Remove llama.cpp submodule and shift new build to top

.github/workflows/release.yaml

@@ -167,12 +167,23 @@ jobs:
           $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
           make -j $cores
         name: make
+      - name: 'gather rocm dependencies'
+        run: |
+          $HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          md "dist\deps\bin\rocblas\library"
+          cp "${HIP_PATH}\bin\hipblas.dll" "dist\deps\bin\"
+          cp "${HIP_PATH}\bin\rocblas.dll" "dist\deps\bin\"
+          cp "${HIP_PATH}\bin\rocblas\library\*" "dist\deps\bin\rocblas\library\"
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-rocm
           path: |
             build/**/*
             dist/windows-amd64/**
+      - uses: actions/upload-artifact@v4
+        with:
+          name: windows-rocm-deps
+          path: dist/deps/*
 
   # CUDA generation step
   generate-windows-cuda:
@@ -246,12 +257,24 @@ jobs:
           $env:OLLAMA_SKIP_CPU_GENERATE="1"
           $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
           make -j $cores
+      - name: 'gather cuda dependencies'
+        run: |
+          $NVIDIA_DIR=(resolve-path 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*\bin\')[0]
+          md "dist\deps"
+          cp "${NVIDIA_DIR}\cudart64_*.dll" "dist\deps\"
+          cp "${NVIDIA_DIR}\cublas64_*.dll" "dist\deps\"
+          cp "${NVIDIA_DIR}\cublasLt64_*.dll" "dist\deps\"
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-cuda-${{ matrix.cuda.version }}
           path: |
             build/**/*
             dist/windows-amd64/**
+      - uses: actions/upload-artifact@v4
+        with:
+          name: windows-cuda-deps-${{ matrix.cuda.version }}
+          path: dist/deps/*
+
 
   # windows arm64 generate, go build, and zip file (no installer)
   # Output of this build is aggregated into the final x86 build
@@ -457,6 +480,15 @@ jobs:
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-cuda-12
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-cuda-deps-11
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-cuda-deps-12
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-rocm-deps
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-rocm

llama/llama.go

@@ -46,8 +46,8 @@ package llama
 #cgo rocm CFLAGS: -DGGML_USE_CUDA -DGGML_USE_HIPBLAS -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
 #cgo rocm CXXFLAGS: -DGGML_USE_CUDA -DGGML_USE_HIPBLAS -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
 #cgo rocm LDFLAGS: -L${SRCDIR} -lggml_rocm -lhipblas -lamdhip64 -lrocblas
-#cgo windows CFLAGS: -Wno-discarded-qualifiers -D_WIN32_WINNT=0x602
-#cgo windows CXXFLAGS: -D_WIN32_WINNT=0x602
+#cgo windows CFLAGS: -Wno-discarded-qualifiers
+#cgo windows CFLAGS: -Wno-discarded-qualifiers
 #cgo windows LDFLAGS: -lmsvcrt
 #cgo windows LDFLAGS: -lmsvcrt -static-libstdc++ -static-libgcc -static
 #cgo windows,amd64 LDFLAGS: -L${SRCDIR}/build/Windows/amd64

llama/make/Makefile.rocm

@@ -17,8 +17,8 @@ else ifeq ($(OS),linux)
 	GPU_LIB_DIR_LINUX := $(HIP_PATH)/lib
 	GPU_COMPILER_LINUX := $(shell X=$$(which hipcc 2>/dev/null) && echo $$X)
 	GPU_COMPILER:=$(GPU_COMPILER_LINUX)
-	ROCM_TRANSITIVE_LIBS_INITIAL = $(sort $(shell ldd $(GPU_LIBS) | grep "=>" | cut -f2 -d= | cut -f2 -d' '  | grep -e rocm -e amdgpu -e libtinfo -e libnuma -e libelf))
-	GPU_TRANSITIVE_LIBS = $(sort $(shell readlink -f $(ROCM_TRANSITIVE_LIBS_INITIAL)) $(ROCM_TRANSITIVE_LIBS_INITIAL))
+	ROCM_TRANSITIVE_LIBS_INITIAL = $(sort $(shell ldd $(ROCM_LIBS) | grep "=>" | cut -f2 -d= | cut -f2 -d' '  | grep -e rocm -e amdgpu -e libtinfo -e libnuma -e libelf))
+	ROCM_TRANSITIVE_LIBS = $(sort $(shell readlink -f $(ROCM_TRANSITIVE_LIBS_INITIAL)) $(ROCM_TRANSITIVE_LIBS_INITIAL))
 endif
 
 # TODO future multi-variant support for ROCm
@@ -38,13 +38,13 @@ GPU_COMPILER_CFLAGS_LINUX = $(CFLAGS) -fPIC -D_GNU_SOURCE
 GPU_COMPILER_CXXFLAGS_WIN = $(CXXFLAGS) -D_WIN32_WINNT=0x602
 GPU_COMPILER_CXXFLAGS_LINUX = $(CXXFLAGS) -fPIC -D_GNU_SOURCE
 
-GPU_LIBS = $(wildcard $(addsuffix .$(SHARED_EXT),$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT))))
+ROCM_LIBS = $(wildcard $(addsuffix .$(SHARED_EXT),$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT))))
 ifeq ($(OS),windows)
 	ROCM_DIST_DEPS_DIR = $(abspath $(SRC_DIR)/../dist/$(OS)-$(ARCH))/lib/ollama
 else ifeq ($(OS),linux)
 	ROCM_DIST_DEPS_DIR = $(abspath $(SRC_DIR)/../dist/$(OS)-$(ARCH)-rocm)/lib/ollama
 endif
-GPU_DIST_DEPS_LIBS= $(sort $(addprefix $(ROCM_DIST_DEPS_DIR)/,$(notdir $(GPU_LIBS)) $(notdir $(GPU_TRANSITIVE_LIBS))))
+ROCM_DIST_DEPS_LIBS= $(sort $(addprefix $(ROCM_DIST_DEPS_DIR)/,$(notdir $(ROCM_LIBS)) $(notdir $(ROCM_TRANSITIVE_LIBS))))
 ROCBLAS_DIST_DEP_MANIFEST = $(ROCM_DIST_DEPS_DIR)/rocblas/library/TensileManifest.txt
 
 ifeq ($(OS),linux)
@@ -90,9 +90,12 @@ GPU_COMPILER_CUFLAGS = \
 include make/gpu.make
 
 # Adjust the rules from gpu.make to handle the ROCm dependencies properly
-$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(ROCBLAS_DIST_DEP_MANIFEST)
+$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(ROCBLAS_DIST_DEP_MANIFEST) $(ROCM_DIST_DEPS_LIBS)
 $(ROCBLAS_DIST_DEP_MANIFEST):
 	@-mkdir -p $(dir $@)
 	@echo "Copying rocblas library..."
 	cd $(GPU_LIB_DIR)/rocblas/library/ && tar cf - . | (cd $(dir $@) && tar xf - )
 	@echo "rocblas library copy complete"
+$(ROCM_DIST_DEPS_LIBS): 
+	@-mkdir -p $(dir $@)
+	$(CP) $(dir $(filter %$(notdir $@),$(ROCM_LIBS) $(ROCM_TRANSITIVE_LIBS)))/$(notdir $@) $(dir $@)

llama/make/cuda.make

@@ -19,9 +19,6 @@ GPU_COMPILER_CFLAGS_WIN = $(CFLAGS) -D_WIN32_WINNT=0x602
 GPU_COMPILER_CFLAGS_LINUX = $(CFLAGS) -Xcompiler -fPIC -D_GNU_SOURCE
 GPU_COMPILER_CXXFLAGS_WIN = $(CXXFLAGS) -D_WIN32_WINNT=0x602
 GPU_COMPILER_CXXFLAGS_LINUX = $(CXXFLAGS) -Xcompiler -fPIC -D_GNU_SOURCE
-GPU_LIBS = $(sort $(wildcard $(addsuffix *.$(SHARED_EXT)*,$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT)))))
-GPU_DIST_DEPS_LIBS= $(sort $(addprefix $(DIST_LIB_DIR)/,$(notdir $(GPU_LIBS))))
-
 ifeq ($(OS),linux)
 	CUDA_PATH?=/usr/local/cuda
 	GPU_COMPILER_FPIC = -fPIC -Wno-unused-function -std=c++11

llama/make/gpu.make

@@ -98,16 +98,13 @@ $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).
 $(RUNNERS_DIST_DIR)/%: $(RUNNERS_BUILD_DIR)/%
 	@-mkdir -p $(dir $@)
 	$(CP) $< $@
-$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(DIST_LIB_DIR)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT) $(GPU_DIST_DEPS_LIBS)
+$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(DIST_LIB_DIR)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT)
 $(DIST_LIB_DIR)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT): $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT)
 	@-mkdir -p $(dir $@)
 	$(CP) $< $@
 $(DIST_GPU_RUNNER_LIB_DEPS): 
 	@-mkdir -p $(dir $@)
 	$(CP) $(GPU_LIB_DIR)/$(notdir $@) $(dir $@)
-$(GPU_DIST_DEPS_LIBS): 
-	@-mkdir -p $(dir $@)
-	$(CP) $(dir $(filter %$(notdir $@),$(GPU_LIBS) $(GPU_TRANSITIVE_LIBS)))/$(notdir $@) $(dir $@)
 
 # Payload targets
 $(RUNNERS_PAYLOAD_DIR)/%/ollama_llama_server.gz: $(RUNNERS_BUILD_DIR)/%/ollama_llama_server