feat(amdgpu): try to build in single binary

Signed-off-by: Ettore Di Giacinto <[email protected]>

.github/workflows/release.yaml

@@ -38,6 +38,31 @@ jobs:
           sudo apt-get install -y cuda-nvcc-${CUDA_VERSION} libcublas-dev-${CUDA_VERSION}
         env:
           CUDA_VERSION: 12-3
+      - name: "Install Hipblas"
+        env:
+          ROCM_VERSION: "5.3"
+          AMDGPU_VERSION: "5.3"
+        run: |
+            set -ex 
+
+            sudo apt-get update
+            sudo DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends ca-certificates curl libnuma-dev gnupg 
+            
+            curl -sL https://repo.radeon.com/rocm/rocm.gpg.key | sudo apt-key add - 
+              
+            printf "deb [arch=amd64] https://repo.radeon.com/rocm/apt/$ROCM_VERSION/ jammy main" | sudo tee /etc/apt/sources.list.d/rocm.list
+            
+            printf "deb [arch=amd64] https://repo.radeon.com/amdgpu/$AMDGPU_VERSION/ubuntu jammy main" | sudo tee /etc/apt/sources.list.d/amdgpu.list
+
+            sudo apt-get update
+
+            sudo DEBIAN_FRONTEND=noninteractive apt-get install -y \
+                hipblas-dev rocm-dev \
+                rocblas-dev
+          
+            sudo apt-get clean
+            sudo rm -rf /var/lib/apt/lists/*
+            sudo ldconfig 
       - name: Cache grpc
         id: cache-grpc
         uses: actions/cache@v4
@@ -61,6 +86,7 @@ jobs:
           go install google.golang.org/protobuf/cmd/[email protected]
           export PATH=$PATH:$GOPATH/bin
           export PATH=/usr/local/cuda/bin:$PATH
+          export PATH=/opt/rocm/bin:$PATH
           GO_TAGS=p2p make dist
       - uses: actions/upload-artifact@v4
         with:

Makefile

@@ -327,6 +327,7 @@ ifeq ($(OS),Darwin)
 	$(info ${GREEN}I Skip CUDA build on MacOS${RESET})
 else
 	$(MAKE) backend-assets/grpc/llama-cpp-cuda
+	$(MAKE) backend-assets/grpc/llama-cpp-hipblas
 endif
 	$(MAKE) build
 	mkdir -p release
@@ -712,6 +713,13 @@ backend-assets/grpc/llama-cpp-cuda: backend-assets/grpc
 	CMAKE_ARGS="$(CMAKE_ARGS) -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off -DLLAMA_CUDA=ON" $(MAKE) VARIANT="llama-cuda" build-llama-cpp-grpc-server
 	cp -rfv backend/cpp/llama-cuda/grpc-server backend-assets/grpc/llama-cpp-cuda
 
+backend-assets/grpc/llama-cpp-hipblas: backend-assets/grpc
+	cp -rf backend/cpp/llama backend/cpp/llama-hipblas
+	$(MAKE) -C backend/cpp/llama-hipblas purge
+	$(info ${GREEN}I llama-cpp build info:hipblas${RESET})
+	BUILD_TYPE="hipblas" $(MAKE) VARIANT="llama-hipblas" build-llama-cpp-grpc-server
+	cp -rfv backend/cpp/llama-hipblas/grpc-server backend-assets/grpc/llama-cpp-hipblas
+
 backend-assets/grpc/llama-cpp-grpc: backend-assets/grpc
 	cp -rf backend/cpp/llama backend/cpp/llama-grpc
 	$(MAKE) -C backend/cpp/llama-grpc purge

pkg/model/initializers.go

@@ -37,6 +37,7 @@ const (
 	LLamaCPPAVX      = "llama-cpp-avx"
 	LLamaCPPFallback = "llama-cpp-fallback"
 	LLamaCPPCUDA     = "llama-cpp-cuda"
+	LLamaCPPHipblas  = "llama-cpp-hipblas"
 	LLamaCPPGRPC     = "llama-cpp-grpc"
 
 	Gpt4AllLlamaBackend = "gpt4all-llama"
@@ -93,7 +94,7 @@ ENTRY:
 	if autoDetect {
 		// if we find the llama.cpp variants, show them of as a single backend (llama-cpp) as later we are going to pick that up
 		// when starting the service
-		foundLCPPAVX, foundLCPPAVX2, foundLCPPFallback, foundLCPPGRPC, foundLCPPCuda := false, false, false, false, false
+		foundLCPPAVX, foundLCPPAVX2, foundLCPPFallback, foundLCPPGRPC, foundLCPPCuda, foundLCPPHipblas := false, false, false, false, false, false
 		if _, ok := backends[LLamaCPP]; !ok {
 			for _, e := range entry {
 				if strings.Contains(e.Name(), LLamaCPPAVX2) && !foundLCPPAVX2 {
@@ -116,6 +117,10 @@ ENTRY:
 					backends[LLamaCPP] = append(backends[LLamaCPP], LLamaCPPCUDA)
 					foundLCPPCuda = true
 				}
+				if strings.Contains(e.Name(), LLamaCPPHipblas) && !foundLCPPHipblas {
+					backends[LLamaCPP] = append(backends[LLamaCPP], LLamaCPPHipblas)
+					foundLCPPHipblas = true
+				}
 			}
 		}
 	}
@@ -169,6 +174,7 @@ ENTRY:
 // selectGRPCProcess selects the GRPC process to start based on system capabilities
 func selectGRPCProcess(backend, assetDir string) string {
 	foundCUDA := false
+	foundAMDGPU := false
 	var grpcProcess string
 
 	// Select backend now just for llama.cpp
@@ -195,10 +201,20 @@ func selectGRPCProcess(backend, assetDir string) string {
 					log.Info().Msgf("GPU device found but no CUDA backend present")
 				}
 			}
+			if strings.Contains(gpu.String(), "amd") {
+				p := backendPath(assetDir, LLamaCPPHipblas)
+				if _, err := os.Stat(p); err == nil {
+					log.Info().Msgf("[%s] attempting to load with HIPBLAS variant", backend)
+					grpcProcess = p
+					foundAMDGPU = true
+				} else {
+					log.Info().Msgf("GPU device found but no HIPBLAS backend present")
+				}
+			}
 		}
 	}
 
-	if foundCUDA {
+	if foundCUDA || foundAMDGPU {
 		return grpcProcess
 	}