RC2.alpha002_linux.run --install
+    ```
+
+    Set Ascend Variables:
+    ```sh
+    echo "source ~/Ascend/ascend-toolkit/set_env.sh" >> ~/.bashrc
+    source ~/.bashrc
+    ```
+
+Upon a successful installation, CANN is enabled for the available ascend devices.
+
+### II. Build llama.cpp
+
+```sh
+cmake -B build -DGGML_CANN=on -DCMAKE_BUILD_TYPE=release
+cmake --build build --config release
+```
+
+### III. Run the inference
+
+1. **Retrieve and prepare model**
+
+    You can refer to the general [*Prepare and Quantize*](../../README.md#prepare-and-quantize) guide for model prepration.
+
+    **Notes**:
+
+      - CANN backend only supports FP16/Q4_0/Q8_0 models currently.
+
+2. **Launch inference**
+
+    There are two device selection modes:
+
+    - Single device: Use one device target specified by the user.
+    - Multiple devices: Automatically choose the devices with the same backend.
+
+    | Device selection | Parameter                              |
+    |:----------------:|:--------------------------------------:|
+    | Single device    | --split-mode none --main-gpu DEVICE_ID |
+    | Multiple devices | --split-mode layer (default)           |
+
+    Examples:
+
+    - Use device 0:
+
+    ```sh
+    ./build/bin/llama-cli -m path_to_model -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
+    ```
+
+    - Use multiple devices:
+
+    ```sh
+    ./build/bin/llama-cli -m path_to_model -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
+    ```
+
+### **GitHub contribution**:
+Please add the **[CANN]** prefix/tag in issues/PRs titles to help the CANN-team check/address them without delay.
+
+
+## TODO
+- Support more models and data types.

docs/backend/SYCL.md

@@ -20,7 +20,7 @@
 **oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:
 
 - **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers.
-- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL - Math Kernel Library)*.
+- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL and oneDNN)*.
 - **oneAPI LevelZero**: A high performance low level interface for fine-grained control over intel iGPUs and dGPUs.
 - **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.
 
@@ -28,10 +28,6 @@
 
 The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based on the cross-platform feature of SYCL, it could support other vendor GPUs: Nvidia GPU (*AMD GPU coming*).
 
-When targeting **Intel CPU**, it is recommended to use llama.cpp for [Intel oneMKL](README.md#intel-onemkl) backend.
-
-It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS, cuBLAS, etc..*. In beginning work, the oneAPI's [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) open-source migration tool (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) was used for this purpose.
-
 ## Recommended Release
 
 The SYCL backend would be broken by some PRs due to no online CI.
@@ -45,6 +41,10 @@ The following release is verified with good quality:
 
 ## News
 
+
+- 2024.8
+  - Use oneDNN as the default GEMM library, improve the compatibility for new Intel GPUs.
+
 - 2024.5
   - Performance is increased: 34 -> 37 tokens/s of llama-2-7b.Q4_0 on Arc770.
   - Arch Linux is verified successfully.
@@ -80,7 +80,14 @@ The following release is verified with good quality:
 
 ### Intel GPU
 
-**Verified devices**
+SYCL backend supports Intel GPU Family:
+
+- Intel Data Center Max Series
+- Intel Flex Series, Arc Series
+- Intel Built-in Arc GPU
+- Intel iGPU in Core CPU (11th Generation Core CPU and newer, refer to [oneAPI supported GPU](https://www.intel.com/content/www/us/en/developer/articles/system-requirements/intel-oneapi-base-toolkit-system-requirements.html#inpage-nav-1-1)).
+
+#### Verified devices
 
 | Intel GPU                     | Status  | Verified Model                        |
 |-------------------------------|---------|---------------------------------------|
@@ -88,7 +95,7 @@ The following release is verified with good quality:
 | Intel Data Center Flex Series | Support | Flex 170                              |
 | Intel Arc Series              | Support | Arc 770, 730M, Arc A750               |
 | Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake       |
-| Intel iGPU                    | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 |
+| Intel iGPU                    | Support | iGPU in 13700k, i5-1250P, i7-1260P, i7-1165G7 |
 
 *Notes:*
 
@@ -189,7 +196,7 @@ Please follow the instructions for downloading and installing the Toolkit for Li
 
 Following guidelines/code snippets assume the default installation values. Otherwise, please make sure the necessary changes are reflected where applicable.
 
-Upon a successful installation, SYCL is enabled for the available intel devices, along with relevant libraries such as oneAPI MKL for intel GPUs.
+Upon a successful installation, SYCL is enabled for the available intel devices, along with relevant libraries such as oneAPI oneDNN for Intel GPUs.
 
 - **Adding support to Nvidia GPUs**
 
@@ -237,12 +244,17 @@ Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA devic
 ### II. Build llama.cpp
 
 #### Intel GPU
+
+```
+./examples/sycl/build.sh
+```
+
+or
+
 ```sh
 # Export relevant ENV variables
 source /opt/intel/oneapi/setvars.sh
 
-# Build LLAMA with MKL BLAS acceleration for intel GPU
-
 # Option 1: Use FP32 (recommended for better performance in most cases)
 cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
@@ -276,23 +288,26 @@ cmake --build build --config Release -j -v
 
 ### III. Run the inference
 
-1. Retrieve and prepare model
+#### Retrieve and prepare model
 
 You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
-2. Enable oneAPI running environment
+##### Check device
+
+1. Enable oneAPI running environment
 
 ```sh
 source /opt/intel/oneapi/setvars.sh
 ```
 
-3. List devices information
+2. List devices information
 
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```sh
 ./build/bin/llama-ls-sycl-device
 ```
+
 This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
 ```
 found 2 SYCL devices:
@@ -304,12 +319,37 @@ found 2 SYCL devices:
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
 ```
 
+#### Choose level-zero devices
 
-4. Launch inference
+|Chosen Device ID|Setting|
+|-|-|
+|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
+|1|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
+|0 & 1|`export ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
+
+#### Execute
+
+Choose one of following methods to run.
+
+1. Script
+
+- Use device 0:
+
+```sh
+./examples/sycl/run-llama2.sh 0
+```
+- Use multiple devices:
+
+```sh
+./examples/sycl/run-llama2.sh
+```
+
+2. Command line
+Launch inference
 
 There are two device selection modes:
 
-- Single device: Use one device target specified by the user.
+- Single device: Use one device assigned by user. Default device id is 0.
 - Multiple devices: Automatically choose the devices with the same backend.
 
 In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR.
@@ -326,11 +366,6 @@ Examples:
 ```sh
 ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
 ```
-or run by script:
-
-```sh
-./examples/sycl/run_llama2.sh 0
-```
 
 - Use multiple devices:
 
@@ -338,12 +373,6 @@ or run by script:
 ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
 ```
 
-Otherwise, you can run the script:
-
-```sh
-./examples/sycl/run_llama2.sh
-```
-
 *Notes:*
 
 - Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
@@ -390,7 +419,7 @@ c. Verify installation
 In the oneAPI command line, run the following to print the available SYCL devices:
 
 ```
-sycl-ls
+sycl-ls.exe
 ```
 
 There should be one or more *level-zero* GPU devices displayed as **[ext_oneapi_level_zero:gpu]**. Below is example of such output detecting an *intel Iris Xe* GPU as a Level-zero SYCL device:
@@ -411,6 +440,18 @@ b. The new Visual Studio will install Ninja as default. (If not, please install
 
 ### II. Build llama.cpp
 
+You could download the release package for Windows directly, which including binary files and depended oneAPI dll files.
+
+Choose one of following methods to build from source code.
+
+1. Script
+
+```sh
+.\examples\sycl\win-build-sycl.bat
+```
+
+2. CMake
+
 On the oneAPI command line window, step into the llama.cpp main directory and run the following:
 
 ```
@@ -425,12 +466,8 @@ cmake -B build -G "Ninja" -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPI
 cmake --build build --config Release -j
 ```
 
-Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former instructions:
-```sh
-.\examples\sycl\win-build-sycl.bat
-```
-
 Or, use CMake presets to build:
+
 ```sh
 cmake --preset x64-windows-sycl-release
 cmake --build build-x64-windows-sycl-release -j --target llama-cli
@@ -442,7 +479,9 @@ cmake --preset x64-windows-sycl-debug
 cmake --build build-x64-windows-sycl-debug -j --target llama-cli
 ```
 
-Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
+3. Visual Studio
+
+You can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
 
 *Notes:*
 
@@ -450,23 +489,25 @@ Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choos
 
 ### III. Run the inference
 
-1. Retrieve and prepare model
+#### Retrieve and prepare model
 
-You can refer to the general [*Prepare and Quantize*](README#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
-2. Enable oneAPI running environment
+##### Check device
+
+1. Enable oneAPI running environment
 
 On the oneAPI command line window, run the following and step into the llama.cpp directory:
 ```
 "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
 ```
 
-3. List devices information
+2. List devices information
 
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```
-build\bin\ls-sycl-device.exe
+build\bin\llama-ls-sycl-device.exe
 ```
 
 This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
@@ -479,9 +520,27 @@ found 2 SYCL devices:
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
 
 ```
+#### Choose level-zero devices
 
+|Chosen Device ID|Setting|
+|-|-|
+|0|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
+|1|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
+|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
 
-4. Launch inference
+#### Execute
+
+Choose one of following methods to run.
+
+1. Script
+
+```
+examples\sycl\win-run-llama2.bat
+```
+
+2. Command line
+
+Launch inference
 
 There are two device selection modes:
 
@@ -508,11 +567,7 @@ build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website ca
 ```
 build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
 ```
-Otherwise, run the following wrapper script:
 
-```
-.\examples\sycl\win-run-llama2.bat
-```
 
 Note:
 
@@ -526,17 +581,18 @@ Or
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```
 
+
 ## Environment Variable
 
 #### Build
 
 | Name               | Value                             | Function                                    |
 |--------------------|-----------------------------------|---------------------------------------------|
-| GGML_SYCL          | ON (mandatory)                    | Enable build with SYCL code path.           |
+| GGML_SYCL          | ON (mandatory)                    | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
 | GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA       | Set the SYCL target device type.            |
 | GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path.      |
-| CMAKE_C_COMPILER   | icx                               | Set *icx* compiler for SYCL code path.      |
-| CMAKE_CXX_COMPILER | icpx *(Linux)*, icx *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
+| CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
+| CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 
 #### Runtime
 
@@ -572,9 +628,18 @@ use 1 SYCL GPUs: [0] with Max compute units:512
   ```
   Otherwise, please double-check the GPU driver installation steps.
 
+- Can I report Ollama issue on Intel GPU to llama.cpp SYCL backend?
+
+  No. We can't support Ollama issue directly, because we aren't familiar with Ollama.
+
+  Sugguest reproducing on llama.cpp and report similar issue to llama.cpp. We will surpport it.
+
+  It's same for other projects including llama.cpp SYCL backend.
+
+
 ### **GitHub contribution**:
 Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.
 
 ## TODO
 
-- Support row layer split for multiple card runs.
+- NA

docs/build.md

@@ -178,7 +178,11 @@ For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](ht
   cmake --build build --config Release
   ```
 
-The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. The following compilation options are also available to tweak performance:
+The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used.
+
+The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. In Windows this setting is available in the NVIDIA control panel as `System Memory Fallback`.
+
+The following compilation options are also available to tweak performance:
 
 | Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
 |-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
@@ -192,6 +196,19 @@ The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/c
 | GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
 | GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                  |
 
+### MUSA
+
+- Using `make`:
+  ```bash
+  make GGML_MUSA=1
+  ```
+- Using `CMake`:
+
+  ```bash
+  cmake -B build -DGGML_MUSA=ON
+  cmake --build build --config Release
+  ```
+
 ### hipBLAS
 
 This provides BLAS acceleration on HIP-supported AMD GPUs.
@@ -335,6 +352,31 @@ cmake --build build --config Release
 # ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
 ```
 
+### CANN
+This provides NPU acceleration using the AI cores of your Ascend NPU. And [CANN](https://www.hiascend.com/en/software/cann) is a hierarchical APIs to help you to quickly build AI applications and service based on Ascend NPU.
+
+For more information about Ascend NPU in [Ascend Community](https://www.hiascend.com/en/).
+
+Make sure to have the CANN toolkit installed. You can download it from here: [CANN Toolkit](https://www.hiascend.com/developer/download/community/result?module=cann)
+
+Go to `llama.cpp` directory and build using CMake.
+```bash
+cmake -B build -DGGML_CANN=on -DCMAKE_BUILD_TYPE=release
+cmake --build build --config release
+```
+
+You can test with:
+
+`./build/llama-cli -m PATH_TO_MODEL -p "Building a website can be done in 10 steps:" -ngl 32`
+
+If the fllowing info is output on screen, you are using `llama.cpp by CANN backend`:
+```bash
+llm_load_tensors:       CANN buffer size = 13313.00 MiB
+llama_new_context_with_model:       CANN compute buffer size =  1260.81 MiB
+```
+
+For detailed info, such as model/device supports, CANN install, please refer to [llama.cpp for CANN](./backend/CANN.md).
+
 ### Android
 
 To read documentation for how to build on Android, [click here](./android.md)

docs/docker.md

@@ -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_DOCKER_ARCH` set to `all`
+- `CUDA_VERSION` set to `12.6.0`
+- `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

@@ -1,7 +1,6 @@
 #include "ggml.h"
 #include "train.h"
 
-#include <vector>
 #include <cassert>
 #include <cstdlib>
 #include <cstring>
@@ -19,7 +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/batched-bench/batched-bench.cpp

@@ -69,7 +69,7 @@ int main(int argc, char ** argv) {
     llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
 
     // ensure enough sequences are available
-    ctx_params.n_seq_max = *std::max_element(n_pl.begin(), n_pl.end());
+    ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 

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

@@ -271,7 +271,7 @@ struct tokenized_prompt {
     size_t max_seq_len;
 
     tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
-        const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+        const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
         tokens_pos = ::llama_tokenize(ctx, pos, add_bos, true);
         tokens_neg = ::llama_tokenize(ctx, neg, add_bos, true);
         max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
@@ -414,9 +414,10 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the model to get hparams
-    llama_model * model;
-    llama_context * ctx;
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // int n_ctx = llama_n_ctx(ctx);
     int n_layers = llama_n_layer(model);
@@ -485,8 +486,8 @@ 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_batch = params.n_pca_batch;
+        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);
     } else {

examples/embedding/README.md

@@ -9,13 +9,13 @@ To get started right away, run the following command, making sure to use the cor
 ### Unix-based systems (Linux, macOS, etc.):
 
 ```bash
-./llama-embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null
+./llama-embedding -m ./path/to/model --pooling mean --log-disable -p "Hello World!" 2>/dev/null
 ```
 
 ### Windows:
 
 ```powershell
-llama-embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$null
+llama-embedding.exe -m ./path/to/model --pooling mean --log-disable -p "Hello World!" 2>$null
 ```
 
 The above command will output space-separated float values.
@@ -50,11 +50,11 @@ The above command will output space-separated float values.
 ### Unix-based systems (Linux, macOS, etc.):
 
 ```bash
-./embedding -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
+./llama-embedding -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --pooling mean --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
 ```
 
 ### Windows:
 
 ```powershell
-embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
+llama-embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --pooling mean --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
 ```

examples/embedding/embedding.cpp

@@ -31,13 +31,24 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke
 }
 
 static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd, int embd_norm) {
+    const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
+    const struct llama_model * model = llama_get_model(ctx);
+
     // clear previous kv_cache values (irrelevant for embeddings)
     llama_kv_cache_clear(ctx);
 
     // run model
     fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
-    if (llama_decode(ctx, batch) < 0) {
-        fprintf(stderr, "%s : failed to decode\n", __func__);
+    if (llama_model_has_encoder(model) && !llama_model_has_decoder(model)) {
+        // encoder-only model
+        if (llama_encode(ctx, batch) < 0) {
+            fprintf(stderr, "%s : failed to encode\n", __func__);
+        }
+    } else if (!llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
+        // decoder-only model
+        if (llama_decode(ctx, batch) < 0) {
+            fprintf(stderr, "%s : failed to decode\n", __func__);
+        }
     }
 
     for (int i = 0; i < batch.n_tokens; i++) {
@@ -45,11 +56,22 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
             continue;
         }
 
-        // try to get sequence embeddings - supported only when pooling_type is not NONE
-        const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
-        GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
+        const float * embd = nullptr;
+        int embd_pos = 0;
 
-        float * out = output + batch.seq_id[i][0] * n_embd;
+        if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+            // try to get token embeddings
+            embd = llama_get_embeddings_ith(ctx, i);
+            embd_pos = i;
+            GGML_ASSERT(embd != NULL && "failed to get token embeddings");
+        } else {
+            // try to get sequence embeddings - supported only when pooling_type is not NONE
+            embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
+            embd_pos = batch.seq_id[i][0];
+            GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
+        }
+
+        float * out = output + embd_pos * n_embd;
         llama_embd_normalize(embd, out, n_embd, embd_norm);
     }
 }
@@ -79,11 +101,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
@@ -93,8 +115,9 @@ int main(int argc, char ** argv) {
     const int n_ctx = llama_n_ctx(ctx);
 
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
-    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
-        fprintf(stderr, "%s: error: pooling type NONE not supported\n", __func__);
+
+    if (llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
+        fprintf(stderr, "%s: error: computing embeddings in encoder-decoder models is not supported\n", __func__);
         return 1;
     }
 
@@ -153,13 +176,23 @@ int main(int argc, char ** argv) {
     const int n_prompts = prompts.size();
     struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
+    // count number of embeddings
+    int n_embd_count = 0;
+    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+        for (int k = 0; k < n_prompts; k++) {
+            n_embd_count += inputs[k].size();
+        }
+    } else {
+        n_embd_count = n_prompts;
+    }
+
     // allocate output
     const int n_embd = llama_n_embd(model);
-    std::vector<float> embeddings(n_prompts * n_embd, 0);
+    std::vector<float> embeddings(n_embd_count * n_embd, 0);
     float * emb = embeddings.data();
 
     // break into batches
-    int p = 0; // number of prompts processed already
+    int e = 0; // number of embeddings already stored
     int s = 0; // number of prompts in current batch
     for (int k = 0; k < n_prompts; k++) {
         // clamp to n_batch tokens
@@ -169,11 +202,11 @@ int main(int argc, char ** argv) {
 
         // encode if at capacity
         if (batch.n_tokens + n_toks > n_batch) {
-            float * out = emb + p * n_embd;
+            float * out = emb + e * n_embd;
             batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
-            llama_batch_clear(batch);
-            p += s;
+            e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s;
             s = 0;
+            llama_batch_clear(batch);
         }
 
         // add to batch
@@ -182,40 +215,63 @@ int main(int argc, char ** argv) {
     }
 
     // final batch
-    float * out = emb + p * n_embd;
+    float * out = emb + e * n_embd;
     batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
 
     if (params.embd_out.empty()) {
-        // print the first part of the embeddings or for a single prompt, the full embedding
         fprintf(stdout, "\n");
-        for (int j = 0; j < n_prompts; j++) {
-            fprintf(stdout, "embedding %d: ", j);
-            for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
-                if (params.embd_normalize == 0) {
-                    fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
-                } else {
-                    fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
-                }
-            }
-            fprintf(stdout, "\n");
-        }
 
-        // print cosine similarity matrix
-        if (n_prompts > 1) {
-            fprintf(stdout, "\n");
-            printf("cosine similarity matrix:\n\n");
-            for (int i = 0; i < n_prompts; i++) {
-                fprintf(stdout, "%6.6s ", prompts[i].c_str());
-            }
-            fprintf(stdout, "\n");
-            for (int i = 0; i < n_prompts; i++) {
-                for (int j = 0; j < n_prompts; j++) {
-                    float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
-                    fprintf(stdout, "%6.2f ", sim);
+        if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+            for (int j = 0; j < n_embd_count; j++) {
+                fprintf(stdout, "embedding %d: ", j);
+                for (int i = 0; i < std::min(3, n_embd); i++) {
+                    if (params.embd_normalize == 0) {
+                        fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
+                    } else {
+                        fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
+                    }
+                }
+                fprintf(stdout, " ... ");
+                for (int i = n_embd - 3; i < n_embd; i++) {
+                    if (params.embd_normalize == 0) {
+                        fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
+                    } else {
+                        fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
+                    }
                 }
-                fprintf(stdout, "%1.10s", prompts[i].c_str());
                 fprintf(stdout, "\n");
             }
+        } else {
+            // print the first part of the embeddings or for a single prompt, the full embedding
+            for (int j = 0; j < n_prompts; j++) {
+                fprintf(stdout, "embedding %d: ", j);
+                for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
+                    if (params.embd_normalize == 0) {
+                        fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
+                    } else {
+                        fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
+                    }
+                }
+                fprintf(stdout, "\n");
+            }
+
+            // print cosine similarity matrix
+            if (n_prompts > 1) {
+                fprintf(stdout, "\n");
+                printf("cosine similarity matrix:\n\n");
+                for (int i = 0; i < n_prompts; i++) {
+                    fprintf(stdout, "%6.6s ", prompts[i].c_str());
+                }
+                fprintf(stdout, "\n");
+                for (int i = 0; i < n_prompts; i++) {
+                    for (int j = 0; j < n_prompts; j++) {
+                        float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+                        fprintf(stdout, "%6.2f ", sim);
+                    }
+                    fprintf(stdout, "%1.10s", prompts[i].c_str());
+                    fprintf(stdout, "\n");
+                }
+            }
         }
     }
 
@@ -233,23 +289,23 @@ int main(int argc, char ** argv) {
             }
             fprintf(stdout, notArray ? "]\n    }" : "]");
             j++;
-            if (j < n_prompts) fprintf(stdout, notArray ? ",\n" : ","); else break;
+            if (j < n_embd_count) fprintf(stdout, notArray ? ",\n" : ","); else break;
         }
         fprintf(stdout, notArray ? "\n  ]" : "]\n");
 
         if (params.embd_out == "json+" && n_prompts > 1) {
             fprintf(stdout, ",\n  \"cosineSimilarity\": [\n");
-            for (int i = 0;;) { // at least two iteration (n_prompts > 1)
+            for (int i = 0;;) { // at least two iteration (n_embd_count > 1)
                 fprintf(stdout, "    [");
-                for (int j = 0;;) { // at least two iteration (n_prompts > 1)
+                for (int j = 0;;) { // at least two iteration (n_embd_count > 1)
                     float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
                     fprintf(stdout, "%6.2f", sim);
                     j++;
-                    if (j < n_prompts) fprintf(stdout, ", "); else break;
+                    if (j < n_embd_count) fprintf(stdout, ", "); else break;
                 }
                 fprintf(stdout, " ]");
                 i++;
-                if (i < n_prompts) fprintf(stdout, ",\n"); else break;
+                if (i < n_embd_count) fprintf(stdout, ",\n"); else break;
             }
             fprintf(stdout, "\n  ]");
         }

examples/eval-callback/eval-callback.cpp

@@ -62,7 +62,7 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
                     } else if (type == GGML_TYPE_I8) {
                         v = (float) *(int8_t *) &data[i];
                     } else {
-                        GGML_ASSERT(false);
+                        GGML_ABORT("fatal error");
                     }
                     printf("%12.4f", v);
                     sum += v;
@@ -127,7 +127,7 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
 }
 
 static bool run(llama_context * ctx, const gpt_params & params) {
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
 
     std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
 
@@ -163,9 +163,10 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_model * model;
-    llama_context * ctx;
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;

examples/export-lora/README.md

@@ -17,9 +17,9 @@ For example:
 
 ```bash
 ./bin/llama-export-lora \
-    -m open-llama-3b-v2-q8_0.gguf \
-    -o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \
-    --lora lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.gguf
+    -m open-llama-3b-v2.gguf \
+    -o open-llama-3b-v2-english2tokipona-chat.gguf \
+    --lora lora-open-llama-3b-v2-english2tokipona-chat-LATEST.gguf
 ```
 
 Multiple LORA adapters can be applied by passing multiple `--lora FNAME` or `--lora-scaled FNAME S` command line parameters:

examples/export-lora/export-lora.cpp

@@ -10,6 +10,12 @@
 
 static bool g_verbose = false;
 
+struct tensor_transformation {
+    struct ggml_tensor * in;
+    struct ggml_tensor * out;
+    bool is_copy;
+};
+
 static std::string get_kv_str(struct gguf_context * ctx_gguf, const std::string & key){
     int id = gguf_find_key(ctx_gguf, key.c_str());
     return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf, id));
@@ -50,20 +56,6 @@ static struct gguf_context * load_gguf(std::string & fname, struct ggml_context
     return ctx_gguf;
 }
 
-static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
-    std::string result;
-    for (size_t pos = 0; ; pos += search.length()) {
-        auto new_pos = s.find(search, pos);
-        if (new_pos == std::string::npos) {
-            result += s.substr(pos, s.size() - pos);
-            break;
-        }
-        result += s.substr(pos, new_pos - pos) + replace;
-        pos = new_pos;
-    }
-    s = std::move(result);
-}
-
 struct file_input {
     struct ggml_context * ctx_meta = nullptr;
     struct gguf_context * ctx_gguf = nullptr;
@@ -135,7 +127,7 @@ struct lora_merge_ctx {
 
     lora_merge_ctx(
             std::string & base_fname,
-            std::vector<std::tuple<std::string, float>> & lora_files,
+            std::vector<llama_lora_adapter_info> & lora_files,
             std::string & outfile,
             int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) {
         fout.exceptions(std::ofstream::failbit); // fail fast on write errors
@@ -144,9 +136,9 @@ struct lora_merge_ctx {
             throw std::runtime_error("split model is not yet supported");
         }
 
-        for (auto lora_inp : lora_files) {
-            auto fname = std::get<0>(lora_inp);
-            auto scale = std::get<1>(lora_inp);
+        for (auto & lora_inp : lora_files) {
+            auto fname = lora_inp.path;
+            auto scale = lora_inp.scale;
             std::unique_ptr<file_input> adapter(new file_input(fname, scale));
             check_metadata_lora(adapter.get());
             adapters.push_back(std::move(adapter));
@@ -212,8 +204,7 @@ struct lora_merge_ctx {
         }
 
         // mapping base tensor to out tensor (same shape with base, but different type)
-        // if out_tensor == nullptr, we only copy it
-        std::vector<std::pair<struct ggml_tensor *, struct ggml_tensor *>> base_to_out_tensors;
+        std::vector<tensor_transformation> trans;
         for (auto & it : base_model.tensors) {
             bool t_a = true;
             bool t_b = true;
@@ -226,14 +217,22 @@ struct lora_merge_ctx {
                 // only copy
                 struct ggml_tensor * cpy_tensor = ggml_dup_tensor(ctx_out_ggml, base_tensor);
                 ggml_set_name(cpy_tensor, base_tensor->name);
-                base_to_out_tensors.push_back(std::make_pair(cpy_tensor, nullptr));
+                trans.push_back({
+                    cpy_tensor,
+                    cpy_tensor,
+                    true,
+                });
                 gguf_add_tensor(ctx_out, cpy_tensor);
             } else if (t_a && t_b) {
                 // need merging
                 struct ggml_tensor * out_tensor = ggml_new_tensor(
                     ctx_out_ggml, get_out_tensor_type(base_tensor), GGML_MAX_DIMS, base_tensor->ne);
                 ggml_set_name(out_tensor, base_tensor->name);
-                base_to_out_tensors.push_back(std::make_pair(base_tensor, out_tensor));
+                trans.push_back({
+                    base_tensor,
+                    out_tensor,
+                    false,
+                });
                 gguf_add_tensor(ctx_out, out_tensor);
             } else {
                 throw std::runtime_error("tensor " + it.first + " missing either lora_a or lora_b");
@@ -248,12 +247,12 @@ struct lora_merge_ctx {
 
         // process base model tensors
         size_t n_merged = 0;
-        for (auto & it : base_to_out_tensors) {
-            if (it.second != nullptr) {
-                merge_tensor(it.first, it.second);
+        for (auto & it : trans) {
+            if (!it.is_copy) {
+                merge_tensor(it.in, it.out);
                 n_merged++;
             } else {
-                copy_tensor(it.first);
+                copy_tensor(it.in);
             }
         }
 
@@ -266,7 +265,7 @@ struct lora_merge_ctx {
         }
 
         printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged);
-        printf("%s : wrote %ld tensors to output file\n", __func__, base_to_out_tensors.size());
+        printf("%s : wrote %ld tensors to output file\n", __func__, trans.size());
     }
 
     void copy_tensor(struct ggml_tensor * base) {
@@ -299,6 +298,10 @@ struct lora_merge_ctx {
         for (size_t i = 0; i < adapters.size(); ++i) {
             auto t_a = adapters[i]->get_tensor(name_lora_a);
             auto t_b = adapters[i]->get_tensor(name_lora_b);
+            // TODO: add support for quantized lora
+            if (ggml_is_quantized(t_a->type) || ggml_is_quantized(t_b->type)) {
+                throw std::runtime_error("quantized LoRA adapters is not supported, please retry with f16 or f32");
+            }
             inp_a[i] = ggml_dup_tensor(ctx, t_a);
             inp_b[i] = ggml_dup_tensor(ctx, t_b);
         }
@@ -407,7 +410,7 @@ int main(int argc, char ** argv) {
 
     g_verbose = (params.verbosity == 1);
     try {
-        lora_merge_ctx ctx(params.model, params.lora_adapter, 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/imatrix/imatrix.cpp

@@ -127,7 +127,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         }
         else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
             fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
-            exit(1); //GGML_ASSERT(false);
+            exit(1); //GGML_ABORT("fatal error");
         }
         if (m_params.verbosity > 1) {
             printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type);
@@ -176,7 +176,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         }
         else if (e.values.size() != (size_t)src1->ne[0]) {
             fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
-            exit(1); //GGML_ASSERT(false);
+            exit(1); //GGML_ABORT("fatal error");
         }
         ++e.ncall;
         if (m_params.verbosity > 1) {
@@ -433,8 +433,8 @@ static void process_logits(
 }
 
 static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
     const int n_ctx = llama_n_ctx(ctx);
 
     auto tim1 = std::chrono::high_resolution_clock::now();
@@ -611,10 +611,10 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_model * model;
-    llama_context * ctx;
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
 
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;

examples/infill/infill.cpp

@@ -179,7 +179,10 @@ int main(int argc, char ** argv) {
 
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    model = llama_init.model;
+    ctx = llama_init.context;
 
     if (model == NULL) {
         LOG_TEE("%s: error: unable to load model\n", __func__);
@@ -200,8 +203,8 @@ int main(int argc, char ** argv) {
         LOG_TEE("\n");
         LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
     }
-    const bool add_bos = llama_should_add_bos_token(model);
-    GGML_ASSERT(llama_add_eos_token(model) != 1);
+    const bool add_bos = llama_add_bos_token(model);
+    GGML_ASSERT(!llama_add_eos_token(model));
     LOG("add_bos: %d\n", add_bos);
 
     std::vector<llama_token> embd_inp;

examples/llama-bench/llama-bench.cpp

@@ -16,6 +16,7 @@
 #include <sstream>
 #include <string>
 #include <vector>
+#include <thread>
 
 #include "ggml.h"
 #include "llama.h"
@@ -27,6 +28,14 @@
 #include "ggml-cann.h"
 #endif
 
+#ifdef _WIN32
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
 // utils
 static uint64_t get_time_ns() {
     using clock = std::chrono::high_resolution_clock;
@@ -96,6 +105,27 @@ static std::string get_cpu_info() {
         }
         fclose(f);
     }
+#elif defined(_WIN32)
+    HKEY hKey;
+    if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
+                     TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
+                     0,
+                     KEY_READ,
+                     &hKey) != ERROR_SUCCESS) {
+        // fail to open registry key
+        return "";
+    }
+    char cpu_brand[256];
+    DWORD cpu_brand_size = sizeof(cpu_brand);
+    if (RegQueryValueExA(hKey,
+                        TEXT("ProcessorNameString"),
+                        NULL,
+                        NULL,
+                        (LPBYTE)cpu_brand,
+                        &cpu_brand_size) == ERROR_SUCCESS) {
+        id.assign(cpu_brand, cpu_brand_size);
+    }
+    RegCloseKey(hKey);
 #endif
     // TODO: other platforms
     return id;
@@ -150,7 +180,7 @@ static const char * output_format_str(output_formats format) {
         case JSON:     return "json";
         case MARKDOWN: return "md";
         case SQL:      return "sql";
-        default: GGML_ASSERT(!"invalid output format");
+        default: GGML_ABORT("invalid output format");
     }
 }
 
@@ -176,7 +206,7 @@ static const char * split_mode_str(llama_split_mode mode) {
         case LLAMA_SPLIT_MODE_NONE:  return "none";
         case LLAMA_SPLIT_MODE_LAYER: return "layer";
         case LLAMA_SPLIT_MODE_ROW:   return "row";
-        default: GGML_ASSERT(!"invalid split mode");
+        default: GGML_ABORT("invalid split mode");
     }
 }
 
@@ -196,6 +226,9 @@ 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;
@@ -207,6 +240,8 @@ 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;
@@ -222,6 +257,9 @@ 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},
@@ -233,6 +271,8 @@ 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,
@@ -252,6 +292,9 @@ 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());
     printf("  -rpc, --rpc <rpc_servers>           (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str());
     printf("  -sm, --split-mode <none|layer|row>  (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
@@ -263,6 +306,8 @@ 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("  --prio <0|1|2|3>                    (default: %d)\n", cmd_params_defaults.prio);
+    printf("  --delay <0...N> (seconds)           (default: %d)\n", cmd_params_defaults.delay);
     printf("  -o, --output <csv|json|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
     printf("  -oe, --output-err <csv|json|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");
@@ -309,6 +354,8 @@ 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];
@@ -404,6 +451,27 @@ 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;
@@ -512,6 +580,18 @@ 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;
@@ -556,6 +636,9 @@ 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;
 }
@@ -569,6 +652,9 @@ 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;
@@ -638,7 +724,10 @@ 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;
@@ -652,6 +741,9 @@ 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,
@@ -678,6 +770,9 @@ 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,
@@ -704,6 +799,9 @@ 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,
@@ -740,6 +838,9 @@ 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;
@@ -766,6 +867,9 @@ 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;
@@ -843,13 +947,14 @@ 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;
     }
@@ -858,7 +963,7 @@ 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" ||
@@ -867,6 +972,7 @@ 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;
         }
@@ -899,7 +1005,8 @@ 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),
@@ -1038,7 +1145,7 @@ struct markdown_printer : public printer {
             return -30;
         }
         if (field == "t/s") {
-            return 16;
+            return 20;
         }
         if (field == "size" || field == "params") {
             return 10;
@@ -1120,6 +1227,15 @@ 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");
         }
@@ -1326,7 +1442,7 @@ static std::unique_ptr<printer> create_printer(output_formats format) {
         case SQL:
             return std::unique_ptr<printer>(new sql_printer());
     }
-    GGML_ASSERT(false);
+    GGML_ABORT("fatal error");
 }
 
 int main(int argc, char ** argv) {
@@ -1354,6 +1470,8 @@ 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);
@@ -1399,6 +1517,28 @@ 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);
@@ -1437,6 +1577,8 @@ 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_batch = UInt32(n_threads)
+        ctx_params.n_threads       = Int32(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/CMakeLists.txt

@@ -36,3 +36,10 @@ set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-cli)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TARGET llama-minicpmv-cli)
+add_executable(${TARGET} minicpmv-cli.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-minicpmv-cli)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/llava/README-minicpmv2.5.md

@@ -0,0 +1,99 @@
+## MiniCPM-Llama3-V 2.5
+
+### Prepare models and code
+
+Download [MiniCPM-Llama3-V-2_5](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5) PyTorch model from huggingface to "MiniCPM-Llama3-V-2_5" folder.
+
+Clone llama.cpp:
+```bash
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
+```
+
+### Usage
+
+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/llava/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
+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
+./llama-quantize ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf Q4_K_M
+```
+
+Build for Linux or Mac
+
+```bash
+make
+make llama-minicpmv-cli
+```
+
+Inference on Linux or Mac
+```
+# run f16 version
+./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+
+# run quantized int4 version
+./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+
+# or run in interactive mode
+./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i
+```
+
+### Android
+
+#### Build on Android device using Termux
+We found that build on Android device would bring better runtime performance, so we recommend to build on device.
+
+[Termux](https://github.com/termux/termux-app#installation) is a terminal app on Android device (no root required).
+
+Install tools in Termux:
+```
+apt update && apt upgrade -y
+apt install git make cmake
+```
+
+It's recommended to move your model inside the `~/` directory for best performance:
+```
+cd storage/downloads
+mv model.gguf ~/
+```
+
+#### Building the Project using Android NDK
+Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
+
+Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
+
+```bash
+mkdir build-android
+cd build-android
+export NDK=/your_ndk_path
+cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
+make
+```
+
+Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
+
+Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
+
+(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
+```
+$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
+$cd /data/data/com.termux/files/home/bin
+$chmod +x ./*
+```
+
+Download models and push them to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
+
+```
+$mv /sdcard/llama.cpp/ggml-model-Q4_K_M.gguf /data/data/com.termux/files/home/model/
+$mv /sdcard/llama.cpp/mmproj-model-f16.gguf /data/data/com.termux/files/home/model/
+```
+
+Now, you can start chatting:
+```
+$cd /data/data/com.termux/files/home/bin
+$./llama-minicpmv-cli -m ../model/ggml-model-Q4_K_M.gguf --mmproj ../model/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+```

examples/llava/README-minicpmv2.6.md

@@ -0,0 +1,107 @@
+## MiniCPM-V 2.6
+
+### Prepare models and code
+
+Download [MiniCPM-V-2_6](https://huggingface.co/openbmb/MiniCPM-V-2_6) PyTorch model from huggingface to "MiniCPM-V-2_6" folder.
+
+Clone llama.cpp:
+```bash
+git clone git@github.com:OpenBMB/llama.cpp.git
+cd llama.cpp
+git checkout minicpmv-main
+```
+
+### Usage of MiniCPM-V 2.6
+
+Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) by us)
+
+```bash
+python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-V-2_6
+python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 3
+python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model
+
+# quantize int4 version
+./llama-quantize ../MiniCPM-V-2_6/model/ggml-model-f16.gguf ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M
+```
+
+Build for Linux or Mac
+
+```bash
+make
+make llama-minicpmv-cli
+```
+
+Inference on Linux or Mac
+```
+# run f16 version
+./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+
+# run quantized int4 version
+./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+
+# or run in interactive mode
+./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i
+```
+
+### Video
+Install FFmpeg
+```
+brew install ffmpeg
+brew install pkg-config
+```
+
+### Android
+
+#### Build on Android device using Termux
+We found that build on Android device would bring better runtime performance, so we recommend to build on device.
+
+[Termux](https://github.com/termux/termux-app#installation) is a terminal app on Android device (no root required).
+
+Install tools in Termux:
+```
+apt update && apt upgrade -y
+apt install git make cmake
+```
+
+It's recommended to move your model inside the `~/` directory for best performance:
+```
+cd storage/downloads
+mv model.gguf ~/
+```
+
+#### Building the Project using Android NDK
+Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
+
+Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
+
+```bash
+mkdir build-android
+cd build-android
+export NDK=/your_ndk_path
+cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
+make
+```
+
+Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
+
+Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
+
+(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
+```
+$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
+$cd /data/data/com.termux/files/home/bin
+$chmod +x ./*
+```
+
+Download models and push them to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
+
+```
+$mv /sdcard/llama.cpp/ggml-model-Q4_K_M.gguf /data/data/com.termux/files/home/model/
+$mv /sdcard/llama.cpp/mmproj-model-f16.gguf /data/data/com.termux/files/home/model/
+```
+
+Now, you can start chatting:
+```
+$cd /data/data/com.termux/files/home/bin
+$./llama-minicpmv-cli -m ../model/ggml-model-Q4_K_M.gguf --mmproj ../model/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+```

examples/llava/clip.cpp

@@ -20,6 +20,10 @@
 #include "ggml-cann.h"
 #endif
 
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb_image.h"
 
@@ -74,26 +78,28 @@ static std::string format(const char * fmt, ...) {
 // key constants
 //
 
-#define KEY_FTYPE          "general.file_type"
-#define KEY_NAME           "general.name"
-#define KEY_DESCRIPTION    "general.description"
-#define KEY_HAS_TEXT_ENC   "clip.has_text_encoder"
-#define KEY_HAS_VIS_ENC    "clip.has_vision_encoder"
-#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector"
-#define KEY_USE_GELU       "clip.use_gelu"
-#define KEY_N_EMBD         "clip.%s.embedding_length"
-#define KEY_N_FF           "clip.%s.feed_forward_length"
-#define KEY_N_BLOCK        "clip.%s.block_count"
-#define KEY_N_HEAD         "clip.%s.attention.head_count"
-#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon"
-#define KEY_PROJ_DIM       "clip.%s.projection_dim"
-#define KEY_TOKENS         "tokenizer.ggml.tokens"
-#define KEY_N_POSITIONS    "clip.text.context_length"
-#define KEY_IMAGE_SIZE     "clip.vision.image_size"
-#define KEY_PATCH_SIZE     "clip.vision.patch_size"
-#define KEY_IMAGE_MEAN     "clip.vision.image_mean"
-#define KEY_IMAGE_STD      "clip.vision.image_std"
-#define KEY_PROJ_TYPE      "clip.projector_type"
+#define KEY_FTYPE               "general.file_type"
+#define KEY_NAME                "general.name"
+#define KEY_DESCRIPTION         "general.description"
+#define KEY_HAS_TEXT_ENC        "clip.has_text_encoder"
+#define KEY_HAS_VIS_ENC         "clip.has_vision_encoder"
+#define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
+#define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
+#define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
+#define KEY_USE_GELU            "clip.use_gelu"
+#define KEY_N_EMBD              "clip.%s.embedding_length"
+#define KEY_N_FF                "clip.%s.feed_forward_length"
+#define KEY_N_BLOCK             "clip.%s.block_count"
+#define KEY_N_HEAD              "clip.%s.attention.head_count"
+#define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
+#define KEY_PROJ_DIM            "clip.%s.projection_dim"
+#define KEY_TOKENS              "tokenizer.ggml.tokens"
+#define KEY_N_POSITIONS         "clip.text.context_length"
+#define KEY_IMAGE_SIZE          "clip.vision.image_size"
+#define KEY_PATCH_SIZE          "clip.vision.patch_size"
+#define KEY_IMAGE_MEAN          "clip.vision.image_mean"
+#define KEY_IMAGE_STD           "clip.vision.image_std"
+#define KEY_PROJ_TYPE           "clip.projector_type"
 
 #define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
 #define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
@@ -127,12 +133,20 @@ static std::string format(const char * fmt, ...) {
 #define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
 #define TN_IMAGE_NEWLINE   "model.image_newline"
 
+#define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
+#define TN_MINICPMV_QUERY "resampler.query"
+#define TN_MINICPMV_PROJ "resampler.proj.weight"
+#define TN_MINICPMV_KV_PROJ "resampler.kv.weight"
+#define TN_MINICPMV_ATTN "resampler.attn.%s.%s"
+#define TN_MINICPMV_LN "resampler.ln_%s.%s"
+
 
 enum projector_type {
     PROJECTOR_TYPE_MLP,
     PROJECTOR_TYPE_MLP_NORM,
     PROJECTOR_TYPE_LDP,
     PROJECTOR_TYPE_LDPV2,
+    PROJECTOR_TYPE_RESAMPLER,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -140,6 +154,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_MLP, "mlp" },
     { PROJECTOR_TYPE_LDP, "ldp" },
     { PROJECTOR_TYPE_LDPV2, "ldpv2"},
+    { PROJECTOR_TYPE_RESAMPLER, "resampler"},
 };
 
 
@@ -200,17 +215,20 @@ 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) {
-    std::string result;
-    for (size_t pos = 0; ; pos += search.length()) {
-        auto new_pos = s.find(search, pos);
-        if (new_pos == std::string::npos) {
-            result += s.substr(pos, s.size() - pos);
-            break;
-        }
-        result += s.substr(pos, new_pos - pos) + replace;
-        pos = new_pos;
+    if (search.empty()) {
+        return;
     }
-    s = std::move(result);
+    std::string builder;
+    builder.reserve(s.length());
+    size_t pos = 0;
+    size_t last_pos = 0;
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
+        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) {
@@ -492,12 +510,34 @@ struct clip_vision_model {
     struct ggml_tensor * mm_model_mlp_2_b;
     struct ggml_tensor * mm_model_peg_0_w;
     struct ggml_tensor * mm_model_peg_0_b;
+
+    // MINICPMV projection
+    struct ggml_tensor * mm_model_pos_embed_k;
+    struct ggml_tensor * mm_model_query;
+    struct ggml_tensor * mm_model_proj;
+    struct ggml_tensor * mm_model_kv_proj;
+    struct ggml_tensor * mm_model_attn_q_w;
+    struct ggml_tensor * mm_model_attn_q_b;
+    struct ggml_tensor * mm_model_attn_k_w;
+    struct ggml_tensor * mm_model_attn_k_b;
+    struct ggml_tensor * mm_model_attn_v_w;
+    struct ggml_tensor * mm_model_attn_v_b;
+    struct ggml_tensor * mm_model_attn_o_w;
+    struct ggml_tensor * mm_model_attn_o_b;
+    struct ggml_tensor * mm_model_ln_q_w;
+    struct ggml_tensor * mm_model_ln_q_b;
+    struct ggml_tensor * mm_model_ln_kv_w;
+    struct ggml_tensor * mm_model_ln_kv_b;
+    struct ggml_tensor * mm_model_ln_post_w;
+    struct ggml_tensor * mm_model_ln_post_b;
 };
 
 struct clip_ctx {
     bool has_text_encoder    = false;
     bool has_vision_encoder  = false;
     bool has_llava_projector = false;
+    bool has_minicpmv_projector = false;
+    int minicpmv_version = 2;
 
     struct clip_vision_model vision_model;
     projector_type proj_type = PROJECTOR_TYPE_MLP;
@@ -522,9 +562,11 @@ struct clip_ctx {
 
     ggml_backend_t backend       = NULL;
     ggml_gallocr_t compute_alloc = NULL;
+
+    struct clip_image_size * load_image_size;
 };
 
-static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs) {
+static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) {
     if (!ctx->has_vision_encoder) {
         LOG_TEE("This gguf file seems to have no vision encoder\n");
         return nullptr;
@@ -533,20 +575,33 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     const auto & model = ctx->vision_model;
     const auto & hparams = model.hparams;
 
-    const int image_size           = hparams.image_size;
+    const int image_size = hparams.image_size;
+    int image_size_width  = image_size;
+    int image_size_height = image_size;
+    if (ctx->has_minicpmv_projector) {
+        if (load_image_size == nullptr) {
+            load_image_size = clip_image_size_init();
+        }
+        LOG_TEE("%s: %d %d\n", __func__, load_image_size->width, load_image_size->height);
+        image_size_width  = load_image_size->width;
+        image_size_height = load_image_size->height;
+        if (is_inf) {
+            image_size_width  = imgs->data->nx;
+            image_size_height = imgs->data->ny;
+        }
+    }
     const int patch_size           = hparams.patch_size;
-    const int num_patches          = ((image_size / patch_size) * (image_size / patch_size));
-    const int num_patches_per_side = image_size / patch_size; GGML_UNUSED(num_patches_per_side);
+    const int num_patches          = ((image_size_width / patch_size) * (image_size_height / patch_size));
     const int num_positions        = num_patches + (ctx->has_class_embedding ? 1 : 0);
     const int hidden_size          = hparams.hidden_size;
     const int n_head               = hparams.n_head;
     const int d_head               = hidden_size / n_head;
-    const int n_layer              = hparams.n_layer;
+    int n_layer                    = hparams.n_layer;
     const float eps                = hparams.eps;
 
     const int batch_size = imgs->size;
 
-    if (ctx->has_llava_projector) {
+    if (ctx->has_llava_projector || ctx->has_minicpmv_projector) {
         GGML_ASSERT(batch_size == 1);
     }
 
@@ -559,7 +614,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     struct ggml_context * ctx0 = ggml_init(params);
     struct ggml_cgraph * gf = ggml_new_graph(ctx0);
 
-    struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size, image_size, 3, batch_size);
+    struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3, batch_size);
     ggml_set_name(inp_raw, "inp_raw");
     ggml_set_input(inp_raw);
 
@@ -572,19 +627,21 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         // inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp));
         inp = ggml_add(ctx0, inp, model.patch_bias);
     }
-
-    // concat class_embeddings and patch_embeddings
     struct ggml_tensor * embeddings = inp;
-    if (ctx->has_class_embedding) {
-        embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
-        ggml_set_name(embeddings, "embeddings");
-        ggml_set_input(embeddings);
-        embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
-                embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
-        embeddings = ggml_acc(ctx0, embeddings, inp,
-                embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
-    }
+    struct ggml_tensor * pos_embed = nullptr;
 
+    if (ctx->has_llava_projector) {
+        // concat class_embeddings and patch_embeddings
+        if (ctx->has_class_embedding) {
+            embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
+            ggml_set_name(embeddings, "embeddings");
+            ggml_set_input(embeddings);
+            embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
+                    embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
+            embeddings = ggml_acc(ctx0, embeddings, inp,
+                    embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
+        }
+    }
 
     struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
     ggml_set_name(positions, "positions");
@@ -593,6 +650,19 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     embeddings =
         ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions));
 
+    if (ctx->has_minicpmv_projector) {
+        int pos_w = image_size_width/patch_size;
+        int pos_h = image_size_height/patch_size;
+        if (ctx->minicpmv_version == 2) {
+            pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 4096, pos_w * pos_h, 1);
+        }
+        else if (ctx->minicpmv_version == 3) {
+            pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 3584, pos_w * pos_h, 1);
+        }
+        ggml_set_name(pos_embed, "pos_embed");
+        ggml_set_input(pos_embed);
+    }
+
     // pre-layernorm
     if (ctx->has_pre_norm) {
         embeddings = ggml_norm(ctx0, embeddings, eps);
@@ -602,6 +672,9 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     }
 
     // loop over layers
+    if (ctx->has_minicpmv_projector) {
+        n_layer += 1;
+    }
     for (int il = 0; il < n_layer - 1; il++) {
         struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states
 
@@ -691,7 +764,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     }
 
     // llava projector
-    {
+    if (ctx->has_llava_projector) {
         embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
 
         struct ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_patches);
@@ -712,8 +785,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             embeddings = ggml_gelu(ctx0, embeddings);
             embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
             embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
-
-        } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
+        }
+        else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
             embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
             embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
             // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false);
@@ -868,6 +941,75 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             peg_0 = ggml_reshape_3d(ctx0, peg_0, peg_0->ne[0], peg_0->ne[1] * peg_0->ne[2], peg_0->ne[3]);
             embeddings = peg_0;
         }
+        else {
+            GGML_ABORT("fatal error");
+        }
+    }
+    // minicpmv projector
+    else if (ctx->has_minicpmv_projector)
+    {
+        if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+            struct ggml_tensor * q = model.mm_model_query;
+            { // layernorm
+                q = ggml_norm(ctx0, q, eps);
+                q = ggml_add(ctx0, ggml_mul(ctx0, q, model.mm_model_ln_q_w), model.mm_model_ln_q_b);
+            }
+            struct ggml_tensor * v = ggml_mul_mat(ctx0, model.mm_model_kv_proj, embeddings);
+            { // layernorm
+                v = ggml_norm(ctx0, v, eps);
+                v = ggml_add(ctx0, ggml_mul(ctx0, v, model.mm_model_ln_kv_w), model.mm_model_ln_kv_b);
+            }
+            struct ggml_tensor * k;
+            { // position
+                // q = ggml_add(ctx0, q, model.mm_model_pos_embed);
+                k = ggml_add(ctx0, v, pos_embed);
+            }
+
+            { // attention
+                int hidden_size = 4096;
+                const int d_head = 128;
+                int n_head = hidden_size/d_head;
+                int num_query = 96;
+                if (ctx->minicpmv_version == 2) {
+                    hidden_size = 4096;
+                    n_head = hidden_size/d_head;
+                    num_query = 96;
+                }
+                else if (ctx->minicpmv_version == 3) {
+                    hidden_size = 3584;
+                    n_head = hidden_size/d_head;
+                    num_query = 64;
+                }
+
+                struct ggml_tensor * Q = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q), model.mm_model_attn_q_b);
+                Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
+                struct ggml_tensor * K = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_k_w, k), model.mm_model_attn_k_b);
+                struct ggml_tensor * V = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_v_w, v), model.mm_model_attn_v_b);
+                // permute
+                Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_query, batch_size);
+                Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
+                Q = ggml_reshape_3d(ctx0, Q, d_head, num_query, n_head * batch_size);
+                K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size);
+                K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
+                K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size);
+                V = ggml_reshape_4d(ctx0, V, d_head, n_head, num_positions, batch_size);
+                V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3));
+                V = ggml_reshape_3d(ctx0, V, num_positions, d_head, n_head * batch_size);
+                struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+                KQ = ggml_soft_max_inplace(ctx0, KQ);
+                struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);
+                KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_query, n_head, batch_size);
+                KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+                KQV = ggml_cont_3d(ctx0, KQV, hidden_size, num_query, batch_size);
+
+                embeddings = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_o_w, KQV), model.mm_model_attn_o_b);
+            }
+            { // layernorm
+                embeddings = ggml_norm(ctx0, embeddings, eps);
+                embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_post_w), model.mm_model_ln_post_b);
+            }
+            embeddings = ggml_mul_mat(ctx0, model.mm_model_proj, embeddings);
+        }
         else {
             GGML_ASSERT(false);
         }
@@ -976,7 +1118,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
     }
 
-    clip_ctx * new_clip = new clip_ctx;
+    clip_ctx * new_clip = new clip_ctx{};
 
     // update projector type
     {
@@ -1010,6 +1152,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     LOG_TEE("%s: CLIP using CANN backend\n", __func__);
 #endif
 
+#ifdef GGML_USE_VULKAN
+    new_clip->backend = ggml_backend_vk_init(0);
+    LOG_TEE("%s: CLIP using Vulkan backend\n", __func__);
+#endif
 
     if (!new_clip->backend) {
         new_clip->backend = ggml_backend_cpu_init();
@@ -1029,7 +1175,18 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             new_clip->has_llava_projector = gguf_get_val_bool(ctx, idx);
         }
 
-        GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search
+        idx = gguf_find_key(ctx, KEY_HAS_MINICPMV_PROJ);
+        if (idx != -1) {
+            new_clip->has_minicpmv_projector = gguf_get_val_bool(ctx, idx);
+        }
+
+        idx = gguf_find_key(ctx, KEY_MINICPMV_VERSION);
+        if (idx != -1) {
+            new_clip->minicpmv_version = gguf_get_val_i32(ctx, idx);
+        }
+
+        // GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search
+
         GGML_ASSERT(new_clip->has_vision_encoder);
         GGML_ASSERT(!new_clip->has_text_encoder);
 
@@ -1040,6 +1197,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             LOG_TEE("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
             LOG_TEE("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
             LOG_TEE("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
+            LOG_TEE("%s: minicpmv_projector:  %d\n", __func__, new_clip->has_minicpmv_projector);
             LOG_TEE("%s: model size:     %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
             LOG_TEE("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
         }
@@ -1281,6 +1439,27 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             vision_model.mm_model_peg_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "weight"));
             vision_model.mm_model_peg_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "bias"));
         }
+        else if (new_clip->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+            // vision_model.mm_model_pos_embed = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD);
+            vision_model.mm_model_pos_embed_k = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD_K);
+            vision_model.mm_model_query = get_tensor(new_clip->ctx_data, TN_MINICPMV_QUERY);
+            vision_model.mm_model_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_PROJ);
+            vision_model.mm_model_kv_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_KV_PROJ);
+            vision_model.mm_model_attn_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "weight"));
+            vision_model.mm_model_attn_k_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "weight"));
+            vision_model.mm_model_attn_v_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "weight"));
+            vision_model.mm_model_attn_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "bias"));
+            vision_model.mm_model_attn_k_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "bias"));
+            vision_model.mm_model_attn_v_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "bias"));
+            vision_model.mm_model_attn_o_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "weight"));
+            vision_model.mm_model_attn_o_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "bias"));
+            vision_model.mm_model_ln_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "weight"));
+            vision_model.mm_model_ln_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "bias"));
+            vision_model.mm_model_ln_kv_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "weight"));
+            vision_model.mm_model_ln_kv_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "bias"));
+            vision_model.mm_model_ln_post_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "weight"));
+            vision_model.mm_model_ln_post_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "bias"));
+        }
         else {
             std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type];
             throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
@@ -1319,7 +1498,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend));
         clip_image_f32_batch batch;
         batch.size = 1;
-        ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch);
+        ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false);
         ggml_gallocr_reserve(new_clip->compute_alloc, gf);
         size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
         LOG_TEE("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
@@ -1328,6 +1507,17 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     return new_clip;
 }
 
+void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size) {
+    ctx_clip->load_image_size = load_image_size;
+}
+
+struct clip_image_size * clip_image_size_init() {
+    struct clip_image_size * load_image_size = new struct clip_image_size();
+    load_image_size->width = 448;
+    load_image_size->height = 448;
+    return load_image_size;
+}
+
 struct clip_image_u8 * clip_image_u8_init() {
     return new clip_image_u8();
 }
@@ -1433,7 +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));
 }
 
@@ -1598,9 +1788,182 @@ static std::vector<clip_image_u8*> divide_to_patches_u8(const clip_image_u8 & im
     return patches;
 }
 
+static int ensure_divide(int length, int patch_size) {
+    return std::max(static_cast<int>(std::round(static_cast<float>(length) / patch_size) * patch_size), patch_size);
+}
+
+static std::pair<int, int> uhd_find_best_resize(std::pair<int, int> original_size, int scale_resolution, int patch_size, bool allow_upscale = false) {
+    int width = original_size.first;
+    int height = original_size.second;
+    if ((width * height > scale_resolution * scale_resolution) || allow_upscale) {
+        float r = static_cast<float>(width) / height;
+        height = static_cast<int>(scale_resolution / std::sqrt(r));
+        width = static_cast<int>(height * r);
+    }
+    int best_width = ensure_divide(width, patch_size);
+    int best_height = ensure_divide(height, patch_size);
+    return std::make_pair(best_width, best_height);
+}
+
+static std::pair<int, int> uhd_get_refine_size(std::pair<int, int> original_size, std::pair<int, int> grid, int scale_resolution, int patch_size, bool allow_upscale = false) {
+    int width, height;
+    std::tie(width, height) = original_size;
+    int grid_x, grid_y;
+    std::tie(grid_x, grid_y) = grid;
+
+    int refine_width = ensure_divide(width, grid_x);
+    int refine_height = ensure_divide(height, grid_y);
+
+    int grid_width = refine_width / grid_x;
+    int grid_height = refine_height / grid_y;
+
+   // auto best_grid_size = find_best_resize(std::make_tuple(grid_width, grid_height), scale_resolution, patch_size, allow_upscale); (old line)
+    auto best_grid_size = uhd_find_best_resize(std::make_pair(grid_width, grid_height), scale_resolution, patch_size, allow_upscale); // (new line) => fixes conversion for make_tuple to make_pair
+    int best_grid_width, best_grid_height;
+    std::tie(best_grid_width, best_grid_height) = best_grid_size;
+
+  //  std::pair<int, int> refine_size = std::make_tuple(best_grid_width * grid_x, best_grid_height * grid_y); (old line)
+    std::pair<int, int> refine_size = std::make_pair(best_grid_width * grid_x, best_grid_height * grid_y); // (new line)
+    return refine_size;
+}
+
+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}) {
+        if (i == 1 || i > max_slice_nums) {
+            continue;
+        }
+        candidate_split_grids_nums.push_back(i);
+    }
+
+    std::vector<std::pair<int, int>> candidate_grids;
+    for (int split_grids_nums : candidate_split_grids_nums) {
+        int m = 1;
+        while (m <= split_grids_nums) {
+            if (split_grids_nums % m == 0) {
+                candidate_grids.emplace_back(m, split_grids_nums / m);
+            }
+            ++m;
+        }
+    }
+
+    std::pair<int, int> best_grid{1, 1};
+    float min_error = std::numeric_limits<float>::infinity();
+    for (const auto& grid : candidate_grids) {
+        float error = std::abs(log_ratio - std::log(1.0 * grid.first / grid.second));
+        if (error < min_error) {
+            best_grid = grid;
+            min_error = error;
+        }
+    }
+    return best_grid;
+}
+
+// inspired from LLaVA-UHD:
+//    -> https://arxiv.org/pdf/2403.11703
+//    -> https://github.com/thunlp/LLaVA-UHD
+//    -> https://github.com/thunlp/LLaVA-UHD/blob/302301bc2175f7e717fb8548516188e89f649753/llava_uhd/train/llava-uhd/slice_logic.py#L118
+static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_image_u8 * img, const int max_slice_nums=9, const int scale_resolution=448, const int patch_size=14) {
+    const std::pair<int, int> original_size={img->nx,img->ny};
+    const int original_width = img->nx;
+    const int original_height = img->ny;
+    const float log_ratio = log(1.0*original_width/original_height);
+    const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution);
+    const int multiple = fmin(ceil(ratio), max_slice_nums);
+
+    std::vector<std::vector<clip_image_u8 *>> images;
+    LOG_TEE("%s: multiple %d\n", __func__, multiple);
+    images.push_back(std::vector<clip_image_u8 *>());
+
+    if (multiple <= 1) {
+        auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size, true);
+        clip_image_u8 * source_image = clip_image_u8_init();
+        bicubic_resize(*img, *source_image, best_size.first, best_size.second);
+        // source_image = image.resize(best_size, Image.Resampling.BICUBIC)
+        images[images.size()-1].push_back(source_image);
+    }
+    else if (multiple > 1) {
+        auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size);
+        clip_image_u8 * source_image = clip_image_u8_init();
+        bicubic_resize(*img, *source_image, best_size.first, best_size.second);
+        // source_image = image.copy().resize(best_resize, Image.Resampling.BICUBIC)
+        LOG_TEE("%s: image_size: %d %d; source_image size: %d %d\n", __func__, img->nx, img->ny, best_size.first, best_size.second);
+        images[images.size()-1].push_back(source_image);
+
+        std::pair<int, int> best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio);
+        LOG_TEE("%s: image_size: %d %d; best_grid: %d %d\n", __func__, img->nx, img->ny, best_grid.first, best_grid.second);
+
+        auto refine_size = uhd_get_refine_size(original_size, best_grid, scale_resolution, patch_size, true);
+        clip_image_u8 * refine_image = clip_image_u8_init();
+        bicubic_resize(*img, *refine_image, refine_size.first, refine_size.second);
+
+        LOG_TEE("%s: refine_image_size: %d %d; refine_size: %d %d\n", __func__, refine_image->nx, refine_image->ny, refine_size.first, refine_size.second);
+
+        // split_to_patches
+        int width = refine_image->nx;
+        int height = refine_image->ny;
+        int grid_x = int(width / best_grid.first);
+        int grid_y = int(height / best_grid.second);
+        for (int patches_i = 0, ic = 0; patches_i < height && ic < best_grid.second; patches_i += grid_y, ic += 1){
+            images.push_back(std::vector<clip_image_u8 *>());
+            for(int patches_j = 0, jc = 0; patches_j < width && jc < best_grid.first; patches_j += grid_x, jc += 1){
+                clip_image_u8 * patch = clip_image_u8_init();
+                patch->nx = grid_x;
+                patch->ny = grid_y;
+                patch->buf.resize(3 * patch->nx * patch->ny);
+                for (int y = patches_i; y < patches_i + grid_y; ++y) {
+                    for (int x = patches_j; x < patches_j + grid_x; ++x) {
+                        const int i = 3 * (y * refine_image->nx + x);
+                        const int j = 3 * ((y-patches_i) * patch->nx + (x-patches_j));
+                        patch->buf[j]   = refine_image->buf[i];
+                        patch->buf[j+1] = refine_image->buf[i+1];
+                        patch->buf[j+2] = refine_image->buf[i+2];
+                    }
+                }
+                images[images.size()-1].push_back(patch);
+            }
+        }
+    }
+    return images;
+}
+
+int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip) {
+    const int max_slice_nums=9;
+    const int scale_resolution=448;
+    const int original_width = ctx_clip->load_image_size->width;
+    const int original_height = ctx_clip->load_image_size->height;
+    const float log_ratio = log(1.0*original_width/original_height);
+    const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution);
+    const int multiple = fmin(ceil(ratio), max_slice_nums);
+    std::pair<int, int> best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio);
+    return best_grid.first;
+}
+
 // returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector
 // res_imgs memory is being allocated here, previous allocations will be freed if found
 bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch * res_imgs) {
+
+    if(clip_is_minicpmv(ctx)){
+        int max_slice_nums = 9;
+        std::vector<std::vector<clip_image_u8 *>> imgs = uhd_slice_image(img, max_slice_nums);
+        res_imgs->size = 0;
+        for (size_t i = 0; i < imgs.size(); ++i){
+            res_imgs->size += imgs[i].size();
+        }
+        res_imgs->data = new clip_image_f32[res_imgs->size];
+        int idx = 0;
+        for (size_t i = 0; i < imgs.size(); ++i) {
+            for (size_t j = 0; j < imgs[i].size(); ++j) {
+                LOG_TEE("%s: %d %d\n", __func__,imgs[i][j]->nx,imgs[i][j]->ny);
+                clip_image_f32 * res = clip_image_f32_init();
+                normalize_image_u8_to_f32(imgs[i][j], res, ctx->image_mean, ctx->image_std);
+                res_imgs->data[idx++] = *res;
+                clip_image_f32_free(res);
+            }
+        }
+        return true;
+    }
+
     bool pad_to_square = true;
     if (!ctx->has_vision_encoder) {
         LOG_TEE("This gguf file seems to have no vision encoder\n");
@@ -1816,11 +2179,104 @@ int clip_n_patches(const struct clip_ctx * ctx) {
 
     if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
         n_patches /= 4;
+    } else if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+        if (ctx->minicpmv_version == 2) {
+            n_patches = 96;
+        }
+        else if (ctx->minicpmv_version == 3) {
+            n_patches = 64;
+        }
     }
 
     return n_patches;
 }
 
+static std::vector<std::vector<std::vector<float>>> get_1d_sincos_pos_embed_from_grid_new(int embed_dim, const std::vector<std::vector<float>> & pos) {
+    assert(embed_dim % 2 == 0);
+    int H = pos.size();
+    int W = pos[0].size();
+
+    std::vector<float> omega(embed_dim / 2);
+    for (int i = 0; i < embed_dim / 2; ++i) {
+        omega[i] = 1.0 / pow(10000.0, static_cast<float>(i) / (embed_dim / 2));
+    }
+
+    std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));
+    for (int h = 0; h < H; ++h) {
+        for (int w = 0; w < W; ++w) {
+            for (int d = 0; d < embed_dim / 2; ++d) {
+                float out_value = pos[h][w] * omega[d];
+                emb[h][w][d] = sin(out_value);
+                emb[h][w][d + embed_dim / 2] = cos(out_value);
+            }
+        }
+    }
+
+    return emb;
+}
+
+static std::vector<std::vector<std::vector<float>>> get_2d_sincos_pos_embed_from_grid(int embed_dim, const std::vector<std::vector<std::vector<float>>> & grid) {
+    assert(embed_dim % 2 == 0);
+    std::vector<std::vector<std::vector<float>>> emb_h = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[0]); // (H, W, D/2)
+    std::vector<std::vector<std::vector<float>>> emb_w = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[1]); // (H, W, D/2)
+
+    int H = emb_h.size();
+    int W = emb_h[0].size();
+    std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));
+
+    for (int h = 0; h < H; ++h) {
+        for (int w = 0; w < W; ++w) {
+            for (int d = 0; d < embed_dim / 2; ++d) {
+                emb[h][w][d] = emb_h[h][w][d];
+                emb[h][w][d + embed_dim / 2] = emb_w[h][w][d];
+            }
+        }
+    }
+    return emb;
+}
+
+static std::vector<std::vector<float>> get_2d_sincos_pos_embed(int embed_dim, const std::pair<int, int> image_size) {
+    int grid_h_size = image_size.first;
+    int grid_w_size = image_size.second;
+
+    std::vector<float> grid_h(grid_h_size);
+    std::vector<float> grid_w(grid_w_size);
+
+    for (int i = 0; i < grid_h_size; ++i) {
+        grid_h[i] = static_cast<float>(i);
+    }
+    for (int i = 0; i < grid_w_size; ++i) {
+        grid_w[i] = static_cast<float>(i);
+    }
+
+    std::vector<std::vector<float>> grid(grid_h_size, std::vector<float>(grid_w_size));
+    for (int h = 0; h < grid_h_size; ++h) {
+        for (int w = 0; w < grid_w_size; ++w) {
+            grid[h][w] = grid_w[w];
+        }
+    }
+    std::vector<std::vector<std::vector<float>>> grid_2d = {grid, grid};
+    for (int h = 0; h < grid_h_size; ++h) {
+        for (int w = 0; w < grid_w_size; ++w) {
+            grid_2d[0][h][w] = grid_h[h];
+            grid_2d[1][h][w] = grid_w[w];
+        }
+    }
+
+    std::vector<std::vector<std::vector<float>>> pos_embed_3d = get_2d_sincos_pos_embed_from_grid(embed_dim, grid_2d);
+
+    int H = image_size.first;
+    int W = image_size.second;
+    std::vector<std::vector<float>> pos_embed_2d(H * W, std::vector<float>(embed_dim));
+    for (int h = 0; h < H; ++h) {
+        for (int w = 0; w < W; ++w) {
+            pos_embed_2d[w * H + h] = pos_embed_3d[h][w];
+        }
+    }
+
+    return pos_embed_2d;
+}
+
 bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
     if (!ctx->has_vision_encoder) {
         LOG_TEE("This gguf file seems to have no vision encoder\n");
@@ -1843,19 +2299,33 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     if (ctx->has_llava_projector) {
         GGML_ASSERT(batch_size == 1); // TODO: support multiple images
     }
+    if (ctx->has_minicpmv_projector) {
+        GGML_ASSERT(batch_size == 1);
+    }
 
     // build the inference graph
-    ggml_cgraph * gf = clip_image_build_graph(ctx, imgs);
+    ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, ctx->load_image_size, true);
     ggml_gallocr_alloc_graph(ctx->compute_alloc, gf);
 
     // set inputs
     const auto & model = ctx->vision_model;
     const auto & hparams = model.hparams;
 
-    const int image_size    = hparams.image_size;
+    const int image_size = hparams.image_size;
+    int image_size_width  = image_size;
+    int image_size_height = image_size;
+    if (ctx->has_minicpmv_projector) {
+        image_size_width  = imgs->data[0].nx;
+        image_size_height = imgs->data[0].ny;
+    }
     const int patch_size    = hparams.patch_size;
-    const int num_patches   = ((image_size / patch_size) * (image_size / patch_size));
+    const int num_patches   = ((image_size_width / patch_size) * (image_size_height / patch_size));
     const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0);
+    if(ctx->load_image_size==nullptr){
+        ctx->load_image_size= clip_image_size_init();
+    }
+    const int pos_w = ctx->load_image_size->width/patch_size;
+    const int pos_h = ctx->load_image_size->height/patch_size;
 
     {
         struct ggml_tensor * inp_raw = ggml_graph_get_tensor(gf, "inp_raw");
@@ -1864,7 +2334,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         for (size_t i = 0; i < imgs->size; i++) {
             const int nx = imgs->data[i].nx;
             const int ny = imgs->data[i].ny;
-            GGML_ASSERT(nx == image_size && ny == image_size);
+            if (!ctx->has_minicpmv_projector) {
+                GGML_ASSERT(nx == image_size && ny == image_size);
+            }
 
             const int n = nx * ny;
 
@@ -1881,37 +2353,87 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         ggml_backend_tensor_set(inp_raw, data, 0, ggml_nbytes(inp_raw));
         free(data);
     }
+    if (ctx->has_minicpmv_projector) {
+        {
+            // inspired from siglip:
+            //    -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit
+            //    -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316
+            struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
+            int* positions_data = (int*)malloc(ggml_nbytes(positions));
+            int bucket_coords_h[70];
+            int bucket_coords_w[70];
+            for (int i = 0; i < pos_h; i++){
+                bucket_coords_h[i] = std::floor(70.0*i/pos_h);
+            }
+            for (int i = 0; i < pos_w; i++){
+                bucket_coords_w[i] = std::floor(70.0*i/pos_w);
+            }
+            for (int i = 0, id = 0; i < pos_h; i++){
+                for (int j = 0; j < pos_w; j++){
+                    positions_data[id++] = bucket_coords_h[i]*70 + bucket_coords_w[j];
+                }
+            }
+            ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+            free(positions_data);
+        }
 
-    {
-        if (ctx->has_class_embedding) {
-            struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
+        {
+            // inspired from resampler of Qwen-VL:
+            //    -> https://huggingface.co/Qwen/Qwen-VL/tree/main
+            //    -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23
+            struct ggml_tensor * pos_embed = ggml_graph_get_tensor(gf, "pos_embed");
+            int embed_dim = 4096;
+            if (ctx->minicpmv_version == 2) {
+                embed_dim = 4096;
+            }
+            else if (ctx->minicpmv_version == 3) {
+                embed_dim = 3584;
+            }
+            auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));
 
-            void* zero_mem = malloc(ggml_nbytes(embeddings));
-            memset(zero_mem, 0, ggml_nbytes(embeddings));
-            ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
-            free(zero_mem);
+            float * pos_embed_data = (float *)malloc(ggml_nbytes(pos_embed));
+            for(int i=0;i<pos_w * pos_h;++i){
+                for(int j=0;j<embed_dim;++j){
+                    pos_embed_data[i*embed_dim+j]=pos_embed_t[i][j];
+                }
+            }
+
+            ggml_backend_tensor_set(pos_embed, pos_embed_data, 0, ggml_nbytes(pos_embed));
+            free(pos_embed_data);
         }
     }
+    else{
+        {
+            if (ctx->has_class_embedding) {
+                struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
 
-    {
-        struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
-
-        int* positions_data = (int*)malloc(ggml_nbytes(positions));
-        for (int i = 0; i < num_positions; i++) {
-            positions_data[i] = i;
+                void* zero_mem = malloc(ggml_nbytes(embeddings));
+                memset(zero_mem, 0, ggml_nbytes(embeddings));
+                ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
+                free(zero_mem);
+            }
         }
-        ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
-        free(positions_data);
-    }
 
-    {
-        struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
-        int* patches_data = (int*)malloc(ggml_nbytes(patches));
-        for (int i = 0; i < num_patches; i++) {
-            patches_data[i] = i + 1;
+        {
+            struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
+
+            int* positions_data = (int*)malloc(ggml_nbytes(positions));
+            for (int i = 0; i < num_positions; i++) {
+                positions_data[i] = i;
+            }
+            ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+            free(positions_data);
+        }
+
+        {
+            struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
+            int* patches_data = (int*)malloc(ggml_nbytes(patches));
+            for (int i = 0; i < num_patches; i++) {
+                patches_data[i] = i + 1;
+            }
+            ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
+            free(patches_data);
         }
-        ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
-        free(patches_data);
     }
 
     if (ggml_backend_is_cpu(ctx->backend)) {
@@ -2081,7 +2603,22 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
     if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
         return ctx->vision_model.mm_3_b->ne[0];
     }
+    if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+        if (ctx->minicpmv_version == 2) {
+            return 4096;
+        }
+        else if (ctx->minicpmv_version == 3) {
+            return 3584;
+        }
+    }
 
     std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];
     throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
 }
+
+int clip_is_minicpmv(const struct clip_ctx * ctx) {
+    if (ctx->has_minicpmv_projector) {
+        return ctx->minicpmv_version;
+    }
+    return 0;
+}

examples/llava/clip.h

@@ -18,14 +18,17 @@
 #    define CLIP_API
 #endif
 
-struct clip_ctx;
-
 #ifdef __cplusplus
 extern "C" {
 #endif
 
 struct clip_ctx;
 
+struct clip_image_size {
+    int width;
+    int height;
+};
+
 struct clip_image_u8_batch {
     struct clip_image_u8 * data;
     size_t size;
@@ -55,6 +58,10 @@ CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
 CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
 CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
 
+CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
+CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
+
+CLIP_API struct clip_image_size * clip_image_size_init();
 CLIP_API struct clip_image_u8  * clip_image_u8_init ();
 CLIP_API struct clip_image_f32 * clip_image_f32_init();
 
@@ -78,6 +85,8 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons
 
 CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
 
+CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
+
 #ifdef __cplusplus
 }
 #endif

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/llava.cpp

@@ -202,6 +202,33 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
     return true;
 }
 
+static clip_image_f32 * only_v2_5_reshape_by_patch(clip_image_f32 * image, int patch_size) {
+    int width = image->nx;
+    int height = image->ny;
+    int num_patches = (height / patch_size) * (width / patch_size);
+    clip_image_f32 * patch = clip_image_f32_init();
+    patch->nx = patch_size * num_patches;
+    patch->ny = patch_size;
+    patch->buf.resize(3 * patch->nx * patch->ny);
+
+    int patch_index = 0;
+
+    for (int i = 0; i < height; i += patch_size) {
+        for (int j = 0; j < width; j += patch_size) {
+            for (int pi = 0; pi < patch_size; ++pi) {
+                for (int pj = 0; pj < patch_size; ++pj) {
+                    int input_index = ((i + pi) * width + (j + pj)) * 3;
+                    int output_index = (pi * patch_size * num_patches + patch_index * patch_size + pj) * 3;
+                    patch->buf[output_index] = image->buf[input_index];
+                    patch->buf[output_index+1] = image->buf[input_index+1];
+                    patch->buf[output_index+2] = image->buf[input_index+2];
+                }
+            }
+            patch_index++;
+        }
+    }
+    return patch;
+}
 
 static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
     // std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
@@ -218,7 +245,51 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
 
     const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
 
-    if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
+    if (clip_is_minicpmv(ctx_clip)) {
+        std::vector<float *> image_embd_v;
+        image_embd_v.resize(img_res_v.size);
+        struct clip_image_size * load_image_size = clip_image_size_init();
+        for (size_t i = 0; i < img_res_v.size; i++) {
+            const int64_t t_img_enc_step_start_us = ggml_time_us();
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip));
+            int patch_size=14;
+            load_image_size->width = img_res_v.data[i].nx;
+            load_image_size->height = img_res_v.data[i].ny;
+            clip_add_load_image_size(ctx_clip, load_image_size);
+            bool encoded = false;
+            int has_minicpmv_projector = clip_is_minicpmv(ctx_clip);
+            if (has_minicpmv_projector == 2) {
+                encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
+            }
+            else if (has_minicpmv_projector == 3) {
+                encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
+            }
+            if (!encoded) {
+                LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                return false;
+            }
+            const int64_t t_img_enc_steop_batch_us = ggml_time_us();
+            LOG_TEE("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
+        }
+        const int64_t t_img_enc_batch_us = ggml_time_us();
+        LOG_TEE("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+
+        int n_img_pos_out = 0;
+        for (size_t i = 0; i < image_embd_v.size(); i++) {
+            std::memcpy(image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), image_embd_v[i], clip_embd_nbytes(ctx_clip));
+            n_img_pos_out += clip_n_patches(ctx_clip);
+        }
+        *n_img_pos = n_img_pos_out;
+        for (size_t i = 0; i < image_embd_v.size(); i++) {
+            free(image_embd_v[i]);
+        }
+        image_embd_v.clear();
+        load_image_size->width = img->nx;
+        load_image_size->height = img->ny;
+        clip_add_load_image_size(ctx_clip, load_image_size);
+        LOG_TEE("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
+    }
+    else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
         // flat / default llava-1.5 type embedding
         *n_img_pos = clip_n_patches(ctx_clip);
         bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
@@ -228,7 +299,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
 
             return false;
         }
-    } else {
+    }
+    else {
         // spatial_unpad llava-1.6 type embedding
         // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
         std::vector<float *> image_embd_v;
@@ -297,7 +369,11 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
 }
 
 bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
-    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model
+    int num_max_patches = 6;
+    if (clip_is_minicpmv(ctx_clip)) {
+        num_max_patches = 10;
+    }
+    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
     if (!image_embd) {
         LOG_TEE("Unable to allocate memory for image embeddings\n");
         return false;

examples/llava/llava.h

@@ -17,12 +17,11 @@
 #    define LLAVA_API
 #endif
 
-struct clip_ctx;
-
 #ifdef __cplusplus
 extern "C" {
 #endif
 
+struct clip_ctx;
 struct llava_image_embed {
     float * embed;
     int n_image_pos;
@@ -37,8 +36,8 @@ LLAVA_API bool llava_image_embed_make_with_clip_img(struct clip_ctx * ctx_clip,
 LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);
 /** build an image embed from a path to an image filename */
 LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path);
-LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
 /** free an embedding made with llava_image_embed_make_* */
+LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
 
 /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */
 LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past);

examples/llava/minicpmv-cli.cpp

@@ -0,0 +1,329 @@
+#include "ggml.h"
+#include "log.h"
+#include "common.h"
+#include "clip.h"
+#include "llava.h"
+#include "llama.h"
+
+#include <cstdio>
+#include <cstdlib>
+#include <vector>
+
+struct llava_context {
+    struct clip_ctx * ctx_clip = NULL;
+    struct llama_context * ctx_llama = NULL;
+    struct llama_model * model = NULL;
+};
+
+static void show_additional_info(int /*argc*/, char ** argv) {
+    LOG_TEE("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    LOG_TEE("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
+}
+
+static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) user_data;
+    LOG_TEE("%s", text);
+}
+
+static struct llama_model * llava_init(gpt_params * params) {
+    llama_backend_init();
+    llama_numa_init(params->numa);
+
+    llama_model_params model_params = llama_model_params_from_gpt_params(*params);
+
+    llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
+    if (model == NULL) {
+        LOG_TEE("%s: error: unable to load model\n" , __func__);
+        return NULL;
+    }
+    return model;
+}
+
+static struct llava_context * llava_init_context(gpt_params * params, llama_model * model) {
+    auto prompt = params->prompt;
+    if (prompt.empty()) {
+        prompt = "describe the image in detail.";
+    }
+
+    llama_context_params ctx_params = llama_context_params_from_gpt_params(*params);
+    if (params->n_ctx < 2048) {
+        // warn user here, "Image processing requires at least 2048 context, setting context to 2048"
+        LOG_TEE("%s: warn: Image processing requires at least 2048 context, setting context to 2048\n" , __func__);
+        ctx_params.n_ctx = 2048;
+    } else {
+        ctx_params.n_ctx = params->n_ctx;
+    }
+
+    llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
+
+    if (ctx_llama == NULL) {
+        LOG_TEE("%s: error: failed to create the llama_context\n" , __func__);
+        return NULL;
+    }
+
+    auto ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
+
+    ctx_llava->ctx_llama = ctx_llama;
+    ctx_llava->model = model;
+    return ctx_llava;
+}
+
+static void llava_free(struct llava_context * ctx_llava) {
+    if (ctx_llava->ctx_clip) {
+        clip_free(ctx_llava->ctx_clip);
+        ctx_llava->ctx_clip = NULL;
+    }
+
+    llama_free(ctx_llava->ctx_llama);
+    llama_free_model(ctx_llava->model);
+    llama_backend_free();
+}
+
+static struct clip_ctx * clip_init_context(gpt_params * params) {
+    const char * clip_path = params->mmproj.c_str();
+
+    auto prompt = params->prompt;
+    if (prompt.empty()) {
+        prompt = "describe the image in detail.";
+    }
+    auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
+    return ctx_clip;
+}
+
+static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
+    int N = (int) tokens.size();
+    for (int i = 0; i < N; i += n_batch) {
+        int n_eval = (int) tokens.size() - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
+            LOG_TEE("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
+            return false;
+        }
+        *n_past += n_eval;
+    }
+    return true;
+}
+
+static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
+    std::vector<llama_token> tokens;
+    tokens.push_back(id);
+    return eval_tokens(ctx_llama, tokens, 1, n_past);
+}
+
+static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
+    std::string              str2     = str;
+    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos, true);
+    return eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
+}
+
+static void process_eval_image_embed(struct llava_context * ctx_llava, const struct llava_image_embed * embeds, int n_batch, int * n_past, int idx) {
+    float * image_embed = (float *)malloc(clip_embd_nbytes(ctx_llava->ctx_clip));
+    std::memcpy(image_embed, embeds->embed + idx * clip_n_patches(ctx_llava->ctx_clip) * clip_n_mmproj_embd(ctx_llava->ctx_clip), clip_embd_nbytes(ctx_llava->ctx_clip));
+
+    auto slice_embed = (llava_image_embed*)malloc(sizeof(llava_image_embed));
+    slice_embed->embed = image_embed;
+    slice_embed->n_image_pos = clip_n_patches(ctx_llava->ctx_clip);
+    llava_eval_image_embed(ctx_llava->ctx_llama, slice_embed, n_batch, n_past);
+    llava_image_embed_free(slice_embed);
+}
+
+static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * embeds, gpt_params * params, int &n_past) {
+    std::string system_prompt;
+    int idx = 0;
+    int num_image_embeds = embeds->n_image_pos / clip_n_patches(ctx_llava->ctx_clip);
+    int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
+    if (has_minicpmv_projector == 2) {
+        system_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n";
+    }
+    else if (has_minicpmv_projector == 3) {
+        system_prompt = "<|im_start|>user\n";
+    }
+    LOG_TEE("%s: image token past: %d\n", __func__, n_past);
+    eval_string(ctx_llava->ctx_llama, (system_prompt+"<image>").c_str(), params->n_batch, &n_past, false);
+    process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
+    eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
+    if (num_image_embeds > 1) {
+        size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
+        eval_string(ctx_llava->ctx_llama, std::string("<slice>").c_str(), params->n_batch, &n_past, false);
+        for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) {
+            for (size_t j = 0; j < num_image_embeds_col; ++j) {
+                eval_string(ctx_llava->ctx_llama, std::string("<image>").c_str(), params->n_batch, &n_past, false);
+                process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
+                eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
+                if (j == num_image_embeds_col - 1) {
+                    eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false);
+                }
+            }
+        }
+        eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
+    }
+    LOG_TEE("%s: image token past: %d\n", __func__, n_past);
+}
+
+static const char * sample(struct llama_sampling_context * ctx_sampling,
+                           struct llama_context * ctx_llama,
+                           int * n_past) {
+    const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
+    llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
+    static std::string ret;
+    if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
+        ret = "</s>";
+    } else {
+        ret = llama_token_to_piece(ctx_llama, id);
+    }
+    eval_id(ctx_llama, id, n_past);
+    return ret.c_str();
+}
+
+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->cpuparams.n_threads, fname.c_str());
+    if (!embeds) {
+        std::cerr << "error: failed to load image " << fname << ". Terminating\n\n";
+        return NULL;
+    }
+
+    // process the prompt
+    if (params->prompt.empty() && params->interactive == false) {
+        LOG_TEE("prompt should be given or interactive mode should be on");
+        return NULL;
+    }
+
+    auto model = llava_init(params);
+    if (model == NULL) {
+        fprintf(stderr, "%s: error: failed to init minicpmv model\n", __func__);
+        return NULL;
+    }
+    const int64_t t_llava_init_start_us = ggml_time_us();
+    auto ctx_llava = llava_init_context(params, model);
+    ctx_llava->ctx_clip = ctx_clip;
+    const int64_t t_llava_init_end_us = ggml_time_us();
+    float t_llava_init_ms = (t_llava_init_end_us - t_llava_init_start_us) / 1000.0;
+    LOG_TEE("\n%s: llava init in %8.2f ms.\n", __func__, t_llava_init_ms);
+
+    const int64_t t_process_image_start_us = ggml_time_us();
+    process_image(ctx_llava, embeds, params, n_past);
+    const int64_t t_process_image_end_us = ggml_time_us();
+    float t_process_image_ms = (t_process_image_end_us - t_process_image_start_us) / 1000.0;
+    LOG_TEE("\n%s: llama process image in %8.2f ms.\n", __func__, t_process_image_ms);
+
+    llava_image_embed_free(embeds);
+    return ctx_llava;
+}
+
+static struct llama_sampling_context * llama_init(struct llava_context * ctx_llava, gpt_params * params, std::string prompt, int &n_past, bool is_first = false){
+    std::string user_prompt = prompt;
+    int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
+    if (!is_first) {
+        if (has_minicpmv_projector == 2) {
+            user_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" + prompt;
+        }
+        else if (has_minicpmv_projector == 3) {
+            user_prompt = "<|im_start|>user\n" + prompt;
+        }
+    }
+
+    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
+    if (has_minicpmv_projector == 2) {
+        eval_string(ctx_llava->ctx_llama, "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", params->n_batch, &n_past, false);
+    }
+    else if (has_minicpmv_projector == 3) {
+        eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
+    }
+
+    // generate the response
+
+    LOG_TEE("\n");
+
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
+    return ctx_sampling;
+}
+
+static const char * llama_loop(struct llava_context * ctx_llava,struct llama_sampling_context * ctx_sampling, int &n_past){
+
+    const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
+    return tmp;
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        show_additional_info(argc, argv);
+        return 1;
+    }
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("llava", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+    llama_log_set(llama_log_callback_logTee, nullptr);
+#endif // LOG_DISABLE_LOGS
+
+    if (params.mmproj.empty() || (params.image.empty())) {
+        gpt_params_print_usage(argc, argv, params);
+        show_additional_info(argc, argv);
+        return 1;
+    }
+
+    for (auto & image : params.image) {
+        int n_past = 0;
+        auto ctx_llava = minicpmv_init(&params, image, n_past);
+
+        if (!params.prompt.empty()) {
+            LOG_TEE("<user>%s\n", params.prompt.c_str());
+            LOG_TEE("<assistant>");
+            auto ctx_sampling = llama_init(ctx_llava, &params, params.prompt.c_str(), n_past, true);
+            const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
+            std::string response = "";
+            bool have_tmp = false;
+            for (int i = 0; i < max_tgt_len; i++) {
+                auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past);
+                response += tmp;
+                if (strcmp(tmp, "</s>") == 0){
+                    if(!have_tmp)continue;
+                    else break;
+                }
+                if (strstr(tmp, "###")) break; // Yi-VL behavior
+                have_tmp = true;
+                printf("%s", tmp);
+                if (strstr(response.c_str(), "<user>")) break; // minicpm-v
+
+                fflush(stdout);
+            }
+            llama_sampling_free(ctx_sampling);
+        }else {
+            while (true) {
+                LOG_TEE("<user>");
+                std::string prompt;
+                std::getline(std::cin, prompt);
+                LOG_TEE("<assistant>");
+                auto ctx_sampling = llama_init(ctx_llava, &params, prompt, n_past, true);
+                const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
+                std::string response = "";
+                for (int i = 0; i < max_tgt_len; i++) {
+                    auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past);
+                    response += tmp;
+                    if (strcmp(tmp, "</s>") == 0) break;
+                    if (strstr(tmp, "###")) break; // Yi-VL behavior
+                    printf("%s", tmp);// mistral llava-1.6
+                    if (strstr(response.c_str(), "<user>")) break; // minicpm-v
+                    fflush(stdout);
+                }
+                llama_sampling_free(ctx_sampling);
+            }
+        }
+        printf("\n");
+        llama_print_timings(ctx_llava->ctx_llama);
+
+        ctx_llava->model = NULL;
+        llava_free(ctx_llava);
+    }
+
+    return 0;
+}

examples/llava/minicpmv-convert-image-encoder-to-gguf.py

@@ -0,0 +1,806 @@
+# coding=utf-8
+# Copyright 2024 Google AI and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch Siglip model. """
+# Copied from  HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit and add tgt_sizes
+
+
+import os
+import math
+import warnings
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn.init import _calculate_fan_in_and_fan_out
+
+from transformers.activations import ACT2FN
+from transformers.modeling_utils import PreTrainedModel
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import (
+    logging,
+)
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+class SiglipVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`SiglipVisionModel`]. It is used to instantiate a
+    Siglip vision encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the vision encoder of the Siglip
+    [google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_channels (`int`, *optional*, defaults to 3):
+            Number of channels in the input images.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 16):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+    Example:
+    ```python
+    >>> from transformers import SiglipVisionConfig, SiglipVisionModel
+    >>> # Initializing a SiglipVisionConfig with google/siglip-base-patch16-224 style configuration
+    >>> configuration = SiglipVisionConfig()
+    >>> # Initializing a SiglipVisionModel (with random weights) from the google/siglip-base-patch16-224 style configuration
+    >>> model = SiglipVisionModel(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "siglip_vision_model"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        image_size=224,
+        patch_size=16,
+        hidden_act="gelu_pytorch_tanh",
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+
+_CHECKPOINT_FOR_DOC = "google/siglip-base-patch16-224"
+
+SIGLIP_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "google/siglip-base-patch16-224",
+    # See all SigLIP models at https://huggingface.co/models?filter=siglip
+]
+
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+def _trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn(
+            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+            "The distribution of values may be incorrect.",
+            stacklevel=2,
+        )
+
+    # Values are generated by using a truncated uniform distribution and
+    # then using the inverse CDF for the normal distribution.
+    # Get upper and lower cdf values
+    l = norm_cdf((a - mean) / std)
+    u = norm_cdf((b - mean) / std)
+
+    # Uniformly fill tensor with values from [l, u], then translate to
+    # [2l-1, 2u-1].
+    tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+    # Use inverse cdf transform for normal distribution to get truncated
+    # standard normal
+    if tensor.dtype in [torch.float16, torch.bfloat16]:
+        # The `erfinv_` op is not (yet?) defined in float16+cpu, bfloat16+gpu
+        og_dtype = tensor.dtype
+        tensor = tensor.to(torch.float32)
+        tensor.erfinv_()
+        tensor = tensor.to(og_dtype)
+    else:
+        tensor.erfinv_()
+
+    # Transform to proper mean, std
+    tensor.mul_(std * math.sqrt(2.0))
+    tensor.add_(mean)
+
+    # Clamp to ensure it's in the proper range
+    if tensor.dtype == torch.float16:
+        # The `clamp_` op is not (yet?) defined in float16+cpu
+        tensor = tensor.to(torch.float32)
+        tensor.clamp_(min=a, max=b)
+        tensor = tensor.to(torch.float16)
+    else:
+        tensor.clamp_(min=a, max=b)
+
+
+def trunc_normal_tf_(
+    tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
+):
+    """Fills the input Tensor with values drawn from a truncated
+    normal distribution. The values are effectively drawn from the
+    normal distribution :math:`\\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \\leq \text{mean} \\leq b`.
+    NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
+    bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
+    and the result is subsquently scaled and shifted by the mean and std args.
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    """
+    with torch.no_grad():
+        _trunc_normal_(tensor, 0, 1.0, a, b)
+        tensor.mul_(std).add_(mean)
+
+
+def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
+    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
+    denom = fan_in
+    if mode == "fan_in":
+        denom = fan_in
+    elif mode == "fan_out":
+        denom = fan_out
+    elif mode == "fan_avg":
+        denom = (fan_in + fan_out) / 2
+
+    variance = scale / denom
+
+    if distribution == "truncated_normal":
+        # constant is stddev of standard normal truncated to (-2, 2)
+        trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
+    elif distribution == "normal":
+        with torch.no_grad():
+            tensor.normal_(std=math.sqrt(variance))
+    elif distribution == "uniform":
+        bound = math.sqrt(3 * variance)
+        with torch.no_grad():
+            tensor.uniform_(-bound, bound)
+    else:
+        raise ValueError(f"invalid distribution {distribution}")
+
+
+def lecun_normal_(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
+
+
+def default_flax_embed_init(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="normal")
+
+class SiglipVisionEmbeddings(nn.Module):
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            padding="valid",
+        )
+
+        self.num_patches_per_side = self.image_size // self.patch_size
+        self.num_patches = self.num_patches_per_side**2
+        self.num_positions = self.num_patches
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+
+class SiglipAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    # Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Siglip
+class SiglipMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPEncoderLayer with CLIP->Siglip
+class SiglipEncoderLayer(nn.Module):
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+        self.self_attn = (
+            SiglipAttention(config)
+        )
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = SiglipMLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+class SiglipPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = SiglipVisionConfig
+    base_model_prefix = "siglip"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+
+        if isinstance(module, SiglipVisionEmbeddings):
+            width = self.config.hidden_size
+            nn.init.normal_(module.position_embedding.weight, std=1 / np.sqrt(width))
+        elif isinstance(module, nn.Embedding):
+            default_flax_embed_init(module.weight)
+        elif isinstance(module, SiglipAttention):
+            nn.init.normal_(module.q_proj.weight)
+            nn.init.normal_(module.k_proj.weight)
+            nn.init.normal_(module.v_proj.weight)
+            nn.init.normal_(module.out_proj.weight)
+            nn.init.zeros_(module.q_proj.bias)
+            nn.init.zeros_(module.k_proj.bias)
+            nn.init.zeros_(module.v_proj.bias)
+            nn.init.zeros_(module.out_proj.bias)
+        elif isinstance(module, SiglipMLP):
+            nn.init.normal_(module.fc1.weight)
+            nn.init.normal_(module.fc2.weight)
+            nn.init.normal_(module.fc1.bias, std=1e-6)
+            nn.init.normal_(module.fc2.bias, std=1e-6)
+        elif isinstance(module, (nn.Linear, nn.Conv2d)):
+            lecun_normal_(module.weight)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+SIGLIP_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+    Parameters:
+        config ([`SiglipVisionConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+SIGLIP_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPEncoder with CLIP->Siglip
+class SiglipEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`SiglipEncoderLayer`].
+    Args:
+        config: SiglipConfig
+    """
+
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([SiglipEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+class SiglipVisionTransformer(SiglipPreTrainedModel):
+    config_class = SiglipVisionConfig
+    main_input_name = "pixel_values"
+    _supports_flash_attn_2 = True
+
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__(config)
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = SiglipVisionEmbeddings(config)
+        self.encoder = SiglipEncoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.embeddings.patch_embedding
+
+import argparse
+import json
+import re
+
+import numpy as np
+from gguf import *
+from transformers.models.idefics2.modeling_idefics2 import Idefics2VisionTransformer, Idefics2VisionConfig
+
+TEXT = "clip.text"
+VISION = "clip.vision"
+
+
+def add_key_str(raw_key: str, arch: str) -> str:
+    return raw_key.format(arch=arch)
+
+
+def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_minicpmv: bool) -> bool:
+    if name in (
+        "logit_scale",
+        "text_model.embeddings.position_ids",
+        "vision_model.embeddings.position_ids",
+    ):
+        return True
+
+    if has_minicpmv and name in ["visual_projection.weight"]:
+        return True
+
+    if name.startswith("v") and not has_vision:
+        return True
+
+    if name.startswith("t") and not has_text:
+        return True
+
+    return False
+
+
+def get_tensor_name(name: str) -> str:
+    if "projection" in name:
+        return name
+    if "mm_projector" in name:
+        name = name.replace("model.mm_projector", "mm")
+        name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1)
+        name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1)
+        return name
+
+    return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
+
+
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True)
+ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16")
+ap.add_argument("--text-only", action="store_true", required=False,
+                help="Save a text-only model. It can't be used to encode images")
+ap.add_argument("--vision-only", action="store_true", required=False,
+                help="Save a vision-only model. It can't be used to encode texts")
+ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
+                help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
+ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
+                help="The clip model is from openclip (for ViT-SO400M type))")
+ap.add_argument("--minicpmv-projector", help="Path to minicpmv.projector file. If specified, save an image encoder for MiniCPM-V models.")
+ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp")
+ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
+# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
+# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
+default_image_mean = [0.48145466, 0.4578275, 0.40821073]
+default_image_std = [0.26862954, 0.26130258, 0.27577711]
+ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
+ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
+ap.add_argument('--minicpmv_version', type=int, help='minicpmv_version: MiniCPM-V-2 use 1; MiniCPM-V-2.5 use 2; MiniCPM-V-2.6 use 3', default=2)
+
+# with proper
+args = ap.parse_args()
+
+
+if args.text_only and args.vision_only:
+    print("--text-only and --image-only arguments cannot be specified at the same time.")
+    exit(1)
+
+if args.use_f32:
+    print("WARNING: Weights for the convolution op is always saved in f16, as the convolution op in GGML does not support 32-bit kernel weights yet.")
+
+# output in the same directory as the model if output_dir is None
+dir_model = args.model_dir
+
+if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
+    vocab = None
+    tokens = None
+else:
+    with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+        vocab = json.load(f)
+        tokens = [key for key in vocab]
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if args.use_f32:
+    ftype = 0
+
+# if args.clip_model_is_vision or args.clip_model_is_openclip:
+#     model = CLIPVisionModel.from_pretrained(dir_model)
+#     processor = None
+# else:
+#     model = CLIPModel.from_pretrained(dir_model)
+#     processor = CLIPProcessor.from_pretrained(dir_model)
+
+minicpmv_version = args.minicpmv_version
+emb_dim = 4096
+if minicpmv_version == 1:
+    emb_dim = 2304
+elif minicpmv_version == 2:
+    emb_dim = 4096
+elif minicpmv_version == 3:
+    emb_dim = 3584
+
+default_vision_config = {
+        "hidden_size": 1152,
+        "image_size": 980,
+        "intermediate_size": 4304,
+        "model_type": "idefics2",
+        "num_attention_heads": 16,
+        "num_hidden_layers": 27,
+        "patch_size": 14,
+    }
+
+vision_config = Idefics2VisionConfig(**default_vision_config)
+model = Idefics2VisionTransformer(vision_config)
+if minicpmv_version == 3:
+    vision_config = SiglipVisionConfig(**default_vision_config)
+    model = SiglipVisionTransformer(vision_config)
+
+processor = None
+# if model.attn_pool is not None:
+#     model.attn_pool = torch.nn.Identity()
+
+# model.blocks = model.blocks[:-1]
+model.load_state_dict(torch.load(os.path.join(dir_model, "minicpmv.clip")))
+
+fname_middle = None
+has_text_encoder = True
+has_vision_encoder = True
+has_minicpmv_projector = False
+
+if args.text_only:
+    fname_middle = "text-"
+    has_vision_encoder = False
+elif args.minicpmv_projector is not None:
+    fname_middle = "mmproj-"
+    has_text_encoder = False
+    has_minicpmv_projector = True
+    minicpmv_version = 3
+elif args.vision_only:
+    fname_middle = "vision-"
+    has_text_encoder = False
+else:
+    fname_middle = ""
+
+output_dir = args.output_dir if args.output_dir is not None else dir_model
+os.makedirs(output_dir, exist_ok=True)
+output_prefix = os.path.basename(output_dir).replace("ggml_", "")
+fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf")
+fout = GGUFWriter(path=fname_out, arch="clip")
+
+fout.add_bool("clip.has_text_encoder", has_text_encoder)
+fout.add_bool("clip.has_vision_encoder", has_vision_encoder)
+fout.add_bool("clip.has_minicpmv_projector", has_minicpmv_projector)
+fout.add_file_type(ftype)
+if args.text_only:
+    fout.add_description("text-only CLIP model")
+elif args.vision_only and not has_minicpmv_projector:
+    fout.add_description("vision-only CLIP model")
+elif has_minicpmv_projector:
+    fout.add_description("image encoder for MiniCPM-V")
+    # add projector type
+    fout.add_string("clip.projector_type", "resampler")
+    fout.add_int32("clip.minicpmv_version", minicpmv_version)
+else:
+    fout.add_description("two-tower CLIP model")
+
+if has_vision_encoder:
+    # vision_model hparams
+    fout.add_uint32("clip.vision.image_size", 448)
+    fout.add_uint32("clip.vision.patch_size", 14)
+    fout.add_uint32(add_key_str(KEY_EMBEDDING_LENGTH, VISION), 1152)
+    fout.add_uint32(add_key_str(KEY_FEED_FORWARD_LENGTH, VISION), 4304)
+    fout.add_uint32("clip.vision.projection_dim", 0)
+    fout.add_uint32(add_key_str(KEY_ATTENTION_HEAD_COUNT, VISION), 16)
+    fout.add_float32(add_key_str(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
+    block_count = 26
+    fout.add_uint32(add_key_str(KEY_BLOCK_COUNT, VISION), block_count)
+
+    if processor is not None:
+        image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean
+        image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std
+    else:
+        image_mean = args.image_mean if args.image_mean is not None else default_image_mean
+        image_std = args.image_std if args.image_std is not None else default_image_std
+    fout.add_array("clip.vision.image_mean", image_mean)
+    fout.add_array("clip.vision.image_std", image_std)
+
+use_gelu = True
+fout.add_bool("clip.use_gelu", use_gelu)
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+    """
+    embed_dim: output dimension for each position
+    pos: a list of positions to be encoded: size (M,)
+    out: (M, D)
+    """
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=np.float32)
+    omega /= embed_dim / 2.
+    omega = 1. / 10000 ** omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product
+
+    emb_sin = np.sin(out)  # (M, D/2)
+    emb_cos = np.cos(out)  # (M, D/2)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
+    return emb
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
+
+    emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)
+    return emb
+
+
+# https://github.com/facebookresearch/mae/blob/efb2a8062c206524e35e47d04501ed4f544c0ae8/util/pos_embed.py#L20
+def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
+    """
+    grid_size: int of the grid height and width
+    return:
+    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
+    """
+    if isinstance(grid_size, int):
+        grid_h_size, grid_w_size = grid_size, grid_size
+    else:
+        grid_h_size, grid_w_size = grid_size[0], grid_size[1]
+
+    grid_h = np.arange(grid_h_size, dtype=np.float32)
+    grid_w = np.arange(grid_w_size, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+
+    grid = grid.reshape([2, 1, grid_h_size, grid_w_size])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if cls_token:
+        pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
+    return pos_embed
+
+def _replace_name_resampler(s, v):
+    if re.match("resampler.pos_embed", s):
+        return {
+            s: v,
+            re.sub("pos_embed", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(emb_dim, (70, 70))),
+        }
+    if re.match("resampler.proj", s):
+        return {
+            re.sub("proj", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(emb_dim, (70, 70))),
+            re.sub("proj", "proj.weight", s): v.transpose(-1, -2).contiguous(),
+        }
+    if re.match("resampler.attn.in_proj_.*", s):
+        return {
+            re.sub("attn.in_proj_", "attn.q.", s): v.chunk(3, dim=0)[0],
+            re.sub("attn.in_proj_", "attn.k.", s): v.chunk(3, dim=0)[1],
+            re.sub("attn.in_proj_", "attn.v.", s): v.chunk(3, dim=0)[2],
+        }
+    return {s: v}
+
+if has_minicpmv_projector:
+    projector = torch.load(args.minicpmv_projector)
+    new_state_dict = {}
+    for k, v in projector.items():
+        kvs = _replace_name_resampler(k, v)
+        for nk, nv in kvs.items():
+            new_state_dict[nk] = nv
+    projector = new_state_dict
+    ftype_cur = 0
+    for name, data in projector.items():
+        name = get_tensor_name(name)
+        data = data.squeeze().numpy()
+
+        n_dims = len(data.shape)
+        if ftype == 1:
+            if name[-7:] == ".weight" and n_dims == 2:
+                print("  Converting to float16")
+                data = data.astype(np.float16)
+                ftype_cur = 1
+            else:
+                print("  Converting to float32")
+                data = data.astype(np.float32)
+                ftype_cur = 0
+        else:
+            if data.dtype != np.float32:
+                print("  Converting to float32")
+                data = data.astype(np.float32)
+                ftype_cur = 0
+
+        fout.add_tensor(name, data)
+        print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
+
+    print("Projector tensors added\n")
+
+def _replace_name(s, v):
+    s = "vision_model." + s
+    if re.match("vision_model.embeddings.position_embedding", s):
+        v = v.unsqueeze(0)
+        return {s: v}
+
+    return {s: v}
+
+state_dict = model.state_dict()
+new_state_dict = {}
+for k, v in state_dict.items():
+    kvs = _replace_name(k, v)
+    for nk, nv in kvs.items():
+        new_state_dict[nk] = nv
+state_dict = new_state_dict
+for name, data in state_dict.items():
+    if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_minicpmv_projector):
+        # we don't need this
+        print(f"skipping parameter: {name}")
+        continue
+
+    name = get_tensor_name(name)
+    data = data.squeeze().numpy()
+
+    n_dims = len(data.shape)
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0
+    if n_dims == 4:
+        print(f"tensor {name} is always saved in f16")
+        data = data.astype(np.float16)
+        ftype_cur = 1
+    elif ftype == 1:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
+    fout.add_tensor(name, data)
+
+
+fout.write_header_to_file()
+fout.write_kv_data_to_file()
+fout.write_tensors_to_file()
+fout.close()
+
+print("Done. Output file: " + fname_out)

examples/llava/minicpmv-surgery.py

@@ -0,0 +1,45 @@
+import argparse
+import os
+import torch
+from transformers import AutoModel, AutoTokenizer
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", help="Path to MiniCPM-V model")
+args = ap.parse_args()
+
+# find the model part that includes the the multimodal projector weights
+model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True)
+checkpoint = model.state_dict()
+
+# get a list of mm tensor names
+mm_tensors = [k for k, v in checkpoint.items() if k.startswith("resampler")]
+
+# store these tensors in a new dictionary and torch.save them
+projector = {name: checkpoint[name].float() for name in mm_tensors}
+torch.save(projector, f"{args.model}/minicpmv.projector")
+
+clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vpm")]
+if len(clip_tensors) > 0:
+    clip = {name.replace("vpm.", ""): checkpoint[name].float() for name in clip_tensors}
+    torch.save(clip, f"{args.model}/minicpmv.clip")
+
+    # added tokens should be removed to be able to convert Mistral models
+    if os.path.exists(f"{args.model}/added_tokens.json"):
+        with open(f"{args.model}/added_tokens.json", "w") as f:
+            f.write("{}\n")
+
+config = model.llm.config
+config.auto_map = {
+    "AutoConfig": "configuration_minicpm.MiniCPMConfig",
+    "AutoModel": "modeling_minicpm.MiniCPMModel",
+    "AutoModelForCausalLM": "modeling_minicpm.MiniCPMForCausalLM",
+    "AutoModelForSeq2SeqLM": "modeling_minicpm.MiniCPMForCausalLM",
+    "AutoModelForSequenceClassification": "modeling_minicpm.MiniCPMForSequenceClassification"
+}
+model.llm.save_pretrained(f"{args.model}/model")
+tok = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
+tok.save_pretrained(f"{args.model}/model")
+
+print("Done!")
+print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
+print(f"Also, use {args.model}/minicpmv.projector to prepare a minicpmv-encoder.gguf file.")

examples/llava/requirements.txt

@@ -2,3 +2,4 @@
 --extra-index-url https://download.pytorch.org/whl/cpu
 pillow~=10.2.0
 torch~=2.2.1
+torchvision~=0.17.1

examples/lookahead/lookahead.cpp

@@ -58,11 +58,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the target model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // Tokenize the prompt
     std::vector<llama_token> inp;

examples/lookup/lookup-create.cpp

@@ -22,11 +22,11 @@ int main(int argc, char ** argv){
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     GGML_ASSERT(model != nullptr);
 
     // tokenize the prompt

examples/lookup/lookup-stats.cpp

@@ -26,11 +26,11 @@ int main(int argc, char ** argv){
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector<llama_token> inp;

examples/lookup/lookup.cpp

@@ -34,11 +34,11 @@ int main(int argc, char ** argv){
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector<llama_token> inp;

examples/main/main.cpp

@@ -207,7 +207,10 @@ int main(int argc, char ** argv) {
 
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    model = llama_init.model;
+    ctx = llama_init.context;
     if (sparams.cfg_scale > 1.f) {
         struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
         ctx_guidance = llama_new_context_with_model(model, lparams);
@@ -218,6 +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);
@@ -264,9 +301,9 @@ int main(int argc, char ** argv) {
         }
     }
 
-    const bool add_bos = llama_should_add_bos_token(model);
+    const bool add_bos = llama_add_bos_token(model);
     if (!llama_model_has_encoder(model)) {
-        GGML_ASSERT(llama_add_eos_token(model) != 1);
+        GGML_ASSERT(!llama_add_eos_token(model));
     }
     LOG("add_bos: %d\n", add_bos);
 
@@ -986,6 +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/parallel/parallel.cpp

@@ -129,11 +129,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the target model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // load the prompts from an external file if there are any
     if (params.prompt.empty()) {

examples/perplexity/perplexity.cpp

@@ -340,8 +340,8 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
 
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
 
     fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
 
@@ -480,8 +480,8 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
 
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
 
     std::ofstream logits_stream;
     if (!params.logits_file.empty()) {
@@ -1733,8 +1733,8 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
     const int n_batch = params.n_batch;
     const int num_batches = (n_ctx + n_batch - 1)/n_batch;
     const int nv = 2*((n_vocab + 1)/2) + 4;
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
 
     std::vector<uint16_t> log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv);
     std::vector<float>    kld_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk);
@@ -2018,11 +2018,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-
     // load the model and apply lora adapter, if any
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;

examples/quantize/README.md

@@ -34,7 +34,7 @@ Run the quantized model:
 
 ```bash
 # start inference on a gguf model
-./llama-cli -m ./models/mymodel/ggml-model-Q4_K_M.gguf -n 128
+./llama-cli -m ./models/mymodel/ggml-model-Q4_K_M.gguf -cnv -p "You are a helpful assistant"
 ```
 
 When running the larger models, make sure you have enough disk space to store all the intermediate files.

examples/quantize/quantize.cpp

@@ -91,7 +91,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 }
 
 // usage:
-//  ./quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
+//  ./llama-quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
 //
 [[noreturn]]
 static void usage(const char * executable) {
@@ -104,7 +104,7 @@ 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/retrieval/retrieval.cpp

@@ -148,11 +148,12 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
+
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
@@ -252,6 +253,8 @@ int main(int argc, char ** argv) {
         chunks[i].tokens.clear();
     }
 
+    struct llama_batch query_batch = llama_batch_init(n_batch, 0, 1);
+
     // start loop, receive query and return top k similar chunks based on cosine similarity
     std::string query;
     while (true) {
@@ -259,7 +262,6 @@ int main(int argc, char ** argv) {
         std::getline(std::cin, query);
         std::vector<int32_t> query_tokens = llama_tokenize(ctx, query, true);
 
-        struct llama_batch query_batch = llama_batch_init(n_batch, 0, 1);
         batch_add_seq(query_batch, query_tokens, 0);
 
         std::vector<float> query_emb(n_embd, 0);
@@ -292,6 +294,7 @@ int main(int argc, char ** argv) {
     }
 
     // clean up
+    llama_batch_free(query_batch);
     llama_print_timings(ctx);
     llama_free(ctx);
     llama_free_model(model);

examples/rpc/README.md

@@ -1,5 +1,9 @@
 ## Overview
 
+> [!IMPORTANT]
+> This example and the RPC backend are currently in a proof-of-concept development stage. As such, the functionality is fragile and
+> insecure. **Never run the RPC server on an open network or in a sensitive environment!**
+
 The `rpc-server` allows  running `ggml` backend on a remote host.
 The RPC backend communicates with one or several instances of `rpc-server` and offloads computations to them.
 This can be used for distributed LLM inference with `llama.cpp` in the following way:

examples/rpc/rpc-server.cpp

@@ -16,7 +16,7 @@
 #include <stdio.h>
 
 struct rpc_server_params {
-    std::string host        = "0.0.0.0";
+    std::string host        = "127.0.0.1";
     int         port        = 50052;
     size_t      backend_mem = 0;
 };
@@ -114,6 +114,17 @@ int main(int argc, char * argv[]) {
         fprintf(stderr, "Invalid parameters\n");
         return 1;
     }
+
+    if (params.host != "127.0.0.1") {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n");
+        fprintf(stderr, "WARNING: Host ('%s') is != '127.0.0.1'\n", params.host.c_str());
+        fprintf(stderr, "         Never expose the RPC server to an open network!\n");
+        fprintf(stderr, "         This is an experimental feature and is not secure!\n");
+        fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n");
+        fprintf(stderr, "\n");
+    }
+
     ggml_backend_t backend = create_backend();
     if (!backend) {
         fprintf(stderr, "Failed to create backend\n");

examples/save-load-state/save-load-state.cpp

@@ -28,10 +28,11 @@ int main(int argc, char ** argv) {
     std::string result2;
 
     // init
-    llama_model * model;
-    llama_context * ctx;
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;
@@ -47,7 +48,7 @@ int main(int argc, char ** argv) {
     // save state (rng, logits, embedding and kv_cache) to file
     {
         std::vector<uint8_t> state_mem(llama_state_get_size(ctx));
-        const size_t written = llama_state_get_data(ctx, state_mem.data());
+        const size_t written = llama_state_get_data(ctx, state_mem.data(), state_mem.size());
 
         FILE *fp_write = fopen("dump_state.bin", "wb");
         fwrite(state_mem.data(), 1, written, fp_write);
@@ -99,13 +100,16 @@ int main(int argc, char ** argv) {
 
     // load state (rng, logits, embedding and kv_cache) from file
     {
-        std::vector<uint8_t> state_mem(llama_state_get_size(ctx2));
+        std::vector<uint8_t> state_mem;
 
         FILE * fp_read = fopen("dump_state.bin", "rb");
+        fseek(fp_read, 0, SEEK_END);
+        state_mem.resize(ftell(fp_read));
+        fseek(fp_read, 0, SEEK_SET);
         const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
         fclose(fp_read);
 
-        if (read != llama_state_set_data(ctx2, state_mem.data())) {
+        if (read != llama_state_set_data(ctx2, state_mem.data(), state_mem.size())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
             llama_free(ctx2);
             llama_free_model(model);
@@ -159,13 +163,16 @@ int main(int argc, char ** argv) {
 
     // load state (rng, logits, embedding and kv_cache) from file
     {
-        std::vector<uint8_t> state_mem(llama_state_get_size(ctx3));
+        std::vector<uint8_t> state_mem;
 
         FILE * fp_read = fopen("dump_state.bin", "rb");
+        fseek(fp_read, 0, SEEK_END);
+        state_mem.resize(ftell(fp_read));
+        fseek(fp_read, 0, SEEK_SET);
         const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
         fclose(fp_read);
 
-        if (read != llama_state_set_data(ctx3, state_mem.data())) {
+        if (read != llama_state_set_data(ctx3, state_mem.data(), state_mem.size())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
             llama_free(ctx3);
             llama_free_model(model);
@@ -182,7 +189,7 @@ int main(int argc, char ** argv) {
     {
         // save kv of seq 0
         std::vector<uint8_t> seq_store(llama_state_seq_get_size(ctx3, 0));
-        const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), 0);
+        const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), seq_store.size(), 0);
         if (ncopy != seq_store.size()) {
             fprintf(stderr, "\n%s : seq copy data length %zd does not match expected length %zd\n", __func__, ncopy, seq_store.size());
             llama_free(ctx3);
@@ -196,7 +203,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s : kv cache cleared\n", __func__);
 
         // restore kv into seq 1
-        const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), 1);
+        const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), seq_store.size(), 1);
         if (nset != seq_store.size()) {
             fprintf(stderr, "\n%s : seq set data length %zd does not match expected length %zd\n", __func__, nset, seq_store.size());
             llama_free(ctx3);

examples/server/README.md

@@ -207,47 +207,12 @@ model:
   -hff,  --hf-file FILE           Hugging Face model file (default: unused)
   -hft,  --hf-token TOKEN         Hugging Face access token (default: value from HF_TOKEN environment variable)
 
-retrieval:
-
-         --context-file FNAME     file to load context from (repeat to specify multiple files)
-         --chunk-size N           minimum length of embedded text chunks (default: 64)
-         --chunk-separator STRING
-                                  separator between chunks (default: '
-                                  ')
-
-passkey:
-
-         --junk N                 number of times to repeat the junk text (default: 250)
-         --pos N                  position of the passkey in the junk text (default: -1)
-
-imatrix:
-
-  -o,    --output FNAME           output file (default: 'imatrix.dat')
-         --output-frequency N     output the imatrix every N iterations (default: 10)
-         --save-frequency N       save an imatrix copy every N iterations (default: 0)
-         --process-output         collect data for the output tensor (default: false)
-         --no-ppl                 do not compute perplexity (default: true)
-         --chunk N                start processing the input from chunk N (default: 0)
-
-bench:
-
-  -pps                            is the prompt shared across parallel sequences (default: false)
-  -npp n0,n1,...                  number of prompt tokens
-  -ntg n0,n1,...                  number of text generation tokens
-  -npl n0,n1,...                  number of parallel prompts
-
-embedding:
-
-         --embd-normalize         normalisation for embendings (default: 2) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
-         --embd-output-format     empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
-         --embd-separator         separator of embendings (default \n) for example "<#sep#>"
-
 server:
 
          --host HOST              ip address to listen (default: 127.0.0.1)
          --port PORT              port to listen (default: 8080)
          --path PATH              path to serve static files from (default: )
-         --embedding(s)           enable embedding endpoint (default: disabled)
+         --embedding(s)           restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)
          --api-key KEY            API key to use for authentication (default: none)
          --api-key-file FNAME     path to file containing API keys (default: none)
          --ssl-key-file FNAME     path to file a PEM-encoded SSL private key
@@ -267,7 +232,8 @@ server:
                                   https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
   -sps,  --slot-prompt-similarity SIMILARITY
                                   how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)
-
+         --lora-init-without-apply
+                                  load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled)
 
 logging:
 
@@ -279,17 +245,53 @@ logging:
          --log-file FNAME         Specify a log filename (without extension)
          --log-new                Create a separate new log file on start. Each log file will have unique name: "<name>.<ID>.log"
          --log-append             Don't truncate the old log file.
-
-cvector:
-
-  -o,    --output FNAME           output file (default: 'control_vector.gguf')
-         --positive-file FNAME    positive prompts file, one prompt per line (default: 'examples/cvector-generator/positive.txt')
-         --negative-file FNAME    negative prompts file, one prompt per line (default: 'examples/cvector-generator/negative.txt')
-         --pca-batch N            batch size used for PCA. Larger batch runs faster, but uses more memory (default: 100)
-         --pca-iter N             number of iterations used for PCA (default: 1000)
-         --method {pca,mean}      dimensionality reduction method to be used (default: pca)
 ```
 
+Available environment variables (if specified, these variables will override parameters specified in arguments):
+
+- `LLAMA_CACHE`: cache directory, used by `--hf-repo`
+- `HF_TOKEN`: Hugging Face access token, used when accessing a gated model with `--hf-repo`
+- `LLAMA_ARG_MODEL`: equivalent to `-m`
+- `LLAMA_ARG_MODEL_URL`: equivalent to `-mu`
+- `LLAMA_ARG_MODEL_ALIAS`: equivalent to `-a`
+- `LLAMA_ARG_HF_REPO`: equivalent to `--hf-repo`
+- `LLAMA_ARG_HF_FILE`: equivalent to `--hf-file`
+- `LLAMA_ARG_THREADS`: equivalent to `-t`
+- `LLAMA_ARG_CTX_SIZE`: equivalent to `-c`
+- `LLAMA_ARG_N_PARALLEL`: equivalent to `-np`
+- `LLAMA_ARG_BATCH`: equivalent to `-b`
+- `LLAMA_ARG_UBATCH`: equivalent to `-ub`
+- `LLAMA_ARG_N_GPU_LAYERS`: equivalent to `-ngl`
+- `LLAMA_ARG_THREADS_HTTP`: equivalent to `--threads-http`
+- `LLAMA_ARG_CHAT_TEMPLATE`: equivalent to `--chat-template`
+- `LLAMA_ARG_N_PREDICT`: equivalent to `-n`
+- `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
 
@@ -411,16 +413,18 @@ node index.js
 
 ## API Endpoints
 
-- **GET** `/health`: Returns the current state of the server:
-  - 503 -> `{"status": "loading model"}` if the model is still being loaded.
-  - 500 -> `{"status": "error"}` if the model failed to load.
-  - 200 -> `{"status": "ok", "slots_idle": 1, "slots_processing": 2 }` if the model is successfully loaded and the server is ready for further requests mentioned below.
-  - 200 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slots are currently available.
-  - 503 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if the query parameter `fail_on_no_slot` is provided and no slots are currently available.
+### GET `/health`: Returns heath check result
 
-  If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set.
+**Response format**
 
-- **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
+- HTTP status code 503
+  - Body: `{"error": {"code": 503, "message": "Loading model", "type": "unavailable_error"}}`
+  - Explanation: the model is still being loaded.
+- HTTP status code 200
+  - Body: `{"status": "ok" }`
+  - Explanation: the model is successfully loaded and the server is ready.
+
+### POST `/completion`: Given a `prompt`, it returns the predicted completion.
 
     *Options:*
 
@@ -498,7 +502,7 @@ node index.js
 
     `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values.
 
-### Result JSON
+**Response format**
 
 - Note: When using streaming mode (`stream`), only `content` and `stop` will be returned until end of completion.
 
@@ -537,7 +541,7 @@ Notice that each `probs` is an array of length `n_probs`.
 - `tokens_evaluated`: Number of tokens evaluated in total from the prompt
 - `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
 
-- **POST** `/tokenize`: Tokenize a given text.
+### POST `/tokenize`: Tokenize a given text
 
     *Options:*
 
@@ -545,13 +549,15 @@ Notice that each `probs` is an array of length `n_probs`.
 
     `add_special`: Boolean indicating if special tokens, i.e. `BOS`, should be inserted.  Default: `false`
 
-- **POST** `/detokenize`: Convert tokens to text.
+### POST `/detokenize`: Convert tokens to text
 
     *Options:*
 
     `tokens`: Set the tokens to detokenize.
 
-- **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
+### POST `/embedding`: Generate embedding of a given text
+
+The same as [the embedding example](../embedding) does.
 
     *Options:*
 
@@ -559,7 +565,9 @@ Notice that each `probs` is an array of length `n_probs`.
 
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `content`. You can determine the place of the image in the content as in the following: `Image: [img-21].\nCaption: This is a picture of a house`. In this case, `[img-21]` will be replaced by the embeddings of the image with id `21` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 21}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
-- **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
+### POST `/infill`: For code infilling.
+
+Takes a prefix and a suffix and returns the predicted completion as stream.
 
     *Options:*
 
@@ -571,7 +579,7 @@ Notice that each `probs` is an array of length `n_probs`.
 
 - **GET** `/props`: Return current server settings.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -589,7 +597,9 @@ Notice that each `probs` is an array of length `n_probs`.
 - `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
 - `chat_template` - the model's original Jinja2 prompt template
 
-- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
+### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
+
+Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
 
     *Options:*
 
@@ -641,7 +651,7 @@ Notice that each `probs` is an array of length `n_probs`.
     }'
     ```
 
-- **POST** `/v1/embeddings`: OpenAI-compatible embeddings API.
+### POST `/v1/embeddings`: OpenAI-compatible embeddings API
 
     *Options:*
 
@@ -675,9 +685,15 @@ Notice that each `probs` is an array of length `n_probs`.
     }'
     ```
 
-- **GET** `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
+### GET `/slots`: Returns the current slots processing state
 
-### Result JSON
+This endpoint can be disabled with `--no-slots`
+
+If query param `?fail_on_no_slot=1` is set, this endpoint will respond with status code 503 if there is no available slots.
+
+**Response format**
+
+Example:
 
 ```json
 [
@@ -738,7 +754,13 @@ Notice that each `probs` is an array of length `n_probs`.
 ]
 ```
 
-- **GET** `/metrics`: [Prometheus](https://prometheus.io/) compatible metrics exporter endpoint if `--metrics` is enabled:
+Possible values for `slot[i].state` are:
+- `0`: SLOT_STATE_IDLE
+- `1`: SLOT_STATE_PROCESSING
+
+### GET `/metrics`: Prometheus compatible metrics exporter
+
+This endpoint is only accessible if `--metrics` is set.
 
 Available metrics:
 - `llamacpp:prompt_tokens_total`: Number of prompt tokens processed.
@@ -750,13 +772,13 @@ Available metrics:
 - `llamacpp:requests_processing`: Number of requests processing.
 - `llamacpp:requests_deferred`: Number of requests deferred.
 
-- **POST** `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file.
+### POST `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file.
 
     *Options:*
 
     `filename`: Name of the file to save the slot's prompt cache. The file will be saved in the directory specified by the `--slot-save-path` server parameter.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -770,13 +792,13 @@ Available metrics:
 }
 ```
 
-- **POST** `/slots/{id_slot}?action=restore`: Restore the prompt cache of the specified slot from a file.
+### POST `/slots/{id_slot}?action=restore`: Restore the prompt cache of the specified slot from a file.
 
     *Options:*
 
     `filename`: Name of the file to restore the slot's prompt cache from. The file should be located in the directory specified by the `--slot-save-path` server parameter.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -790,9 +812,9 @@ Available metrics:
 }
 ```
 
-- **POST** `/slots/{id_slot}?action=erase`: Erase the prompt cache of the specified slot.
+### POST `/slots/{id_slot}?action=erase`: Erase the prompt cache of the specified slot.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -801,6 +823,46 @@ Available metrics:
 }
 ```
 
+### GET `/lora-adapters`: Get list of all LoRA adapters
+
+This endpoint returns the loaded LoRA adapters. You can add adapters using `--lora` when starting the server, for example: `--lora my_adapter_1.gguf --lora my_adapter_2.gguf ...`
+
+By default, all adapters will be loaded with scale set to 1. To initialize all adapters scale to 0, add `--lora-init-without-apply`
+
+If an adapter is disabled, the scale will be set to 0.
+
+**Response format**
+
+```json
+[
+    {
+        "id": 0,
+        "path": "my_adapter_1.gguf",
+        "scale": 0.0
+    },
+    {
+        "id": 1,
+        "path": "my_adapter_2.gguf",
+        "scale": 0.0
+    }
+]
+```
+
+### POST `/lora-adapters`: Set list of LoRA adapters
+
+To disable an adapter, either remove it from the list below, or set scale to 0.
+
+**Request format**
+
+To know the `id` of the adapter, use GET `/lora-adapters`
+
+```json
+[
+  {"id": 0, "scale": 0.2},
+  {"id": 1, "scale": 0.8}
+]
+```
+
 ## More examples
 
 ### Change system prompt on runtime

examples/server/server.cpp

@@ -15,6 +15,8 @@
 // Change JSON_ASSERT from assert() to GGML_ASSERT:
 #define JSON_ASSERT GGML_ASSERT
 #include "json.hpp"
+// mime type for sending response
+#define MIMETYPE_JSON "application/json; charset=utf-8"
 
 // auto generated files (update with ./deps.sh)
 #include "colorthemes.css.hpp"
@@ -67,7 +69,6 @@ enum slot_command {
 enum server_state {
     SERVER_STATE_LOADING_MODEL,  // Server is starting up, model not fully loaded yet
     SERVER_STATE_READY,          // Server is ready and model is loaded
-    SERVER_STATE_ERROR           // An error occurred, load_model failed
 };
 
 enum server_task_type {
@@ -78,6 +79,7 @@ enum server_task_type {
     SERVER_TASK_TYPE_SLOT_SAVE,
     SERVER_TASK_TYPE_SLOT_RESTORE,
     SERVER_TASK_TYPE_SLOT_ERASE,
+    SERVER_TASK_TYPE_SET_LORA,
 };
 
 struct server_task {
@@ -622,6 +624,7 @@ struct server_response {
 struct server_context {
     llama_model * model = nullptr;
     llama_context * ctx = nullptr;
+    std::vector<llama_lora_adapter_container> lora_adapters;
 
     gpt_params params;
 
@@ -629,6 +632,7 @@ struct server_context {
 
     bool clean_kv_cache = true;
     bool add_bos_token  = true;
+    bool has_eos_token  = false;
 
     int32_t n_ctx; // total context for all clients / slots
 
@@ -677,7 +681,11 @@ struct server_context {
         // dedicate one sequence to the system prompt
         params.n_parallel += 1;
 
-        std::tie(model, ctx) = llama_init_from_gpt_params(params);
+        llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+        model = llama_init.model;
+        ctx = llama_init.context;
+        lora_adapters = llama_init.lora_adapters;
         params.n_parallel -= 1; // but be sneaky about it
         if (model == nullptr) {
             LOG_ERROR("unable to load model", {{"model", params.model}});
@@ -686,8 +694,8 @@ struct server_context {
 
         n_ctx = llama_n_ctx(ctx);
 
-        add_bos_token = llama_should_add_bos_token(model);
-        GGML_ASSERT(llama_add_eos_token(model) != 1);
+        add_bos_token = llama_add_bos_token(model);
+        has_eos_token = !llama_add_eos_token(model);
 
         return true;
     }
@@ -747,13 +755,13 @@ struct server_context {
         default_generation_settings_for_props = get_formated_generation(slots.front());
         default_generation_settings_for_props["seed"] = -1;
 
-        // the update_slots() logic will always submit a maximum of n_batch tokens
+        // the update_slots() logic will always submit a maximum of n_batch or n_parallel tokens
         // note that n_batch can be > n_ctx (e.g. for non-causal attention models such as BERT where the KV cache is not used)
         {
             const int32_t n_batch = llama_n_batch(ctx);
 
             // only a single seq_id per token is needed
-            batch = llama_batch_init(n_batch, 0, 1);
+            batch = llama_batch_init(std::max(n_batch, params.n_parallel), 0, 1);
         }
 
         metrics.init();
@@ -900,7 +908,7 @@ struct server_context {
 
         slot.params.stream             = json_value(data, "stream",            false);
         slot.params.cache_prompt       = json_value(data, "cache_prompt",      false);
-        slot.params.n_predict          = json_value(data, "n_predict",         default_params.n_predict);
+        slot.params.n_predict          = json_value(data, "n_predict",         json_value(data, "max_tokens", default_params.n_predict));
         slot.sparams.top_k             = json_value(data, "top_k",             default_sparams.top_k);
         slot.sparams.top_p             = json_value(data, "top_p",             default_sparams.top_p);
         slot.sparams.min_p             = json_value(data, "min_p",             default_sparams.min_p);
@@ -969,6 +977,8 @@ struct server_context {
                 (prompt->is_array() &&  prompt->size() == 1 && prompt->at(0).is_string()) ||
                 (prompt->is_array() && !prompt->empty()     && prompt->at(0).is_number_integer())) {
                 slot.prompt = *prompt;
+            } else if (prompt->is_array() && prompt->size() == 1 && prompt->at(0).is_array()) {
+                slot.prompt = prompt->at(0);
             } else {
                 send_error(task, "\"prompt\" must be a string or an array of integers", ERROR_TYPE_INVALID_REQUEST);
                 return false;
@@ -1023,7 +1033,7 @@ struct server_context {
         {
             slot.sparams.logit_bias.clear();
 
-            if (json_value(data, "ignore_eos", false)) {
+            if (json_value(data, "ignore_eos", false) && has_eos_token) {
                 slot.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
             }
 
@@ -1128,28 +1138,19 @@ struct server_context {
         if (!system_prompt.empty()) {
             system_tokens = ::llama_tokenize(ctx, system_prompt, true);
 
-            llama_batch_clear(batch);
-
-            for (int i = 0; i < (int)system_tokens.size(); ++i) {
-                llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
-            }
-
             const int32_t n_batch = llama_n_batch(ctx);
+            const int32_t n_tokens_prompt = system_tokens.size();
 
-            for (int32_t i = 0; i < batch.n_tokens; i += n_batch) {
-                const int32_t n_tokens = std::min(params.n_batch, batch.n_tokens - i);
-                llama_batch batch_view = {
-                    n_tokens,
-                    batch.token    + i,
-                    nullptr,
-                    batch.pos      + i,
-                    batch.n_seq_id + i,
-                    batch.seq_id   + i,
-                    batch.logits   + i,
-                    0, 0, 0, // unused
-                };
+            for (int32_t i = 0; i < n_tokens_prompt; i += n_batch) {
+                const int32_t n_tokens = std::min(n_batch, n_tokens_prompt - i);
 
-                if (llama_decode(ctx, batch_view) != 0) {
+                llama_batch_clear(batch);
+
+                for (int32_t j = 0; j < n_tokens; ++j) {
+                    llama_batch_add(batch, system_tokens[i + j], i + j, { 0 }, false);
+                }
+
+                if (llama_decode(ctx, batch) != 0) {
                     LOG_ERROR("llama_decode() failed", {});
                     return;
                 }
@@ -1322,7 +1323,7 @@ struct server_context {
 
         return json {
             {"n_ctx",                     slot.n_ctx},
-            {"n_predict",                 slot.n_predict},
+            {"n_predict",                 slot.n_predict},     // Server configured n_predict
             {"model",                     params.model_alias},
             {"seed",                      slot.sparams.seed},
             {"temperature",               slot.sparams.temp},
@@ -1344,7 +1345,7 @@ struct server_context {
             {"mirostat_eta",              slot.sparams.mirostat_eta},
             {"penalize_nl",               slot.sparams.penalize_nl},
             {"stop",                      slot.params.antiprompt},
-            {"n_predict",                 slot.params.n_predict}, // TODO: fix duplicate key n_predict
+            {"max_tokens",                slot.params.n_predict}, // User configured n_predict
             {"n_keep",                    slot.params.n_keep},
             {"n_discard",                 slot.params.n_discard},
             {"ignore_eos",                ignore_eos},
@@ -1847,6 +1848,16 @@ struct server_context {
                     };
                     queue_results.send(result);
                 } break;
+            case SERVER_TASK_TYPE_SET_LORA:
+                {
+                    llama_lora_adapters_apply(ctx, lora_adapters);
+                    server_task_result result;
+                    result.id = task.id;
+                    result.stop = true;
+                    result.error = false;
+                    result.data = json{{ "success", true }};
+                    queue_results.send(result);
+                } break;
         }
     }
 
@@ -2028,7 +2039,7 @@ struct server_context {
                         slot.t_start_generation = 0;
 
                         if (slot.infill) {
-                            const bool add_bos = llama_should_add_bos_token(model);
+                            const bool add_bos = llama_add_bos_token(model);
                             bool suff_rm_leading_spc = true;
                             if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
                                 params.input_suffix.erase(0, 1);
@@ -2496,6 +2507,9 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    // parse arguments from environment variables
+    gpt_params_parse_from_env(params);
+
     // TODO: not great to use extern vars
     server_log_json = params.log_json;
     server_verbose = params.verbosity > 0;
@@ -2520,8 +2534,8 @@ int main(int argc, char ** argv) {
     });
 
     LOG_INFO("system info", {
-        {"n_threads",       params.n_threads},
-        {"n_threads_batch", params.n_threads_batch},
+        {"n_threads",       params.cpuparams.n_threads},
+        {"n_threads_batch", params.cpuparams_batch.n_threads},
         {"total_threads",   std::thread::hardware_concurrency()},
         {"system_info",     llama_print_system_info()},
     });
@@ -2546,19 +2560,19 @@ int main(int argc, char ** argv) {
     svr->set_default_headers({{"Server", "llama.cpp"}});
 
     // CORS preflight
-    svr->Options(R"(.*)", [](const httplib::Request & req, httplib::Response & res) {
-        res.set_header("Access-Control-Allow-Origin",      req.get_header_value("Origin"));
+    svr->Options(R"(.*)", [](const httplib::Request &, httplib::Response & res) {
+        // Access-Control-Allow-Origin is already set by middleware
         res.set_header("Access-Control-Allow-Credentials", "true");
         res.set_header("Access-Control-Allow-Methods",     "POST");
         res.set_header("Access-Control-Allow-Headers",     "*");
-        return res.set_content("", "application/json; charset=utf-8");
+        return res.set_content("", "text/html"); // blank response, no data
     });
 
     svr->set_logger(log_server_request);
 
     auto res_error = [](httplib::Response & res, json error_data) {
         json final_response {{"error", error_data}};
-        res.set_content(final_response.dump(), "application/json; charset=utf-8");
+        res.set_content(final_response.dump(-1, ' ', false, json::error_handler_t::replace), MIMETYPE_JSON);
         res.status = json_value(error_data, "code", 500);
     };
 
@@ -2588,11 +2602,6 @@ int main(int argc, char ** argv) {
     svr->set_read_timeout (params.timeout_read);
     svr->set_write_timeout(params.timeout_write);
 
-    if (!svr->bind_to_port(params.hostname, params.port)) {
-        fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", params.hostname.c_str(), params.port);
-        return 1;
-    }
-
     std::unordered_map<std::string, std::string> log_data;
 
     log_data["hostname"] = params.hostname;
@@ -2608,35 +2617,6 @@ int main(int argc, char ** argv) {
     // Necessary similarity of prompt for slot selection
     ctx_server.slot_prompt_similarity = params.slot_prompt_similarity;
 
-    // load the model
-    if (!ctx_server.load_model(params)) {
-        state.store(SERVER_STATE_ERROR);
-        return 1;
-    } else {
-        ctx_server.init();
-        state.store(SERVER_STATE_READY);
-    }
-
-    LOG_INFO("model loaded", {});
-
-    const auto model_meta = ctx_server.model_meta();
-
-    // if a custom chat template is not supplied, we will use the one that comes with the model (if any)
-    if (params.chat_template.empty()) {
-        if (!ctx_server.validate_model_chat_template()) {
-            LOG_WARNING("The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {});
-            params.chat_template = "chatml";
-        }
-    }
-
-    // print sample chat example to make it clear which template is used
-    {
-        LOG_INFO("chat template", {
-            {"chat_example", llama_chat_format_example(ctx_server.model, params.chat_template)},
-            {"built_in",     params.chat_template.empty()},
-        });
-    }
-
     //
     // Middlewares
     //
@@ -2680,8 +2660,6 @@ int main(int argc, char ** argv) {
         }
 
         // API key is invalid or not provided
-        // TODO: make another middleware for CORS related logic
-        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
         res_error(res, format_error_response("Invalid API Key", ERROR_TYPE_AUTHENTICATION));
 
         LOG_WARNING("Unauthorized: Invalid API Key", {});
@@ -2689,8 +2667,21 @@ int main(int argc, char ** argv) {
         return false;
     };
 
+    auto middleware_server_state = [&res_error, &state](const httplib::Request &, httplib::Response & res) {
+        server_state current_state = state.load();
+        if (current_state == SERVER_STATE_LOADING_MODEL) {
+            res_error(res, format_error_response("Loading model", ERROR_TYPE_UNAVAILABLE));
+            return false;
+        }
+        return true;
+    };
+
     // register server middlewares
-    svr->set_pre_routing_handler([&middleware_validate_api_key](const httplib::Request & req, httplib::Response & res) {
+    svr->set_pre_routing_handler([&middleware_validate_api_key, &middleware_server_state](const httplib::Request & req, httplib::Response & res) {
+        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
+        if (!middleware_server_state(req, res)) {
+            return httplib::Server::HandlerResponse::Handled;
+        }
         if (!middleware_validate_api_key(req, res)) {
             return httplib::Server::HandlerResponse::Handled;
         }
@@ -2701,62 +2692,15 @@ int main(int argc, char ** argv) {
     // Route handlers (or controllers)
     //
 
-    const auto handle_health = [&](const httplib::Request & req, httplib::Response & res) {
-        server_state current_state = state.load();
-        switch (current_state) {
-            case SERVER_STATE_READY:
-                {
-                    // request slots data using task queue
-                    server_task task;
-                    task.id   = ctx_server.queue_tasks.get_new_id();
-                    task.type = SERVER_TASK_TYPE_METRICS;
-                    task.id_target = -1;
-
-                    ctx_server.queue_results.add_waiting_task_id(task.id);
-                    ctx_server.queue_tasks.post(task);
-
-                    // get the result
-                    server_task_result result = ctx_server.queue_results.recv(task.id);
-                    ctx_server.queue_results.remove_waiting_task_id(task.id);
-
-                    const int n_idle_slots       = result.data.at("idle");
-                    const int n_processing_slots = result.data.at("processing");
-
-                    json health = {
-                        {"status",           "ok"},
-                        {"slots_idle",       n_idle_slots},
-                        {"slots_processing", n_processing_slots}
-                    };
-
-                    res.status = 200; // HTTP OK
-                    if (params.endpoint_slots && req.has_param("include_slots")) {
-                        health["slots"] = result.data.at("slots");
-                    }
-
-                    if (n_idle_slots == 0) {
-                        health["status"] = "no slot available";
-                        if (req.has_param("fail_on_no_slot")) {
-                            res.status = 503; // HTTP Service Unavailable
-                        }
-                    }
-
-                    res.set_content(health.dump(), "application/json");
-                    break;
-                }
-            case SERVER_STATE_LOADING_MODEL:
-                {
-                    res_error(res, format_error_response("Loading model", ERROR_TYPE_UNAVAILABLE));
-                } break;
-            case SERVER_STATE_ERROR:
-                {
-                    res_error(res, format_error_response("Model failed to load", ERROR_TYPE_SERVER));
-                } break;
-        }
+    const auto handle_health = [&](const httplib::Request &, httplib::Response & res) {
+        // error and loading states are handled by middleware
+        json health = {{"status", "ok"}};
+        res.set_content(health.dump(), "application/json");
     };
 
-    const auto handle_slots = [&](const httplib::Request &, httplib::Response & res) {
+    const auto handle_slots = [&](const httplib::Request & req, httplib::Response & res) {
         if (!params.endpoint_slots) {
-            res_error(res, format_error_response("This server does not support slots endpoint.", ERROR_TYPE_NOT_SUPPORTED));
+            res_error(res, format_error_response("This server does not support slots endpoint. Start it without `--no-slots`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
 
@@ -2774,13 +2718,22 @@ int main(int argc, char ** argv) {
         server_task_result result = ctx_server.queue_results.recv(task.id);
         ctx_server.queue_results.remove_waiting_task_id(task.id);
 
-        res.set_content(result.data.at("slots").dump(), "application/json");
+        // optionally return "fail_on_no_slot" error
+        const int n_idle_slots = result.data.at("idle");
+        if (req.has_param("fail_on_no_slot")) {
+            if (n_idle_slots == 0) {
+                res_error(res, format_error_response("no slot available", ERROR_TYPE_UNAVAILABLE));
+                return;
+            }
+        }
+
+        res.set_content(result.data.at("slots").dump(), MIMETYPE_JSON);
         res.status = 200; // HTTP OK
     };
 
     const auto handle_metrics = [&](const httplib::Request &, httplib::Response & res) {
         if (!params.endpoint_metrics) {
-            res_error(res, format_error_response("This server does not support metrics endpoint.", ERROR_TYPE_NOT_SUPPORTED));
+            res_error(res, format_error_response("This server does not support metrics endpoint. Start it with `--metrics`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
 
@@ -2905,7 +2858,7 @@ int main(int argc, char ** argv) {
         if (result.error) {
             res_error(res, result.data);
         } else {
-            res.set_content(result.data.dump(), "application/json");
+            res.set_content(result.data.dump(), MIMETYPE_JSON);
         }
     };
 
@@ -2935,7 +2888,7 @@ int main(int argc, char ** argv) {
         if (result.error) {
             res_error(res, result.data);
         } else {
-            res.set_content(result.data.dump(), "application/json");
+            res.set_content(result.data.dump(), MIMETYPE_JSON);
         }
     };
 
@@ -2955,13 +2908,11 @@ int main(int argc, char ** argv) {
         if (result.error) {
             res_error(res, result.data);
         } else {
-            res.set_content(result.data.dump(), "application/json");
+            res.set_content(result.data.dump(), MIMETYPE_JSON);
         }
     };
 
     const auto handle_slots_action = [&res_error, &handle_slots_save, &handle_slots_restore, &handle_slots_erase](const httplib::Request & req, httplib::Response & res) {
-        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
-
         std::string id_slot_str = req.path_params.at("id_slot");
         int id_slot;
 
@@ -2985,7 +2936,7 @@ int main(int argc, char ** argv) {
         }
     };
 
-    const auto handle_props = [&ctx_server](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_props = [&ctx_server](const httplib::Request &, httplib::Response & res) {
         std::string template_key = "tokenizer.chat_template", curr_tmpl;
         int32_t tlen = llama_model_meta_val_str(ctx_server.model, template_key.c_str(), nullptr, 0);
         if (tlen > 0) {
@@ -2994,7 +2945,6 @@ int main(int argc, char ** argv) {
                 curr_tmpl = std::string(curr_tmpl_buf.data(), tlen);
             }
         }
-        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
         json data = {
             { "system_prompt",               ctx_server.system_prompt.c_str() },
             { "default_generation_settings", ctx_server.default_generation_settings_for_props },
@@ -3002,7 +2952,7 @@ int main(int argc, char ** argv) {
             { "chat_template",               curr_tmpl.c_str() }
         };
 
-        res.set_content(data.dump(), "application/json; charset=utf-8");
+        res.set_content(data.dump(), MIMETYPE_JSON);
     };
 
     const auto handle_completions = [&ctx_server, &res_error](const httplib::Request & req, httplib::Response & res) {
@@ -3011,8 +2961,6 @@ int main(int argc, char ** argv) {
             return;
         }
 
-        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
-
         json data = json::parse(req.body);
 
         const int id_task = ctx_server.queue_tasks.get_new_id();
@@ -3023,7 +2971,7 @@ int main(int argc, char ** argv) {
         if (!json_value(data, "stream", false)) {
             server_task_result result = ctx_server.queue_results.recv(id_task);
             if (!result.error && result.stop) {
-                res.set_content(result.data.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8");
+                res.set_content(result.data.dump(-1, ' ', false, json::error_handler_t::replace), MIMETYPE_JSON);
             } else {
                 res_error(res, result.data);
             }
@@ -3086,9 +3034,7 @@ int main(int argc, char ** argv) {
         }
     };
 
-    const auto handle_models = [&params, &model_meta](const httplib::Request & req, httplib::Response & res) {
-        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
-
+    const auto handle_models = [&params, &ctx_server](const httplib::Request &, httplib::Response & res) {
         json models = {
             {"object", "list"},
             {"data", {
@@ -3097,12 +3043,12 @@ int main(int argc, char ** argv) {
                      {"object",   "model"},
                      {"created",  std::time(0)},
                      {"owned_by", "llamacpp"},
-                     {"meta",     model_meta}
+                     {"meta",     ctx_server.model_meta()}
                  },
              }}
         };
 
-        res.set_content(models.dump(), "application/json; charset=utf-8");
+        res.set_content(models.dump(), MIMETYPE_JSON);
     };
 
     const auto handle_chat_completions = [&ctx_server, &params, &res_error](const httplib::Request & req, httplib::Response & res) {
@@ -3110,8 +3056,6 @@ int main(int argc, char ** argv) {
             res_error(res, format_error_response("This server does not support chat completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
-
-        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
         json data = oaicompat_completion_params_parse(ctx_server.model, json::parse(req.body), params.chat_template);
 
         const int id_task = ctx_server.queue_tasks.get_new_id();
@@ -3126,7 +3070,7 @@ int main(int argc, char ** argv) {
             if (!result.error && result.stop) {
                 json result_oai = format_final_response_oaicompat(data, result.data, completion_id);
 
-                res.set_content(result_oai.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8");
+                res.set_content(result_oai.dump(-1, ' ', false, json::error_handler_t::replace), MIMETYPE_JSON);
             } else {
                 res_error(res, result.data);
             }
@@ -3188,8 +3132,6 @@ int main(int argc, char ** argv) {
             return;
         }
 
-        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
-
         json data = json::parse(req.body);
 
         const int id_task = ctx_server.queue_tasks.get_new_id();
@@ -3200,7 +3142,7 @@ int main(int argc, char ** argv) {
         if (!json_value(data, "stream", false)) {
             server_task_result result = ctx_server.queue_results.recv(id_task);
             if (!result.error && result.stop) {
-                res.set_content(result.data.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8");
+                res.set_content(result.data.dump(-1, ' ', false, json::error_handler_t::replace), MIMETYPE_JSON);
             } else {
                 res_error(res, result.data);
             }
@@ -3248,7 +3190,6 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_tokenize = [&ctx_server](const httplib::Request & req, httplib::Response & res) {
-        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
         const json body = json::parse(req.body);
 
         std::vector<llama_token> tokens;
@@ -3257,11 +3198,10 @@ int main(int argc, char ** argv) {
             tokens = ctx_server.tokenize(body.at("content"), add_special);
         }
         const json data = format_tokenizer_response(tokens);
-        return res.set_content(data.dump(), "application/json; charset=utf-8");
+        return res.set_content(data.dump(), MIMETYPE_JSON);
     };
 
     const auto handle_detokenize = [&ctx_server](const httplib::Request & req, httplib::Response & res) {
-        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
         const json body = json::parse(req.body);
 
         std::string content;
@@ -3271,12 +3211,10 @@ int main(int argc, char ** argv) {
         }
 
         const json data = format_detokenized_response(content);
-        return res.set_content(data.dump(), "application/json; charset=utf-8");
+        return res.set_content(data.dump(), MIMETYPE_JSON);
     };
 
     const auto handle_embeddings = [&ctx_server, &res_error](const httplib::Request & req, httplib::Response & res) {
-        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
-
         const json body = json::parse(req.body);
         bool is_openai = false;
 
@@ -3322,7 +3260,53 @@ int main(int argc, char ** argv) {
         json root = is_openai
             ? format_embeddings_response_oaicompat(body, responses)
             : responses[0];
-        return res.set_content(root.dump(), "application/json; charset=utf-8");
+        return res.set_content(root.dump(), MIMETYPE_JSON);
+    };
+
+    const auto handle_lora_adapters_list = [&](const httplib::Request &, httplib::Response & res) {
+        json result = json::array();
+        for (size_t i = 0; i < ctx_server.lora_adapters.size(); ++i) {
+            auto & la = ctx_server.lora_adapters[i];
+            result.push_back({
+                {"id", i},
+                {"path", la.path},
+                {"scale", la.scale},
+            });
+        }
+        res.set_content(result.dump(), MIMETYPE_JSON);
+        res.status = 200; // HTTP OK
+    };
+
+    const auto handle_lora_adapters_apply = [&](const httplib::Request & req, httplib::Response & res) {
+        const std::vector<json> body = json::parse(req.body);
+        int max_idx = ctx_server.lora_adapters.size();
+
+        // clear existing value
+        for (auto & la : ctx_server.lora_adapters) {
+            la.scale = 0.0f;
+        }
+
+        // set value
+        for (auto entry : body) {
+            int id      = entry.at("id");
+            float scale = entry.at("scale");
+            if (0 <= id && id < max_idx) {
+                ctx_server.lora_adapters[id].scale = scale;
+            } else {
+                throw std::runtime_error("invalid adapter id");
+            }
+        }
+
+        server_task task;
+        task.type = SERVER_TASK_TYPE_SET_LORA;
+        const int id_task = ctx_server.queue_tasks.post(task);
+        ctx_server.queue_results.add_waiting_task_id(id_task);
+
+        server_task_result result = ctx_server.queue_results.recv(id_task);
+        ctx_server.queue_results.remove_waiting_task_id(id_task);
+
+        res.set_content(result.data.dump(), MIMETYPE_JSON);
+        res.status = 200; // HTTP OK
     };
 
     auto handle_static_file = [](unsigned char * content, size_t len, const char * mime_type) {
@@ -3363,7 +3347,6 @@ int main(int argc, char ** argv) {
 
     // register API routes
     svr->Get ("/health",              handle_health);
-    svr->Get ("/slots",               handle_slots);
     svr->Get ("/metrics",             handle_metrics);
     svr->Get ("/props",               handle_props);
     svr->Get ("/v1/models",           handle_models);
@@ -3378,6 +3361,11 @@ int main(int argc, char ** argv) {
     svr->Post("/v1/embeddings",       handle_embeddings);
     svr->Post("/tokenize",            handle_tokenize);
     svr->Post("/detokenize",          handle_detokenize);
+    // LoRA adapters hotswap
+    svr->Get ("/lora-adapters",       handle_lora_adapters_list);
+    svr->Post("/lora-adapters",       handle_lora_adapters_apply);
+    // Save & load slots
+    svr->Get ("/slots",               handle_slots);
     if (!params.slot_save_path.empty()) {
         // only enable slot endpoints if slot_save_path is set
         svr->Post("/slots/:id_slot",  handle_slots_action);
@@ -3393,35 +3381,75 @@ int main(int argc, char ** argv) {
     log_data["n_threads_http"] =  std::to_string(params.n_threads_http);
     svr->new_task_queue = [&params] { return new httplib::ThreadPool(params.n_threads_http); };
 
-    LOG_INFO("HTTP server listening", log_data);
+    // clean up function, to be called before exit
+    auto clean_up = [&svr]() {
+        svr->stop();
+        llama_backend_free();
+    };
 
-    // run the HTTP server in a thread - see comment below
-    std::thread t([&]() {
-        if (!svr->listen_after_bind()) {
-            state.store(SERVER_STATE_ERROR);
-            return 1;
+    // bind HTTP listen port, run the HTTP server in a thread
+    if (!svr->bind_to_port(params.hostname, params.port)) {
+        LOG_ERROR("couldn't bind HTTP server socket", {
+            {"hostname", params.hostname},
+            {"port", params.port},
+        });
+        clean_up();
+        LOG_ERROR("exiting due to HTTP server error", {});
+        return 1;
+    }
+    std::thread t([&]() { svr->listen_after_bind(); });
+    svr->wait_until_ready();
+
+    LOG_INFO("HTTP server is listening", log_data);
+
+    // load the model
+    LOG_INFO("loading model", log_data);
+    if (!ctx_server.load_model(params)) {
+        clean_up();
+        t.join();
+        LOG_ERROR("exiting due to model loading error", {});
+        return 1;
+    } else {
+        ctx_server.init();
+        state.store(SERVER_STATE_READY);
+
+        LOG_INFO("model loaded", {});
+
+        // if a custom chat template is not supplied, we will use the one that comes with the model (if any)
+        if (params.chat_template.empty()) {
+            if (!ctx_server.validate_model_chat_template()) {
+                LOG_WARNING("The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {});
+                params.chat_template = "chatml";
+            }
         }
 
-        return 0;
-    });
+        // print sample chat example to make it clear which template is used
+        {
+            LOG_INFO("chat template", {
+                {"chat_example", llama_chat_format_example(ctx_server.model, params.chat_template)},
+                {"built_in",     params.chat_template.empty()},
+            });
+        }
 
-    ctx_server.queue_tasks.on_new_task(std::bind(
-        &server_context::process_single_task, &ctx_server, std::placeholders::_1));
-    ctx_server.queue_tasks.on_finish_multitask(std::bind(
-        &server_context::on_finish_multitask, &ctx_server, std::placeholders::_1));
-    ctx_server.queue_tasks.on_update_slots(std::bind(
-        &server_context::update_slots, &ctx_server));
-    ctx_server.queue_results.on_multitask_update(std::bind(
-        &server_queue::update_multitask,
-        &ctx_server.queue_tasks,
-        std::placeholders::_1,
-        std::placeholders::_2,
-        std::placeholders::_3
-    ));
+        ctx_server.queue_tasks.on_new_task(std::bind(
+            &server_context::process_single_task, &ctx_server, std::placeholders::_1));
+        ctx_server.queue_tasks.on_finish_multitask(std::bind(
+            &server_context::on_finish_multitask, &ctx_server, std::placeholders::_1));
+        ctx_server.queue_tasks.on_update_slots(std::bind(
+            &server_context::update_slots, &ctx_server));
+        ctx_server.queue_results.on_multitask_update(std::bind(
+            &server_queue::update_multitask,
+            &ctx_server.queue_tasks,
+            std::placeholders::_1,
+            std::placeholders::_2,
+            std::placeholders::_3
+        ));
 
-    shutdown_handler = [&](int) {
-        ctx_server.queue_tasks.terminate();
-    };
+        shutdown_handler = [&](int) {
+            ctx_server.queue_tasks.terminate();
+        };
+        ctx_server.queue_tasks.start_loop();
+    }
 
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
     struct sigaction sigint_action;
@@ -3437,12 +3465,8 @@ int main(int argc, char ** argv) {
     SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
 
-    ctx_server.queue_tasks.start_loop();
-
-    svr->stop();
+    clean_up();
     t.join();
 
-    llama_backend_free();
-
     return 0;
 }

examples/server/tests/features/lora.feature

@@ -0,0 +1,36 @@
+@llama.cpp
+@lora
+Feature: llama.cpp server
+
+  Background: Server startup
+    Given a server listening on localhost:8080
+    And   a model url https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/stories15M_MOE-F16.gguf
+    And   a model file stories15M_MOE-F16.gguf
+    And   a model alias stories15M_MOE
+    And   a lora adapter file from https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/moe_shakespeare15M.gguf
+    And   42 as server seed
+    And   1024 as batch size
+    And   1024 as ubatch size
+    And   2048 KV cache size
+    And   64 max tokens to predict
+    And   0.0 temperature
+    Then  the server is starting
+    Then  the server is healthy
+
+  Scenario: Completion LoRA disabled
+    Given switch off lora adapter 0
+    Given a prompt:
+    """
+    Look in thy glass
+    """
+    And   a completion request with no api error
+    Then  64 tokens are predicted matching little|girl|three|years|old
+
+  Scenario: Completion LoRA enabled
+    Given switch on lora adapter 0
+    Given a prompt:
+    """
+    Look in thy glass
+    """
+    And   a completion request with no api error
+    Then  64 tokens are predicted matching eye|love|glass|sun

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

@@ -7,6 +7,7 @@ import subprocess
 import sys
 import threading
 import time
+import requests
 from collections.abc import Sequence
 from contextlib import closing
 from re import RegexFlag
@@ -70,6 +71,7 @@ def step_server_config(context, server_fqdn: str, server_port: str):
     context.user_api_key = None
     context.response_format = None
     context.temperature = None
+    context.lora_file = None
 
     context.tasks_result = []
     context.concurrent_tasks = []
@@ -82,6 +84,12 @@ def step_download_hf_model(context, hf_file: str, hf_repo: str):
     context.model_hf_file = hf_file
     context.model_file = os.path.basename(hf_file)
 
+@step('a lora adapter file from {lora_file_url}')
+def step_download_lora_file(context, lora_file_url: str):
+    file_name = lora_file_url.split('/').pop()
+    context.lora_file = f'../../../{file_name}'
+    with open(context.lora_file, 'wb') as f:
+        f.write(requests.get(lora_file_url).content)
 
 @step('a model file {model_file}')
 def step_model_file(context, model_file: str):
@@ -197,27 +205,20 @@ def step_start_server(context):
 async def step_wait_for_the_server_to_be_started(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str):
     match expecting_status:
         case 'healthy':
-            await wait_for_health_status(context, context.base_url, 200, 'ok',
-                                         timeout=30)
+            await wait_for_slots_status(context, context.base_url, 200,
+                                        timeout=30)
 
         case 'ready' | 'idle':
-            await wait_for_health_status(context, context.base_url, 200, 'ok',
-                                         timeout=30,
-                                         params={'fail_on_no_slot': 0, 'include_slots': 0},
-                                         slots_idle=context.n_slots,
-                                         slots_processing=0,
-                                         expected_slots=[{'id': slot_id, 'state': 0}
-                                                         for slot_id in
-                                                         range(context.n_slots if context.n_slots else 1)])
+            await wait_for_slots_status(context, context.base_url, 200,
+                                        timeout=30,
+                                        params={'fail_on_no_slot': 1},
+                                        slots_idle=context.n_slots,
+                                        slots_processing=0)
         case 'busy':
-            await wait_for_health_status(context, context.base_url, 503,
-                                         'no slot available',
-                                         params={'fail_on_no_slot': 0, 'include_slots': 0},
-                                         slots_idle=0,
-                                         slots_processing=context.n_slots,
-                                         expected_slots=[{'id': slot_id, 'state': 1}
-                                                         for slot_id in
-                                                         range(context.n_slots if context.n_slots else 1)])
+            await wait_for_slots_status(context, context.base_url, 503,
+                                        params={'fail_on_no_slot': 1},
+                                        slots_idle=0,
+                                        slots_processing=context.n_slots)
         case _:
             assert False, "unknown status"
 
@@ -849,6 +850,17 @@ async def step_erase_slot(context, slot_id):
             context.response = response
 
 
+@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 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:
+            context.response = response
+            print([{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}])
+
+
 @step('the server responds with status code {status_code:d}')
 def step_server_responds_with_status_code(context, status_code):
     assert context.response.status == status_code
@@ -1168,17 +1180,15 @@ async def gather_tasks_results(context):
     return n_completions
 
 
-async def wait_for_health_status(context,
-                                 base_url,
-                                 expected_http_status_code,
-                                 expected_health_status,
-                                 timeout=3,
-                                 params=None,
-                                 slots_idle=None,
-                                 slots_processing=None,
-                                 expected_slots=None):
+async def wait_for_slots_status(context,
+                                base_url,
+                                expected_http_status_code,
+                                timeout=3,
+                                params=None,
+                                slots_idle=None,
+                                slots_processing=None):
     if context.debug:
-        print(f"Starting checking for health for expected_health_status={expected_health_status}")
+        print(f"Starting checking for health for expected_http_status_code={expected_http_status_code}")
     interval = 0.5
     counter = 0
     if 'GITHUB_ACTIONS' in os.environ:
@@ -1186,26 +1196,19 @@ async def wait_for_health_status(context,
 
     async with aiohttp.ClientSession() as session:
         while True:
-            async with await session.get(f'{base_url}/health', params=params) as health_response:
-                status_code = health_response.status
-                health = await health_response.json()
+            async with await session.get(f'{base_url}/slots', params=params) as slots_response:
+                status_code = slots_response.status
+                slots = await slots_response.json()
                 if context.debug:
-                    print(f"HEALTH - response for expected health status='{expected_health_status}' on "
-                          f"'{base_url}/health'?{params} is {health}\n")
-                if (status_code == expected_http_status_code
-                        and health['status'] == expected_health_status
-                        and (slots_idle is None or health['slots_idle'] == slots_idle)
-                        and (slots_processing is None or health['slots_processing'] == slots_processing)):
-                    if expected_slots is not None:
-                        assert_slots_status(health['slots'], expected_slots)
-                    return
-                if (status_code == expected_http_status_code
-                        and health['status'] == expected_health_status
-                        and (slots_idle is None or health['slots_idle'] == slots_idle)
-                        and (slots_processing is None or health['slots_processing'] == slots_processing)):
-                    if expected_slots is not None:
-                        assert_slots_status(health['slots'], expected_slots)
+                    print(f"slots responses {slots}\n")
+                if status_code == 503 and status_code == expected_http_status_code:
                     return
+                if status_code == 200 and status_code == expected_http_status_code:
+                    n_slots_idle = sum(1 if slot["state"] == 0 else 0 for slot in slots)
+                    n_slots_processing = sum(1 if slot["state"] != 0 else 0 for slot in slots)
+                    if ((slots_idle is None or slots_idle == n_slots_idle)
+                        and (slots_processing is None or slots_processing == n_slots_processing)):
+                        return
             await asyncio.sleep(interval)
 
             counter += interval
@@ -1219,7 +1222,7 @@ async def wait_for_health_status(context,
                         if n_completions > 0:
                             return
 
-                assert False, f'{expected_health_status} timeout exceeded {counter}s>={timeout}'
+                assert False, f'slots check timeout exceeded {counter}s>={timeout}'
 
 
 def assert_embeddings(embeddings):
@@ -1326,6 +1329,8 @@ def start_server_background(context):
         server_args.extend(['--grp-attn-w', context.n_ga_w])
     if context.debug:
         server_args.append('--verbose')
+    if context.lora_file:
+        server_args.extend(['--lora', context.lora_file])
     if 'SERVER_LOG_FORMAT_JSON' not in os.environ:
         server_args.extend(['--log-format', "text"])
 

examples/server/tests/requirements.txt

@@ -4,3 +4,4 @@ huggingface_hub~=0.20.3
 numpy~=1.26.4
 openai~=1.30.3
 prometheus-client~=0.20.0
+requests~=2.32.3

examples/server/utils.hpp

@@ -355,24 +355,6 @@ static json oaicompat_completion_params_parse(
 
     llama_params["__oaicompat"] = true;
 
-    // Map OpenAI parameters to llama.cpp parameters
-    //
-    // For parameters that are defined by the OpenAI documentation (e.g.
-    // temperature), we explicitly specify OpenAI's intended default; we
-    // need to do that because sometimes OpenAI disagrees with llama.cpp
-    //
-    // https://platform.openai.com/docs/api-reference/chat/create
-    llama_sampling_params default_sparams;
-    llama_params["model"]             = json_value(body,   "model",             std::string("unknown"));
-    llama_params["frequency_penalty"] = json_value(body,   "frequency_penalty", 0.0);
-    llama_params["logit_bias"]        = json_value(body,   "logit_bias",        json::object());
-    llama_params["n_predict"]         = json_value(body,   "max_tokens",        -1);
-    llama_params["presence_penalty"]  = json_value(body,   "presence_penalty",  0.0);
-    llama_params["seed"]              = json_value(body,   "seed",              LLAMA_DEFAULT_SEED);
-    llama_params["stream"]            = json_value(body,   "stream",            false);
-    llama_params["temperature"]       = json_value(body,   "temperature",       1.0);
-    llama_params["top_p"]             = json_value(body,   "top_p",             1.0);
-
     // Apply chat template to the list of messages
     llama_params["prompt"] = format_chat(model, chat_template, body.at("messages"));
 

examples/simple/README.md

@@ -3,7 +3,7 @@
 The purpose of this example is to demonstrate a minimal usage of llama.cpp for generating text with a given prompt.
 
 ```bash
-./simple -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is"
+./llama-simple -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is"
 
 ...
 

examples/speculative/speculative.cpp

@@ -66,16 +66,21 @@ int main(int argc, char ** argv) {
     llama_context * ctx_dft = NULL;
 
     // load the target model
-    std::tie(model_tgt, ctx_tgt) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init_tgt = llama_init_from_gpt_params(params);
+    model_tgt = llama_init_tgt.model;
+    ctx_tgt = llama_init_tgt.context;
 
     // load the draft model
     params.model = params.model_draft;
     params.n_gpu_layers = params.n_gpu_layers_draft;
-    if (params.n_threads_draft > 0) {
-        params.n_threads = params.n_threads_draft;
+    if (params.draft_cpuparams.n_threads > 0) {
+        params.cpuparams.n_threads = params.draft_cpuparams.n_threads;
     }
-    params.n_threads_batch = params.n_threads_batch_draft;
-    std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params);
+
+    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;
 
     const bool vocab_type_tgt = llama_vocab_type(model_tgt);
     LOG("vocab_type tgt: %d\n", vocab_type_tgt);

examples/sycl/README.md

@@ -12,9 +12,9 @@ This example program provides the tools for llama.cpp for SYCL on Intel GPU.
 
 List all SYCL devices with ID, compute capability, max work group size, ect.
 
-1. Build the llama.cpp for SYCL for all targets.
+1. Build the llama.cpp for SYCL for the specified target *(using GGML_SYCL_TARGET)*.
 
-2. Enable oneAPI running environment
+2. Enable oneAPI running environment *(if GGML_SYCL_TARGET is set to INTEL -default-)*
 
 ```
 source /opt/intel/oneapi/setvars.sh
@@ -29,19 +29,13 @@ source /opt/intel/oneapi/setvars.sh
 Check the ID in startup log, like:
 
 ```
-found 4 SYCL devices:
-  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
-    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
-  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
-    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
-  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+found 2 SYCL devices:
+|  |                   |                                       |       |Max    |        |Max  |Global |                     |
+|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |
+|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|
+|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
+| 0| [level_zero:gpu:0]|                Intel Arc A770 Graphics|    1.3|    512|    1024|   32| 16225M|            1.3.29138|
+| 1| [level_zero:gpu:1]|                 Intel UHD Graphics 750|    1.3|     32|     512|   32| 62631M|            1.3.29138|
 
 ```
 
-|Attribute|Note|
-|-|-|
-|compute capability 1.3|Level-zero running time, recommended |
-|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|

examples/sycl/win-run-llama2.bat

@@ -6,4 +6,4 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
 
-.\build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0
+.\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0

examples/tokenize/tokenize.cpp

@@ -163,7 +163,7 @@ static void write_utf8_cstr_to_stdout(const char * str, bool & invalid_utf8) {
                 printf(">");
                 return;
             }
-            GGML_ASSERT(false && "MultiByteToWideChar() failed in an unexpected way.");
+            GGML_ABORT("MultiByteToWideChar() failed in an unexpected way.");
         }
 
         LPWSTR wstr = (LPWSTR) calloc(length_needed+1, sizeof(*wstr));
@@ -362,7 +362,7 @@ int main(int raw_argc, char ** raw_argv) {
         prompt = stdin_buffer.str();
     }
 
-    const bool model_wants_add_bos = llama_should_add_bos_token(model);
+    const bool model_wants_add_bos = llama_add_bos_token(model);
     const bool add_bos = model_wants_add_bos && !no_bos;
     const bool parse_special = !no_parse_special;
 

flake.lock

@@ -5,11 +5,11 @@
         "nixpkgs-lib": "nixpkgs-lib"
       },
       "locked": {
-        "lastModified": 1719994518,
-        "narHash": "sha256-pQMhCCHyQGRzdfAkdJ4cIWiw+JNuWsTX7f0ZYSyz0VY=",
+        "lastModified": 1722555600,
+        "narHash": "sha256-XOQkdLafnb/p9ij77byFQjDf5m5QYl9b2REiVClC+x4=",
         "owner": "hercules-ci",
         "repo": "flake-parts",
-        "rev": "9227223f6d922fee3c7b190b2cc238a99527bbb7",
+        "rev": "8471fe90ad337a8074e957b69ca4d0089218391d",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1721379653,
-        "narHash": "sha256-8MUgifkJ7lkZs3u99UDZMB4kbOxvMEXQZ31FO3SopZ0=",
+        "lastModified": 1724224976,
+        "narHash": "sha256-Z/ELQhrSd7bMzTO8r7NZgi9g5emh+aRKoCdaAv5fiO0=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "1d9c2c9b3e71b9ee663d11c5d298727dace8d374",
+        "rev": "c374d94f1536013ca8e92341b540eba4c22f9c62",
         "type": "github"
       },
       "original": {
@@ -36,14 +36,14 @@
     },
     "nixpkgs-lib": {
       "locked": {
-        "lastModified": 1719876945,
-        "narHash": "sha256-Fm2rDDs86sHy0/1jxTOKB1118Q0O3Uc7EC0iXvXKpbI=",
+        "lastModified": 1722555339,
+        "narHash": "sha256-uFf2QeW7eAHlYXuDktm9c25OxOyCoUOQmh5SZ9amE5Q=",
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/5daf0514482af3f97abaefc78a6606365c9108e2.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
       },
       "original": {
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/5daf0514482af3f97abaefc78a6606365c9108e2.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
       }
     },
     "root": {

ggml/CMakeLists.txt

@@ -50,9 +50,15 @@ else()
     set(GGML_BLAS_VENDOR_DEFAULT "Generic")
 endif()
 
+if (CMAKE_CROSSCOMPILING)
+    set(GGML_NATIVE_DEFAULT OFF)
+else()
+    set(GGML_NATIVE_DEFAULT ON)
+endif()
+
 # general
 option(GGML_STATIC "ggml: static link libraries"         OFF)
-option(GGML_NATIVE "ggml: enable -march=native flag"     ON)
+option(GGML_NATIVE "ggml: enable -march=native flag"     ${GGML_NATIVE_DEFAULT})
 option(GGML_LTO    "ggml: enable link time optimization" OFF)
 option(GGML_CCACHE "ggml: use ccache if available"       ON)
 
@@ -70,7 +76,7 @@ option(GGML_SANITIZE_ADDRESS   "ggml: enable address sanitizer"   OFF)
 option(GGML_SANITIZE_UNDEFINED "ggml: enable undefined sanitizer" OFF)
 
 # instruction set specific
-if (GGML_NATIVE)
+if (GGML_NATIVE OR NOT GGML_NATIVE_DEFAULT)
     set(INS_ENB OFF)
 else()
     set(INS_ENB ON)
@@ -107,6 +113,7 @@ set(GGML_BLAS_VENDOR ${GGML_BLAS_VENDOR_DEFAULT} CACHE STRING
 option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"                             OFF)
 
 option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
+option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
 option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
 option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
@@ -122,13 +129,13 @@ option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
 option(GGML_CUDA_USE_GRAPHS                 "ggml: use CUDA graphs (llama.cpp only)"          OFF)
 
-option(GGML_CURL                            "ggml: use libcurl to download model from an URL" OFF)
 option(GGML_HIPBLAS                         "ggml: use hipBLAS"                               OFF)
 option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)
 option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug output"         OFF)
+option(GGML_VULKAN_PERF                     "ggml: enable Vulkan perf output"                 OFF)
 option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"                  OFF)
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
 option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)
@@ -200,6 +207,7 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-alloc.h
     include/ggml-backend.h
     include/ggml-blas.h
+    include/ggml-cann.h
     include/ggml-cuda.h
     include/ggml.h
     include/ggml-kompute.h

ggml/include/ggml-alloc.h

@@ -7,8 +7,8 @@ extern "C" {
 #endif
 
 typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
-typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
-typedef struct ggml_backend * ggml_backend_t;
+typedef struct      ggml_backend_buffer * ggml_backend_buffer_t;
+typedef struct             ggml_backend * ggml_backend_t;
 
 // Tensor allocator
 struct ggml_tallocr {

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-cuda.h

@@ -6,6 +6,9 @@
 #ifdef GGML_USE_HIPBLAS
 #define GGML_CUDA_NAME "ROCm"
 #define GGML_CUBLAS_NAME "hipBLAS"
+#elif defined(GGML_USE_MUSA)
+#define GGML_CUDA_NAME "MUSA"
+#define GGML_CUBLAS_NAME "muBLAS"
 #else
 #define GGML_CUDA_NAME "CUDA"
 #define GGML_CUBLAS_NAME "cuBLAS"

ggml/include/ggml-metal.h

@@ -50,6 +50,8 @@ GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void
 
 GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
 
+GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
+
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
 
 // helper to check if the device supports a specific family

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
@@ -244,6 +246,8 @@
 #define GGML_EXIT_SUCCESS 0
 #define GGML_EXIT_ABORTED 1
 
+#define GGML_ROPE_TYPE_NEOX 2
+
 #define GGUF_MAGIC "GGUF"
 
 #define GGUF_VERSION 3
@@ -254,18 +258,8 @@
 
 #define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
 
-#define GGML_ASSERT(x) \
-    do { \
-        if (!(x)) { \
-            fflush(stdout); \
-            fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
-            ggml_print_backtrace(); \
-            abort(); \
-        } \
-    } while (0)
-
 #ifndef NDEBUG
-#define GGML_UNREACHABLE() GGML_ASSERT(!"statement should not be reached")
+#define GGML_UNREACHABLE() do { fprintf(stderr, "statement should be unreachable\n"); abort(); } while(0)
 #elif defined(__GNUC__)
 #define GGML_UNREACHABLE() __builtin_unreachable()
 #elif defined(_MSC_VER)
@@ -274,6 +268,17 @@
 #define GGML_UNREACHABLE() ((void) 0)
 #endif
 
+#ifdef __cplusplus
+#define GGML_NORETURN [[noreturn]]
+#elif defined(_MSC_VER)
+#define GGML_NORETURN __declspec(noreturn)
+#else
+#define GGML_NORETURN _Noreturn
+#endif
+
+#define GGML_ABORT(...) ggml_abort(__FILE__, __LINE__, __VA_ARGS__)
+#define GGML_ASSERT(x) if (!(x)) GGML_ABORT("GGML_ASSERT(%s) failed", #x)
+
 // used to copy the number of elements and stride in bytes of tensors into local variables.
 // main purpose is to reduce code duplication and improve readability.
 //
@@ -322,6 +327,9 @@
 extern "C" {
 #endif
 
+    GGML_NORETURN GGML_ATTRIBUTE_FORMAT(3, 4)
+    GGML_API void ggml_abort(const char * file, int line, const char * fmt, ...);
+
     enum ggml_status {
         GGML_STATUS_ALLOC_FAILED = -2,
         GGML_STATUS_FAILED = -1,
@@ -345,6 +353,7 @@ extern "C" {
     GGML_API ggml_bf16_t ggml_fp32_to_bf16(float);
     GGML_API float       ggml_bf16_to_fp32(ggml_bf16_t);  // consider just doing << 16
     GGML_API void        ggml_bf16_to_fp32_row(const ggml_bf16_t *, float *, int64_t);
+    GGML_API void        ggml_fp32_to_bf16_row_ref(const float *, ggml_bf16_t *, int64_t);
     GGML_API void        ggml_fp32_to_bf16_row(const float *, ggml_bf16_t *, int64_t);
 
     struct ggml_object;
@@ -446,6 +455,8 @@ 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,
@@ -483,9 +494,11 @@ 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,
@@ -501,6 +514,7 @@ extern "C" {
         GGML_OP_WIN_UNPART,
         GGML_OP_GET_REL_POS,
         GGML_OP_ADD_REL_POS,
+        GGML_OP_RWKV_WKV,
 
         GGML_OP_UNARY,
 
@@ -535,6 +549,7 @@ extern "C" {
         GGML_UNARY_OP_SILU,
         GGML_UNARY_OP_HARDSWISH,
         GGML_UNARY_OP_HARDSIGMOID,
+        GGML_UNARY_OP_EXP,
 
         GGML_UNARY_OP_COUNT,
     };
@@ -617,6 +632,29 @@ 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 {
@@ -624,6 +662,7 @@ 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;
@@ -636,8 +675,11 @@ extern "C" {
         GGML_CGRAPH_EVAL_ORDER_COUNT
     };
 
+    typedef uint32_t ggml_bitset_t;
+
     struct ggml_hash_set {
         size_t size;
+        ggml_bitset_t * used;
         struct ggml_tensor ** keys;
     };
 
@@ -651,7 +693,7 @@ extern "C" {
         struct ggml_tensor ** grads;
         struct ggml_tensor ** leafs;
 
-        struct ggml_hash_set visited_hash_table;
+        struct ggml_hash_set visited_hash_set;
 
         enum ggml_cgraph_eval_order order;
     };
@@ -698,8 +740,6 @@ extern "C" {
     GGML_API int64_t ggml_cycles(void);
     GGML_API int64_t ggml_cycles_per_ms(void);
 
-    GGML_API void    ggml_print_backtrace(void);
-
     // accepts a UTF-8 path, even on Windows
     GGML_API FILE *  ggml_fopen(const char * fname, const char * mode);
 
@@ -961,6 +1001,22 @@ 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,
@@ -1111,6 +1167,14 @@ 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,
@@ -1134,16 +1198,17 @@ extern "C" {
 
     // group normalize along ne0*ne1*n_groups
     // used in stable-diffusion
-    // TODO: eps is hardcoded to 1e-6 for now
     GGML_API struct ggml_tensor * ggml_group_norm(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_groups);
+            int                   n_groups,
+            float                 eps);
 
     GGML_API struct ggml_tensor * ggml_group_norm_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_groups);
+            int                   n_groups,
+            float                 eps);
 
     // a - x
     // b - dy
@@ -1446,11 +1511,10 @@ extern "C" {
             struct ggml_tensor  * b);
 
     // rotary position embedding
-    // if mode & 1 == 1, skip n_past elements (NOT SUPPORTED)
-    // if mode & 2 == 1, GPT-NeoX style
+    // if (mode & 1) - skip n_past elements (NOT SUPPORTED)
+    // if (mode & GGML_ROPE_TYPE_NEOX) - GPT-NeoX style
     //
     // b is an int32 vector with size a->ne[2], it contains the positions
-    // c is freq factors (e.g. phi3-128k), (optional)
     GGML_API struct ggml_tensor * ggml_rope(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1467,6 +1531,7 @@ extern "C" {
             int                   mode);
 
     // custom RoPE
+    // c is freq factors (e.g. phi3-128k), (optional)
     GGML_API struct ggml_tensor * ggml_rope_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1557,34 +1622,49 @@ 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  * b,
-            int                  s0,
-            int                  s1,
-            int                  p0,
-            int                  p1,
-            int                  d0,
-            int                  d1,
-            bool                 is_2D,
-            enum ggml_type       dst_type);
+            struct ggml_tensor  * a,  // convolution kernel
+            struct ggml_tensor  * b,  // data
+            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,
+            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  * b,
-            int                  s0,
-            int                  s1,
-            int                  p0,
-            int                  p1,
-            int                  d0,
-            int                  d1);
+            struct ggml_tensor  * a,  // convolution kernel
+            struct ggml_tensor  * b,  // data
+            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
 
     GGML_API struct ggml_tensor * ggml_conv_1d(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b,
+            struct ggml_tensor  * a,   // convolution kernel
+            struct ggml_tensor  * b,   // data
             int                   s0,  // stride
             int                   p0,  // padding
             int                   d0); // dilation
@@ -1593,29 +1673,29 @@ 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  * b,
-            int                   s,
-            int                   d);
+            struct ggml_tensor  * a,  // convolution kernel
+            struct ggml_tensor  * b,  // data
+            int                   s,  // stride
+            int                   d); // dilation
 
     GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b,
-            int                   s0,
-            int                   p0,
-            int                   d0);
+            struct ggml_tensor  * a,   // convolution kernel
+            struct ggml_tensor  * b,   // data
+            int                   s0,  // stride
+            int                   p0,  // padding
+            int                   d0); // dilation
 
     GGML_API struct ggml_tensor * ggml_conv_2d(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b,
-            int                   s0,
-            int                   s1,
-            int                   p0,
-            int                   p1,
-            int                   d0,
-            int                   d1);
+            struct ggml_tensor  * a,   // convolution kernel
+            struct ggml_tensor  * b,   // data
+            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
 
 
     // kernel size is a->ne[0] x a->ne[1]
@@ -1677,6 +1757,18 @@ 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
@@ -1751,7 +1843,8 @@ 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,
@@ -1768,10 +1861,8 @@ extern "C" {
 
     GGML_API struct ggml_tensor * ggml_ssm_conv(
             struct ggml_context * ctx,
-            struct ggml_tensor  * s,
-            struct ggml_tensor  * x,
-            struct ggml_tensor  * c,
-            struct ggml_tensor  * sq);
+            struct ggml_tensor  * sx,
+            struct ggml_tensor  * c);
 
     GGML_API struct ggml_tensor * ggml_ssm_scan(
             struct ggml_context * ctx,
@@ -1780,8 +1871,7 @@ extern "C" {
             struct ggml_tensor  * dt,
             struct ggml_tensor  * A,
             struct ggml_tensor  * B,
-            struct ggml_tensor  * C,
-            struct ggml_tensor  * sq);
+            struct ggml_tensor  * C);
 
     // partition into non-overlapping windows with padding if needed
     // example:
@@ -1833,6 +1923,15 @@ 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 *);
@@ -2003,10 +2102,23 @@ 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 enum ggml_status  ggml_graph_compute         (      struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+    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

@@ -139,6 +139,17 @@ if (GGML_METAL)
         )
 endif()
 
+if (GGML_MUSA)
+    set(CMAKE_C_COMPILER clang)
+    set(CMAKE_C_EXTENSIONS OFF)
+    set(CMAKE_CXX_COMPILER clang++)
+    set(CMAKE_CXX_EXTENSIONS OFF)
+
+    set(GGML_CUDA ON)
+
+    list(APPEND GGML_CDEF_PUBLIC GGML_USE_MUSA)
+endif()
+
 if (GGML_OPENMP)
     find_package(OpenMP)
     if (OpenMP_FOUND)
@@ -147,6 +158,11 @@ if (GGML_OPENMP)
         add_compile_definitions(GGML_USE_OPENMP)
 
         set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+
+        if (GGML_MUSA)
+            set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} "/usr/lib/llvm-10/include/openmp")
+            set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} "/usr/lib/llvm-10/lib/libomp.so")
+        endif()
     else()
         message(WARNING "OpenMP not found")
     endif()
@@ -249,7 +265,13 @@ endif()
 if (GGML_CUDA)
     cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
 
-    find_package(CUDAToolkit)
+    if (GGML_MUSA)
+        list(APPEND CMAKE_MODULE_PATH "/usr/local/musa/cmake/")
+        find_package(MUSAToolkit)
+        set(CUDAToolkit_FOUND ${MUSAToolkit_FOUND})
+    else()
+        find_package(CUDAToolkit)
+    endif()
 
     if (CUDAToolkit_FOUND)
         message(STATUS "CUDA found")
@@ -268,7 +290,11 @@ if (GGML_CUDA)
         endif()
         message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
 
-        enable_language(CUDA)
+        if (GGML_MUSA)
+            set(CMAKE_CUDA_COMPILER ${MUSAToolkit_MCC_EXECUTABLE})
+        else()
+            enable_language(CUDA)
+        endif()
 
         file(GLOB   GGML_HEADERS_CUDA "ggml-cuda/*.cuh")
         list(APPEND GGML_HEADERS_CUDA "../include/ggml-cuda.h")
@@ -332,21 +358,40 @@ if (GGML_CUDA)
             add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
         endif()
 
+        if (GGML_MUSA)
+            set_source_files_properties(${GGML_SOURCES_CUDA} PROPERTIES LANGUAGE CXX)
+            foreach(SOURCE ${GGML_SOURCES_CUDA})
+                set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS "-x musa -mtgpu --cuda-gpu-arch=mp_22")
+            endforeach()
+        endif()
+
         if (GGML_STATIC)
             if (WIN32)
                 # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
                 set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
             else ()
-                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+                if (GGML_MUSA)
+                    set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musart_static MUSA::mublas_static)
+                else()
+                    set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+                endif()
             endif()
         else()
-            set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
+            if (GGML_MUSA)
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musart MUSA::mublas)
+            else()
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
+            endif()
         endif()
 
         if (GGML_CUDA_NO_VMM)
             # No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
         else()
-            set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
+            if (GGML_MUSA)
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musa_driver) # required by muDeviceGetAttribute(), muMemGetAllocationGranularity(...), ...
+            else()
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
+            endif()
         endif()
     else()
         message(WARNING "CUDA not found")
@@ -504,6 +549,13 @@ if (GGML_SYCL)
     file(GLOB   GGML_SOURCES_SYCL "ggml-sycl/*.cpp")
     list(APPEND GGML_SOURCES_SYCL "ggml-sycl.cpp")
 
+    find_package(DNNL)
+    message("-- DNNL found:" ${DNNL_FOUND})
+    if (GGML_SYCL_TARGET STREQUAL "INTEL")
+        add_compile_definitions(GGML_SYCL_DNNL=${DNNL_FOUND})
+    else()
+        add_compile_definitions(GGML_SYCL_DNNL=0)
+    endif()
     if (WIN32)
         find_package(IntelSYCL REQUIRED)
         find_package(MKL REQUIRED)
@@ -516,6 +568,9 @@ if (GGML_SYCL)
             set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} -fsycl pthread m dl onemkl)
         endif()
     endif()
+    if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
+        list(APPEND GGML_EXTRA_LIBS DNNL::dnnl)
+    endif()
 endif()
 
 if (GGML_RPC)
@@ -557,6 +612,10 @@ if (GGML_VULKAN)
             add_compile_definitions(GGML_VULKAN_MEMORY_DEBUG)
         endif()
 
+        if (GGML_VULKAN_PERF)
+            add_compile_definitions(GGML_VULKAN_PERF)
+        endif()
+
         if (GGML_VULKAN_VALIDATE)
             add_compile_definitions(GGML_VULKAN_VALIDATE)
         endif()
@@ -804,11 +863,6 @@ if (GGML_CANN)
                 ${CANN_INSTALL_DIR}/acllib/include
             )
 
-            # TODO: find libs
-            link_directories(
-                ${CANN_INSTALL_DIR}/lib64
-            )
-
             add_subdirectory(ggml-cann/kernels)
             list(APPEND CANN_LIBRARIES
                 ascendcl
@@ -827,6 +881,7 @@ if (GGML_CANN)
 
             set(GGML_EXTRA_LIBS     ${GGML_EXTRA_LIBS}     ${CANN_LIBRARIES} )
             set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} ${CANN_INCLUDE_DIRS})
+            set(GGML_EXTRA_LIBDIRS  ${GGML_EXTRA_LIBDIRS}  ${CANN_INSTALL_DIR}/lib64)
             list(APPEND GGML_CDEF_PUBLIC GGML_USE_CANN)
         endif()
     else()
@@ -857,8 +912,10 @@ function(get_flags CCID CCVER)
         set(C_FLAGS   -Wdouble-promotion)
         set(CXX_FLAGS -Wno-array-bounds)
 
-        if (CCVER VERSION_GREATER_EQUAL 7.1.0)
-            list(APPEND CXX_FLAGS -Wno-format-truncation)
+        if (NOT GGML_MUSA)
+            if (CCVER VERSION_GREATER_EQUAL 7.1.0)
+                list(APPEND CXX_FLAGS -Wno-format-truncation)
+            endif()
         endif()
         if (CCVER VERSION_GREATER_EQUAL 8.1.0)
             list(APPEND CXX_FLAGS -Wextra-semi)
@@ -1190,7 +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()
 
@@ -1264,6 +1321,7 @@ endif()
 target_compile_definitions(ggml PUBLIC  ${GGML_CDEF_PUBLIC})
 target_include_directories(ggml PUBLIC ../include)
 target_include_directories(ggml PRIVATE . ${GGML_EXTRA_INCLUDES})
+target_link_directories(ggml PRIVATE ${GGML_EXTRA_LIBDIRS})
 target_compile_features   (ggml PRIVATE c_std_11) # don't bump
 
 target_link_libraries(ggml PRIVATE Threads::Threads ${GGML_EXTRA_LIBS})

ggml/src/ggml-aarch64.c

@@ -16,6 +16,8 @@
 
 #if defined(__GNUC__)
 #pragma GCC diagnostic ignored "-Woverlength-strings"
+#elif defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
 #define UNUSED GGML_UNUSED
@@ -335,33 +337,18 @@ 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) {
-        return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4);
-    }
-    else {
-        assert(false);
-        return 0;
-    }
+    UNUSED(quant_weights);
+    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4);
 }
 
 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) {
-        return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8);
-    }
-    else {
-        assert(false);
-        return 0;
-    }
+    UNUSED(quant_weights);
+    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8);
 }
 
 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) {
-        return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
-    }
-    else {
-        assert(false);
-        return 0;
-    }
+    UNUSED(quant_weights);
+    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
 }
 
 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) {
@@ -384,8 +371,8 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE)
-    if (svcntw() == 8) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -496,8 +483,8 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE)
-    if (svcntw() == 8) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -614,7 +601,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
-    if (svcntw() == 8) {
+    if (ggml_sve_cnt_b == QK8_0) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;
@@ -680,12 +667,12 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
         return;
     }
     else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
-        GGML_ASSERT((ggml_cpu_has_sve() && (svcntw() == 8)) &&
+        GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                     "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
                     "performance");
     }
     else if (ggml_cpu_has_neon()) {
-        GGML_ASSERT(((ggml_cpu_has_sve() && (svcntw() == 8)) || ggml_cpu_has_matmul_int8()) &&
+        GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
                     "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
                     "quantization format for optimal performance");
     }
@@ -745,8 +732,8 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
-    if (svcntw() == 8) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -1266,8 +1253,8 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
-    if (svcntw() == 8) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -1728,7 +1715,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
-    if (svcntw() == 8) {
+    if (ggml_sve_cnt_b == QK8_0) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;
@@ -2139,12 +2126,12 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
         return;
     }
     else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
-        GGML_ASSERT((ggml_cpu_has_sve() && (svcntw() == 8)) &&
+        GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
                     "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
                     "performance");
     }
     else if (ggml_cpu_has_neon()) {
-        GGML_ASSERT(((ggml_cpu_has_sve() && (svcntw() == 8)) || ggml_cpu_has_matmul_int8()) &&
+        GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
                     "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
                     "quantization format for optimal performance");
     }

ggml/src/ggml-alloc.c

@@ -91,8 +91,7 @@ void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tenso
     if (talloc->offset + size > ggml_backend_buffer_get_size(talloc->buffer)) {
         fprintf(stderr, "%s: not enough space in the buffer to allocate %s (needed %zu, available %zu)\n",
                 __func__, tensor->name, size, ggml_backend_buffer_get_size(talloc->buffer) - talloc->offset);
-        GGML_ASSERT(!"not enough space in the buffer");
-        return;
+        GGML_ABORT("not enough space in the buffer");
     }
 
     void * addr = (char *)ggml_backend_buffer_get_base(talloc->buffer) + talloc->offset;
@@ -133,7 +132,7 @@ static void add_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset,
             return;
         }
     }
-    GGML_ASSERT(!"out of allocated_tensors");
+    GGML_ABORT("out of allocated_tensors");
 }
 static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, const struct ggml_tensor * tensor) {
     for (int i = 0; i < 1024; i++) {
@@ -142,8 +141,7 @@ static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offs
             return;
         }
     }
-    fprintf(stderr, "tried to free tensor %s not found\n", tensor->name);
-    GGML_ASSERT(!"tensor not found");
+    GGML_ABORT("tried to free tensor %s not found\n", tensor->name);
 }
 #endif
 
@@ -176,8 +174,7 @@ static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t siz
             // this should never happen
             fprintf(stderr, "%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
                     __func__, size, max_avail);
-            GGML_ASSERT(!"not enough space in the buffer");
-            GGML_UNREACHABLE();
+            GGML_ABORT("not enough space in the buffer");
         }
     }
 
@@ -443,7 +440,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
         }
     }
 
-    free(galloc->hash_set.keys);
+    ggml_hash_set_free(&galloc->hash_set);
     free(galloc->hash_values);
     free(galloc->bufts);
     free(galloc->buffers);
@@ -456,7 +453,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
 typedef struct ggml_gallocr * ggml_gallocr_t;
 
 static struct hash_node * ggml_gallocr_hash_get(ggml_gallocr_t galloc, struct ggml_tensor * t) {
-    size_t i = ggml_hash_find_or_insert(galloc->hash_set, t);
+    size_t i = ggml_hash_find_or_insert(&galloc->hash_set, t);
     return &galloc->hash_values[i];
 }
 
@@ -565,8 +562,8 @@ static int get_node_buffer_id(const int * node_buffer_ids, int i) {
 
 static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
     // clear hash tables
-    memset(galloc->hash_set.keys, 0, galloc->hash_set.size * sizeof(struct ggml_tensor *));
-    memset(galloc->hash_values,   0, galloc->hash_set.size * sizeof(struct hash_node));
+    ggml_hash_set_reset(&galloc->hash_set);
+    memset(galloc->hash_values, 0, sizeof(struct hash_node) * galloc->hash_set.size);
 
     // allocate leafs
     // these may be tensors that the application is not using in the graph, but may still want to allocate for other purposes
@@ -671,21 +668,19 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
 }
 
 bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
-    size_t hash_size = graph->visited_hash_table.size;
+    size_t min_hash_size = graph->n_nodes + graph->n_leafs;
+    // add 25% margin to avoid hash collisions
+    min_hash_size += min_hash_size / 4;
 
     // initialize hash table
-    if (galloc->hash_set.size < hash_size) {
-        free(galloc->hash_set.keys);
-        free(galloc->hash_values);
-        galloc->hash_set.size = hash_size;
-        galloc->hash_set.keys = calloc(hash_size, sizeof(struct ggml_tensor *));
-        galloc->hash_values   = calloc(hash_size, sizeof(struct hash_node));
+    if (galloc->hash_set.size < min_hash_size) {
+        ggml_hash_set_free(&galloc->hash_set);
+        galloc->hash_set = ggml_hash_set_new(min_hash_size);
         GGML_ASSERT(galloc->hash_set.keys != NULL);
+
+        free(galloc->hash_values);
+        galloc->hash_values = malloc(sizeof(struct hash_node) * galloc->hash_set.size);
         GGML_ASSERT(galloc->hash_values != NULL);
-    } else {
-        // reset hash table
-        memset(galloc->hash_set.keys, 0, sizeof(struct ggml_tensor *) * galloc->hash_set.size);
-        memset(galloc->hash_values,   0, sizeof(struct hash_node) * galloc->hash_set.size);
     }
 
     // reset allocators
@@ -817,8 +812,7 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
 }
 
 static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
-    ggml_backend_buffer_type_t buft = talloc->buffer_id != -1 ? galloc->bufts[talloc->buffer_id] : NULL;
-    size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(buft, node);
+    size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
     return talloc->size_max >= node_size;
 }
 

ggml/src/ggml-backend.c

@@ -351,15 +351,10 @@ void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t b
     }
 
     // an async copy would normally happen after all the queued operations on both backends are completed
-    // sync src, set_async dst
-    if (ggml_backend_buffer_is_host(src->buffer)) {
-        ggml_backend_synchronize(backend_src);
-        ggml_backend_tensor_set_async(backend_dst, dst, src->data, 0, ggml_nbytes(src));
-    } else {
-        ggml_backend_synchronize(backend_src);
-        ggml_backend_tensor_copy(src, dst);
-        ggml_backend_synchronize(backend_dst);
-    }
+    // to simulate the same behavior, we need to synchronize both backends first, and do a blocking copy
+    ggml_backend_synchronize(backend_src);
+    ggml_backend_synchronize(backend_dst);
+    ggml_backend_tensor_copy(src, dst);
 }
 
 // events
@@ -727,9 +722,11 @@ ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
 #endif
 
 struct ggml_backend_cpu_context {
-    int n_threads;
-    void * work_data;
-    size_t work_size;
+    int                 n_threads;
+    ggml_threadpool_t   threadpool;
+
+    void *              work_data;
+    size_t              work_size;
 
     ggml_abort_callback abort_callback;
     void *              abort_callback_data;
@@ -764,7 +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) {
@@ -801,7 +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);
@@ -878,6 +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;
@@ -908,6 +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));
 
@@ -1023,10 +1033,6 @@ static bool ggml_is_view_op(enum ggml_op op) {
 #define GGML_SCHED_MAX_BACKENDS 16
 #endif
 
-#ifndef GGML_SCHED_MAX_SPLITS
-#define GGML_SCHED_MAX_SPLITS 2048
-#endif
-
 #ifndef GGML_SCHED_MAX_SPLIT_INPUTS
 #define GGML_SCHED_MAX_SPLIT_INPUTS GGML_MAX_SRC
 #endif
@@ -1055,11 +1061,10 @@ struct ggml_backend_sched {
     ggml_backend_buffer_type_t bufts[GGML_SCHED_MAX_BACKENDS];
     ggml_gallocr_t galloc;
 
-    // hash keys of the nodes in the graph
-    struct ggml_hash_set    hash_set;
-    // hash values
-    int * tensor_backend_id;
-    struct ggml_tensor * (* tensor_copies)[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
+    // hash map of the nodes in the graph
+    struct ggml_hash_set  hash_set;
+    int                 * hv_tensor_backend_ids; // [hash_set.size]
+    struct ggml_tensor ** hv_tensor_copies;      // [hash_set.size][n_backends][n_copies]
 
     int * node_backend_ids; // [graph_size]
     int * leaf_backend_ids; // [graph_size]
@@ -1068,7 +1073,7 @@ struct ggml_backend_sched {
     int * prev_leaf_backend_ids; // [graph_size]
 
     // copy of the graph with modified inputs
-    struct ggml_cgraph * graph;
+    struct ggml_cgraph graph;
 
     // graph splits
     struct ggml_backend_sched_split * splits;
@@ -1087,19 +1092,16 @@ struct ggml_backend_sched {
     ggml_backend_sched_eval_callback callback_eval;
     void * callback_eval_user_data;
 
-    bool debug;
+    char * context_buffer;
+    size_t context_buffer_size;
 
-    // align context_buffer to GGML_MEM_ALIGN
-#ifdef _MSC_VER
-    __declspec(align(GGML_MEM_ALIGN))
-#else
-    __attribute__((aligned(GGML_MEM_ALIGN)))
-#endif
-    char context_buffer[GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + sizeof(struct ggml_cgraph)];
+    bool debug;
 };
 
-#define hash_id(tensor) ggml_hash_find_or_insert(sched->hash_set, tensor)
-#define tensor_backend_id(tensor) sched->tensor_backend_id[hash_id(tensor)]
+#define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
+#define tensor_backend_id(tensor) sched->hv_tensor_backend_ids[hash_id(tensor)]
+#define tensor_id_copy(id, backend_id, copy_id) sched->hv_tensor_copies[(id) * sched->n_backends * sched->n_copies + (backend_id) * sched->n_copies + (copy_id)]
+#define tensor_copy(tensor, backend_id, copy_id) tensor_id_copy(hash_id(tensor), backend_id, copy_id)
 
 // returns the priority of the backend, lower id is higher priority
 static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backend_t backend) {
@@ -1134,7 +1136,8 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, co
 }
 
 #if 0
-static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
+#define GGML_SCHED_MAX_SPLITS_DEBUG 4096
+static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
 #define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)
 #define GET_CAUSE(node) causes[hash_id(node)]
 #else
@@ -1169,7 +1172,6 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
         return cur_backend_id;
     }
 
-    // assign nodes that use weights to the backend of the weights
     // operations with weights are preferably run on the same backend as the weights
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         const struct ggml_tensor * src = tensor->src[i];
@@ -1275,7 +1277,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
     sched->is_reset = false;
 
     struct ggml_init_params params = {
-        /* .mem_size =   */ sizeof(sched->context_buffer),
+        /* .mem_size =   */ sched->context_buffer_size,
         /* .mem_buffer = */ sched->context_buffer,
         /* .no_alloc =   */ true
     };
@@ -1284,39 +1286,43 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 
     sched->ctx = ggml_init(params);
     if (sched->ctx == NULL) {
-        fprintf(stderr, "%s: failed to initialize context\n", __func__);
-        GGML_ASSERT(false);
+        GGML_ABORT("%s: failed to initialize context\n", __func__);
     }
 
     // pass 1: assign backends to ops with pre-allocated inputs
     for (int i = 0; i < graph->n_leafs; i++) {
         struct ggml_tensor * leaf = graph->leafs[i];
         int * leaf_backend_id = &tensor_backend_id(leaf);
-        if (*leaf_backend_id != -1) {
-            // do not overwrite user assignments
-            continue;
+        // do not overwrite user assignments
+        if (*leaf_backend_id == -1) {
+            *leaf_backend_id = ggml_backend_sched_backend_id_from_cur(sched, leaf);
         }
-        *leaf_backend_id = ggml_backend_sched_backend_id_from_cur(sched, leaf);
     }
 
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
         int * node_backend_id = &tensor_backend_id(node);
-        if (*node_backend_id != -1) {
-            // do not overwrite user assignments
-            continue;
-        }
-        *node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);
-        // src
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
+        // do not overwrite user assignments
+        if (*node_backend_id == -1) {
+            *node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);
+
+#if 0
+            // src
+            if (node->op == GGML_OP_NONE) {
                 continue;
             }
-            int * src_backend_id = &tensor_backend_id(src);
-            if (*src_backend_id == -1) {
-                *src_backend_id = ggml_backend_sched_backend_id_from_cur(sched, src);
+
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * src = node->src[j];
+                if (src == NULL) {
+                    continue;
+                }
+                int * src_backend_id = &tensor_backend_id(src);
+                if (*src_backend_id == -1) {
+                    *src_backend_id = ggml_backend_sched_backend_id_from_cur(sched, src);
+                }
             }
+#endif
         }
     }
 
@@ -1488,12 +1494,13 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         }
     }
 
-    // pass 4: split graph, find tensors that need to be copied
+    // pass 5: split graph, find tensors that need to be copied
     {
         int i_split = 0;
         struct ggml_backend_sched_split * split = &sched->splits[0];
         // find the backend of the first split, skipping view ops
-        for (int i = 0; i < graph->n_nodes; i++) {
+        int i = 0;
+        for (; i < graph->n_nodes; i++) {
             struct ggml_tensor * node = graph->nodes[i];
             if (!ggml_is_view_op(node->op)) {
                 split->backend_id = tensor_backend_id(node);
@@ -1502,9 +1509,8 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         }
         split->i_start = 0;
         split->n_inputs = 0;
-        memset(split->inputs, 0, sizeof(split->inputs)); //HACK
         int cur_backend_id = split->backend_id;
-        for (int i = 0; i < graph->n_nodes; i++) {
+        for (; i < graph->n_nodes; i++) {
             struct ggml_tensor * node = graph->nodes[i];
 
             if (ggml_is_view_op(node->op)) {
@@ -1513,7 +1519,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 
             const int node_backend_id = tensor_backend_id(node);
 
-            GGML_ASSERT(node_backend_id != -1); // all nodes should be assigned by now
+            assert(node_backend_id != -1); // all nodes should be assigned by now
 
             // check if we should start a new split based on the sources of the current node
             bool need_new_split = false;
@@ -1527,7 +1533,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                     // by starting a new split, the memory of the previously offloaded weights can be reused
                     if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
                         int src_backend_id = tensor_backend_id(src);
-                        if (src_backend_id != -1 && src_backend_id != cur_backend_id) {
+                        if (src_backend_id != cur_backend_id) {
                             need_new_split = true;
                             break;
                         }
@@ -1536,9 +1542,9 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                     // FIXME: count the number of inputs instead of only checking when full
                     if (split->n_inputs == GGML_SCHED_MAX_SPLIT_INPUTS) {
                         const size_t id = hash_id(src);
-                        int src_backend_id = sched->tensor_backend_id[id];
+                        int src_backend_id = sched->hv_tensor_backend_ids[id];
                         bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
-                        if (src_backend_id != cur_backend_id && sched->tensor_copies[hash_id(src)][cur_backend_id][0] == NULL && !supported) {
+                        if (src_backend_id != cur_backend_id && tensor_id_copy(id, cur_backend_id, 0) == NULL && !supported) {
                             //printf("starting new split because of too many inputs: node %s, input %s\n", node->name, src->name);
                             need_new_split = true;
                             break;
@@ -1555,7 +1561,6 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                     sched->splits = realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split));
                     GGML_ASSERT(sched->splits != NULL);
                 }
-                GGML_ASSERT(i_split < GGML_SCHED_MAX_SPLITS);
                 split = &sched->splits[i_split];
                 split->backend_id = node_backend_id;
                 split->i_start = i;
@@ -1570,12 +1575,12 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                     continue;
                 }
 
-                const int src_backend_id = tensor_backend_id(src);
+                size_t src_id = hash_id(src);
+                const int src_backend_id = sched->hv_tensor_backend_ids[src_id];
                 assert(src_backend_id != -1); // all inputs should be assigned by now
 
                 if (src->flags & GGML_TENSOR_FLAG_INPUT && sched->n_copies > 1) {
-                    size_t id = hash_id(src);
-                    if (sched->tensor_copies[id][src_backend_id][0] == NULL) {
+                    if (tensor_id_copy(src_id, src_backend_id, 0) == NULL) {
                         ggml_backend_t backend = sched->backends[src_backend_id];
                         for (int c = 0; c < sched->n_copies; c++) {
                             struct ggml_tensor * tensor_copy;
@@ -1589,7 +1594,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                                 ggml_set_input(tensor_copy);
                                 ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
                             }
-                            sched->tensor_copies[id][src_backend_id][c] = tensor_copy;
+                            tensor_id_copy(src_id, src_backend_id, c) = tensor_copy;
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }
                         int n_graph_inputs = sched->n_graph_inputs++;
@@ -1598,11 +1603,9 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                     }
                 }
 
-                bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
-                if (src_backend_id != cur_backend_id && !supported) {
+                if (src_backend_id != cur_backend_id && !ggml_backend_sched_buffer_supported(sched, src, cur_backend_id)) {
                     // create a copy of the input in the split's backend
-                    const size_t id = hash_id(src);
-                    if (sched->tensor_copies[id][cur_backend_id][0] == NULL) {
+                    if (tensor_id_copy(src_id, cur_backend_id, 0) == NULL) {
                         ggml_backend_t backend = sched->backends[cur_backend_id];
                         for (int c = 0; c < sched->n_copies; c++) {
                             struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
@@ -1611,14 +1614,14 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                                 ggml_set_input(tensor_copy);
                                 ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
                             }
-                            sched->tensor_copies[id][cur_backend_id][c] = tensor_copy;
+                            tensor_id_copy(src_id, cur_backend_id, c) = tensor_copy;
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }
                         int n_inputs = split->n_inputs++;
                         GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
                         split->inputs[n_inputs] = src;
                     }
-                    node->src[j] = sched->tensor_copies[id][cur_backend_id][sched->cur_copy];
+                    node->src[j] = tensor_id_copy(src_id, cur_backend_id, sched->cur_copy);
                 }
             }
         }
@@ -1630,7 +1633,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         ggml_backend_sched_print_assignments(sched, graph);
     }
 
-    // swap node_backend_ids and leaf_backend_ids and prevs
+    // swap node_backend_ids and leaf _backend_ids with prevs
     {
         int * tmp = sched->node_backend_ids;
         sched->node_backend_ids = sched->prev_node_backend_ids;
@@ -1641,9 +1644,19 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         sched->prev_leaf_backend_ids = tmp;
     }
 
-    // create copies of the graph for each split
-    // TODO: avoid this copy
-    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2, false);
+    int graph_size = graph->n_nodes + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2;
+    if (sched->graph.size < graph_size) {
+        sched->graph.size = graph_size;
+        sched->graph.nodes = realloc(sched->graph.nodes, graph_size * sizeof(struct ggml_tensor *));
+        sched->graph.leafs = realloc(sched->graph.leafs, graph_size * sizeof(struct ggml_tensor *));
+        GGML_ASSERT(sched->graph.nodes != NULL);
+        GGML_ASSERT(sched->graph.leafs != NULL);
+    }
+    sched->graph.n_nodes = 0;
+    sched->graph.n_leafs = 0;
+
+    struct ggml_cgraph * graph_copy = &sched->graph;
+
     for (int i = 0; i < sched->n_splits; i++) {
         struct ggml_backend_sched_split * split = &sched->splits[i];
         split->graph = ggml_graph_view(graph, split->i_start, split->i_end);
@@ -1654,12 +1667,12 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 
             struct ggml_tensor * input = split->inputs[j];
             const size_t input_id = hash_id(input);
-            struct ggml_tensor * input_cpy = sched->tensor_copies[input_id][split->backend_id][sched->cur_copy];
+            struct ggml_tensor * input_cpy = tensor_id_copy(input_id, split->backend_id, sched->cur_copy);
 
             // add a dependency to the input source so that it is not freed before the copy is done
             struct ggml_tensor * input_dep = ggml_view_tensor(sched->ctx, input);
             input_dep->src[0] = input;
-            sched->node_backend_ids[graph_copy->n_nodes] = sched->tensor_backend_id[input_id];
+            sched->node_backend_ids[graph_copy->n_nodes] = sched->hv_tensor_backend_ids[input_id];
             graph_copy->nodes[graph_copy->n_nodes++] = input_dep;
 
             // add a dependency to the input copy so that it is allocated at the start of the split
@@ -1681,7 +1694,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
             size_t id = hash_id(input);
             int backend_id = tensor_backend_id(input);
             for (int c = 0; c < sched->n_copies; c++) {
-                struct ggml_tensor * input_cpy = sched->tensor_copies[id][backend_id][c];
+                struct ggml_tensor * input_cpy = tensor_id_copy(id, backend_id, c);
                 sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
                 graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
             }
@@ -1694,7 +1707,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                 struct ggml_tensor * input = split->inputs[j];
                 size_t id = hash_id(input);
                 for (int c = 0; c < sched->n_copies; c++) {
-                    struct ggml_tensor * input_cpy = sched->tensor_copies[id][backend_id][c];
+                    struct ggml_tensor * input_cpy = tensor_id_copy(id, backend_id, c);
                     sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
                     graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
                 }
@@ -1708,13 +1721,11 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         sched->leaf_backend_ids[graph_copy->n_leafs] = tensor_backend_id(leaf);
         graph_copy->leafs[graph_copy->n_leafs++] = leaf;
     }
-
-    sched->graph = graph_copy;
 }
 
 static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
     bool backend_ids_changed = false;
-    for (int i = 0; i < sched->graph->n_nodes; i++) {
+    for (int i = 0; i < sched->graph.n_nodes; i++) {
         if (sched->node_backend_ids[i] != sched->prev_node_backend_ids[i] &&
             sched->bufts[sched->node_backend_ids[i]] != sched->bufts[sched->prev_node_backend_ids[i]]) {
             backend_ids_changed = true;
@@ -1722,7 +1733,7 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
         }
     }
     if (!backend_ids_changed) {
-        for (int i = 0; i < sched->graph->n_leafs; i++) {
+        for (int i = 0; i < sched->graph.n_leafs; i++) {
             if (sched->leaf_backend_ids[i] != sched->prev_leaf_backend_ids[i] &&
                 sched->bufts[sched->leaf_backend_ids[i]] != sched->bufts[sched->prev_leaf_backend_ids[i]]) {
                 backend_ids_changed = true;
@@ -1732,14 +1743,14 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
     }
 
     // allocate graph
-    if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, sched->graph)) {
+    if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
         // the re-allocation may cause the split inputs to be moved to a different address
         ggml_backend_sched_synchronize(sched);
 #ifndef NDEBUG
-        fprintf(stderr, "%s: failed to allocate graph, reserving\n", __func__);
+        fprintf(stderr, "%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
 #endif
-        ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
-        if (!ggml_gallocr_alloc_graph(sched->galloc, sched->graph)) {
+        ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
+        if (!ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
             fprintf(stderr, "%s: failed to allocate graph\n", __func__);
             return false;
         }
@@ -1760,7 +1771,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
         for (int j = 0; j < split->n_inputs; j++) {
             ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[j]);
             struct ggml_tensor * input = split->inputs[j];
-            struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split_backend_id][sched->cur_copy];
+            struct ggml_tensor * input_cpy = tensor_copy(input, split_backend_id, sched->cur_copy);
 
             if (input->flags & GGML_TENSOR_FLAG_INPUT) {
                 // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
@@ -1777,7 +1788,17 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 } else {
                     ggml_backend_synchronize(split_backend);
                 }
-                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
+                // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
+                // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
+                if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
+                    ggml_backend_synchronize(input_backend);
+                    if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                        ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                    } else {
+                        ggml_backend_synchronize(split_backend);
+                    }
+                    ggml_backend_tensor_copy(input, input_cpy);
+                }
             }
         }
 
@@ -1846,21 +1867,24 @@ ggml_backend_sched_t ggml_backend_sched_new(
     struct ggml_backend_sched * sched = calloc(1, sizeof(struct ggml_backend_sched));
 
     sched->debug = getenv("GGML_SCHED_DEBUG") != NULL;
+    sched->n_backends = n_backends;
+    sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
 
     // initialize hash table
-    sched->hash_set          = ggml_hash_set_new(graph_size);
-    sched->tensor_backend_id = calloc(sched->hash_set.size, sizeof(sched->tensor_backend_id[0]));
-    sched->tensor_copies     = calloc(sched->hash_set.size, sizeof(sched->tensor_copies[0]));
+    // FIXME: needs to be size*2 to account for leafs (do it in graph_split instead)
+    sched->hash_set    = ggml_hash_set_new(graph_size);
+    sched->hv_tensor_backend_ids = malloc(sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
+    sched->hv_tensor_copies      = malloc(sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *));
 
-    const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2;
-    sched->node_backend_ids  = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
-    sched->leaf_backend_ids  = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
+    const size_t ggml_sched_max_splits = graph_size; // at most there is one split for each node in the graph
+    const size_t nodes_size = graph_size + ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2;
+    sched->node_backend_ids = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
+    sched->leaf_backend_ids = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
     sched->prev_node_backend_ids = calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0]));
     sched->prev_leaf_backend_ids = calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0]));
 
-    sched->n_backends = n_backends;
-
-    sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
+    sched->context_buffer_size = ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false);
+    sched->context_buffer = malloc(sched->context_buffer_size);
 
     const int initial_splits_capacity = 16;
     sched->splits = calloc(initial_splits_capacity, sizeof(sched->splits[0]));
@@ -1895,37 +1919,37 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
     }
     ggml_gallocr_free(sched->galloc);
     ggml_free(sched->ctx);
+    ggml_hash_set_free(&sched->hash_set);
     free(sched->splits);
-    free(sched->hash_set.keys);
-    free(sched->tensor_backend_id);
-    free(sched->tensor_copies);
+    free(sched->hv_tensor_backend_ids);
+    free(sched->hv_tensor_copies);
     free(sched->node_backend_ids);
     free(sched->leaf_backend_ids);
     free(sched->prev_node_backend_ids);
     free(sched->prev_leaf_backend_ids);
+    free(sched->context_buffer);
+    free(sched->graph.nodes);
+    free(sched->graph.leafs);
     free(sched);
 }
 
 void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
     // reset state for the next run
     if (!sched->is_reset) {
-        size_t hash_size = sched->hash_set.size;
-        memset(sched->hash_set.keys,      0, sizeof(sched->hash_set.keys[0])     * hash_size); // NOLINT
-        memset(sched->tensor_backend_id, -1, sizeof(sched->tensor_backend_id[0]) * hash_size);
-        memset(sched->tensor_copies,      0, sizeof(sched->tensor_copies[0])     * hash_size);
-
+        ggml_hash_set_reset(&sched->hash_set);
+        memset(sched->hv_tensor_backend_ids, -1, sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
+        memset(sched->hv_tensor_copies,       0, sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *));
         sched->is_reset = true;
     }
     sched->is_alloc = false;
 }
 
 bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
-    GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes);
+    GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);
 
     ggml_backend_sched_split_graph(sched, measure_graph);
 
-    // TODO: extract this to a separate function
-    if (!ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
+    if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
         return false;
     }
 
@@ -1936,10 +1960,11 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
 }
 
 bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
-    GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes);
+    GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + graph->n_leafs);
 
     ggml_backend_sched_split_graph(sched, graph);
 
+
     if (!ggml_backend_sched_alloc_splits(sched)) {
         return false;
     }
@@ -2009,6 +2034,7 @@ void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct gg
     GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
     tensor_backend_id(node) = backend_index;
     SET_CAUSE(node, "usr");
+    sched->is_reset = false;
 }
 
 ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node) {
@@ -2051,9 +2077,9 @@ static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set,
     GGML_ASSERT(src != NULL);
     GGML_ASSERT(src->data && "graph must be allocated");
 
-    size_t id = ggml_hash_insert(hash_set, src);
-    if (id == GGML_HASHTABLE_ALREADY_EXISTS) {
-        return node_copies[ggml_hash_find(hash_set, src)];
+    size_t id = ggml_hash_insert(&hash_set, src);
+    if (id == GGML_HASHSET_ALREADY_EXISTS) {
+        return node_copies[ggml_hash_find(&hash_set, src)];
     }
 
     struct ggml_tensor * dst = ggml_dup_tensor_layout(src->data && !src->view_src ? ctx_allocated : ctx_unallocated, src);
@@ -2078,7 +2104,7 @@ static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set,
     return dst;
 }
 
-static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_tensor ** node_copies, bool * node_init, struct ggml_tensor * src) {
+static void graph_copy_init_tensor(struct ggml_hash_set * hash_set, struct ggml_tensor ** node_copies, bool * node_init, struct ggml_tensor * src) {
     size_t id = ggml_hash_find(hash_set, src);
     if (node_init[id]) {
         return;
@@ -2105,10 +2131,7 @@ static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_te
 }
 
 struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
-    struct ggml_hash_set hash_set = {
-        /* .size = */ graph->visited_hash_table.size,
-        /* .keys = */ calloc(graph->visited_hash_table.size, sizeof(hash_set.keys[0])) // NOLINT
-    };
+    struct ggml_hash_set hash_set = ggml_hash_set_new(graph->visited_hash_set.size);
     struct ggml_tensor ** node_copies = calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT
     bool * node_init = calloc(hash_set.size, sizeof(node_init[0]));
 
@@ -2123,7 +2146,7 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s
 
     if (ctx_allocated == NULL || ctx_unallocated == NULL) {
         fprintf(stderr, "failed to allocate context for graph copy\n");
-        free(hash_set.keys);
+        ggml_hash_set_free(&hash_set);
         free(node_copies);
         free(node_init);
         ggml_free(ctx_allocated);
@@ -2146,7 +2169,7 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s
     ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx_allocated, backend);
     if (buffer == NULL) {
         fprintf(stderr, "failed to allocate buffer for graph copy\n");
-        free(hash_set.keys);
+        ggml_hash_set_free(&hash_set);
         free(node_copies);
         free(node_init);
         ggml_free(ctx_allocated);
@@ -2164,19 +2187,19 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s
     // copy data and init views
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
-        graph_copy_init_tensor(hash_set, node_copies, node_init, node);
+        graph_copy_init_tensor(&hash_set, node_copies, node_init, node);
     }
 
     // build graph copy
     struct ggml_cgraph * graph_copy = ggml_new_graph_custom(ctx_allocated, graph->size, false);
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
-        struct ggml_tensor * node_copy = node_copies[ggml_hash_find(hash_set, node)];
+        struct ggml_tensor * node_copy = node_copies[ggml_hash_find(&hash_set, node)];
         graph_copy->nodes[i] = node_copy;
     }
     graph_copy->n_nodes = graph->n_nodes;
 
-    free(hash_set.keys);
+    ggml_hash_set_free(&hash_set);
     free(node_copies);
     free(node_init);
 

ggml/src/ggml-blas.cpp

@@ -275,8 +275,7 @@ GGML_CALL static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t
                 break;
 
             default:
-                fprintf(stderr, "%s: unsupported op %s\n", __func__, ggml_op_desc(node));
-                GGML_ASSERT(false);
+                GGML_ABORT("%s: unsupported op %s\n", __func__, ggml_op_desc(node));
         }
     }
 

ggml/src/ggml-cann.cpp

@@ -120,7 +120,7 @@ static void ggml_cann_log(enum ggml_log_level level, const char* format, ...) {
             file, line);
     GGML_CANN_LOG_ERROR("  %s\n", stmt);
     // abort with GGML_ASSERT to get a stack trace
-    GGML_ASSERT(!"CANN error");
+    GGML_ABORT("CANN error");
 }
 
 /**
@@ -342,7 +342,7 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
         // memory should always buffered. these memory may still needed by
         // tasks in stream.
         // TODO, fix me.
-        GGML_ASSERT(!"Cann buffer pool full, increase MAX_CANN_BUFFERS\n");
+        GGML_ABORT("Cann buffer pool full, increase MAX_CANN_BUFFERS\n");
     }
 };
 
@@ -627,7 +627,6 @@ GGML_CALL static void* ggml_backend_cann_buffer_get_base(
 GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
                                                        const void* src,
                                                        void* dst) {
-    GGML_ASSERT(tensor->op == GGML_OP_NONE);
 
     int64_t n_elems = ggml_nelements(tensor);
     int64_t groups = n_elems / QK4_0;
@@ -679,7 +678,6 @@ GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
  */
 GGML_CALL static void ggml_backend_cann_transform_back_q4_0(
     const ggml_tensor* tensor, void* src, void* dst) {
-    GGML_ASSERT(tensor->op == GGML_OP_NONE);
 
     int64_t n_elems = ggml_nelements(tensor);
     int64_t groups = n_elems / QK4_0;
@@ -898,11 +896,10 @@ GGML_CALL static void ggml_backend_cann_buffer_init_tensor(
  * @param size Size of the data to be copied, in bytes.
  */
 GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
-    ggml_backend_buffer_t buffer, ggml_tensor* tensor, const void* data,
+    ggml_backend_buffer_t buffer, ggml_tensor *tensor, const void *data,
     size_t offset, size_t size) {
-    // GGML_ASSERT(size == ggml_nbytes(tensor));
-    ggml_backend_cann_buffer_context* ctx =
-        (ggml_backend_cann_buffer_context*)buffer->context;
+    ggml_backend_cann_buffer_context *ctx =
+        (ggml_backend_cann_buffer_context *)buffer->context;
 
     ggml_cann_set_device(ctx->device);
     // TODO: refer to cann(#6017), it use thread's default stream.
@@ -910,22 +907,21 @@ GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
     // Why aclrtSynchronizeDevice?
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpy(tensor->data, size, (const char*)data + offset,
-                              size, ACL_MEMCPY_HOST_TO_DEVICE));
+        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
+                              ACL_MEMCPY_HOST_TO_DEVICE));
     } else {
-        void* transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, (const char*)data + offset,
-                                    transform_buffer);
+        void *transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, data, transform_buffer);
 
 #ifndef NDEBUG
-        void* check_buffer = malloc(size);
+        void *check_buffer = malloc(size);
         ggml_backend_cann_transform_back(tensor, transform_buffer,
                                          check_buffer);
-        GGML_ASSERT(memcmp((const char*)data + offset, check_buffer, size) ==
-                    0);
+        GGML_ASSERT(memcmp(data, check_buffer, size) == 0);
         free(check_buffer);
 #endif
-        ACL_CHECK(aclrtMemcpy(tensor->data, size, transform_buffer, size,
+        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size,
+                              transform_buffer, size,
                               ACL_MEMCPY_HOST_TO_DEVICE));
         free(transform_buffer);
     }
@@ -947,21 +943,20 @@ GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
 GGML_CALL static void ggml_backend_cann_buffer_get_tensor(
     ggml_backend_buffer_t buffer, const ggml_tensor* tensor, void* data,
     size_t offset, size_t size) {
-    GGML_ASSERT(size == ggml_nbytes(tensor));
     ggml_backend_cann_buffer_context* ctx =
         (ggml_backend_cann_buffer_context*)buffer->context;
 
     ggml_cann_set_device(ctx->device);
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpy((char*)data + offset, size, tensor->data, size,
+        ACL_CHECK(aclrtMemcpy(data, size, (char*)tensor->data + offset, size,
                               ACL_MEMCPY_DEVICE_TO_HOST));
     } else {
         void* transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpy(transform_buffer, size, tensor->data, size,
+        ACL_CHECK(aclrtMemcpy(transform_buffer, size,
+                              (char*)tensor->data + offset, size,
                               ACL_MEMCPY_DEVICE_TO_HOST));
-        ggml_backend_cann_transform_back(tensor, transform_buffer,
-                                         (char*)data + offset);
+        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
         free(transform_buffer);
     }
 }
@@ -1450,42 +1445,41 @@ ggml_backend_cann_get_default_buffer_type(ggml_backend_t backend) {
  * @param size Size of the data to copy in bytes.
  */
 GGML_CALL static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
-                                                         ggml_tensor* tensor,
-                                                         const void* data,
+                                                         ggml_tensor *tensor,
+                                                         const void *data,
                                                          size_t offset,
                                                          size_t size) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
+    ggml_backend_cann_context *cann_ctx =
+        (ggml_backend_cann_context *)backend->context;
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync(
-            tensor->data, size, (const char*)data + offset, size,
-            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
+        ACL_CHECK(aclrtMemcpyAsync((char *)tensor->data + offset, size, data,
+                                   size, ACL_MEMCPY_HOST_TO_DEVICE,
+                                   cann_ctx->stream()));
     } else {
-        void* transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, (const char*)data + offset,
-                                    transform_buffer);
+        void *transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, data, transform_buffer);
 
 #ifndef NDEBUG
-        void* check_buffer = malloc(size);
+        void *check_buffer = malloc(size);
         ggml_backend_cann_transform_back(tensor, transform_buffer,
                                          check_buffer);
-        GGML_ASSERT(memcmp((const char*)data + offset, check_buffer, size));
+        GGML_ASSERT(memcmp(data, check_buffer, size));
         free(check_buffer);
 #endif
-        ACL_CHECK(aclrtMemcpyAsync(tensor->data, size, transform_buffer, size,
-                                   ACL_MEMCPY_HOST_TO_DEVICE,
-                                   cann_ctx->stream()));
+        ACL_CHECK(aclrtMemcpyAsync(
+            (char *)tensor->data + offset, size, transform_buffer, size,
+            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
         ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
         free(transform_buffer);
     }
 }
 
 GGML_CALL static void ggml_backend_cann_get_tensor_async(
-    ggml_backend_t backend, const ggml_tensor* tensor, void* data,
+    ggml_backend_t backend, const ggml_tensor *tensor, void *data,
     size_t offset, size_t size) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
+    ggml_backend_cann_context *cann_ctx =
+        (ggml_backend_cann_context *)backend->context;
     ggml_backend_buffer_t buf =
         tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
@@ -1493,17 +1487,16 @@ GGML_CALL static void ggml_backend_cann_get_tensor_async(
                 "unsupported buffer type");
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync((char*)data + offset, size, tensor->data,
+        ACL_CHECK(aclrtMemcpyAsync(data, size, (char *)tensor->data + offset,
                                    size, ACL_MEMCPY_DEVICE_TO_HOST,
                                    cann_ctx->stream()));
     } else {
-        void* transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpyAsync(transform_buffer, size, tensor->data, size,
-                                   ACL_MEMCPY_DEVICE_TO_HOST,
-                                   cann_ctx->stream()));
+        void *transform_buffer = malloc(size);
+        ACL_CHECK(aclrtMemcpyAsync(
+            transform_buffer, size, (char *)tensor->data + offset, size,
+            ACL_MEMCPY_DEVICE_TO_HOST, cann_ctx->stream()));
         ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        ggml_backend_cann_transform_back(tensor, transform_buffer,
-                                         (char*)data + offset);
+        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
         free(transform_buffer);
     }
 }
@@ -1559,23 +1552,18 @@ GGML_CALL static bool ggml_backend_cann_cpy_tensor_async(
             return false;
         }
 
+        // need open both directions for memcpyasync between devices.
+        ggml_cann_set_device(cann_ctx_dst->device);
+        ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_src->device, 0));
         ggml_cann_set_device(cann_ctx_src->device);
         ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_dst->device, 0));
+
         ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
                                    ACL_MEMCPY_DEVICE_TO_DEVICE,
-                                   cann_ctx_dst->stream()));
+                                   cann_ctx_src->stream()));
 
-        // record event on src stream
-        if (!cann_ctx_src->copy_event) {
-            ACL_CHECK(aclrtCreateEvent(&cann_ctx_src->copy_event));
-        }
-
-        ACL_CHECK(
-            aclrtRecordEvent(cann_ctx_src->copy_event, cann_ctx_src->stream()));
-
-        // wait on dst stream for the copy to complete
-        ACL_CHECK(aclrtStreamWaitEvent(cann_ctx_dst->stream(),
-                                       cann_ctx_src->copy_event));
+        //TODO: workaround for Event didn`t work here.
+        aclrtSynchronizeStream(cann_ctx_src->stream());
     } else {
         // src and dst are on the same backend
         ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
@@ -1671,10 +1659,13 @@ GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
             }
         case GGML_OP_MUL_MAT: {
             switch (op->src[0]->type) {
-                // case GGML_TYPE_Q4_0:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_F32:
                 case GGML_TYPE_Q8_0:
+                    // TODO: fix me
+                    // Current groupsize should not be greater than k-1 in
+                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize().
+                case GGML_TYPE_Q4_0:
                     return true;
                 default:
                     return false;
@@ -1699,6 +1690,7 @@ GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
                 case GGML_TYPE_F32:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_0:
                     return true;
                 default:
                     return false;
@@ -1763,8 +1755,8 @@ static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
  *
  * This function determines whether the CANN backend supports the given backend
  * buffer type by comparing the device context of the backend and buffer type.
- * It returns true if the device associated with the buffer type matches the
- * device associated with the backend.
+ * It returns true if the devices are same between the backend context and
+ * buffer type context.
  *
  * @param backend Pointer to the CANN backend.
  * @param buft Pointer to the backend buffer type to check.
@@ -1773,9 +1765,14 @@ static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
  */
 GGML_CALL static bool ggml_backend_cann_supports_buft(
     ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
-    return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
-
-    GGML_UNUSED(backend);
+    if (ggml_backend_buft_is_cann(buft)) {
+        ggml_backend_cann_context * cann_ctx =
+                        (ggml_backend_cann_context *)backend->context;
+        ggml_backend_cann_buffer_type_context * buft_ctx =
+                        (ggml_backend_cann_buffer_type_context *)buft->context;
+        return buft_ctx->device == cann_ctx->device;
+    }
+    return false;
 }
 
 /**
@@ -1874,7 +1871,7 @@ static void ggml_backend_cann_event_wait(ggml_backend_t backend,
         ACL_CHECK(aclrtStreamWaitEvent(cann_ctx->stream(),
                                        (aclrtEvent)event->context));
     } else {
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 

ggml/src/ggml-cann/acl_tensor.cpp

@@ -37,6 +37,10 @@ aclDataType ggml_cann_type_mapping(ggml_type type) {
             return ACL_INT16;
         case GGML_TYPE_I32:
             return ACL_INT32;
+        case GGML_TYPE_Q4_0:
+            return ACL_INT4;
+        case GGML_TYPE_Q8_0:
+            return ACL_INT8;
         default:
             return ACL_DT_UNDEFINED;
     }
@@ -89,33 +93,6 @@ bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1) {
     return false;
 }
 
-aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
-                                   size_t type_size, int64_t* ne, size_t* nb,
-                                   int64_t dims, aclFormat format,
-                                   size_t offset) {
-    int64_t tmp_ne[GGML_MAX_DIMS * 2];
-    int64_t tmp_stride[GGML_MAX_DIMS * 2];
-
-    memcpy(tmp_ne, ne, dims * sizeof(int64_t));
-    for (int i = 0; i < dims; i++) {
-        tmp_stride[i] = nb[i] / type_size;
-    }
-
-    std::reverse(tmp_ne, tmp_ne + dims);
-    std::reverse(tmp_stride, tmp_stride + dims);
-
-    int64_t acl_storage_len = 0;
-    for (int i = 0; i < dims; i++) {
-        acl_storage_len += (ne[i] - 1) * nb[i];
-    }
-
-    aclTensor* acl_tensor =
-        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
-                        format, &acl_storage_len, 1, data_ptr);
-
-    return acl_tensor;
-}
-
 int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0,
                                   const ggml_tensor* src1,
                                   int64_t* bcast_src0_ne,

ggml/src/ggml-cann/acl_tensor.h

@@ -23,6 +23,9 @@
 #ifndef CANN_ACL_TENSOR_H
 #define CANN_ACL_TENSOR_H
 
+#include <algorithm>
+#include <cstring>
+
 #include <aclnn/aclnn_base.h>
 #include "common.h"
 
@@ -65,7 +68,8 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = null
                              size_t offset = 0);
 
 /**
- * @brief   Creates an ACL tensor from provided parameters.
+ * @brief   Template for creating an ACL tensor from provided parameters. typename TYPE
+ *          should be size_t or float.
  *
  * @details This function creates an ACL tensor using the provided data pointer,
  *          data type, dimensions, strides, format, offset, and additional parameters.
@@ -83,10 +87,34 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = null
  * @param   offset      Offset in bytes for the ACL tensor data. Defaults to 0.
  * @return  Pointer to the created ACL tensor.
  */
+template<typename TYPE>
 aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
-                             size_t type_size, int64_t* ne, size_t* nb,
-                             int64_t dims, aclFormat format = ACL_FORMAT_ND,
-                             size_t offset = 0);
+                                   TYPE type_size, int64_t* ne, TYPE* nb,
+                                   int64_t dims,
+                                   aclFormat format = ACL_FORMAT_ND,
+                                   size_t offset = 0) {
+    int64_t tmp_ne[GGML_MAX_DIMS * 2];
+    int64_t tmp_stride[GGML_MAX_DIMS * 2];
+
+    memcpy(tmp_ne, ne, dims * sizeof(int64_t));
+    for (int i = 0; i < dims; i++) {
+        tmp_stride[i] = nb[i] / type_size;
+    }
+
+    std::reverse(tmp_ne, tmp_ne + dims);
+    std::reverse(tmp_stride, tmp_stride + dims);
+
+    int64_t acl_storage_len = 0;
+    for (int i = 0; i < dims; i++) {
+        acl_storage_len += (ne[i] - 1) * nb[i];
+    }
+
+    aclTensor* acl_tensor =
+        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
+                        format, &acl_storage_len, 1, data_ptr);
+
+    return acl_tensor;
+}
 
 /**
  * @brief   Checks if tensors require broadcasting based on their shapes.

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -464,9 +464,11 @@ void ggml_cann_group_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     aclTensor* acl_src = ggml_cann_create_tensor(src);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 
-    const float eps = 1e-6f;  // TODO: make this a parameter
     int n_groups = dst->op_params[0];
 
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
     void* workspaceAddr = nullptr;
@@ -844,7 +846,7 @@ void ggml_cann_pool2d(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
             ggml_cann_max_pool2d(ctx, dst);
             break;
         case GGML_OP_POOL_COUNT:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 }
@@ -910,6 +912,13 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)dst->extra)->ne);
             return;
         }
+        if (dst->type == GGML_TYPE_Q4_0) {
+            aclrtlaunch_ascendc_quantize_f16_to_q4_0(
+                24, ctx.stream(), src->data, dst->data,
+                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
+                ((ggml_tensor*)dst->extra)->ne);
+            return;
+        }
         if (dst->type == GGML_TYPE_F16) {
             if (ggml_are_same_shape(src, dst)) {
                 cann_copy(ctx, acl_src, acl_dst);
@@ -931,9 +940,9 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                         ((ggml_tensor*)dst->extra)->nb);
                     return;
                 }
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
             }
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
         if (dst->type == GGML_TYPE_F32) {
             if (ggml_are_same_shape(src, dst)) {
@@ -955,12 +964,12 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                         ((ggml_tensor*)dst->extra)->nb);
                     return;
                 }
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
             }
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
         // TODO
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     } else if (src->type == GGML_TYPE_F32) {
         // TODO: if (src0->type == dst->type && ne00 == ne0 && nb00 == type_size
         //          && nb0 == type_size)
@@ -971,6 +980,13 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)dst->extra)->ne);
             return;
         }
+        if (dst->type == GGML_TYPE_Q4_0) {
+            aclrtlaunch_ascendc_quantize_f32_to_q4_0(
+                24, ctx.stream(), src->data, dst->data,
+                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
+                ((ggml_tensor*)dst->extra)->ne);
+            return;
+        }
         if (dst->type == GGML_TYPE_F32) {
             if (ggml_are_same_shape(src, dst)) {
                 cann_copy(ctx, acl_src, acl_dst);
@@ -991,10 +1007,10 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                         ((ggml_tensor*)dst->extra)->nb);
                     return;
                 }
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
             } else {
                 // TODO: dst not contiguous
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
             }
         }
         if (dst->type == GGML_TYPE_F16) {
@@ -1017,11 +1033,11 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                         ((ggml_tensor*)dst->extra)->nb);
                     return;
                 }
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
             }
         }
         // TODO
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     } else {
         if (ggml_are_same_shape(src, dst)) {
             cann_copy(ctx, acl_src, acl_dst);
@@ -1029,7 +1045,7 @@ void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
             ACL_CHECK(aclDestroyTensor(acl_dst));
             return;
         }
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
@@ -1312,6 +1328,111 @@ aclnnStatus aclnnIm2col(void* workspace, uint64_t workspaceSize,
 #ifdef __cplusplus
 }
 #endif
+
+static void ggml_cann_im2col_2d_post_process(ggml_backend_cann_context& ctx,
+                                             ggml_tensor* dst,
+                                             ggml_tensor* src1,
+                                             aclTensor* tmp_cast_tensor,
+                                             aclTensor* tmp_im2col_tensor) {
+    // Permute: [N, IC * KH * KW, OW * OH] -> [N, OW * OH, IC * KH * KW]
+    int64_t dst_ne[] = {dst->ne[0], dst->ne[1] * dst->ne[2], dst->ne[3]};
+    size_t dst_nb[] = {dst->nb[0], dst->nb[1], dst->nb[3]};
+    aclTensor* acl_dst =
+        ggml_cann_create_tensor(dst, dst_ne, dst_nb, GGML_MAX_DIMS - 1);
+
+    int64_t permute_dim[] = {0, 2, 1};
+    if (src1->type != dst->type) {
+        aclnn_permute(ctx, tmp_cast_tensor, acl_dst, permute_dim, 3);
+    } else {
+        aclnn_permute(ctx, tmp_im2col_tensor, acl_dst, permute_dim, 3);
+    }
+
+    // release
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+static void ggml_cann_im2col_1d_post_process(
+    ggml_backend_cann_context& ctx, ggml_tensor* dst, ggml_tensor* src1,
+    aclTensor* tmp_cast_tensor, aclTensor* tmp_im2col_tensor,
+    const std::vector<int64_t>& im2col_op_params) {
+    // get params
+    const int64_t KH = im2col_op_params[0];
+    const int64_t KW = im2col_op_params[1];
+    const int64_t IW = im2col_op_params[2];
+    const int64_t IC = im2col_op_params[3];
+    const int64_t N = im2col_op_params[4];
+    const int64_t OH = im2col_op_params[5];
+    const int64_t OW = im2col_op_params[6];
+    const int64_t s0 = im2col_op_params[7];
+    const int64_t p0 = im2col_op_params[8];
+    const int64_t d0 = im2col_op_params[9];
+    const int64_t n_bytes_factor = im2col_op_params[10];
+
+    // Permute: [N, IC * KH * KW, OW * OH] ->
+    // [N, OW * OH * n_bytes_factor, IC * KH * KW]
+    aclTensor* tmp_permute_tensor = nullptr;
+    ggml_cann_pool_alloc tmp_permute_allocator(ctx.pool());
+    tmp_permute_allocator.alloc(ggml_nbytes(dst) * n_bytes_factor);
+    void* tmp_permute_buffer = tmp_permute_allocator.get();
+
+    int64_t tmp_permute_ne[] = {IC * KH * KW, OW * OH * n_bytes_factor, N};
+    size_t tmp_permute_nb[GGML_MAX_DIMS - 1];
+    tmp_permute_nb[0] = ggml_type_size(dst->type);
+    for (int i = 1; i < GGML_MAX_DIMS - 1; i++) {
+        tmp_permute_nb[i] = tmp_permute_nb[i - 1] * tmp_permute_ne[i - 1];
+    }
+
+    tmp_permute_tensor = ggml_cann_create_tensor(
+        tmp_permute_buffer, ggml_cann_type_mapping(dst->type),
+        ggml_type_size(dst->type), tmp_permute_ne, tmp_permute_nb,
+        GGML_MAX_DIMS - 1, ACL_FORMAT_ND);
+
+    int64_t permute_dim[] = {0, 2, 1};
+    if (src1->type != dst->type) {
+        aclnn_permute(ctx, tmp_cast_tensor, tmp_permute_tensor, permute_dim, 3);
+    } else {
+        aclnn_permute(ctx, tmp_im2col_tensor, tmp_permute_tensor, permute_dim,
+                      3);
+    }
+
+    // number of times the kernel moves in W dimension
+    const int n_step_w = (IW + 2 * p0 - d0 * (KW - 1) - 1) / s0 + 1;
+    size_t offset;
+    void *cur_dst_buffer = dst->data, *cur_permute_buffer = tmp_permute_buffer;
+
+    // memory copy with offset to restore 1D im2col from 2d
+    if (IC > 1) {
+        offset = IC * KH * KW * n_step_w * ggml_type_size(dst->type);
+        size_t size_cpy = KH * KW * ggml_type_size(dst->type);
+
+        for (int c = 0; c < IC; c++) {
+            cur_permute_buffer = (char*)tmp_permute_buffer + offset +
+                                 KH * KW * c * ggml_type_size(dst->type);
+            cur_dst_buffer = (char*)dst->data +
+                             c * KH * KW * n_step_w * ggml_type_size(dst->type);
+
+            for (int i = 0; i < n_step_w; i++) {
+                ACL_CHECK(aclrtMemcpyAsync(
+                    cur_dst_buffer, size_cpy, cur_permute_buffer, size_cpy,
+                    ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
+                cur_dst_buffer =
+                    (char*)cur_dst_buffer + KH * KW * ggml_type_size(dst->type);
+                cur_permute_buffer = (char*)cur_permute_buffer +
+                                     KH * KW * IC * ggml_type_size(dst->type);
+            }
+        }
+    } else {
+        offset = KH * KW * n_step_w *
+                 ggml_type_size(dst->type);  // equal to ggml_nbytes(dst)
+        ACL_CHECK(aclrtMemcpyAsync(dst->data, offset,
+                                   (char*)tmp_permute_buffer + offset, offset,
+                                   ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
+    }
+
+    // release
+    ACL_CHECK(aclDestroyTensor(tmp_permute_tensor));
+}
+
 void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src0 = dst->src[0];  // kernel
     ggml_tensor* src1 = dst->src[1];  // input
@@ -1320,21 +1441,23 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
 
-    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
-    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
-    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
-    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
-    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
-    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
-    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
-
     GGML_TENSOR_BINARY_OP_LOCALS;
 
-    const int64_t N = is_2D ? ne13 : ne12;
-    const int64_t IC = is_2D ? ne12 : ne11;
+    // aclnnIm2col only works on 2D. set s1, p1, d1 to 1 to perform 2D
+    // im2col and do post-processing to restore it to 1D.
+    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t s1 = is_2D ? ((const int32_t*)(dst->op_params))[1] : 1;
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t p1 = is_2D ? ((const int32_t*)(dst->op_params))[3] : 1;
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t d1 = is_2D ? ((const int32_t*)(dst->op_params))[5] : 1;
 
-    const int64_t KH = is_2D ? ne01 : 1;
+    const int64_t N = ne13;
+    const int64_t IC = ne12;
+    const int64_t KH = ne01;
     const int64_t KW = ne00;
+    const int64_t IW = ne10;
 
     const int64_t OH = is_2D ? ne2 : 1;
     const int64_t OW = ne1;
@@ -1342,9 +1465,12 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
     GGML_ASSERT(nb10 == sizeof(float));
 
-    // im2col: [N,C,H,W] -> [N, IC * KH * KW, OW * OH]
+    // memory allocated increased to 3x when is_2D == false
+    const int64_t n_bytes_factor = is_2D ? 1 : 3;
+
+    // im2col: [N,C,H,W] -> [N, IC * KH * KW, OW * OH * n_bytes_factor]
     aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
-    int64_t tmp_im2col_ne[] = {OW * OH, IC * KH * KW, N};
+    int64_t tmp_im2col_ne[] = {OW * OH * n_bytes_factor, IC * KH * KW, N};
     size_t tmp_im2col_nb[GGML_MAX_DIMS - 1];
 
     tmp_im2col_nb[0] = ggml_type_size(src1->type);
@@ -1356,8 +1482,10 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     // If dst is f16, tmp_buffer is f32, we need alloc src.typesize *
     // dst.elemcount.
     ggml_cann_pool_alloc im2col_allocator(
-        ctx.pool(), ggml_nelements(dst) * ggml_element_size(src1));
+        ctx.pool(),
+        ggml_nelements(dst) * ggml_element_size(src1) * n_bytes_factor);
     void* tmp_im2col_buffer = im2col_allocator.get();
+
     aclTensor* tmp_im2col_tensor = ggml_cann_create_tensor(
         tmp_im2col_buffer, ggml_cann_type_mapping(src1->type),
         ggml_type_size(src1->type), tmp_im2col_ne, tmp_im2col_nb,
@@ -1380,8 +1508,9 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                                           paddings, strides, tmp_im2col_tensor,
                                           &workspaceSize, &executor));
 
+    ggml_cann_pool_alloc workspace_allocator(ctx.pool());
     if (workspaceSize > 0) {
-        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspace_allocator.alloc(workspaceSize);
         workspaceAddr = workspace_allocator.get();
     }
 
@@ -1391,9 +1520,10 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     // Cast if dst is f16.
     aclTensor* tmp_cast_tensor = nullptr;
     ggml_cann_pool_alloc tmp_cast_allocator(ctx.pool());
+    void* tmp_cast_buffer = nullptr;
     if (src1->type != dst->type) {
-        tmp_cast_allocator.alloc(ggml_nbytes(dst));
-        void* tmp_cast_buffer = tmp_cast_allocator.get();
+        tmp_cast_allocator.alloc(ggml_nbytes(dst) * n_bytes_factor);
+        tmp_cast_buffer = tmp_cast_allocator.get();
         size_t temp_cast_nb[GGML_MAX_DIMS - 1];
         temp_cast_nb[0] = ggml_type_size(dst->type);
         for (int i = 1; i < GGML_MAX_DIMS - 1; i++) {
@@ -1408,24 +1538,21 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                    ggml_cann_type_mapping(dst->type));
     }
 
-    // Permute: [N, IC * KH * KW, OW * OH] -> [N, OW * OH, IC * KH * KW]
-    int64_t dst_ne[] = {dst->ne[0], dst->ne[1] * dst->ne[2], dst->ne[3]};
-    size_t dst_nb[] = {dst->nb[0], dst->nb[1], dst->nb[3]};
-    aclTensor* acl_dst =
-        ggml_cann_create_tensor(dst, dst_ne, dst_nb, GGML_MAX_DIMS - 1);
-
-    int64_t permute_dim[] = {0, 2, 1};
-    if (src1->type != dst->type) {
-        aclnn_permute(ctx, tmp_cast_tensor, acl_dst, permute_dim, 3);
+    // post-processing
+    if (is_2D) {
+        ggml_cann_im2col_2d_post_process(ctx, dst, src1, tmp_cast_tensor,
+                                         tmp_im2col_tensor);
     } else {
-        aclnn_permute(ctx, tmp_im2col_tensor, acl_dst, permute_dim, 3);
+        std::vector<int64_t> im2col_op_params = {
+            KH, KW, IW, IC, N, OH, OW, s0, p0, d0, n_bytes_factor};
+        ggml_cann_im2col_1d_post_process(ctx, dst, src1, tmp_cast_tensor,
+                                         tmp_im2col_tensor, im2col_op_params);
     }
 
     // release
     ACL_CHECK(aclDestroyTensor(acl_src1));
     ACL_CHECK(aclDestroyTensor(tmp_im2col_tensor));
     ACL_CHECK(aclDestroyTensor(tmp_cast_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_dst));
     ACL_CHECK(aclDestroyIntArray(kernel_size));
     ACL_CHECK(aclDestroyIntArray(dilations));
     ACL_CHECK(aclDestroyIntArray(paddings));
@@ -2219,7 +2346,7 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)dst->extra)->nb);
             break;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 }
@@ -2352,21 +2479,33 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
  * @param dst The destination tensor where the result of the matrix
  * multiplication will be stored.
  */
-static void ggml_cann_mul_mat_q8_0(ggml_backend_cann_context& ctx,
-                                   ggml_tensor* dst) {
+static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
+                                   ggml_tensor* dst,
+                                   const enum ggml_type type) {
     ggml_tensor* src0 = dst->src[0];  // weight
     ggml_tensor* src1 = dst->src[1];  // input
 
     // The shape of the weight is NCHW. Matrix multiplication uses HW dims. HC
     // is regarded as batch. weight need transpose.
     int64_t weight_ne[] = {src0->ne[1], src0->ne[0]};
-    size_t weight_elem_size = sizeof(uint8_t);
-    size_t weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size};
+    float weight_elem_size;
+    if (type == GGML_TYPE_Q4_0) {
+        weight_elem_size = float(sizeof(uint8_t)) / 2;
+    }
+    else if (type == GGML_TYPE_Q8_0) {
+        weight_elem_size = float(sizeof(uint8_t));
+    }
+    else {
+        GGML_ABORT("Only support Q4_0 and Q8_0 MUL_MAT");
+    }
+    float weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size};
+
     // size of one matrix is element_size * height * width.
     size_t weight_stride = weight_elem_size * src0->ne[0] * src0->ne[1];
     size_t weight_size = weight_stride * src0->ne[2] * src0->ne[3];
 
     // scale stored at the end of weight. Also need transpose.
+    GGML_ASSERT(QK4_0 == QK8_0);
     int64_t scale_ne[] = {src0->ne[1], src0->ne[0] / QK8_0};
     size_t scale_elem_size = sizeof(uint16_t);
     size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size,
@@ -2381,10 +2520,10 @@ static void ggml_cann_mul_mat_q8_0(ggml_backend_cann_context& ctx,
     size_t input_nb[] = {input_elem_size, input_elem_size * src1->ne[0]};
     size_t input_stride = input_elem_size * src1->ne[0] * src1->ne[1];
 
+    ggml_cann_pool_alloc input_alloctor(ctx.pool());
     if (src1->type != GGML_TYPE_F16) {
         aclTensor* acl_src1_tensor = ggml_cann_create_tensor(src1);
-        ggml_cann_pool_alloc input_alloctor(
-            ctx.pool(), ggml_nelements(src1) * input_elem_size);
+        input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
         input_buffer = input_alloctor.get();
 
         int64_t* input_cast_ne = src1->ne;
@@ -2430,8 +2569,9 @@ static void ggml_cann_mul_mat_q8_0(ggml_backend_cann_context& ctx,
                 (char*)input_buffer + batch1 * input_stride, ACL_FLOAT16,
                 input_elem_size, input_ne, input_nb, 2);
             aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
-                (char*)src0->data + batch0 * weight_stride, ACL_INT8,
-                weight_elem_size, weight_ne, weight_nb, 2);
+                (char*)src0->data + batch0 * weight_stride,
+                ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
+                weight_nb, 2);
             aclTensor* acl_scale_tensor = ggml_cann_create_tensor(
                 scale_offset + batch0 * scale_stride, ACL_FLOAT16,
                 scale_elem_size, scale_ne, scale_nb, 2);
@@ -2485,14 +2625,12 @@ void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         case GGML_TYPE_F16:
             ggml_cann_mat_mul_fp(ctx, dst);
             break;
-        // case GGML_TYPE_Q4_0:
-        //     ggml_cann_mul_mat_q4_0(ctx, dst);
-        //     break;
+        case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q8_0:
-            ggml_cann_mul_mat_q8_0(ctx, dst);
+            ggml_cann_mul_mat_quant(ctx, dst, type);
             break;
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
             break;
     }
 }
@@ -2743,7 +2881,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast,
                              beta_slow, corr_dims);
 
-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
     // init cos/sin cache
     ggml_cann_pool_alloc sin_allocator(

ggml/src/ggml-cann/kernels/CMakeLists.txt

@@ -9,6 +9,7 @@ file(GLOB SRC_FILES
     get_row_q8_0.cpp
     quantize_f32_q8_0.cpp
     quantize_f16_q8_0.cpp
+    quantize_float_to_q4_0.cpp
     dup.cpp
 )
 
@@ -29,4 +30,4 @@ ascendc_library(ascendc_kernels STATIC
     ${SRC_FILES}
 )
 
-#ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)
+# ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)