Merge branch 'main' into rebase_ollama_main

Dockerfile

@@ -49,6 +49,15 @@ RUN mkdir /tmp/scratch && \
     (cd /tmp/scratch/ && tar czvf /go/src/github.com/jmorganca/ollama/dist/deps/ollama-linux-amd64-rocm.tgz . )
 
 
+FROM --platform=linux/amd64 intel/oneapi-basekit:2024.0.1-devel-rockylinux9 AS oneapi-build-amd64
+ARG CMAKE_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
+COPY --from=llm-code / /go/src/github.com/jmorganca/ollama/
+WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
+ARG CGO_CFLAGS
+RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+
 FROM --platform=linux/amd64 intel/oneapi-basekit:2024.0.1-devel-rockylinux9 AS oneapi-build-amd64
 ARG CMAKE_VERSION
 COPY ./scripts/rh_linux_deps.sh /

gpu/gpu.go

@@ -59,6 +59,11 @@ var OneapiLinuxGlobs = []string{
 	"/usr/lib*/libze_intel_gpu.so*",
 }
 
+var OneapiLinuxGlobs = []string{
+	"/usr/lib/x86_64-linux-gnu/libze_intel_gpu.so*",
+	"/usr/lib*/libze_intel_gpu.so*",
+}
+
 // Note: gpuMutex must already be held
 func initGPUHandles() {
 
@@ -105,6 +110,16 @@ func initGPUHandles() {
 			return
 		}
 	}
+
+	oneapiLibPaths := FindGPULibs(oneapiMgmtName, oneapiMgmtPatterns)
+	if len(oneapiLibPaths) > 0 {
+		oneapi := LoadOneapiMgmt(oneapiLibPaths)
+		if oneapi != nil {
+			slog.Info("Intel GPU detected")
+			gpuHandles.oneapi = oneapi
+			return
+		}
+	}
 }
 
 func GetGPUInfo() GpuInfo {
@@ -170,6 +185,28 @@ func GetGPUInfo() GpuInfo {
 		if resp.Library != "" {
 			return resp
 		}
+	} else if gpuHandles.oneapi != nil && (cpuVariant != "" || runtime.GOARCH != "amd64") {
+		C.oneapi_check_vram(*gpuHandles.oneapi, &memInfo)
+		if memInfo.err != nil {
+			slog.Info(fmt.Sprintf("error looking up OneAPI GPU memory: %s", C.GoString(memInfo.err)))
+			C.free(unsafe.Pointer(memInfo.err))
+		} else if memInfo.igpu_index >= 0 && memInfo.count == 1 {
+			// Only one GPU detected and it appears to be an integrated GPU - skip it
+			slog.Info("OneAPI unsupported integrated GPU detected")
+		} else if memInfo.count > 0 {
+			if memInfo.igpu_index >= 0 {
+				// We have multiple GPUs reported, and one of them is an integrated GPU
+				// so we have to set the env var to bypass it
+				// If the user has specified their own SYCL_DEVICE_ALLOWLIST, don't clobber it
+				val := os.Getenv("SYCL_DEVICE_ALLOWLIST")
+				if val == "" {
+					val = "DeviceType:gpu"
+					os.Setenv("SYCL_DEVICE_ALLOWLIST", val)
+				}
+				slog.Info(fmt.Sprintf("oneAPI integrated GPU detected - SYCL_DEVICE_ALLOWLIST=%s", val))
+			}
+			resp.Library = "oneapi"
+		}
 	}
 	if resp.Library == "" {
 		C.cpu_check_ram(&memInfo)
@@ -315,6 +352,23 @@ func LoadOneapiMgmt(rocmLibPaths []string) *C.oneapi_handle_t {
 	return nil
 }
 
+func LoadOneapiMgmt(rocmLibPaths []string) *C.oneapi_handle_t {
+	var resp C.oneapi_init_resp_t
+	resp.oh.verbose = getVerboseState()
+	for _, libPath := range rocmLibPaths {
+		lib := C.CString(libPath)
+		defer C.free(unsafe.Pointer(lib))
+		C.oneapi_init(lib, &resp)
+		if resp.err != nil {
+			slog.Info(fmt.Sprintf("Unable to load oneAPI management library %s: %s", libPath, C.GoString(resp.err)))
+			C.free(unsafe.Pointer(resp.err))
+		} else {
+			return &resp.oh
+		}
+	}
+	return nil
+}
+
 func getVerboseState() C.uint16_t {
 	if debug := os.Getenv("OLLAMA_DEBUG"); debug != "" {
 		return C.uint16_t(1)

llm/generate/gen_linux.sh

@@ -227,4 +227,25 @@ if [ -d "${ONEAPI_ROOT}" ]; then
     compress_libs
 fi
 
+if [ -z "${ONEAPI_ROOT}" ]; then
+    # Try the default location in case it exists
+    ONEAPI_ROOT=/opt/intel/oneapi
+fi
+
+if [ -d "${ONEAPI_ROOT}" ]; then
+    echo "OneAPI libraries detected - building dynamic OneAPI library"
+    init_vars
+    source ${ONEAPI_ROOT}/setvars.sh --force # set up environment variables for oneAPI
+    CC=icx
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL=ON -DLLAMA_SYCL_F16=OFF"
+    BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/oneapi"
+    EXTRA_LIBS="-fsycl -Wl,-rpath,${ONEAPI_ROOT}/compiler/latest/lib,-rpath,${ONEAPI_ROOT}/mkl/latest/lib,-rpath,${ONEAPI_ROOT}/tbb/latest/lib,-rpath,${ONEAPI_ROOT}/compiler/latest/opt/oclfpga/linux64/lib -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
+    DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
+    build
+
+    # Note: the OneAPI libs and runtime library files are too large to embed, so we depend on
+    #       them being present at runtime on the host
+    compress_libs
+fi
+
 cleanup

scripts/build_docker.sh

@@ -84,3 +84,13 @@ docker build \
     -f Dockerfile \
     -t ollama/ollama:$VERSION-oneapi \
     .
+
+docker build \
+    --load \
+    --platform=linux/amd64 \
+    --build-arg=VERSION \
+    --build-arg=GOFLAGS \
+    --target runtime-oneapi \
+    -f Dockerfile \
+    -t ollama/ollama:$VERSION-oneapi \
+    .