Using dp4a ptx intrinsics for an improved Mul8MAT perf [By Alcpz]

.devops/llama-cli-cann.Dockerfile

@@ -1,44 +0,0 @@
-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-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
+    apt-get install -y build-essential git libcurl4-openssl-dev curl
 
 WORKDIR /app
 
@@ -16,7 +16,7 @@ RUN make -j$(nproc) llama-server
 FROM ubuntu:$UBUNTU_VERSION AS runtime
 
 RUN apt-get update && \
-    apt-get install -y libcurl4-openssl-dev libgomp1 curl
+    apt-get install -y libcurl4-openssl-dev libgomp1
 
 COPY --from=build /app/llama-server /llama-server
 

.devops/nix/package.nix

@@ -126,9 +126,16 @@ let
     ++ optionals useMetalKit [ MetalKit ];
 
   cudaBuildInputs = with cudaPackages; [
-    cuda_cudart
-    cuda_cccl # <nv/target>
-    libcublas
+    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
   ];
 
   rocmBuildInputs = with rocmPackages; [

.ecrc

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

.github/workflows/bench.yml

@@ -1,6 +1,3 @@
-# TODO: there have been some issues with the workflow, so disabling for now
-#       https://github.com/ggerganov/llama.cpp/issues/7893
-#
 # Benchmark
 name: Benchmark
 
@@ -132,8 +129,6 @@ jobs:
 
       - name: Server bench
         id: server_bench
-        env:
-            HEAD_REF: ${{ github.head_ref || github.ref_name }}
         run: |
           set -eux
 
@@ -142,7 +137,7 @@ jobs:
           python bench.py \
               --runner-label ${{ env.RUNNER_LABEL }} \
               --name ${{ github.job }} \
-              --branch $HEAD_REF \
+              --branch ${{ github.head_ref || github.ref_name }} \
               --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 -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF ..
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=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 -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
+          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=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 -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -696,20 +696,22 @@ jobs:
     strategy:
       matrix:
         include:
-          - build: 'noavx-x64'
-            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'
+          - 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'
+          - build: 'avx2-x64'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'avx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx512-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'openblas-x64'
-            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"'
+            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"'
           - build: 'kompute-x64'
-            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'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=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_RPC=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=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'

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

@@ -6,13 +6,15 @@ 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

@@ -79,6 +79,7 @@ models-mnt
 !models/ggml-vocab-*.gguf*
 
 # Zig
+
 zig-out/
 zig-cache/
 
@@ -129,6 +130,3 @@ poetry.toml
 
 # Scripts
 !/scripts/install-oneapi.bat
-
-# Test models for lora adapters
-/lora-tests

CMakeLists.txt

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

@@ -28,7 +28,6 @@
     { "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,
@@ -61,8 +60,6 @@
     { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] },
 
     { "name": "x64-windows-sycl-debug"  , "inherits": [ "sycl-base", "debug"   ] },
-    { "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" ] }
+    { "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] }
   ]
 }

CONTRIBUTING.md

@@ -5,7 +5,6 @@
   - 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

@@ -19,7 +19,6 @@ BUILD_TARGETS = \
 	llama-imatrix \
 	llama-infill \
 	llama-llava-cli \
-	llama-minicpmv-cli\
 	llama-lookahead \
 	llama-lookup \
 	llama-lookup-create \
@@ -763,10 +762,6 @@ 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
@@ -893,16 +888,15 @@ 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 -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}
+	$(eval TEMP_ASSEMBLY=$(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}
 endif
 endif # GGML_METAL
 
@@ -1211,7 +1205,6 @@ 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
@@ -1458,20 +1451,15 @@ 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/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
-
-llama-minicpmv-cli: examples/llava/minicpmv-cli.cpp \
-	examples/llava/llava.cpp \
-	examples/llava/llava.h \
 	examples/llava/clip.cpp \
-	examples/llava/clip.h \
+	examples/llava/llava.h \
+	examples/llava/llava.cpp \
 	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
+	$(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)
 
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift
@@ -1617,41 +1605,42 @@ 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.o
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
+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)
 	@echo "NOTICE: The 'main' binary is deprecated. Please use 'llama-cli' instead."
 
-server: examples/deprecation-warning/deprecation-warning.o
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
+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)
 	@echo "NOTICE: The 'server' binary is deprecated. Please use 'llama-server' instead."
 
-quantize: examples/deprecation-warning/deprecation-warning.o
+quantize: examples/deprecation-warning/deprecation-warning.cpp
 ifneq (,$(wildcard quantize))
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -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.o
+perplexity: examples/deprecation-warning/deprecation-warning.cpp
 ifneq (,$(wildcard perplexity))
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -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.o
+embedding: examples/deprecation-warning/deprecation-warning.cpp
 ifneq (,$(wildcard embedding))
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -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

@@ -95,18 +95,8 @@ 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))
 
@@ -155,7 +145,6 @@ 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)
@@ -188,12 +177,10 @@ 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.
@@ -426,7 +413,6 @@ Please refer to [Build llama.cpp locally](./docs/build.md)
 | [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

@@ -13,9 +13,6 @@
 # # 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>"
@@ -43,7 +40,7 @@ if [ ! -z ${GG_BUILD_METAL} ]; then
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=1"
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then
@@ -55,10 +52,6 @@ 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
@@ -114,7 +107,7 @@ 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$(nproc)                                  ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j                                          ) 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
 
@@ -145,7 +138,7 @@ 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$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j                                            ) 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
@@ -273,6 +266,7 @@ function gg_sum_ctest_with_model_release {
 }
 
 # open_llama_7b_v2
+# requires: GG_BUILD_CUDA
 
 function gg_run_open_llama_7b_v2 {
     cd ${SRC}
@@ -296,8 +290,8 @@ function gg_run_open_llama_7b_v2 {
 
     set -e
 
-    (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
+    (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
 
     python3 ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
@@ -431,7 +425,7 @@ 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$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
@@ -541,6 +535,7 @@ function gg_sum_pythia_1_4b {
 }
 
 # pythia_2_8b
+# requires: GG_BUILD_CUDA
 
 function gg_run_pythia_2_8b {
     cd ${SRC}
@@ -561,8 +556,8 @@ function gg_run_pythia_2_8b {
 
     set -e
 
-    (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
+    (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
 
     python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
@@ -697,7 +692,7 @@ 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$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
@@ -766,7 +761,7 @@ 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} ] && [ -z ${GG_BUILD_VULKAN} ]; then
+        if [ -z ${GG_BUILD_CUDA} ]; then
             test $ret -eq 0 && gg_run pythia_1_4b
         else
             test $ret -eq 0 && gg_run pythia_2_8b

common/common.cpp

@@ -77,41 +77,6 @@
 
 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
 //
@@ -145,34 +110,8 @@ int32_t cpu_get_num_physical_cores() {
     if (result == 0) {
         return num_physical_cores;
     }
-#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;
+#elif defined(_WIN32)
+    //TODO: Implement
 #endif
     unsigned int n_threads = std::thread::hardware_concurrency();
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
@@ -255,6 +194,12 @@ int32_t cpu_get_num_math() {
 // 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
@@ -302,9 +247,7 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
 
     gpt_params_handle_model_default(params);
 
-    if (params.hf_token.empty()) {
-        get_env("HF_TOKEN", params.hf_token);
-    }
+    gpt_params_handle_hf_token(params);
 
     if (params.escape) {
         string_process_escapes(params.prompt);
@@ -324,25 +267,6 @@ 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_THREADS",          params.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);
-}
-
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
     const auto params_org = params; // the example can modify the default params
 
@@ -760,24 +684,14 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
     }
     if (arg == "--lora") {
         CHECK_ARG
-        params.lora_adapters.push_back({
-            std::string(argv[i]),
-            1.0,
-        });
+        params.lora_adapter.emplace_back(argv[i], 1.0f);
         return true;
     }
     if (arg == "--lora-scaled") {
         CHECK_ARG
-        std::string lora_adapter = argv[i];
+        const char* lora_adapter = argv[i];
         CHECK_ARG
-        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;
+        params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i]));
         return true;
     }
     if (arg == "--control-vector") {
@@ -901,7 +815,7 @@ 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 == "--n-gpu-layers-draft") {
+    if (arg == "-ngld" || arg == "--gpu-layers-draft" || arg == "--gpu-layers-draft") {
         CHECK_ARG
         params.n_gpu_layers_draft = std::stoi(argv[i]);
         if (!llama_supports_gpu_offload()) {
@@ -1720,7 +1634,7 @@ 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)",         "restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled" });
+    options.push_back({ "server",      "       --embedding(s)",         "enable embedding endpoint (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" });
@@ -1740,7 +1654,6 @@ 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" });
@@ -1803,13 +1716,7 @@ std::string gpt_params_get_system_info(const gpt_params & params) {
     if (params.n_threads_batch != -1) {
         os << " (n_threads_batch = " << params.n_threads_batch << ")";
     }
-#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();
 }
@@ -1859,23 +1766,6 @@ 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;
@@ -2149,8 +2039,8 @@ std::string fs_get_cache_file(const std::string & filename) {
 //
 // Model utils
 //
-struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
-    llama_init_result iparams;
+
+std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params) {
     auto mparams = llama_model_params_from_gpt_params(params);
 
     llama_model * model = nullptr;
@@ -2165,7 +2055,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
 
     if (model == NULL) {
         fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
-        return iparams;
+        return std::make_tuple(nullptr, nullptr);
     }
 
     auto cparams = llama_context_params_from_gpt_params(params);
@@ -2174,7 +2064,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
     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 iparams;
+        return std::make_tuple(nullptr, nullptr);
     }
 
     if (!params.control_vectors.empty()) {
@@ -2185,7 +2075,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         if (cvec.n_embd == -1) {
             llama_free(lctx);
             llama_free_model(model);
-            return iparams;
+            return std::make_tuple(nullptr, nullptr);
         }
 
         int err = llama_control_vector_apply(lctx,
@@ -2197,26 +2087,21 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         if (err) {
             llama_free(lctx);
             llama_free_model(model);
-            return iparams;
+            return std::make_tuple(nullptr, nullptr);
         }
     }
 
-    // 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());
+    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__);
             llama_free(lctx);
             llama_free_model(model);
-            return iparams;
+            return std::make_tuple(nullptr, nullptr);
         }
-        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);
+        llama_lora_adapter_set(lctx, adapter, lora_scale);
     }
 
     if (params.ignore_eos) {
@@ -2244,26 +2129,13 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
             tmp.clear();
             tmp.push_back(decoder_start_token_id);
         }
-        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_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);
     }
 
-    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);
-        }
-    }
+    return std::make_tuple(model, lctx);
 }
 
 struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params) {
@@ -2790,6 +2662,12 @@ 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
 //
@@ -3282,18 +3160,19 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
     }
 
     fprintf(stream, "lora:\n");
-    for (auto & la : params.lora_adapters) {
-        if (la.scale == 1.0f) {
-            fprintf(stream, "  - %s\n", la.path.c_str());
+    for (std::tuple<std::string, float> la : params.lora_adapter) {
+        if (std::get<1>(la) != 1.0f) {
+            continue;
         }
+        fprintf(stream, "  - %s\n", std::get<0>(la).c_str());
     }
     fprintf(stream, "lora_scaled:\n");
-    for (auto & la : params.lora_adapters) {
-        if (la.scale != 1.0f) {
-            fprintf(stream, "  - %s: %f\n", la.path.c_str(), la.scale);
+    for (std::tuple<std::string, float> la : params.lora_adapter) {
+        if (std::get<1>(la) == 1.0f) {
+            continue;
         }
+        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);

common/common.h

@@ -33,15 +33,6 @@
 
 #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;
@@ -135,8 +126,8 @@ 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;
 
-    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
+    // TODO: avoid tuple, use struct
+    std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
 
     std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
 
@@ -267,7 +258,7 @@ struct gpt_params {
     std::string lora_outfile = "ggml-lora-merged-f16.gguf";
 };
 
-void gpt_params_parse_from_env(gpt_params & params);
+void gpt_params_handle_hf_token(gpt_params & params);
 void gpt_params_handle_model_default(gpt_params & params);
 
 bool gpt_params_parse_ex   (int argc, char ** argv, gpt_params & params);
@@ -286,8 +277,6 @@ 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;
@@ -319,13 +308,8 @@ std::string fs_get_cache_file(const std::string & filename);
 // Model utils
 //
 
-struct llama_init_result {
-    struct llama_model   * model   = nullptr;
-    struct llama_context * context = nullptr;
-    std::vector<llama_lora_adapter_container> lora_adapters;
-};
-
-struct llama_init_result    llama_init_from_gpt_params(gpt_params & params);
+// TODO: avoid tuplue, use struct
+std::tuple<struct llama_model *, struct llama_context *> 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);
@@ -333,9 +317,6 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
 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);
@@ -380,6 +361,10 @@ 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

@@ -369,9 +369,6 @@ 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

@@ -63,7 +63,6 @@ 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
@@ -71,7 +70,7 @@ 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, is_lora: bool = False):
+                 split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False):
         if type(self) is Model:
             raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
 
@@ -93,7 +92,6 @@ 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:
@@ -253,7 +251,12 @@ class Model:
 
         return [(self.map_tensor_name(name), data_torch)]
 
-    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
+    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:
         del name, new_name, bid, n_dims  # unused
 
         return False
@@ -282,47 +285,54 @@ 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_qtype: gguf.GGMLQuantizationType | bool = self.tensor_force_quant(name, new_name, bid, n_dims)
+                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)
 
                 # Most of the codebase that takes in 1D tensors or norms only handles F32 tensors
-                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
+                extra_f32 = any(cond for cond in (
+                    extra_f32,
+                    n_dims == 1,
+                    new_name.endswith("_norm.weight"),
+                ))
+
                 # 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,
-                        )
-                    )
-                    or not name.endswith(".weight")
-                ):
-                    data_qtype = gguf.GGMLQuantizationType.F32
+                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,
+                ))
 
-                # 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:
+                # 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:
-                        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)
+                    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)
+                    data_qtype = gguf.GGMLQuantizationType.F32
 
                 shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape
 
@@ -593,15 +603,6 @@ 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")
@@ -905,7 +906,7 @@ class GPTNeoXModel(Model):
         return tensors
 
 
-@Model.register("BloomForCausalLM", "BloomModel")
+@Model.register("BloomForCausalLM")
 class BloomModel(Model):
     model_arch = gguf.MODEL_ARCH.BLOOM
 
@@ -1572,7 +1573,7 @@ class LlamaModel(Model):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
+                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)
@@ -1595,8 +1596,7 @@ class LlamaModel(Model):
                         smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
                         rope_factors.append(1 / ((1 - smooth) / factor + smooth))
 
-                if not self.is_lora:
-                    self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
+                self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
 
         super().prepare_tensors()
 
@@ -1765,7 +1765,7 @@ class DbrxModel(Model):
 
         return [(new_name, data_torch)]
 
-    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
+    def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
         del name, new_name, bid  # unused
 
         return n_dims > 1
@@ -2143,9 +2143,8 @@ class Phi3MiniModel(Model):
         if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
             raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
 
-        if not self.is_lora:
-            self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG]  + ".weight", np.array(long_factors, dtype=np.float32))
-            self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
+        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")
@@ -2507,112 +2506,6 @@ 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
@@ -2716,7 +2609,7 @@ class StarCoder2Model(Model):
     model_arch = gguf.MODEL_ARCH.STARCODER2
 
 
-@Model.register("MambaForCausalLM", "MambaLMHeadModel", "FalconMambaForCausalLM")
+@Model.register("MambaForCausalLM", "MambaLMHeadModel")
 class MambaModel(Model):
     model_arch = gguf.MODEL_ARCH.MAMBA
 
@@ -2747,10 +2640,7 @@ 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
 
@@ -2758,13 +2648,12 @@ 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.block_count)
+        self.gguf_writer.add_block_count(self.hparams["n_layer"])
         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
@@ -2791,6 +2680,19 @@ 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):
@@ -3325,145 +3227,6 @@ 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
@@ -3735,122 +3498,9 @@ class ChatGLMModel(Model):
         name = name.removeprefix("transformer.")
         return [(self.map_tensor_name(name), data_torch)]
 
-
-@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):
     _tensor_type = torch.Tensor

convert_hf_to_gguf_update.py

@@ -94,9 +94,6 @@ 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= gguf.GGMLQuantizationType(dtype)
+        self.dtype= 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

@@ -386,7 +386,6 @@ 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

@@ -1,259 +0,0 @@
-# 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 and oneDNN)*.
+- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL - Math Kernel Library)*.
 - **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.
 
@@ -28,6 +28,10 @@
 
 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.
@@ -41,10 +45,6 @@ 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.
@@ -80,14 +80,7 @@ The following release is verified with good quality:
 
 ### Intel GPU
 
-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
+**Verified devices**
 
 | Intel GPU                     | Status  | Verified Model                        |
 |-------------------------------|---------|---------------------------------------|
@@ -95,7 +88,7 @@ SYCL backend supports Intel GPU Family:
 | 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 13700k, i5-1250P, i7-1260P, i7-1165G7 |
+| Intel iGPU                    | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 |
 
 *Notes:*
 
@@ -196,7 +189,7 @@ 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 oneDNN for Intel GPUs.
+Upon a successful installation, SYCL is enabled for the available intel devices, along with relevant libraries such as oneAPI MKL for intel GPUs.
 
 - **Adding support to Nvidia GPUs**
 
@@ -244,17 +237,12 @@ 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
 
@@ -288,26 +276,23 @@ cmake --build build --config Release -j -v
 
 ### III. Run the inference
 
-#### Retrieve and prepare model
+1. 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.
 
-##### Check device
-
-1. Enable oneAPI running environment
+2. Enable oneAPI running environment
 
 ```sh
 source /opt/intel/oneapi/setvars.sh
 ```
 
-2. List devices information
+3. 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:
@@ -319,37 +304,12 @@ 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|`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
+4. Launch inference
 
 There are two device selection modes:
 
-- Single device: Use one device assigned by user. Default device id is 0.
+- Single device: Use one device target specified by the user.
 - 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.
@@ -366,6 +326,11 @@ 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:
 
@@ -373,6 +338,12 @@ ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Bui
 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:
@@ -419,7 +390,7 @@ c. Verify installation
 In the oneAPI command line, run the following to print the available SYCL devices:
 
 ```
-sycl-ls.exe
+sycl-ls
 ```
 
 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:
@@ -440,18 +411,6 @@ 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:
 
 ```
@@ -466,8 +425,12 @@ cmake -B build -G "Ninja" -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPI
 cmake --build build --config Release -j
 ```
 
-Or, use CMake presets to build:
+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
@@ -479,9 +442,7 @@ cmake --preset x64-windows-sycl-debug
 cmake --build build-x64-windows-sycl-debug -j --target llama-cli
 ```
 
-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.
+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.
 
 *Notes:*
 
@@ -489,25 +450,23 @@ You can use Visual Studio to open llama.cpp folder as a CMake project. Choose th
 
 ### III. Run the inference
 
-#### Retrieve and prepare model
+1. 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.
+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.
 
-##### Check device
-
-1. Enable oneAPI running environment
+2. 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
 ```
 
-2. List devices information
+3. List devices information
 
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```
-build\bin\llama-ls-sycl-device.exe
+build\bin\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:
@@ -520,27 +479,9 @@ 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"`|
 
-#### Execute
-
-Choose one of following methods to run.
-
-1. Script
-
-```
-examples\sycl\win-run-llama2.bat
-```
-
-2. Command line
-
-Launch inference
+4. Launch inference
 
 There are two device selection modes:
 
@@ -567,7 +508,11 @@ 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:
 
@@ -581,18 +526,17 @@ 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.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
+| GGML_SYCL          | ON (mandatory)                    | Enable build with SYCL code path.           |
 | 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` *(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. |
+| 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. |
 
 #### Runtime
 
@@ -628,18 +572,9 @@ 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
 
-- NA
+- Support row layer split for multiple card runs.

docs/build.md

@@ -178,11 +178,7 @@ 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 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:
+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:
 
 | Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
 |-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
@@ -352,31 +348,6 @@ 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)

examples/baby-llama/baby-llama.cpp

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

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 = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());
+    ctx_params.n_seq_max = *std::max_element(n_pl.begin(), n_pl.end());
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 

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_add_bos_token(llama_get_model(ctx));
+        const bool add_bos = llama_should_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());
@@ -414,10 +414,9 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the model to get hparams
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
-
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    llama_model * model;
+    llama_context * ctx;
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
 
     // int n_ctx = llama_n_ctx(ctx);
     int n_layers = llama_n_layer(model);

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 --pooling mean --log-disable -p "Hello World!" 2>/dev/null
+./llama-embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null
 ```
 
 ### Windows:
 
 ```powershell
-llama-embedding.exe -m ./path/to/model --pooling mean --log-disable -p "Hello World!" 2>$null
+llama-embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$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
-./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
+./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
 ```
 
 ### Windows:
 
 ```powershell
-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
+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
 ```

examples/embedding/embedding.cpp

@@ -31,24 +31,13 @@ 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_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__);
-        }
+    if (llama_decode(ctx, batch) < 0) {
+        fprintf(stderr, "%s : failed to decode\n", __func__);
     }
 
     for (int i = 0; i < batch.n_tokens; i++) {
@@ -56,22 +45,11 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
             continue;
         }
 
-        const float * embd = nullptr;
-        int embd_pos = 0;
+        // 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");
 
-        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;
+        float * out = output + batch.seq_id[i][0] * n_embd;
         llama_embd_normalize(embd, out, n_embd, embd_norm);
     }
 }
@@ -101,11 +79,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    // load the model
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    llama_model * model;
+    llama_context * ctx;
 
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
@@ -115,9 +93,8 @@ 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 (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__);
+    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+        fprintf(stderr, "%s: error: pooling type NONE not supported\n", __func__);
         return 1;
     }
 
@@ -176,23 +153,13 @@ 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_embd_count * n_embd, 0);
+    std::vector<float> embeddings(n_prompts * n_embd, 0);
     float * emb = embeddings.data();
 
     // break into batches
-    int e = 0; // number of embeddings already stored
+    int p = 0; // number of prompts processed already
     int s = 0; // number of prompts in current batch
     for (int k = 0; k < n_prompts; k++) {
         // clamp to n_batch tokens
@@ -202,11 +169,11 @@ int main(int argc, char ** argv) {
 
         // encode if at capacity
         if (batch.n_tokens + n_toks > n_batch) {
-            float * out = emb + e * n_embd;
+            float * out = emb + p * n_embd;
             batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
-            e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s;
-            s = 0;
             llama_batch_clear(batch);
+            p += s;
+            s = 0;
         }
 
         // add to batch
@@ -215,62 +182,39 @@ int main(int argc, char ** argv) {
     }
 
     // final batch
-    float * out = emb + e * n_embd;
+    float * out = emb + p * 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");
-
-        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]);
-                    }
+        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, " ... ");
-                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, "\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");
             }
+            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());
+        // 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");
-                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");
-                }
             }
         }
     }
@@ -289,23 +233,23 @@ int main(int argc, char ** argv) {
             }
             fprintf(stdout, notArray ? "]\n    }" : "]");
             j++;
-            if (j < n_embd_count) fprintf(stdout, notArray ? ",\n" : ","); else break;
+            if (j < n_prompts) 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_embd_count > 1)
+            for (int i = 0;;) { // at least two iteration (n_prompts > 1)
                 fprintf(stdout, "    [");
-                for (int j = 0;;) { // at least two iteration (n_embd_count > 1)
+                for (int j = 0;;) { // at least two iteration (n_prompts > 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_embd_count) fprintf(stdout, ", "); else break;
+                    if (j < n_prompts) fprintf(stdout, ", "); else break;
                 }
                 fprintf(stdout, " ]");
                 i++;
-                if (i < n_embd_count) fprintf(stdout, ",\n"); else break;
+                if (i < n_prompts) fprintf(stdout, ",\n"); else break;
             }
             fprintf(stdout, "\n  ]");
         }

examples/eval-callback/eval-callback.cpp

@@ -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_add_bos_token(llama_get_model(ctx));
+    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
 
     std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
 
@@ -163,10 +163,9 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
-
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    llama_model * model;
+    llama_context * ctx;
+    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;

examples/export-lora/README.md

@@ -17,9 +17,9 @@ For example:
 
 ```bash
 ./bin/llama-export-lora \
-    -m open-llama-3b-v2.gguf \
-    -o open-llama-3b-v2-english2tokipona-chat.gguf \
-    --lora lora-open-llama-3b-v2-english2tokipona-chat-LATEST.gguf
+    -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
 ```
 
 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

@@ -10,12 +10,6 @@
 
 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));
@@ -56,6 +50,20 @@ 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;
@@ -127,7 +135,7 @@ struct lora_merge_ctx {
 
     lora_merge_ctx(
             std::string & base_fname,
-            std::vector<llama_lora_adapter_info> & lora_files,
+            std::vector<std::tuple<std::string, float>> & 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
@@ -136,9 +144,9 @@ struct lora_merge_ctx {
             throw std::runtime_error("split model is not yet supported");
         }
 
-        for (auto & lora_inp : lora_files) {
-            auto fname = lora_inp.path;
-            auto scale = lora_inp.scale;
+        for (auto lora_inp : lora_files) {
+            auto fname = std::get<0>(lora_inp);
+            auto scale = std::get<1>(lora_inp);
             std::unique_ptr<file_input> adapter(new file_input(fname, scale));
             check_metadata_lora(adapter.get());
             adapters.push_back(std::move(adapter));
@@ -204,7 +212,8 @@ struct lora_merge_ctx {
         }
 
         // mapping base tensor to out tensor (same shape with base, but different type)
-        std::vector<tensor_transformation> trans;
+        // if out_tensor == nullptr, we only copy it
+        std::vector<std::pair<struct ggml_tensor *, struct ggml_tensor *>> base_to_out_tensors;
         for (auto & it : base_model.tensors) {
             bool t_a = true;
             bool t_b = true;
@@ -217,22 +226,14 @@ 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);
-                trans.push_back({
-                    cpy_tensor,
-                    cpy_tensor,
-                    true,
-                });
+                base_to_out_tensors.push_back(std::make_pair(cpy_tensor, nullptr));
                 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);
-                trans.push_back({
-                    base_tensor,
-                    out_tensor,
-                    false,
-                });
+                base_to_out_tensors.push_back(std::make_pair(base_tensor, out_tensor));
                 gguf_add_tensor(ctx_out, out_tensor);
             } else {
                 throw std::runtime_error("tensor " + it.first + " missing either lora_a or lora_b");
@@ -247,12 +248,12 @@ struct lora_merge_ctx {
 
         // process base model tensors
         size_t n_merged = 0;
-        for (auto & it : trans) {
-            if (!it.is_copy) {
-                merge_tensor(it.in, it.out);
+        for (auto & it : base_to_out_tensors) {
+            if (it.second != nullptr) {
+                merge_tensor(it.first, it.second);
                 n_merged++;
             } else {
-                copy_tensor(it.in);
+                copy_tensor(it.first);
             }
         }
 
@@ -265,7 +266,7 @@ 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__, trans.size());
+        printf("%s : wrote %ld tensors to output file\n", __func__, base_to_out_tensors.size());
     }
 
     void copy_tensor(struct ggml_tensor * base) {
@@ -298,10 +299,6 @@ 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);
         }
@@ -410,7 +407,7 @@ int main(int argc, char ** argv) {
 
     g_verbose = (params.verbosity == 1);
     try {
-        lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.n_threads);
+        lora_merge_ctx ctx(params.model, params.lora_adapter, params.lora_outfile, params.n_threads);
         ctx.run_merge();
     } catch (const std::exception & err) {
         fprintf(stderr, "%s\n", err.what());

examples/imatrix/imatrix.cpp

@@ -433,8 +433,8 @@ static void process_logits(
 }
 
 static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
-    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
+    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 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_init_result llama_init = llama_init_from_gpt_params(params);
+    llama_model * model;
+    llama_context * ctx;
 
-    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;

examples/infill/infill.cpp

@@ -179,10 +179,7 @@ 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__);
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
-
-    model = llama_init.model;
-    ctx = llama_init.context;
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
 
     if (model == NULL) {
         LOG_TEE("%s: error: unable to load model\n", __func__);
@@ -203,8 +200,8 @@ 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_add_bos_token(model);
-    GGML_ASSERT(!llama_add_eos_token(model));
+    const bool add_bos = llama_should_add_bos_token(model);
+    GGML_ASSERT(llama_add_eos_token(model) != 1);
     LOG("add_bos: %d\n", add_bos);
 
     std::vector<llama_token> embd_inp;

examples/llama-bench/llama-bench.cpp

@@ -27,14 +27,6 @@
 #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;
@@ -104,27 +96,6 @@ 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;

examples/llava/CMakeLists.txt

@@ -36,10 +36,3 @@ 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

@@ -1,99 +0,0 @@
-## 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/minicpmv/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
-python ./examples/minicpmv/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2
-python ./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

@@ -1,107 +0,0 @@
-## 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

@@ -20,10 +20,6 @@
 #include "ggml-cann.h"
 #endif
 
-#ifdef GGML_USE_VULKAN
-#include "ggml-vulkan.h"
-#endif
-
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb_image.h"
 
@@ -78,28 +74,26 @@ 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_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_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_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
 #define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
@@ -133,20 +127,12 @@ 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,
 };
 
@@ -154,7 +140,6 @@ 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"},
 };
 
 
@@ -215,20 +200,17 @@ static std::string gguf_data_to_str(enum gguf_type type, const void * data, int
 }
 
 static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
-    if (search.empty()) {
-        return;
+    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;
     }
-    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);
+    s = std::move(result);
 }
 
 static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
@@ -510,34 +492,12 @@ 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;
@@ -562,11 +522,9 @@ 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, struct clip_image_size * load_image_size, bool is_inf = false) {
+static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs) {
     if (!ctx->has_vision_encoder) {
         LOG_TEE("This gguf file seems to have no vision encoder\n");
         return nullptr;
@@ -575,33 +533,20 @@ 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;
-    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 image_size           = hparams.image_size;
     const int patch_size           = hparams.patch_size;
-    const int num_patches          = ((image_size_width / patch_size) * (image_size_height / 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_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;
-    int n_layer                    = hparams.n_layer;
+    const int n_layer              = hparams.n_layer;
     const float eps                = hparams.eps;
 
     const int batch_size = imgs->size;
 
-    if (ctx->has_llava_projector || ctx->has_minicpmv_projector) {
+    if (ctx->has_llava_projector) {
         GGML_ASSERT(batch_size == 1);
     }
 
@@ -614,7 +559,7 @@ 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_width, image_size_height, 3, batch_size);
+    struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size, image_size, 3, batch_size);
     ggml_set_name(inp_raw, "inp_raw");
     ggml_set_input(inp_raw);
 
@@ -627,22 +572,20 @@ 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);
     }
-    struct ggml_tensor * embeddings = inp;
-    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]);
-        }
+    // 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 * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
     ggml_set_name(positions, "positions");
     ggml_set_input(positions);
@@ -650,19 +593,6 @@ 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);
@@ -672,9 +602,6 @@ 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
 
@@ -764,7 +691,7 @@ 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);
@@ -785,8 +712,8 @@ 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);
@@ -945,75 +872,6 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             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);
-        }
-    }
 
     // build the graph
     ggml_build_forward_expand(gf, embeddings);
@@ -1118,7 +976,7 @@ 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
     {
@@ -1152,10 +1010,6 @@ 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();
@@ -1175,18 +1029,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             new_clip->has_llava_projector = gguf_get_val_bool(ctx, idx);
         }
 
-        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_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);
 
@@ -1197,7 +1040,6 @@ 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);
         }
@@ -1439,27 +1281,6 @@ 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()));
@@ -1498,7 +1319,7 @@ 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, nullptr, false);
+        ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch);
         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);
@@ -1507,17 +1328,6 @@ 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();
 }
@@ -1788,186 +1598,9 @@ 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;
-}
-
-inline int clip(int x, int lower, int upper) {
-    return std::max(lower, std::min(x, upper));
-}
-
-static std::pair<int, int> uhd_best_grid(const int max_slice_nums, const int multiple, const float log_ratio) {
-    std::vector<int> candidate_split_grids_nums;
-    for (int i : {multiple - 1, multiple, multiple + 1}) {
-        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");
@@ -2183,104 +1816,11 @@ 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");
@@ -2303,33 +1843,19 @@ 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, ctx->load_image_size, true);
+    ggml_cgraph * gf = clip_image_build_graph(ctx, imgs);
     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;
-    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 image_size    = hparams.image_size;
     const int patch_size    = hparams.patch_size;
-    const int num_patches   = ((image_size_width / patch_size) * (image_size_height / patch_size));
+    const int num_patches   = ((image_size / patch_size) * (image_size / 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");
@@ -2338,9 +1864,7 @@ 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;
-            if (!ctx->has_minicpmv_projector) {
-                GGML_ASSERT(nx == image_size && ny == image_size);
-            }
+            GGML_ASSERT(nx == image_size && ny == image_size);
 
             const int n = nx * ny;
 
@@ -2357,87 +1881,37 @@ 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);
-        }
 
-        {
-            // 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));
+    {
+        if (ctx->has_class_embedding) {
+            struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
 
-            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);
+            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);
         }
     }
-    else{
-        {
-            if (ctx->has_class_embedding) {
-                struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
 
-                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);
-            }
+    {
+        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 * 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);
+    {
+        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);
     }
 
     if (ggml_backend_is_cpu(ctx->backend)) {
@@ -2607,22 +2081,7 @@ 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

@@ -18,17 +18,14 @@
 #    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;
@@ -58,10 +55,6 @@ 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();
 
@@ -85,8 +78,6 @@ 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.cpp

@@ -202,33 +202,6 @@ 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
@@ -245,51 +218,7 @@ 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 (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) {
+    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
@@ -299,8 +228,7 @@ 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;
@@ -369,11 +297,7 @@ 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) {
-    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
+    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // 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

@@ -17,11 +17,12 @@
 #    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;
@@ -36,8 +37,8 @@ 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);
-/** free an embedding made with llava_image_embed_make_* */
 LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
+/** free an embedding made with llava_image_embed_make_* */
 
 /** 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

@@ -1,329 +0,0 @@
-#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->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

@@ -1,806 +0,0 @@
-# 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

@@ -1,45 +0,0 @@
-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

@@ -2,4 +2,3 @@
 --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);
 
-    // load the target model
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
 
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    // load the target model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
 
     // 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);
 
-    // load the model
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
 
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
     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);
 
-    // load the model
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
 
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
 
     // 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);
 
-    // load the model
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
 
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
 
     // tokenize the prompt
     std::vector<llama_token> inp;

examples/main/main.cpp

@@ -207,10 +207,7 @@ 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__);
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
-
-    model = llama_init.model;
-    ctx = llama_init.context;
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
     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);
@@ -267,9 +264,9 @@ int main(int argc, char ** argv) {
         }
     }
 
-    const bool add_bos = llama_add_bos_token(model);
+    const bool add_bos = llama_should_add_bos_token(model);
     if (!llama_model_has_encoder(model)) {
-        GGML_ASSERT(!llama_add_eos_token(model));
+        GGML_ASSERT(llama_add_eos_token(model) != 1);
     }
     LOG("add_bos: %d\n", add_bos);
 

examples/parallel/parallel.cpp

@@ -129,11 +129,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    // load the target model
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
 
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    // load the target model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
 
     // 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_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
+    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
 
     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_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
+    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
 
     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_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
+    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
 
     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);
 
-    // load the model and apply lora adapter, if any
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    llama_model * model;
+    llama_context * ctx;
 
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    // load the model and apply lora adapter, if any
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
     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 -cnv -p "You are a helpful assistant"
+./llama-cli -m ./models/mymodel/ggml-model-Q4_K_M.gguf -n 128
 ```
 
 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:
-//  ./llama-quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
+//  ./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 quantized model in the same shards as input\n");
+    printf("  --keep-split: will generate quatized model in the same shards as input");
     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

@@ -148,12 +148,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
+    llama_model * model;
+    llama_context * ctx;
+
     // load the model
-    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 == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
@@ -253,8 +252,6 @@ 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) {
@@ -262,6 +259,7 @@ 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);
@@ -294,7 +292,6 @@ 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

@@ -1,9 +1,5 @@
 ## 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        = "127.0.0.1";
+    std::string host        = "0.0.0.0";
     int         port        = 50052;
     size_t      backend_mem = 0;
 };
@@ -114,17 +114,6 @@ 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

@@ -28,11 +28,10 @@ int main(int argc, char ** argv) {
     std::string result2;
 
     // init
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
-
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    llama_model * model;
+    llama_context * ctx;
 
+    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;

examples/server/README.md

@@ -207,12 +207,47 @@ 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)           restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)
+         --embedding(s)           enable embedding endpoint (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
@@ -232,8 +267,7 @@ 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:
 
@@ -245,27 +279,17 @@ 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`
-- `LLAMA_ARG_THREADS`
-- `LLAMA_ARG_CTX_SIZE`
-- `LLAMA_ARG_N_PARALLEL`
-- `LLAMA_ARG_BATCH`
-- `LLAMA_ARG_UBATCH`
-- `LLAMA_ARG_N_GPU_LAYERS`
-- `LLAMA_ARG_THREADS_HTTP`
-- `LLAMA_ARG_CHAT_TEMPLATE`
-- `LLAMA_ARG_N_PREDICT`
-- `LLAMA_ARG_ENDPOINT_METRICS`
-- `LLAMA_ARG_ENDPOINT_SLOTS`
-- `LLAMA_ARG_EMBEDDINGS`
-- `LLAMA_ARG_FLASH_ATTN`
-- `LLAMA_ARG_DEFRAG_THOLD`
 
 ## Build
 
@@ -387,18 +411,16 @@ node index.js
 
 ## API Endpoints
 
-### GET `/health`: Returns heath check result
+- **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.
 
-**Response format**
+  If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set.
 
-- 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.
+- **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
 
     *Options:*
 
@@ -476,7 +498,7 @@ 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.
 
-**Response format**
+### Result JSON
 
 - Note: When using streaming mode (`stream`), only `content` and `stop` will be returned until end of completion.
 
@@ -515,7 +537,7 @@ 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:*
 
@@ -523,15 +545,13 @@ 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
-
-The same as [the embedding example](../embedding) does.
+- **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
 
     *Options:*
 
@@ -539,9 +559,7 @@ The same as [the embedding example](../embedding) does.
 
     `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:*
 
@@ -553,7 +571,7 @@ Takes a prefix and a suffix and returns the predicted completion as stream.
 
 - **GET** `/props`: Return current server settings.
 
-**Response format**
+### Result JSON
 
 ```json
 {
@@ -571,9 +589,7 @@ Takes a prefix and a suffix and returns the predicted completion as stream.
 - `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:*
 
@@ -625,7 +641,7 @@ Given a ChatML-formatted json description in `messages`, it returns the predicte
     }'
     ```
 
-### POST `/v1/embeddings`: OpenAI-compatible embeddings API
+- **POST** `/v1/embeddings`: OpenAI-compatible embeddings API.
 
     *Options:*
 
@@ -659,15 +675,9 @@ Given a ChatML-formatted json description in `messages`, it returns the predicte
     }'
     ```
 
-### GET `/slots`: Returns the current slots processing state
+- **GET** `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
 
-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:
+### Result JSON
 
 ```json
 [
@@ -728,13 +738,7 @@ Example:
 ]
 ```
 
-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.
+- **GET** `/metrics`: [Prometheus](https://prometheus.io/) compatible metrics exporter endpoint if `--metrics` is enabled:
 
 Available metrics:
 - `llamacpp:prompt_tokens_total`: Number of prompt tokens processed.
@@ -746,13 +750,13 @@ 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.
 
-**Response format**
+### Result JSON
 
 ```json
 {
@@ -766,13 +770,13 @@ 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.
 
-**Response format**
+### Result JSON
 
 ```json
 {
@@ -786,9 +790,9 @@ 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.
 
-**Response format**
+### Result JSON
 
 ```json
 {
@@ -797,46 +801,6 @@ 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

@@ -15,8 +15,6 @@
 // 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"
@@ -69,6 +67,7 @@ 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 {
@@ -79,7 +78,6 @@ 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 {
@@ -624,7 +622,6 @@ 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;
 
@@ -632,7 +629,6 @@ 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
 
@@ -681,11 +677,7 @@ struct server_context {
         // dedicate one sequence to the system prompt
         params.n_parallel += 1;
 
-        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;
+        std::tie(model, ctx) = llama_init_from_gpt_params(params);
         params.n_parallel -= 1; // but be sneaky about it
         if (model == nullptr) {
             LOG_ERROR("unable to load model", {{"model", params.model}});
@@ -694,8 +686,8 @@ struct server_context {
 
         n_ctx = llama_n_ctx(ctx);
 
-        add_bos_token = llama_add_bos_token(model);
-        has_eos_token = !llama_add_eos_token(model);
+        add_bos_token = llama_should_add_bos_token(model);
+        GGML_ASSERT(llama_add_eos_token(model) != 1);
 
         return true;
     }
@@ -755,13 +747,13 @@ 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 or n_parallel tokens
+        // the update_slots() logic will always submit a maximum of n_batch 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(std::max(n_batch, params.n_parallel), 0, 1);
+            batch = llama_batch_init(n_batch, 0, 1);
         }
 
         metrics.init();
@@ -908,7 +900,7 @@ 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",         json_value(data, "max_tokens", default_params.n_predict));
+        slot.params.n_predict          = json_value(data, "n_predict",         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);
@@ -977,8 +969,6 @@ 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;
@@ -1033,7 +1023,7 @@ struct server_context {
         {
             slot.sparams.logit_bias.clear();
 
-            if (json_value(data, "ignore_eos", false) && has_eos_token) {
+            if (json_value(data, "ignore_eos", false)) {
                 slot.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
             }
 
@@ -1138,19 +1128,28 @@ 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 < n_tokens_prompt; i += n_batch) {
-                const int32_t n_tokens = std::min(n_batch, n_tokens_prompt - i);
+            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
+                };
 
-                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) {
+                if (llama_decode(ctx, batch_view) != 0) {
                     LOG_ERROR("llama_decode() failed", {});
                     return;
                 }
@@ -1323,7 +1322,7 @@ struct server_context {
 
         return json {
             {"n_ctx",                     slot.n_ctx},
-            {"n_predict",                 slot.n_predict},     // Server configured n_predict
+            {"n_predict",                 slot.n_predict},
             {"model",                     params.model_alias},
             {"seed",                      slot.sparams.seed},
             {"temperature",               slot.sparams.temp},
@@ -1345,7 +1344,7 @@ struct server_context {
             {"mirostat_eta",              slot.sparams.mirostat_eta},
             {"penalize_nl",               slot.sparams.penalize_nl},
             {"stop",                      slot.params.antiprompt},
-            {"max_tokens",                slot.params.n_predict}, // User configured n_predict
+            {"n_predict",                 slot.params.n_predict}, // TODO: fix duplicate key n_predict
             {"n_keep",                    slot.params.n_keep},
             {"n_discard",                 slot.params.n_discard},
             {"ignore_eos",                ignore_eos},
@@ -1848,16 +1847,6 @@ 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;
         }
     }
 
@@ -2039,7 +2028,7 @@ struct server_context {
                         slot.t_start_generation = 0;
 
                         if (slot.infill) {
-                            const bool add_bos = llama_add_bos_token(model);
+                            const bool add_bos = llama_should_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);
@@ -2507,9 +2496,6 @@ 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;
@@ -2560,19 +2546,19 @@ int main(int argc, char ** argv) {
     svr->set_default_headers({{"Server", "llama.cpp"}});
 
     // CORS preflight
-    svr->Options(R"(.*)", [](const httplib::Request &, httplib::Response & res) {
-        // Access-Control-Allow-Origin is already set by middleware
+    svr->Options(R"(.*)", [](const httplib::Request & req, httplib::Response & res) {
+        res.set_header("Access-Control-Allow-Origin",      req.get_header_value("Origin"));
         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("", "text/html"); // blank response, no data
+        return res.set_content("", "application/json; charset=utf-8");
     });
 
     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(), MIMETYPE_JSON);
+        res.set_content(final_response.dump(), "application/json; charset=utf-8");
         res.status = json_value(error_data, "code", 500);
     };
 
@@ -2602,6 +2588,11 @@ 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;
@@ -2617,6 +2608,35 @@ 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
     //
@@ -2660,6 +2680,8 @@ 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", {});
@@ -2667,21 +2689,8 @@ 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, &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;
-        }
+    svr->set_pre_routing_handler([&middleware_validate_api_key](const httplib::Request & req, httplib::Response & res) {
         if (!middleware_validate_api_key(req, res)) {
             return httplib::Server::HandlerResponse::Handled;
         }
@@ -2692,15 +2701,62 @@ int main(int argc, char ** argv) {
     // Route handlers (or controllers)
     //
 
-    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_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_slots = [&](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_slots = [&](const httplib::Request &, httplib::Response & res) {
         if (!params.endpoint_slots) {
-            res_error(res, format_error_response("This server does not support slots endpoint. Start it without `--no-slots`", ERROR_TYPE_NOT_SUPPORTED));
+            res_error(res, format_error_response("This server does not support slots endpoint.", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
 
@@ -2718,22 +2774,13 @@ 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);
 
-        // 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.set_content(result.data.at("slots").dump(), "application/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. Start it with `--metrics`", ERROR_TYPE_NOT_SUPPORTED));
+            res_error(res, format_error_response("This server does not support metrics endpoint.", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
 
@@ -2858,7 +2905,7 @@ int main(int argc, char ** argv) {
         if (result.error) {
             res_error(res, result.data);
         } else {
-            res.set_content(result.data.dump(), MIMETYPE_JSON);
+            res.set_content(result.data.dump(), "application/json");
         }
     };
 
@@ -2888,7 +2935,7 @@ int main(int argc, char ** argv) {
         if (result.error) {
             res_error(res, result.data);
         } else {
-            res.set_content(result.data.dump(), MIMETYPE_JSON);
+            res.set_content(result.data.dump(), "application/json");
         }
     };
 
@@ -2908,11 +2955,13 @@ int main(int argc, char ** argv) {
         if (result.error) {
             res_error(res, result.data);
         } else {
-            res.set_content(result.data.dump(), MIMETYPE_JSON);
+            res.set_content(result.data.dump(), "application/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;
 
@@ -2936,7 +2985,7 @@ int main(int argc, char ** argv) {
         }
     };
 
-    const auto handle_props = [&ctx_server](const httplib::Request &, httplib::Response & res) {
+    const auto handle_props = [&ctx_server](const httplib::Request & req, 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) {
@@ -2945,6 +2994,7 @@ 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 },
@@ -2952,7 +3002,7 @@ int main(int argc, char ** argv) {
             { "chat_template",               curr_tmpl.c_str() }
         };
 
-        res.set_content(data.dump(), MIMETYPE_JSON);
+        res.set_content(data.dump(), "application/json; charset=utf-8");
     };
 
     const auto handle_completions = [&ctx_server, &res_error](const httplib::Request & req, httplib::Response & res) {
@@ -2961,6 +3011,8 @@ 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();
@@ -2971,7 +3023,7 @@ 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), MIMETYPE_JSON);
+                res.set_content(result.data.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8");
             } else {
                 res_error(res, result.data);
             }
@@ -3034,7 +3086,9 @@ int main(int argc, char ** argv) {
         }
     };
 
-    const auto handle_models = [&params, &ctx_server](const httplib::Request &, httplib::Response & res) {
+    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"));
+
         json models = {
             {"object", "list"},
             {"data", {
@@ -3043,12 +3097,12 @@ int main(int argc, char ** argv) {
                      {"object",   "model"},
                      {"created",  std::time(0)},
                      {"owned_by", "llamacpp"},
-                     {"meta",     ctx_server.model_meta()}
+                     {"meta",     model_meta}
                  },
              }}
         };
 
-        res.set_content(models.dump(), MIMETYPE_JSON);
+        res.set_content(models.dump(), "application/json; charset=utf-8");
     };
 
     const auto handle_chat_completions = [&ctx_server, &params, &res_error](const httplib::Request & req, httplib::Response & res) {
@@ -3056,6 +3110,8 @@ 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();
@@ -3070,7 +3126,7 @@ 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), MIMETYPE_JSON);
+                res.set_content(result_oai.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8");
             } else {
                 res_error(res, result.data);
             }
@@ -3132,6 +3188,8 @@ 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();
@@ -3142,7 +3200,7 @@ 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), MIMETYPE_JSON);
+                res.set_content(result.data.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8");
             } else {
                 res_error(res, result.data);
             }
@@ -3190,6 +3248,7 @@ 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;
@@ -3198,10 +3257,11 @@ 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(), MIMETYPE_JSON);
+        return res.set_content(data.dump(), "application/json; charset=utf-8");
     };
 
     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;
@@ -3211,10 +3271,12 @@ int main(int argc, char ** argv) {
         }
 
         const json data = format_detokenized_response(content);
-        return res.set_content(data.dump(), MIMETYPE_JSON);
+        return res.set_content(data.dump(), "application/json; charset=utf-8");
     };
 
     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;
 
@@ -3260,53 +3322,7 @@ int main(int argc, char ** argv) {
         json root = is_openai
             ? format_embeddings_response_oaicompat(body, responses)
             : responses[0];
-        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
+        return res.set_content(root.dump(), "application/json; charset=utf-8");
     };
 
     auto handle_static_file = [](unsigned char * content, size_t len, const char * mime_type) {
@@ -3347,6 +3363,7 @@ 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);
@@ -3361,11 +3378,6 @@ 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);
@@ -3381,76 +3393,36 @@ 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); };
 
-    // clean up function, to be called before exit
-    auto clean_up = [&svr]() {
-        svr->stop();
-        llama_backend_free();
+    LOG_INFO("HTTP server listening", log_data);
+
+    // 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;
+        }
+
+        return 0;
+    });
+
+    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();
     };
 
-    // 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";
-            }
-        }
-
-        // 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
-        ));
-
-        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;
     sigint_action.sa_handler = signal_handler;
@@ -3465,8 +3437,12 @@ int main(int argc, char ** argv) {
     SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
 
-    clean_up();
+    ctx_server.queue_tasks.start_loop();
+
+    svr->stop();
     t.join();
 
+    llama_backend_free();
+
     return 0;
 }

examples/server/tests/features/lora.feature

@@ -1,36 +0,0 @@
-@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

@@ -7,7 +7,6 @@ import subprocess
 import sys
 import threading
 import time
-import requests
 from collections.abc import Sequence
 from contextlib import closing
 from re import RegexFlag
@@ -71,7 +70,6 @@ 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 = []
@@ -84,12 +82,6 @@ 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):
@@ -205,20 +197,27 @@ 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_slots_status(context, context.base_url, 200,
-                                        timeout=30)
+            await wait_for_health_status(context, context.base_url, 200, 'ok',
+                                         timeout=30)
 
         case 'ready' | 'idle':
-            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)
+            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)])
         case 'busy':
-            await wait_for_slots_status(context, context.base_url, 503,
-                                        params={'fail_on_no_slot': 1},
-                                        slots_idle=0,
-                                        slots_processing=context.n_slots)
+            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)])
         case _:
             assert False, "unknown status"
 
@@ -850,17 +849,6 @@ 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
@@ -1180,15 +1168,17 @@ async def gather_tasks_results(context):
     return n_completions
 
 
-async def wait_for_slots_status(context,
-                                base_url,
-                                expected_http_status_code,
-                                timeout=3,
-                                params=None,
-                                slots_idle=None,
-                                slots_processing=None):
+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):
     if context.debug:
-        print(f"Starting checking for health for expected_http_status_code={expected_http_status_code}")
+        print(f"Starting checking for health for expected_health_status={expected_health_status}")
     interval = 0.5
     counter = 0
     if 'GITHUB_ACTIONS' in os.environ:
@@ -1196,19 +1186,26 @@ async def wait_for_slots_status(context,
 
     async with aiohttp.ClientSession() as session:
         while True:
-            async with await session.get(f'{base_url}/slots', params=params) as slots_response:
-                status_code = slots_response.status
-                slots = await slots_response.json()
+            async with await session.get(f'{base_url}/health', params=params) as health_response:
+                status_code = health_response.status
+                health = await health_response.json()
                 if context.debug:
-                    print(f"slots responses {slots}\n")
-                if status_code == 503 and status_code == expected_http_status_code:
+                    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)
                     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
@@ -1222,7 +1219,7 @@ async def wait_for_slots_status(context,
                         if n_completions > 0:
                             return
 
-                assert False, f'slots check timeout exceeded {counter}s>={timeout}'
+                assert False, f'{expected_health_status} timeout exceeded {counter}s>={timeout}'
 
 
 def assert_embeddings(embeddings):
@@ -1329,8 +1326,6 @@ 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

@@ -4,4 +4,3 @@ 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

@@ -355,6 +355,24 @@ 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
-./llama-simple -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is"
+./simple -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is"
 
 ...
 

examples/speculative/speculative.cpp

@@ -66,9 +66,7 @@ int main(int argc, char ** argv) {
     llama_context * ctx_dft = NULL;
 
     // load the target model
-    llama_init_result llama_init_tgt = llama_init_from_gpt_params(params);
-    model_tgt = llama_init_tgt.model;
-    ctx_tgt = llama_init_tgt.context;
+    std::tie(model_tgt, ctx_tgt) = llama_init_from_gpt_params(params);
 
     // load the draft model
     params.model = params.model_draft;
@@ -77,9 +75,7 @@ int main(int argc, char ** argv) {
         params.n_threads = params.n_threads_draft;
     }
     params.n_threads_batch = params.n_threads_batch_draft;
-    llama_init_result llama_init_dft = llama_init_from_gpt_params(params);
-    model_dft = llama_init_dft.model;
-    ctx_dft = llama_init_dft.context;
+    std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params);
 
     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 the specified target *(using GGML_SYCL_TARGET)*.
+1. Build the llama.cpp for SYCL for all targets.
 
-2. Enable oneAPI running environment *(if GGML_SYCL_TARGET is set to INTEL -default-)*
+2. Enable oneAPI running environment
 
 ```
 source /opt/intel/oneapi/setvars.sh
@@ -29,13 +29,19 @@ source /opt/intel/oneapi/setvars.sh
 Check the ID in startup log, like:
 
 ```
-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|
+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
 
 ```
 
+|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\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0
+.\build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0

examples/tokenize/tokenize.cpp

@@ -362,7 +362,7 @@ int main(int raw_argc, char ** raw_argv) {
         prompt = stdin_buffer.str();
     }
 
-    const bool model_wants_add_bos = llama_add_bos_token(model);
+    const bool model_wants_add_bos = llama_should_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": 1722555600,
-        "narHash": "sha256-XOQkdLafnb/p9ij77byFQjDf5m5QYl9b2REiVClC+x4=",
+        "lastModified": 1719994518,
+        "narHash": "sha256-pQMhCCHyQGRzdfAkdJ4cIWiw+JNuWsTX7f0ZYSyz0VY=",
         "owner": "hercules-ci",
         "repo": "flake-parts",
-        "rev": "8471fe90ad337a8074e957b69ca4d0089218391d",
+        "rev": "9227223f6d922fee3c7b190b2cc238a99527bbb7",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1723637854,
-        "narHash": "sha256-med8+5DSWa2UnOqtdICndjDAEjxr5D7zaIiK4pn0Q7c=",
+        "lastModified": 1721379653,
+        "narHash": "sha256-8MUgifkJ7lkZs3u99UDZMB4kbOxvMEXQZ31FO3SopZ0=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "c3aa7b8938b17aebd2deecf7be0636000d62a2b9",
+        "rev": "1d9c2c9b3e71b9ee663d11c5d298727dace8d374",
         "type": "github"
       },
       "original": {
@@ -36,14 +36,14 @@
     },
     "nixpkgs-lib": {
       "locked": {
-        "lastModified": 1722555339,
-        "narHash": "sha256-uFf2QeW7eAHlYXuDktm9c25OxOyCoUOQmh5SZ9amE5Q=",
+        "lastModified": 1719876945,
+        "narHash": "sha256-Fm2rDDs86sHy0/1jxTOKB1118Q0O3Uc7EC0iXvXKpbI=",
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/5daf0514482af3f97abaefc78a6606365c9108e2.tar.gz"
       },
       "original": {
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/5daf0514482af3f97abaefc78a6606365c9108e2.tar.gz"
       }
     },
     "root": {

ggml/CMakeLists.txt

@@ -129,13 +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)
@@ -207,7 +207,6 @@ 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-metal.h

@@ -50,8 +50,6 @@ 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

@@ -244,8 +244,6 @@
 #define GGML_EXIT_SUCCESS 0
 #define GGML_EXIT_ABORTED 1
 
-#define GGML_ROPE_TYPE_NEOX 2
-
 #define GGUF_MAGIC "GGUF"
 
 #define GGUF_VERSION 3
@@ -351,7 +349,6 @@ 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;
@@ -1142,17 +1139,16 @@ 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,
-            float                 eps);
+            int                   n_groups);
 
     GGML_API struct ggml_tensor * ggml_group_norm_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_groups,
-            float                 eps);
+            int                   n_groups);
 
     // a - x
     // b - dy
@@ -1455,10 +1451,11 @@ extern "C" {
             struct ggml_tensor  * b);
 
     // rotary position embedding
-    // if (mode & 1) - skip n_past elements (NOT SUPPORTED)
-    // if (mode & GGML_ROPE_TYPE_NEOX) - GPT-NeoX style
+    // if mode & 1 == 1, skip n_past elements (NOT SUPPORTED)
+    // if mode & 2 == 1, 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,
@@ -1475,7 +1472,6 @@ 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,
@@ -1760,8 +1756,7 @@ extern "C" {
             struct ggml_tensor  * v,
             struct ggml_tensor  * mask,
             float                 scale,
-            float                 max_bias,
-            float                 logit_softcap);
+            float                 max_bias);
 
     GGML_API void ggml_flash_attn_ext_set_prec(
             struct ggml_tensor * a,
@@ -1778,8 +1773,10 @@ extern "C" {
 
     GGML_API struct ggml_tensor * ggml_ssm_conv(
             struct ggml_context * ctx,
-            struct ggml_tensor  * sx,
-            struct ggml_tensor  * c);
+            struct ggml_tensor  * s,
+            struct ggml_tensor  * x,
+            struct ggml_tensor  * c,
+            struct ggml_tensor  * sq);
 
     GGML_API struct ggml_tensor * ggml_ssm_scan(
             struct ggml_context * ctx,
@@ -1788,7 +1785,8 @@ extern "C" {
             struct ggml_tensor  * dt,
             struct ggml_tensor  * A,
             struct ggml_tensor  * B,
-            struct ggml_tensor  * C);
+            struct ggml_tensor  * C,
+            struct ggml_tensor  * sq);
 
     // partition into non-overlapping windows with padding if needed
     // example:

ggml/src/CMakeLists.txt

@@ -549,13 +549,6 @@ 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)
@@ -568,9 +561,6 @@ 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)
@@ -612,10 +602,6 @@ 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()
@@ -863,6 +849,11 @@ 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
@@ -881,7 +872,6 @@ 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()

ggml/src/ggml-aarch64.c

@@ -16,8 +16,6 @@
 
 #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
@@ -337,18 +335,33 @@ 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) {
-    UNUSED(quant_weights);
-    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4);
+    if (!quant_weights) {
+        return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4);
+    }
+    else {
+        assert(false);
+        return 0;
+    }
 }
 
 size_t quantize_q4_0_4x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    UNUSED(quant_weights);
-    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8);
+    if (!quant_weights) {
+        return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8);
+    }
+    else {
+        assert(false);
+        return 0;
+    }
 }
 
 size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    UNUSED(quant_weights);
-    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
+    if (!quant_weights) {
+        return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
+    }
+    else {
+        assert(false);
+        return 0;
+    }
 }
 
 void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
@@ -371,8 +384,8 @@ 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 (ggml_sve_cnt_b == QK8_0) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+    if (svcntw() == 8) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -483,8 +496,8 @@ 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 (ggml_sve_cnt_b == QK8_0) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+    if (svcntw() == 8) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -601,7 +614,7 @@ 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 (ggml_sve_cnt_b == QK8_0) {
+    if (svcntw() == 8) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;
@@ -667,12 +680,12 @@ 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() && (ggml_sve_cnt_b == QK8_0)) &&
+        GGML_ASSERT((ggml_cpu_has_sve() && (svcntw() == 8)) &&
                     "__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() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
+        GGML_ASSERT(((ggml_cpu_has_sve() && (svcntw() == 8)) || 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");
     }
@@ -732,8 +745,8 @@ 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 (ggml_sve_cnt_b == QK8_0) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+    if (svcntw() == 8) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -1253,8 +1266,8 @@ 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 (ggml_sve_cnt_b == QK8_0) {
-        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+    if (svcntw() == 8) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
                     "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
     }
 #endif
@@ -1715,7 +1728,7 @@ 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 (ggml_sve_cnt_b == QK8_0) {
+    if (svcntw() == 8) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;
@@ -2126,12 +2139,12 @@ 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() && (ggml_sve_cnt_b == QK8_0)) &&
+        GGML_ASSERT((ggml_cpu_has_sve() && (svcntw() == 8)) &&
                     "__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() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
+        GGML_ASSERT(((ggml_cpu_has_sve() && (svcntw() == 8)) || 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-backend.c

@@ -351,10 +351,15 @@ 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
-    // 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);
+    // 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);
+    }
 }
 
 // events
@@ -1018,6 +1023,10 @@ 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
@@ -1121,8 +1130,7 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, co
 }
 
 #if 0
-#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
+static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS*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
@@ -1546,6 +1554,7 @@ 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;
@@ -1773,17 +1782,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 } else {
                     ggml_backend_synchronize(split_backend);
                 }
-                // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
-                // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
-                if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
-                    ggml_backend_synchronize(input_backend);
-                    if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
-                        ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
-                    } else {
-                        ggml_backend_synchronize(split_backend);
-                    }
-                    ggml_backend_tensor_copy(input, input_cpy);
-                }
+                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
             }
         }
 
@@ -1861,14 +1860,13 @@ ggml_backend_sched_t ggml_backend_sched_new(
     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 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;
+    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->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_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;

ggml/src/ggml-cann.cpp

@@ -627,6 +627,7 @@ 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;
@@ -678,6 +679,7 @@ 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;
@@ -896,10 +898,11 @@ 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_backend_cann_buffer_context *ctx =
-        (ggml_backend_cann_buffer_context *)buffer->context;
+    // 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);
     // TODO: refer to cann(#6017), it use thread's default stream.
@@ -907,21 +910,22 @@ GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
     // Why aclrtSynchronizeDevice?
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
-                              ACL_MEMCPY_HOST_TO_DEVICE));
+        ACL_CHECK(aclrtMemcpy(tensor->data, size, (const char*)data + offset,
+                              size, ACL_MEMCPY_HOST_TO_DEVICE));
     } else {
-        void *transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, data, transform_buffer);
+        void* transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, (const char*)data + offset,
+                                    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(data, check_buffer, size) == 0);
+        GGML_ASSERT(memcmp((const char*)data + offset, check_buffer, size) ==
+                    0);
         free(check_buffer);
 #endif
-        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size,
-                              transform_buffer, size,
+        ACL_CHECK(aclrtMemcpy(tensor->data, size, transform_buffer, size,
                               ACL_MEMCPY_HOST_TO_DEVICE));
         free(transform_buffer);
     }
@@ -943,20 +947,21 @@ 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(data, size, (char*)tensor->data + offset, size,
+        ACL_CHECK(aclrtMemcpy((char*)data + offset, size, tensor->data, size,
                               ACL_MEMCPY_DEVICE_TO_HOST));
     } else {
         void* transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpy(transform_buffer, size,
-                              (char*)tensor->data + offset, size,
+        ACL_CHECK(aclrtMemcpy(transform_buffer, size, tensor->data, size,
                               ACL_MEMCPY_DEVICE_TO_HOST));
-        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
+        ggml_backend_cann_transform_back(tensor, transform_buffer,
+                                         (char*)data + offset);
         free(transform_buffer);
     }
 }
@@ -1445,41 +1450,42 @@ 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((char *)tensor->data + offset, size, data,
-                                   size, ACL_MEMCPY_HOST_TO_DEVICE,
-                                   cann_ctx->stream()));
+        ACL_CHECK(aclrtMemcpyAsync(
+            tensor->data, size, (const char*)data + offset, size,
+            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
     } else {
-        void *transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, data, transform_buffer);
+        void* transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, (const char*)data + offset,
+                                    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(data, check_buffer, size));
+        GGML_ASSERT(memcmp((const char*)data + offset, check_buffer, size));
         free(check_buffer);
 #endif
-        ACL_CHECK(aclrtMemcpyAsync(
-            (char *)tensor->data + offset, size, transform_buffer, size,
-            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
+        ACL_CHECK(aclrtMemcpyAsync(tensor->data, 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;
 
@@ -1487,16 +1493,17 @@ GGML_CALL static void ggml_backend_cann_get_tensor_async(
                 "unsupported buffer type");
 
     if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync(data, size, (char *)tensor->data + offset,
+        ACL_CHECK(aclrtMemcpyAsync((char*)data + offset, size, tensor->data,
                                    size, ACL_MEMCPY_DEVICE_TO_HOST,
                                    cann_ctx->stream()));
     } else {
-        void *transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpyAsync(
-            transform_buffer, size, (char *)tensor->data + offset, size,
-            ACL_MEMCPY_DEVICE_TO_HOST, cann_ctx->stream()));
+        void* transform_buffer = malloc(size);
+        ACL_CHECK(aclrtMemcpyAsync(transform_buffer, size, tensor->data, size,
+                                   ACL_MEMCPY_DEVICE_TO_HOST,
+                                   cann_ctx->stream()));
         ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
+        ggml_backend_cann_transform_back(tensor, transform_buffer,
+                                         (char*)data + offset);
         free(transform_buffer);
     }
 }
@@ -1659,13 +1666,10 @@ 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;
@@ -1690,7 +1694,6 @@ 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;

ggml/src/ggml-cann/acl_tensor.cpp

@@ -37,10 +37,6 @@ 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;
     }
@@ -93,6 +89,33 @@ 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

@@ -23,9 +23,6 @@
 #ifndef CANN_ACL_TENSOR_H
 #define CANN_ACL_TENSOR_H
 
-#include <algorithm>
-#include <cstring>
-
 #include <aclnn/aclnn_base.h>
 #include "common.h"
 
@@ -68,8 +65,7 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = null
                              size_t offset = 0);
 
 /**
- * @brief   Template for creating an ACL tensor from provided parameters. typename TYPE
- *          should be size_t or float.
+ * @brief   Creates an ACL tensor from provided parameters.
  *
  * @details This function creates an ACL tensor using the provided data pointer,
  *          data type, dimensions, strides, format, offset, and additional parameters.
@@ -87,34 +83,10 @@ 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,
-                                   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;
-}
+                             size_t type_size, int64_t* ne, size_t* nb,
+                             int64_t dims, aclFormat format = ACL_FORMAT_ND,
+                             size_t offset = 0);
 
 /**
  * @brief   Checks if tensors require broadcasting based on their shapes.

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -464,11 +464,9 @@ 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;
@@ -912,13 +910,6 @@ 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);
@@ -980,13 +971,6 @@ 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);
@@ -1328,111 +1312,6 @@ 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
@@ -1441,23 +1320,21 @@ 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;
 
-    // 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 N = is_2D ? ne13 : ne12;
+    const int64_t IC = is_2D ? ne12 : ne11;
 
-    const int64_t N = ne13;
-    const int64_t IC = ne12;
-    const int64_t KH = ne01;
+    const int64_t KH = is_2D ? ne01 : 1;
     const int64_t KW = ne00;
-    const int64_t IW = ne10;
 
     const int64_t OH = is_2D ? ne2 : 1;
     const int64_t OW = ne1;
@@ -1465,12 +1342,9 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
     GGML_ASSERT(nb10 == sizeof(float));
 
-    // 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]
+    // im2col: [N,C,H,W] -> [N, IC * KH * KW, OW * OH]
     aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
-    int64_t tmp_im2col_ne[] = {OW * OH * n_bytes_factor, IC * KH * KW, N};
+    int64_t tmp_im2col_ne[] = {OW * OH, IC * KH * KW, N};
     size_t tmp_im2col_nb[GGML_MAX_DIMS - 1];
 
     tmp_im2col_nb[0] = ggml_type_size(src1->type);
@@ -1482,10 +1356,8 @@ 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) * n_bytes_factor);
+        ctx.pool(), ggml_nelements(dst) * ggml_element_size(src1));
     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,
@@ -1508,9 +1380,8 @@ 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) {
-        workspace_allocator.alloc(workspaceSize);
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
         workspaceAddr = workspace_allocator.get();
     }
 
@@ -1520,10 +1391,9 @@ 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) * n_bytes_factor);
-        tmp_cast_buffer = tmp_cast_allocator.get();
+        tmp_cast_allocator.alloc(ggml_nbytes(dst));
+        void* 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++) {
@@ -1538,21 +1408,24 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                    ggml_cann_type_mapping(dst->type));
     }
 
-    // post-processing
-    if (is_2D) {
-        ggml_cann_im2col_2d_post_process(ctx, dst, src1, tmp_cast_tensor,
-                                         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 {
-        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);
+        aclnn_permute(ctx, tmp_im2col_tensor, acl_dst, permute_dim, 3);
     }
 
     // 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));
@@ -2479,33 +2352,21 @@ 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_quant(ggml_backend_cann_context& ctx,
-                                   ggml_tensor* dst,
-                                   const enum ggml_type type) {
+static void ggml_cann_mul_mat_q8_0(ggml_backend_cann_context& ctx,
+                                   ggml_tensor* dst) {
     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]};
-    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_t weight_elem_size = sizeof(uint8_t);
+    size_t 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,
@@ -2520,10 +2381,10 @@ static void ggml_cann_mul_mat_quant(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);
-        input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
+        ggml_cann_pool_alloc input_alloctor(
+            ctx.pool(), ggml_nelements(src1) * input_elem_size);
         input_buffer = input_alloctor.get();
 
         int64_t* input_cast_ne = src1->ne;
@@ -2569,9 +2430,8 @@ static void ggml_cann_mul_mat_quant(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,
-                ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
-                weight_nb, 2);
+                (char*)src0->data + batch0 * weight_stride, ACL_INT8,
+                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);
@@ -2625,9 +2485,11 @@ 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:
+        // case GGML_TYPE_Q4_0:
+        //     ggml_cann_mul_mat_q4_0(ctx, dst);
+        //     break;
         case GGML_TYPE_Q8_0:
-            ggml_cann_mul_mat_quant(ctx, dst, type);
+            ggml_cann_mul_mat_q8_0(ctx, dst);
             break;
         default:
             GGML_ABORT("fatal error");
@@ -2881,7 +2743,7 @@ 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 & GGML_ROPE_TYPE_NEOX;
+    const bool is_neox = mode & 2;
 
     // init cos/sin cache
     ggml_cann_pool_alloc sin_allocator(

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

@@ -9,7 +9,6 @@ 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
 )
 
@@ -30,4 +29,4 @@ ascendc_library(ascendc_kernels STATIC
     ${SRC_FILES}
 )
 
-# ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)
+#ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)

ggml/src/ggml-cann/kernels/ascendc_kernels.h

@@ -8,8 +8,6 @@
 
 #include "aclrtlaunch_ascendc_quantize_f32_q8_0.h"
 #include "aclrtlaunch_ascendc_quantize_f16_q8_0.h"
-#include "aclrtlaunch_ascendc_quantize_f16_to_q4_0.h"
-#include "aclrtlaunch_ascendc_quantize_f32_to_q4_0.h"
 
 #include "aclrtlaunch_ascendc_dup_by_rows_fp16.h"
 #include "aclrtlaunch_ascendc_dup_by_rows_fp32.h"

ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp

@@ -1,278 +0,0 @@
-#include "kernel_operator.h"
-
-using namespace AscendC;
-
-#define BUFFER_NUM 2
-#define Group_Size 32
-
-template <typename SRC_T>
-class QUANTIZE_FLOAT_TO_Q4_0 {
-   public:
-    __aicore__ inline QUANTIZE_FLOAT_TO_Q4_0() {}
-    __aicore__ inline void init(GM_ADDR input, GM_ADDR output,
-                                int64_t *input_ne_ub, size_t *input_nb_ub,
-                                int64_t *output_ne_ub) {
-        // TODO: fix test_case CPY(type_src=f16,type_dst=q4_0,ne=[256,4,4,4],
-        //                         permute=[0,0,0,0]):
-        // [CPY] NMSE = 0.000008343 > 0.000001000 FAIL
-        int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
-
-        // input stride of data elements
-        for (int i = 0; i < 4; i++) {
-            input_ne[i] = input_ne_ub[i];
-            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
-            output_ne[i] = output_ne_ub[i];
-        }
-
-        // output stride of data elements
-        output_stride[0] = 1;
-        for (int i = 1; i < 4; i++) {
-            output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
-        }
-
-        // scale saved one by one after data:. [group1_scale, group2_scale, ...]
-        scale_ne = input_ne;
-        scale_stride[0] = 1;
-        scale_stride[1] = input_ne[0] / Group_Size;
-        for (int i = 2; i < 4; i++) {
-            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
-        }
-
-        // split input tensor by rows.
-        uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
-        dr = nr / op_block_num;
-
-        uint64_t tails = nr % op_block_num;
-        if (op_block_idx < tails) {
-            dr += 1;
-            ir = dr * op_block_idx;
-        } else {
-            ir = dr * op_block_idx + tails;
-        }
-
-        group_size_in_row = scale_stride[1];
-        int64_t scale_offset = output_ne[0] * output_ne[1] * output_ne[2] *
-                              output_ne[3] * sizeof(uint8_t) / 2;
-
-        input_gm.SetGlobalBuffer((__gm__ SRC_T *)input);
-        output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
-        scale_gm.SetGlobalBuffer((__gm__ half *)(output + scale_offset + ir *
-                                                 group_size_in_row *
-                                                 sizeof(half)));
-
-        pipe.InitBuffer(input_queue, BUFFER_NUM, Group_Size * sizeof(SRC_T));
-        pipe.InitBuffer(output_queue, BUFFER_NUM,
-                            Group_Size * sizeof(int8_t) / 2);
-        pipe.InitBuffer(cast_queue , 1, Group_Size * sizeof(float));
-        pipe.InitBuffer(work_queue, 1, Group_Size * sizeof(float));
-        pipe.InitBuffer(max_queue, 1, Group_Size * sizeof(float));
-        pipe.InitBuffer(min_queue, 1, Group_Size * sizeof(float));
-        pipe.InitBuffer(scale_queue, 1, Group_Size / 2 * sizeof(half));
-        pipe.InitBuffer(int8_queue, 1, Group_Size * sizeof(int8_t));
-        pipe.InitBuffer(half_queue, 1, Group_Size * sizeof(half));
-    }
-
-    __aicore__ inline void copy_in(uint32_t offset) {
-        LocalTensor<SRC_T> input_local = input_queue.AllocTensor<SRC_T>();
-        DataCopy(input_local, input_gm[offset], Group_Size);
-        input_queue.EnQue(input_local);
-    }
-
-    __aicore__ inline void copy_out(uint32_t offset) {
-        // reinterpretcast Group_Size(32) * int4b_t to Group_Size / 2 * int8_t,
-        // and using DataCopyPad to avoid 32 bits align.
-        LocalTensor<int4b_t> output_local = output_queue.DeQue<int4b_t>();
-        LocalTensor<int8_t> output_int8_local =
-                                    output_local.ReinterpretCast<int8_t>();
-
-        DataCopyExtParams dataCopyParams;
-        dataCopyParams.blockCount = 1;
-        dataCopyParams.blockLen = Group_Size / 2  * sizeof(int8_t);
-        DataCopyPad(output_gm[offset], output_int8_local, dataCopyParams);
-
-        output_queue.FreeTensor(output_local);
-    }
-
-    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
-                                         LocalTensor<float> input_local) {
-        DataCopy(cast_local, input_local, Group_Size);
-    }
-
-    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
-                                         LocalTensor<half> input_local) {
-        Cast(cast_local, input_local, RoundMode::CAST_NONE, Group_Size);
-    }
-
-    __aicore__ inline half calculate_group(int64_t row, int64_t group) {
-        const int64_t i3 = row / (input_ne[1] * input_ne[2]);
-        const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
-        const int64_t i1 =
-            row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
-
-        const int64_t input_offset = i1 * input_stride[1] +
-                                     i2 * input_stride[2] +
-                                     i3 * input_stride[3] + Group_Size * group;
-
-        // output_offset is stride for output_gm which datatype is int8_t and
-        // divided by 2 is needed for int4b_t.
-        const int64_t output_offset = (i1 * output_stride[1] +
-                                       i2 * output_stride[2] +
-                                       i3 * output_stride[3] +
-                                       Group_Size * group) / 2;
-        copy_in(input_offset);
-
-        LocalTensor<SRC_T> input_local = input_queue.DeQue<SRC_T>();
-        LocalTensor<int4b_t> output_local = output_queue.AllocTensor<int4b_t>();
-        LocalTensor<float> cast_local = cast_queue.AllocTensor<float>();
-        LocalTensor<float> work_local = work_queue.AllocTensor<float>();
-        LocalTensor<float> max_local = max_queue.AllocTensor<float>();
-        LocalTensor<float> min_local = min_queue.AllocTensor<float>();
-        LocalTensor<int8_t> int8_local = int8_queue.AllocTensor<int8_t>();
-        LocalTensor<half> half_local = half_queue.AllocTensor<half>();
-
-        input_to_cast(cast_local, input_local);
-
-        ReduceMax(max_local, cast_local, work_local, Group_Size);
-        ReduceMin(min_local, cast_local, work_local, Group_Size);
-        const float max_value = max_local.GetValue(0);
-        const float min_value = min_local.GetValue(0);
-        float d = max_value;
-        if (min_value < 0 && (-1 * min_value) > max_value) {
-            d = min_value;
-        }
-
-        d = d / (-8);
-        if (d != 0) {
-            Muls(cast_local, cast_local, 1.0f / d, Group_Size);
-        }
-
-        // range: [-8,8] -> [0.5,16.5] -> [0,16] -> [0,15] -> [-8,7]
-        float scalar = 8.5f;
-        Adds(cast_local, cast_local, scalar, Group_Size);
-        Cast(cast_local, cast_local, RoundMode::CAST_FLOOR, Group_Size);
-        scalar = 15.0f;
-        Mins(cast_local, cast_local, scalar, Group_Size);
-        scalar = -8.0f;
-        Adds(cast_local, cast_local, scalar, Group_Size);
-
-        // float->half->int4b
-        Cast(half_local, cast_local, RoundMode::CAST_NONE, Group_Size);
-        Cast(output_local, half_local, RoundMode::CAST_NONE, Group_Size);
-
-        output_queue.EnQue(output_local);
-        copy_out(output_offset);
-
-        input_queue.FreeTensor(input_local);
-        work_queue.FreeTensor(work_local);
-        max_queue.FreeTensor(max_local);
-        min_queue.FreeTensor(min_local);
-        int8_queue.FreeTensor(int8_local);
-        half_queue.FreeTensor(half_local);
-        cast_queue.FreeTensor(cast_local);
-        return (half)d;
-    }
-
-    __aicore__ inline void calculate() {
-        LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
-        uint32_t scale_local_offset = 0;
-        uint32_t scale_global_offset = 0;
-        for (int64_t i = ir; i < ir + dr; i++) {
-            for (int64_t j = 0; j < group_size_in_row; j++) {
-                half scale = calculate_group(i, j);
-                scale_local.SetValue(scale_local_offset++, scale);
-                // Copy Group_Size/2 length data each time.
-                if (scale_local_offset == Group_Size / 2) {
-                    scale_local_offset = 0;
-                    // TODO: OPTIMIZE ME
-                    pipe_barrier(PIPE_ALL);
-                    DataCopy(scale_gm[scale_global_offset], scale_local,
-                                      Group_Size / 2);
-                    pipe_barrier(PIPE_ALL);
-                    scale_global_offset += Group_Size / 2;
-                }
-            }
-        }
-
-        if (scale_local_offset != 0) {
-            pipe_barrier(PIPE_ALL);
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = scale_local_offset * sizeof(half);
-            DataCopyPad(scale_gm[scale_global_offset], scale_local,
-                        dataCopyParams);
-            pipe_barrier(PIPE_ALL);
-        }
-        scale_queue.FreeTensor(scale_local);
-    }
-
-   private:
-    int64_t input_ne[4];
-    size_t input_stride[4];
-
-    int64_t *scale_ne;
-    size_t scale_stride[4];
-
-    int64_t output_ne[4];
-    size_t output_stride[4];
-
-    int64_t group_size_in_row;
-
-    int64_t ir;
-    int64_t dr;
-
-    TPipe pipe;
-    GlobalTensor<SRC_T> input_gm;
-    GlobalTensor<half> scale_gm;
-    GlobalTensor<int8_t> output_gm;
-    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
-    TQue<QuePosition::VECIN, BUFFER_NUM> work_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> max_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> min_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> scale_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> cast_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> int8_queue;
-    TQue<QuePosition::VECOUT, BUFFER_NUM> half_queue;
-};
-
-template <typename T>
-__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
-    auto gm_ptr = (__gm__ uint8_t *)gm;
-    auto ub_ptr = (uint8_t *)(ub);
-    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
-        *ub_ptr = *gm_ptr;
-    }
-}
-
-extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0(
-    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-
-    QUANTIZE_FLOAT_TO_Q4_0<half> op;
-    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
-    op.calculate();
-}
-
-extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
-    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
-    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
-    int64_t input_ne_ub[4];
-    size_t input_nb_ub[4];
-    int64_t output_ne_ub[4];
-
-    copy_to_ub(input_ne_gm, input_ne_ub, 32);
-    copy_to_ub(input_nb_gm, input_nb_ub, 32);
-    copy_to_ub(output_ne_gm, output_ne_ub, 32);
-
-    QUANTIZE_FLOAT_TO_Q4_0<float> op;
-    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
-    op.calculate();
-}

ggml/src/ggml-cuda.cu

@@ -130,22 +130,7 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
     }
     return res;
 #else
-
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
-    cudaError_t err;
-    if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
-    {
-        err = cudaMallocManaged(ptr, size);
-    }
-    else
-    {
-        err = cudaMalloc(ptr, size);
-    }
-    return err;
-#else
     return cudaMalloc(ptr, size);
-#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
-
 #endif
 }
 
@@ -1501,7 +1486,7 @@ static void ggml_cuda_op_mul_mat(
         }
 
         // If src0 is on a temporary compute buffers (partial offloading) there may be some padding that needs to be cleared:
-        if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
+        if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
             const int64_t nbytes_data    = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
             const int64_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
             CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
@@ -1900,9 +1885,10 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buffer_is_cuda_split(src0->buffer);
 
-    bool use_dequantize_mul_mat_vec = ggml_cuda_dmmv_type_supported(src0->type)
+    bool use_dequantize_mul_mat_vec = (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src0->ne[0] % (GGML_CUDA_DMMV_X*2) == 0 && src1->ne[1] == 1;
+        && src0->ne[0] % GGML_CUDA_DMMV_X == 0 && src0->ne[0] >= GGML_CUDA_DMMV_X*2
+        && src1->ne[1] == 1;
     bool          use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
@@ -2358,35 +2344,33 @@ GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend,
 }
 
 GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
+    GGML_ASSERT(ggml_backend_is_cuda(backend_src) || ggml_backend_is_cuda(backend_dst));
+
     ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
     ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
 
-    if (!ggml_backend_is_cuda(backend_src) || !ggml_backend_is_cuda(backend_dst)) {
+    if (!ggml_backend_buffer_is_cuda(src->buffer)) {
         return false;
     }
 
-    if (!ggml_backend_buffer_is_cuda(src->buffer) || !ggml_backend_buffer_is_cuda(dst->buffer)) {
+    if (!ggml_backend_buffer_is_cuda(dst->buffer)) {
         return false;
     }
 
-    // device -> device copy
+    // device -> device
     ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *)backend_src->context;
     ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *)backend_dst->context;
 
-    ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
-    ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
-
-    if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
-#ifndef NDEBUG
-        GGML_CUDA_LOG_WARN("%s: backend and buffer devices do not match\n", __func__);
-#endif
-        return false;
-    }
-
     if (backend_src != backend_dst) {
+        ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
+        ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
+
+        GGML_ASSERT(cuda_ctx_src->device == buf_ctx_src->device);
+        GGML_ASSERT(cuda_ctx_dst->device == buf_ctx_dst->device);
+
         // copy on src stream
         if (cuda_ctx_src->device == cuda_ctx_dst->device) {
-            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
+            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream()));
         } else {
 #ifdef GGML_CUDA_NO_PEER_COPY
             return false;
@@ -2395,7 +2379,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_
 #endif
         }
 
-        // record event on src stream after the copy
+        // record event on src stream
         if (!cuda_ctx_src->copy_event) {
             ggml_cuda_set_device(cuda_ctx_src->device);
             CUDA_CHECK(cudaEventCreateWithFlags(&cuda_ctx_src->copy_event, cudaEventDisableTiming));
@@ -2407,7 +2391,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_
         CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx_dst->stream(), cuda_ctx_src->copy_event, 0));
     } else {
         // src and dst are on the same backend
-        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
+        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream()));
     }
     return true;
 }
@@ -2744,12 +2728,11 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_MUL_MAT_ID:
             {
                 struct ggml_tensor * a = op->src[0];
-                struct ggml_tensor * b = op->src[1];
-                if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16) {
-                    return false;
-                }
-                if (op->op == GGML_OP_MUL_MAT && a->ne[3] != b->ne[3]) {
-                    return false;
+                if (op->op == GGML_OP_MUL_MAT) {
+                    struct ggml_tensor * b = op->src[1];
+                    if (a->ne[3] != b->ne[3]) {
+                        return false;
+                    }
                 }
                 switch (a->type) {
                     case GGML_TYPE_F32:
@@ -2880,7 +2863,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
             return true;
         case GGML_OP_FLASH_ATTN_EXT:
 #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-            return (op->src[0]->ne[0] == 64 && op->src[1]->type == GGML_TYPE_F16) || op->src[0]->ne[0] == 128;
+            return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128;
 #else
             if (op->src[0]->ne[0] == 128) {
                 return true;

ggml/src/ggml-cuda/dmmv.cu

@@ -500,7 +500,7 @@ static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, cons
 }
 
 static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
     const dim3 block_nums(block_num_y, 1, 1);
@@ -510,7 +510,7 @@ static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y,
 }
 
 static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -519,7 +519,7 @@ static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y,
 }
 
 static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -528,7 +528,7 @@ static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y,
 }
 
 static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -537,7 +537,7 @@ static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y,
 }
 
 static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -588,7 +588,7 @@ static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, f
 }
 
 static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
@@ -672,12 +672,3 @@ void ggml_cuda_op_dequantize_mul_mat_vec(
     GGML_UNUSED(src1_ncols);
     GGML_UNUSED(src1_padded_row_size);
 }
-
-bool ggml_cuda_dmmv_type_supported(ggml_type src0_type) {
-    return src0_type == GGML_TYPE_Q4_0 || src0_type == GGML_TYPE_Q4_1 ||
-        src0_type == GGML_TYPE_Q5_0 || src0_type == GGML_TYPE_Q5_1 ||
-        src0_type == GGML_TYPE_Q8_0 || src0_type == GGML_TYPE_Q2_K ||
-        src0_type == GGML_TYPE_Q3_K || src0_type == GGML_TYPE_Q4_K ||
-        src0_type == GGML_TYPE_Q5_K || src0_type == GGML_TYPE_Q6_K ||
-        src0_type == GGML_TYPE_F16;
-}

ggml/src/ggml-cuda/dmmv.cuh

@@ -16,5 +16,3 @@ void ggml_cuda_op_dequantize_mul_mat_vec(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
     const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
     const int64_t src1_padded_row_size, cudaStream_t stream);
-
-bool ggml_cuda_dmmv_type_supported(ggml_type src0_type);

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

@@ -22,7 +22,6 @@ typedef void (* fattn_kernel_t)(
         const float m0,
         const float m1,
         const uint32_t n_head_log2,
-        const float logit_softcap,
         const int ne00,
         const int ne01,
         const int ne02,
@@ -658,17 +657,11 @@ void launch_fattn(
     const dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
     const int  shmem = 0;
 
-    float scale         = 1.0f;
-    float max_bias      = 0.0f;
-    float logit_softcap = 0.0f;
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
 
-    memcpy(&scale,         (float *) KQV->op_params + 0, sizeof(float));
-    memcpy(&max_bias,      (float *) KQV->op_params + 1, sizeof(float));
-    memcpy(&logit_softcap, (float *) KQV->op_params + 2, sizeof(float));
-
-    if (logit_softcap != 0.0f) {
-        scale /= logit_softcap;
-    }
+    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
 
     const uint32_t n_head      = Q->ne[2];
     const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
@@ -682,7 +675,7 @@ void launch_fattn(
         V_data,
         mask ? ((const char *) mask->data) : nullptr,
         (parallel_blocks) == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
-        scale, max_bias, m0, m1, n_head_log2, logit_softcap,
+        scale, max_bias, m0, m1, n_head_log2,
         Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
         K->ne[0], K->ne[1], K->ne[2], K->ne[3],
         mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,

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

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

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

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

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

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

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

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

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

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

ggml/src/ggml-cuda/fattn.cu

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

ggml/src/ggml-cuda/norm.cu

@@ -142,7 +142,8 @@ static void norm_f32_cuda(const float * x, float * dst, const int ncols, const i
     }
 }
 
-static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const float eps, const int group_size, const int ne_elements, cudaStream_t stream) {
+static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const int group_size, const int ne_elements, cudaStream_t stream) {
+    static const float eps = 1e-6f;
     if (group_size < 1024) {
         const dim3 block_dims(WARP_SIZE, 1, 1);
         group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
@@ -195,12 +196,8 @@ void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     int num_groups = dst->op_params[0];
-
-    float eps;
-    memcpy(&eps, dst->op_params + 1, sizeof(float));
-
     int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
-    group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], eps, group_size, ggml_nelements(src0), stream);
+    group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], group_size, ggml_nelements(src0), stream);
 }
 
 void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cuda/rope.cu

@@ -226,7 +226,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
 
-    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_neox = mode & 2;
 
     const int32_t * pos = (const int32_t *) src1_d;
 

ggml/src/ggml-impl.h

@@ -80,9 +80,8 @@ static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
 /**
  * Converts float32 to brain16.
  *
- * This is binary identical with Google Brain float conversion.
- * Floats shall round to nearest even, and NANs shall be quiet.
- * Subnormals aren't flushed to zero, except perhaps when used.
+ * This function is binary identical to AMD Zen4 VCVTNEPS2BF16.
+ * Subnormals shall be flushed to zero, and NANs will be quiet.
  * This code should vectorize nicely if using modern compilers.
  */
 static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
@@ -96,6 +95,10 @@ static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
         h.bits = (u.i >> 16) | 64; /* force to quiet */
         return h;
     }
+    if (!(u.i & 0x7f800000)) { /* subnormal */
+        h.bits = (u.i & 0x80000000) >> 16; /* flush to zero */
+        return h;
+    }
     h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
     return h;
 }
@@ -143,7 +146,6 @@ extern "C" {
 
 #if defined(__ARM_FEATURE_SVE)
 #include <arm_sve.h>
-#include <sys/prctl.h>
 #endif
 
 // 16-bit float

ggml/src/ggml-metal.m

@@ -82,8 +82,6 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_RMS_NORM,
     GGML_METAL_KERNEL_TYPE_GROUP_NORM,
     GGML_METAL_KERNEL_TYPE_NORM,
-    GGML_METAL_KERNEL_TYPE_SSM_CONV_F32,
-    GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,
@@ -212,7 +210,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_COUNT
 };
 
-struct ggml_backend_metal_context {
+struct ggml_metal_context {
     int n_cb;
 
     id<MTLDevice>       device;
@@ -226,10 +224,6 @@ struct ggml_backend_metal_context {
     bool support_simdgroup_mm;
 
     bool should_capture_next_compute;
-
-    // abort ggml_metal_graph_compute if callback returns true
-    ggml_abort_callback abort_callback;
-    void *              abort_callback_data;
 };
 
 // MSL code
@@ -295,7 +289,7 @@ static void * ggml_metal_host_malloc(size_t n) {
     return data;
 }
 
-static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) {
+static struct ggml_metal_context * ggml_metal_init(int n_cb) {
     GGML_METAL_LOG_INFO("%s: allocating\n", __func__);
 
 #if TARGET_OS_OSX && !GGML_METAL_NDEBUG
@@ -312,7 +306,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) {
     GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
 
     // Configure context
-    struct ggml_backend_metal_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_context));
+    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
     ctx->device = device;
     ctx->n_cb   = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
     ctx->queue  = [ctx->device newCommandQueue];
@@ -544,8 +538,6 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM,                      rms_norm,                       ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM,                    group_norm,                     ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM,                          norm,                           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_CONV_F32,                  ssm_conv_f32,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32,                  ssm_scan_f32,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,                mul_mv_f32_f32,                 ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,                mul_mv_f16_f16,                 ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,                mul_mv_f16_f32,                 ctx->support_simdgroup_reduction);
@@ -676,7 +668,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) {
     return ctx;
 }
 
-static void ggml_metal_free(struct ggml_backend_metal_context * ctx) {
+static void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_LOG_INFO("%s: deallocating\n", __func__);
 
     for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) {
@@ -742,7 +734,7 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs
     return nil;
 }
 
-static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx, const struct ggml_tensor * op) {
+static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const struct ggml_tensor * op) {
     for (size_t i = 0, n = 3; i < n; ++i) {
         if (op->src[i] != NULL && op->src[i]->type == GGML_TYPE_BF16) {
             return false;
@@ -807,9 +799,6 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx
                 return false;
             }
             return ctx->support_simdgroup_mm; // TODO: over-restricted for vec-kernels
-        case GGML_OP_SSM_CONV:
-        case GGML_OP_SSM_SCAN:
-            return true;
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             return ctx->support_simdgroup_reduction &&
@@ -856,7 +845,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx
 }
 
 static enum ggml_status ggml_metal_graph_compute(
-        struct ggml_backend_metal_context * ctx,
+        struct ggml_metal_context * ctx,
                struct ggml_cgraph * gf) {
 
     @autoreleasepool {
@@ -889,11 +878,8 @@ static enum ggml_status ggml_metal_graph_compute(
         id<MTLCommandBuffer> command_buffer  = [ctx->queue commandBufferWithUnretainedReferences];
         command_buffer_builder[cb_idx] = command_buffer;
 
-        // always enqueue the first two command buffers
-        // enqueue all of the command buffers if we don't need to abort
-        if (cb_idx < 2 || ctx->abort_callback == NULL) {
-            [command_buffer enqueue];
-        }
+        // enqueue the command buffers in order to specify their execution order
+        [command_buffer enqueue];
     }
 
     const id<MTLCommandBuffer> *command_buffers = command_buffer_builder;
@@ -1545,121 +1531,6 @@ static enum ggml_status ggml_metal_graph_compute(
                             [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         }
                     } break;
-                case GGML_OP_SSM_CONV:
-                    {
-                        GGML_ASSERT(src0t == GGML_TYPE_F32);
-                        GGML_ASSERT(src1t == GGML_TYPE_F32);
-
-                        GGML_ASSERT(ggml_is_contiguous(src0));
-                        GGML_ASSERT(ggml_is_contiguous(src1));
-
-                        id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_CONV_F32].pipeline;
-
-                        [encoder setComputePipelineState:pipeline];
-                        [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
-                        [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
-                        [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
-                        [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:3];
-                        [encoder setBytes:&ne01    length:sizeof(ne01) atIndex:4];
-                        [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:5];
-                        [encoder setBytes:&nb00    length:sizeof(nb00) atIndex:6];
-                        [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:7];
-                        [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:8];
-                        [encoder setBytes:&ne10    length:sizeof(ne10) atIndex:9];
-                        [encoder setBytes:&ne11    length:sizeof(ne11) atIndex:10];
-                        [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:11];
-                        [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:12];
-                        [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:13];
-                        [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:14];
-                        [encoder setBytes:&ne2     length:sizeof(ne2)  atIndex:15];
-                        [encoder setBytes:&nb0     length:sizeof(nb0)  atIndex:16];
-                        [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:17];
-                        [encoder setBytes:&nb2     length:sizeof(nb2)  atIndex:18];
-
-                        [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne1, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                    } break;
-                case GGML_OP_SSM_SCAN:
-                    {
-                        struct ggml_tensor * src3 = gf->nodes[i]->src[3];
-                        struct ggml_tensor * src4 = gf->nodes[i]->src[4];
-                        struct ggml_tensor * src5 = gf->nodes[i]->src[5];
-
-                        GGML_ASSERT(src3);
-                        GGML_ASSERT(src4);
-                        GGML_ASSERT(src5);
-
-                        size_t offs_src3 = 0;
-                        size_t offs_src4 = 0;
-                        size_t offs_src5 = 0;
-
-                        id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
-                        id<MTLBuffer> id_src4 = src4 ? ggml_metal_get_buffer(src4, &offs_src4) : nil;
-                        id<MTLBuffer> id_src5 = src5 ? ggml_metal_get_buffer(src5, &offs_src5) : nil;
-
-                        const int64_t  ne30 = src3->ne[0]; GGML_UNUSED(ne30);
-                        const int64_t  ne31 = src3->ne[1]; GGML_UNUSED(ne31);
-
-                        const uint64_t nb30 = src3->nb[0];
-                        const uint64_t nb31 = src3->nb[1];
-
-                        const int64_t  ne40 = src4->ne[0]; GGML_UNUSED(ne40);
-                        const int64_t  ne41 = src4->ne[1]; GGML_UNUSED(ne41);
-                        const int64_t  ne42 = src4->ne[2]; GGML_UNUSED(ne42);
-
-                        const uint64_t nb40 = src4->nb[0];
-                        const uint64_t nb41 = src4->nb[1];
-                        const uint64_t nb42 = src4->nb[2];
-
-                        const int64_t  ne50 = src5->ne[0]; GGML_UNUSED(ne50);
-                        const int64_t  ne51 = src5->ne[1]; GGML_UNUSED(ne51);
-                        const int64_t  ne52 = src5->ne[2]; GGML_UNUSED(ne52);
-
-                        const uint64_t nb50 = src5->nb[0];
-                        const uint64_t nb51 = src5->nb[1];
-                        const uint64_t nb52 = src5->nb[2];
-
-                        const int64_t d_state      = ne00;
-                        const int64_t d_inner      = ne01;
-                        const int64_t n_seq_tokens = ne11;
-                        const int64_t n_seqs       = ne02;
-
-                        id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline;
-
-                        [encoder setComputePipelineState:pipeline];
-                        [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                        [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                        [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2];
-                        [encoder setBuffer:id_src3 offset:offs_src3 atIndex:3];
-                        [encoder setBuffer:id_src4 offset:offs_src4 atIndex:4];
-                        [encoder setBuffer:id_src5 offset:offs_src5 atIndex:5];
-                        [encoder setBuffer:id_dst  offset:offs_dst  atIndex:6];
-
-                        [encoder setBytes:&d_state      length:sizeof(d_state)      atIndex:7];
-                        [encoder setBytes:&d_inner      length:sizeof(d_inner)      atIndex:8];
-                        [encoder setBytes:&n_seq_tokens length:sizeof(n_seq_tokens) atIndex:9];
-                        [encoder setBytes:&n_seqs       length:sizeof(n_seqs)       atIndex:10];
-
-                        [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:11];
-                        [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:12];
-                        [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:13];
-                        [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
-                        [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
-                        [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
-                        [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17];
-                        [encoder setBytes:&nb20 length:sizeof(nb20) atIndex:18];
-                        [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:19];
-                        [encoder setBytes:&nb22 length:sizeof(nb22) atIndex:20];
-                        [encoder setBytes:&nb30 length:sizeof(nb30) atIndex:21];
-                        [encoder setBytes:&nb31 length:sizeof(nb31) atIndex:22];
-                        [encoder setBytes:&nb40 length:sizeof(nb40) atIndex:23];
-                        [encoder setBytes:&nb41 length:sizeof(nb41) atIndex:24];
-                        [encoder setBytes:&nb42 length:sizeof(nb42) atIndex:25];
-                        [encoder setBytes:&nb50 length:sizeof(nb50) atIndex:26];
-                        [encoder setBytes:&nb51 length:sizeof(nb51) atIndex:27];
-                        [encoder setBytes:&nb52 length:sizeof(nb52) atIndex:28];
-
-                        [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                    } break;
                 case GGML_OP_MUL_MAT:
                     {
                         GGML_ASSERT(ne00 == ne10);
@@ -2358,8 +2229,10 @@ static enum ggml_status ggml_metal_graph_compute(
                         GGML_ASSERT(ne00 % 4 == 0);
                         GGML_ASSERT(ggml_is_contiguous(src0));
 
-                        float eps;
-                        memcpy(&eps, dst->op_params + 1, sizeof(float));
+                        //float eps;
+                        //memcpy(&eps, dst->op_params, sizeof(float));
+
+                        const float eps = 1e-6f; // TODO: temporarily hardcoded
 
                         const int32_t n_groups = ((int32_t *) dst->op_params)[0];
 
@@ -2435,7 +2308,7 @@ static enum ggml_status ggml_metal_graph_compute(
                         memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
                         memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
 
-                        const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+                        const bool is_neox = mode & 2;
 
                         id<MTLComputePipelineState> pipeline = nil;
 
@@ -2746,14 +2619,9 @@ static enum ggml_status ggml_metal_graph_compute(
 
                         float scale;
                         float max_bias;
-                        float logit_softcap;
-                        memcpy(&scale,         ((int32_t *) dst->op_params) + 0, sizeof(scale));
-                        memcpy(&max_bias,      ((int32_t *) dst->op_params) + 1, sizeof(max_bias));
-                        memcpy(&logit_softcap, ((int32_t *) dst->op_params) + 2, sizeof(logit_softcap));
 
-                        if (logit_softcap != 0.0f) {
-                            scale /= logit_softcap;
-                        }
+                        memcpy(&scale,    ((int32_t *) dst->op_params) + 0, sizeof(scale));
+                        memcpy(&max_bias, ((int32_t *) dst->op_params) + 1, sizeof(max_bias));
 
                         const uint32_t n_head      = src0->ne[2];
                         const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
@@ -2804,31 +2672,30 @@ static enum ggml_status ggml_metal_graph_compute(
                         } else {
                             [encoder setBuffer:id_src0     offset:offs_src0           atIndex:3];
                         }
-                        [encoder setBuffer:id_dst        offset:offs_dst              atIndex:4];
-                        [encoder setBytes:&ne01          length:sizeof( int64_t)      atIndex:5];
-                        [encoder setBytes:&ne02          length:sizeof( int64_t)      atIndex:6];
-                        [encoder setBytes:&ne03          length:sizeof( int64_t)      atIndex:7];
-                        [encoder setBytes:&nb01          length:sizeof(uint64_t)      atIndex:8];
-                        [encoder setBytes:&nb02          length:sizeof(uint64_t)      atIndex:9];
-                        [encoder setBytes:&nb03          length:sizeof(uint64_t)      atIndex:10];
-                        [encoder setBytes:&ne11          length:sizeof( int64_t)      atIndex:11];
-                        [encoder setBytes:&ne12          length:sizeof( int64_t)      atIndex:12];
-                        [encoder setBytes:&ne13          length:sizeof( int64_t)      atIndex:13];
-                        [encoder setBytes:&nb11          length:sizeof(uint64_t)      atIndex:14];
-                        [encoder setBytes:&nb12          length:sizeof(uint64_t)      atIndex:15];
-                        [encoder setBytes:&nb13          length:sizeof(uint64_t)      atIndex:16];
-                        [encoder setBytes:&nb21          length:sizeof(uint64_t)      atIndex:17];
-                        [encoder setBytes:&nb22          length:sizeof(uint64_t)      atIndex:18];
-                        [encoder setBytes:&nb23          length:sizeof(uint64_t)      atIndex:19];
-                        [encoder setBytes:&nb31          length:sizeof(uint64_t)      atIndex:20];
-                        [encoder setBytes:&ne1           length:sizeof( int64_t)      atIndex:21];
-                        [encoder setBytes:&ne2           length:sizeof( int64_t)      atIndex:22];
-                        [encoder setBytes:&scale         length:sizeof(   float)      atIndex:23];
-                        [encoder setBytes:&max_bias      length:sizeof(   float)      atIndex:24];
-                        [encoder setBytes:&m0            length:sizeof(m0)            atIndex:25];
-                        [encoder setBytes:&m1            length:sizeof(m1)            atIndex:26];
-                        [encoder setBytes:&n_head_log2   length:sizeof(n_head_log2)   atIndex:27];
-                        [encoder setBytes:&logit_softcap length:sizeof(logit_softcap) atIndex:28];
+                        [encoder setBuffer:id_dst      offset:offs_dst            atIndex:4];
+                        [encoder setBytes:&ne01        length:sizeof( int64_t)    atIndex:5];
+                        [encoder setBytes:&ne02        length:sizeof( int64_t)    atIndex:6];
+                        [encoder setBytes:&ne03        length:sizeof( int64_t)    atIndex:7];
+                        [encoder setBytes:&nb01        length:sizeof(uint64_t)    atIndex:8];
+                        [encoder setBytes:&nb02        length:sizeof(uint64_t)    atIndex:9];
+                        [encoder setBytes:&nb03        length:sizeof(uint64_t)    atIndex:10];
+                        [encoder setBytes:&ne11        length:sizeof( int64_t)    atIndex:11];
+                        [encoder setBytes:&ne12        length:sizeof( int64_t)    atIndex:12];
+                        [encoder setBytes:&ne13        length:sizeof( int64_t)    atIndex:13];
+                        [encoder setBytes:&nb11        length:sizeof(uint64_t)    atIndex:14];
+                        [encoder setBytes:&nb12        length:sizeof(uint64_t)    atIndex:15];
+                        [encoder setBytes:&nb13        length:sizeof(uint64_t)    atIndex:16];
+                        [encoder setBytes:&nb21        length:sizeof(uint64_t)    atIndex:17];
+                        [encoder setBytes:&nb22        length:sizeof(uint64_t)    atIndex:18];
+                        [encoder setBytes:&nb23        length:sizeof(uint64_t)    atIndex:19];
+                        [encoder setBytes:&nb31        length:sizeof(uint64_t)    atIndex:20];
+                        [encoder setBytes:&ne1         length:sizeof( int64_t)    atIndex:21];
+                        [encoder setBytes:&ne2         length:sizeof( int64_t)    atIndex:22];
+                        [encoder setBytes:&scale       length:sizeof(   float)    atIndex:23];
+                        [encoder setBytes:&max_bias    length:sizeof(   float)    atIndex:24];
+                        [encoder setBytes:&m0          length:sizeof(m0)          atIndex:25];
+                        [encoder setBytes:&m1          length:sizeof(m1)          atIndex:26];
+                        [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:27];
 
                         if (!use_vec_kernel) {
                             // half8x8 kernel
@@ -2962,9 +2829,7 @@ static enum ggml_status ggml_metal_graph_compute(
 
         [encoder endEncoding];
 
-        if (cb_idx < 2 || ctx->abort_callback == NULL) {
-            [command_buffer commit];
-        }
+        [command_buffer commit];
     });
 
     // Wait for completion and check status of each command buffer
@@ -2984,23 +2849,6 @@ static enum ggml_status ggml_metal_graph_compute(
 
             return GGML_STATUS_FAILED;
         }
-
-        id<MTLCommandBuffer> next_buffer = (i + 1 < n_cb ? command_buffers[i + 1] : nil);
-        if (!next_buffer) {
-            continue;
-        }
-
-        bool next_queued = ([next_buffer status] != MTLCommandBufferStatusNotEnqueued);
-        if (next_queued) {
-            continue;
-        }
-
-        if (ctx->abort_callback && ctx->abort_callback(ctx->abort_callback_data)) {
-            GGML_METAL_LOG_INFO("%s: command buffer %d aborted", __func__, i);
-            return GGML_STATUS_ABORTED;
-        }
-
-        [next_buffer commit];
     }
 
     if (should_capture) {
@@ -3304,7 +3152,7 @@ GGML_CALL static const char * ggml_backend_metal_name(ggml_backend_t backend) {
 }
 
 GGML_CALL static void ggml_backend_metal_free(ggml_backend_t backend) {
-    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
+    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
     ggml_metal_free(ctx);
     free(backend);
 }
@@ -3316,13 +3164,13 @@ GGML_CALL static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffe
 }
 
 GGML_CALL static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context;
+    struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
 
     return ggml_metal_graph_compute(metal_ctx, cgraph);
 }
 
 GGML_CALL static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context;
+    struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
 
     return ggml_metal_supports_op(metal_ctx, op);
 }
@@ -3367,9 +3215,9 @@ static ggml_guid_t ggml_backend_metal_guid(void) {
 }
 
 ggml_backend_t ggml_backend_metal_init(void) {
-    struct ggml_backend_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
+    struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
+
     if (ctx == NULL) {
-        GGML_METAL_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
         return NULL;
     }
 
@@ -3391,24 +3239,15 @@ bool ggml_backend_is_metal(ggml_backend_t backend) {
 void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
     GGML_ASSERT(ggml_backend_is_metal(backend));
 
-    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
+    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
 
     ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
 }
 
-void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data) {
-    GGML_ASSERT(ggml_backend_is_metal(backend));
-
-    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
-
-    ctx->abort_callback = abort_callback;
-    ctx->abort_callback_data = user_data;
-}
-
 bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
     GGML_ASSERT(ggml_backend_is_metal(backend));
 
-    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
+    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
 
     return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
 }
@@ -3416,7 +3255,7 @@ bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
 void ggml_backend_metal_capture_next_compute(ggml_backend_t backend) {
     GGML_ASSERT(ggml_backend_is_metal(backend));
 
-    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
+    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
     ctx->should_capture_next_compute = true;
 }