first pass at linux gpu support (#454)

* linux gpu support
* handle multiple gpus
* add cuda docker image (#488)
---------
Co-authored-by: Michael Yang <mxyng@pm.me>

Dockerfile.cuda

+FROM nvidia/cuda:12.2.0-devel-ubuntu22.04
+
+WORKDIR /go/src/github.com/jmorganca/ollama
+RUN apt-get update && apt-get install -y git build-essential cmake
+ADD https://dl.google.com/go/go1.21.1.linux-amd64.tar.gz /tmp/go1.21.1.tar.gz
+RUN mkdir -p /usr/local && tar xz -C /usr/local </tmp/go1.21.1.tar.gz
+
+COPY . .
+RUN /usr/local/go/bin/go generate ./... && /usr/local/go/bin/go build -ldflags '-linkmode external -extldflags "-static"' .
+
+FROM nvidia/cuda:12.2.0-runtime-ubuntu22.04
+ENV OLLAMA_HOST 0.0.0.0
+
+ARG USER=ollama
+ARG GROUP=ollama
+RUN groupadd $GROUP && useradd -m -g $GROUP $USER
+
+COPY --from=0 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
+
+USER $USER:$GROUP
+ENTRYPOINT ["/bin/ollama"]
+CMD ["serve"]

api/types.go

@@ -291,7 +291,7 @@ func DefaultOptions() Options {
 		NumCtx:             2048,
 		NumKeep:            -1,
 		NumBatch:           512,
-		NumGPU:             1,
+		NumGPU:             -1, // -1 here indicates that NumGPU should be set dynamically
 		NumGQA:             1,
 		LowVRAM:            false,
 		F16KV:              true,

docs/development.md

@@ -6,6 +6,10 @@
 
 Install required tools:
 
+- cmake version 3.24 or higher
+- go version 1.20 or higher
+- gcc version 11.4.0 or higher
+
 ```
 brew install go cmake gcc
 ```
@@ -27,3 +31,9 @@ Now you can run `ollama`:
 ```
 ./ollama
 ```
+
+## Building on Linux with GPU support
+
+- Install cmake and nvidia-cuda-toolkit
+- run `go generate ./...`
+- run `go build .`

llm/llama.cpp/generate.go

@@ -4,12 +4,14 @@
 package llm
 
 //go:generate git submodule init
-//go:generate git submodule update --force ggml gguf
-//go:generate git -C ggml apply ../ggml_patch/0001-add-detokenize-endpoint.patch
-//go:generate git -C ggml apply ../ggml_patch/0002-34B-model-support.patch
-//go:generate git -C ggml apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
-//go:generate git -C ggml apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
-//go:generate cmake --fresh -S ggml -B ggml/build/cpu -DLLAMA_K_QUANTS=on
+
+//go:generate git submodule update --force ggml
+//go:generate -command git-apply git -C ggml apply
+//go:generate git-apply ../ggml_patch/0001-add-detokenize-endpoint.patch
+//go:generate git-apply ../ggml_patch/0002-34B-model-support.patch
+//go:generate cmake -S ggml -B ggml/build/cpu -DLLAMA_K_QUANTS=on
 //go:generate cmake --build ggml/build/cpu --target server --config Release
+
+//go:generate git submodule update --force gguf
 //go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_K_QUANTS=on
 //go:generate cmake --build gguf/build/cpu --target server --config Release

llm/llama.cpp/generate_darwin_amd64.go

 package llm
 
 //go:generate git submodule init
-//go:generate git submodule update --force ggml gguf
-//go:generate git -C ggml apply ../ggml_patch/0001-add-detokenize-endpoint.patch
-//go:generate git -C ggml apply ../ggml_patch/0002-34B-model-support.patch
-//go:generate git -C ggml apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
-//go:generate git -C ggml apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
-//go:generate cmake --fresh -S ggml -B ggml/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
+
+//go:generate git submodule update --force ggml
+//go:generate -command git-apply git -C ggml apply
+//go:generate git-apply ../ggml_patch/0001-add-detokenize-endpoint.patch
+//go:generate git-apply ../ggml_patch/0002-34B-model-support.patch
+//go:generate git-apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
+//go:generate git-apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
+//go:generate cmake -S ggml -B ggml/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
 //go:generate cmake --build ggml/build/cpu --target server --config Release
-//go:generate cmake --fresh -S gguf -B gguf/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
+
+//go:generate git submodule update --force gguf
+//go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
 //go:generate cmake --build gguf/build/cpu --target server --config Release

llm/llama.cpp/generate_darwin_arm64.go

 package llm
 
 //go:generate git submodule init
-//go:generate git submodule update --force ggml gguf
-//go:generate git -C ggml apply ../ggml_patch/0001-add-detokenize-endpoint.patch
-//go:generate git -C ggml apply ../ggml_patch/0002-34B-model-support.patch
-//go:generate git -C ggml apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
-//go:generate git -C ggml apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
-//go:generate cmake --fresh -S ggml -B ggml/build/gpu -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
+
+//go:generate git submodule update --force ggml
+//go:generate -command git-apply git -C ggml apply
+//go:generate git-apply ../ggml_patch/0001-add-detokenize-endpoint.patch
+//go:generate git-apply ../ggml_patch/0002-34B-model-support.patch
+//go:generate git-apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
+//go:generate git-apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
+//go:generate cmake -S ggml -B ggml/build/gpu -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
 //go:generate cmake --build ggml/build/gpu --target server --config Release
+
+//go:generate git submodule update --force gguf
 //go:generate cmake -S gguf -B gguf/build/gpu -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
 //go:generate cmake --build gguf/build/gpu --target server --config Release

llm/llama.cpp/generate_linux.go

+package llm
+
+//go:generate git submodule init
+
+//go:generate git submodule update --force ggml
+//go:generate -command git-apply git -C ggml apply
+//go:generate git-apply ../ggml_patch/0001-add-detokenize-endpoint.patch
+//go:generate git-apply ../ggml_patch/0002-34B-model-support.patch
+//go:generate cmake -S ggml -B ggml/build/gpu -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on 
+//go:generate cmake --build ggml/build/gpu --target server --config Release
+
+//go:generate git submodule update --force gguf
+//go:generate cmake -S gguf -B gguf/build/gpu -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on 
+//go:generate cmake --build gguf/build/gpu --target server --config Release

llm/llama.go

@@ -58,6 +58,12 @@ func chooseRunner(gpuPath, cpuPath string) string {
 		if llamaPath == osPath(gpuPath) {
 			files = append(files, "ggml-metal.metal")
 		}
+	case "linux":
+		// check if there is a GPU available
+		if _, err := CheckVRAM(); errors.Is(err, errNoGPU) {
+			// this error was logged on start-up, so we don't need to log it again
+			llamaPath = osPath(cpuPath)
+		}
 	}
 
 	for _, f := range files {
@@ -218,6 +224,72 @@ type llama struct {
 	Running
 }
 
+var errNoGPU = errors.New("nvidia-smi command failed")
+
+// CheckVRAM returns the available VRAM in MiB on Linux machines with NVIDIA GPUs
+func CheckVRAM() (int, error) {
+	cmd := exec.Command("nvidia-smi", "--query-gpu=memory.total", "--format=csv,noheader,nounits")
+	var stdout bytes.Buffer
+	cmd.Stdout = &stdout
+	err := cmd.Run()
+	if err != nil {
+		return 0, errNoGPU
+	}
+
+	var total int
+	scanner := bufio.NewScanner(&stdout)
+	for scanner.Scan() {
+		line := scanner.Text()
+		vram, err := strconv.Atoi(line)
+		if err != nil {
+			return 0, fmt.Errorf("failed to parse available VRAM: %v", err)
+		}
+
+		total += vram
+	}
+
+	return total, nil
+}
+
+func NumGPU(opts api.Options) int {
+	if opts.NumGPU != -1 {
+		return opts.NumGPU
+	}
+	n := 1 // default to enable metal on macOS
+	if runtime.GOOS == "linux" {
+		vram, err := CheckVRAM()
+		if err != nil {
+			if err.Error() != "nvidia-smi command failed" {
+				log.Print(err.Error())
+			}
+			// nvidia driver not installed or no nvidia GPU found
+			return 0
+		}
+		// TODO: this is a very rough heuristic, better would be to calculate this based on number of layers and context size
+		switch {
+		case vram < 500:
+			log.Printf("WARNING: Low VRAM detected, disabling GPU")
+			n = 0
+		case vram < 1000:
+			n = 4
+		case vram < 2000:
+			n = 8
+		case vram < 4000:
+			n = 12
+		case vram < 8000:
+			n = 16
+		case vram < 12000:
+			n = 24
+		case vram < 16000:
+			n = 32
+		default:
+			n = 48
+		}
+		log.Printf("%d MB VRAM available, loading %d GPU layers", vram, n)
+	}
+	return n
+}
+
 func newLlama(model string, adapters []string, runner ModelRunner, opts api.Options) (*llama, error) {
 	if _, err := os.Stat(model); err != nil {
 		return nil, err
@@ -237,7 +309,7 @@ func newLlama(model string, adapters []string, runner ModelRunner, opts api.Opti
 		"--rope-freq-base", fmt.Sprintf("%f", opts.RopeFrequencyBase),
 		"--rope-freq-scale", fmt.Sprintf("%f", opts.RopeFrequencyScale),
 		"--batch-size", fmt.Sprintf("%d", opts.NumBatch),
-		"--n-gpu-layers", fmt.Sprintf("%d", opts.NumGPU),
+		"--n-gpu-layers", fmt.Sprintf("%d", NumGPU(opts)),
 		"--embedding",
 	}
 
@@ -305,7 +377,7 @@ func newLlama(model string, adapters []string, runner ModelRunner, opts api.Opti
 func waitForServer(llm *llama) error {
 	// wait for the server to start responding
 	start := time.Now()
-	expiresAt := time.Now().Add(30 * time.Second)
+	expiresAt := time.Now().Add(45 * time.Second)
 	ticker := time.NewTicker(200 * time.Millisecond)
 
 	log.Print("waiting for llama.cpp server to start responding")

server/routes.go

@@ -12,6 +12,7 @@ import (
 	"os/signal"
 	"path/filepath"
 	"reflect"
+	"runtime"
 	"strconv"
 	"strings"
 	"sync"
@@ -548,6 +549,13 @@ func Serve(ln net.Listener, origins []string) error {
 		os.Exit(0)
 	}()
 
+	if runtime.GOOS == "linux" {
+		// check compatibility to log warnings
+		if _, err := llm.CheckVRAM(); err != nil {
+			log.Printf("Warning: GPU support not enabled, you may need to install GPU drivers: %v", err)
+		}
+	}
+
 	return s.Serve(ln)
 }