feat(sycl): Add sycl support (#1647)

Author: mudler

.github/workflows/image-pr.yml

@@ -75,6 +75,13 @@ jobs:
             ffmpeg: 'true'
             image-type: 'core'
             runs-on: 'ubuntu-latest'
+          - build-type: 'sycl_f16'
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: 'sycl-f16-ffmpeg-core'
+            ffmpeg: 'true'
+            image-type: 'core'
+            runs-on: 'ubuntu-latest'
           - build-type: 'cublas'
             cuda-major-version: "12"
             cuda-minor-version: "1"

.github/workflows/image.yml

@@ -122,6 +122,34 @@ jobs:
             ffmpeg: 'true'
             image-type: 'core'
             runs-on: 'ubuntu-latest'
+          - build-type: 'sycl_f16'
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: 'sycl-f16-core'
+            ffmpeg: 'false'
+            image-type: 'core'
+            runs-on: 'ubuntu-latest'
+          - build-type: 'sycl_f32'
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: 'sycl-f32-core'
+            ffmpeg: 'false'
+            image-type: 'core'
+            runs-on: 'ubuntu-latest'
+          - build-type: 'sycl_f16'
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: 'sycl-f16-ffmpeg-core'
+            ffmpeg: 'true'
+            image-type: 'core'
+            runs-on: 'ubuntu-latest'
+          - build-type: 'sycl_f32'
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: 'sycl-f32-ffmpeg-core'
+            ffmpeg: 'true'
+            image-type: 'core'
+            runs-on: 'ubuntu-latest'
           - build-type: 'cublas'
             cuda-major-version: "11"
             cuda-minor-version: "7"

Dockerfile

@@ -2,7 +2,6 @@ ARG GO_VERSION=1.21-bullseye
 ARG IMAGE_TYPE=extras
 # extras or core
 
-
 FROM golang:$GO_VERSION as requirements-core
 
 ARG BUILD_TYPE
@@ -38,6 +37,11 @@ RUN if [ "${BUILD_TYPE}" = "cublas" ]; then \
     apt-get update && \
     apt-get install -y cuda-nvcc-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libcublas-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libcusparse-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libcusolver-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION}  && apt-get clean \
     ; fi
+# oneapi requirements
+RUN if [ "${BUILD_TYPE}" = "sycl_f16" ] || [ "${BUILD_TYPE}" = "sycl_f32" ]; then \
+    wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/163da6e4-56eb-4948-aba3-debcec61c064/l_BaseKit_p_2024.0.1.46_offline.sh && \
+    sh ./l_BaseKit_p_2024.0.1.46_offline.sh \
+    ; fi
 ENV PATH /usr/local/cuda/bin:${PATH}
 
 # OpenBLAS requirements and stable diffusion

Makefile

@@ -111,6 +111,14 @@ ifeq ($(BUILD_TYPE),hipblas)
 	CGO_LDFLAGS += -O3 --rtlib=compiler-rt -unwindlib=libgcc -lhipblas -lrocblas --hip-link
 endif
 
+ifeq ($(BUILD_TYPE),sycl_f16)
+	CMAKE_ARGS+=-DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
+endif
+
+ifeq ($(BUILD_TYPE),sycl_f32)
+	CMAKE_ARGS+=-DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+endif
+
 ifeq ($(BUILD_TYPE),metal)
 	CGO_LDFLAGS+=-framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
 	export LLAMA_METAL=1

backend/cpp/llama/Makefile

@@ -3,6 +3,7 @@ LLAMA_VERSION?=
 
 CMAKE_ARGS?=
 BUILD_TYPE?=
+ONEAPI_VARS?=/opt/intel/oneapi/setvars.sh
 
 # If build type is cublas, then we set -DLLAMA_CUBLAS=ON to CMAKE_ARGS automatically
 ifeq ($(BUILD_TYPE),cublas)
@@ -49,5 +50,10 @@ clean:
 	rm -rf grpc-server
 
 grpc-server: llama.cpp llama.cpp/examples/grpc-server
+ifneq (,$(findstring sycl,$(BUILD_TYPE)))
+	source $(ONEAPI_VARS) && \
+	cd llama.cpp && mkdir -p build && cd build && cmake .. $(CMAKE_ARGS) && cmake --build . --config Release		
+else
 	cd llama.cpp && mkdir -p build && cd build && cmake .. $(CMAKE_ARGS) && cmake --build . --config Release
+endif
 	cp llama.cpp/build/bin/grpc-server .

entrypoint.sh

@@ -13,6 +13,10 @@ if [ -n "$EXTRA_BACKENDS" ]; then
 	done
 fi
 
+if [ -e "/opt/intel/oneapi/setvars.sh"]; then
+	source /opt/intel/oneapi/setvars.sh
+fi
+
 if [ "$REBUILD" != "false" ]; then
 	rm -rf ./local-ai
 	make build -j${BUILD_PARALLELISM:-1}