feat: add llama-master backend

.github/workflows/bump_deps.yaml

@@ -12,6 +12,9 @@ jobs:
           - repository: "go-skynet/go-llama.cpp"
             variable: "GOLLAMA_VERSION"
             branch: "master"
+          - repository: "go-skynet/go-llama.cpp"
+            variable: "GOLLAMA_MASTER_VERSION"
+            branch: "master"
           - repository: "go-skynet/go-ggml-transformers.cpp"
             variable: "GOGGMLTRANSFORMERS_VERSION"
             branch: "master"

Makefile

@@ -7,6 +7,7 @@ BINARY_NAME=local-ai
 # Temporarly pinned to https://github.com/go-skynet/go-llama.cpp/pull/124
 GOLLAMA_VERSION?=cb8d7cd4cb95725a04504a9e3a26dd72a12b69ac
 
+GOLLAMA_MASTER_VERSION?=6c97625cca76aa5fca98a5a138ee4e5fe4797ecb
 # Temporary set a specific version of llama.cpp
 # containing: https://github.com/ggerganov/llama.cpp/pull/1773 and
 # rebased on top of master.
@@ -204,17 +205,25 @@ ifneq ($(LLAMA_CPP_REPO),)
 	cd go-llama && rm -rf llama.cpp && git clone $(LLAMA_CPP_REPO) llama.cpp && cd llama.cpp && git checkout -b build $(LLAMA_CPP_VERSION) && git submodule update --init --recursive --depth 1
 endif
 
+go-llama-master:
+	git clone --recurse-submodules https://github.com/go-skynet/go-llama.cpp go-llama-master
+	cd go-llama-master && git checkout -b build $(GOLLAMA_MASTER_VERSION) && git submodule update --init --recursive --depth 1
+
 go-llama/libbinding.a: go-llama
 	$(MAKE) -C go-llama BUILD_TYPE=$(BUILD_TYPE) libbinding.a
 
+go-llama-master/libbinding.a: go-llama-master
+	$(MAKE) -C go-llama-master BUILD_TYPE=$(BUILD_TYPE) libbinding.a
+
 go-piper/libpiper_binding.a:
 	$(MAKE) -C go-piper libpiper_binding.a example/main
 
-get-sources: go-llama go-ggllm go-ggml-transformers gpt4all go-piper go-rwkv whisper.cpp go-bert bloomz go-stable-diffusion
+get-sources: go-llama go-ggllm go-llama-master go-ggml-transformers gpt4all go-piper go-rwkv whisper.cpp go-bert bloomz go-stable-diffusion
 	touch $@
 
 replace:
 	$(GOCMD) mod edit -replace github.com/go-skynet/go-llama.cpp=$(shell pwd)/go-llama
+	$(GOCMD) mod edit -replace github.com/go-skynet/go-llama.cpp-master=$(shell pwd)/go-llama-master
 	$(GOCMD) mod edit -replace github.com/nomic-ai/gpt4all/gpt4all-bindings/golang=$(shell pwd)/gpt4all/gpt4all-bindings/golang
 	$(GOCMD) mod edit -replace github.com/go-skynet/go-ggml-transformers.cpp=$(shell pwd)/go-ggml-transformers
 	$(GOCMD) mod edit -replace github.com/donomii/go-rwkv.cpp=$(shell pwd)/go-rwkv
@@ -232,6 +241,7 @@ prepare-sources: get-sources replace
 rebuild: ## Rebuilds the project
 	$(GOCMD) clean -cache
 	$(MAKE) -C go-llama clean
+	$(MAKE) -C go-llama-master clean
 	$(MAKE) -C gpt4all/gpt4all-bindings/golang/ clean
 	$(MAKE) -C go-ggml-transformers clean
 	$(MAKE) -C go-rwkv clean
@@ -361,6 +371,10 @@ backend-assets/grpc/llama: backend-assets/grpc go-llama/libbinding.a
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-llama LIBRARY_PATH=$(shell pwd)/go-llama \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/llama ./cmd/grpc/llama/
 
+backend-assets/grpc/llama-master: backend-assets/grpc go-llama-master/libbinding.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-llama-master LIBRARY_PATH=$(shell pwd)/go-llama-master \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/llama-master ./cmd/grpc/llama-master/
+
 backend-assets/grpc/gpt4all: backend-assets/grpc backend-assets/gpt4all gpt4all/gpt4all-bindings/golang/libgpt4all.a
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/gpt4all/gpt4all-bindings/golang/ LIBRARY_PATH=$(shell pwd)/gpt4all/gpt4all-bindings/golang/ \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/gpt4all ./cmd/grpc/gpt4all/
@@ -424,4 +438,4 @@ backend-assets/grpc/whisper: backend-assets/grpc whisper.cpp/libwhisper.a
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/whisper.cpp LIBRARY_PATH=$(shell pwd)/whisper.cpp \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/whisper ./cmd/grpc/whisper/
 
-grpcs: prepare backend-assets/grpc/langchain-huggingface backend-assets/grpc/falcon-ggml backend-assets/grpc/bert-embeddings backend-assets/grpc/falcon backend-assets/grpc/bloomz backend-assets/grpc/llama backend-assets/grpc/gpt4all backend-assets/grpc/dolly backend-assets/grpc/gpt2 backend-assets/grpc/gptj backend-assets/grpc/gptneox backend-assets/grpc/mpt backend-assets/grpc/replit backend-assets/grpc/starcoder backend-assets/grpc/rwkv backend-assets/grpc/whisper $(OPTIONAL_GRPC)
+grpcs: prepare backend-assets/grpc/langchain-huggingface backend-assets/grpc/llama-master backend-assets/grpc/falcon-ggml backend-assets/grpc/bert-embeddings backend-assets/grpc/falcon backend-assets/grpc/bloomz backend-assets/grpc/llama backend-assets/grpc/gpt4all backend-assets/grpc/dolly backend-assets/grpc/gpt2 backend-assets/grpc/gptj backend-assets/grpc/gptneox backend-assets/grpc/mpt backend-assets/grpc/replit backend-assets/grpc/starcoder backend-assets/grpc/rwkv backend-assets/grpc/whisper $(OPTIONAL_GRPC)

api/api_test.go

@@ -555,9 +555,10 @@ var _ = Describe("API test", func() {
 		})
 
 		It("returns errors", func() {
+			backends := len(model.AutoLoadBackends)
 			_, err := client.CreateCompletion(context.TODO(), openai.CompletionRequest{Model: "foomodel", Prompt: "abcdedfghikl"})
 			Expect(err).To(HaveOccurred())
-			Expect(err.Error()).To(ContainSubstring("error, status code: 500, message: could not load model - all backends returned error: 12 errors occurred:"))
+			Expect(err.Error()).To(ContainSubstring(fmt.Sprintf("error, status code: 500, message: could not load model - all backends returned error: %d errors occurred:", backends)))
 		})
 		It("transcribes audio", func() {
 			if runtime.GOOS != "linux" {

cmd/grpc/llama-master/main.go

@@ -0,0 +1,25 @@
+package main
+
+// GRPC Falcon server
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	llama "github.com/go-skynet/LocalAI/pkg/grpc/llm/llama-master"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &llama.LLM{}); err != nil {
+		panic(err)
+	}
+}

go.mod

@@ -39,6 +39,7 @@ require (
 require (
 	github.com/dlclark/regexp2 v1.8.1 // indirect
 	github.com/dsnet/compress v0.0.2-0.20210315054119-f66993602bf5 // indirect
+	github.com/go-skynet/go-llama.cpp-master v0.0.0-20230703203849-ffa57fbc3a12 // indirect
 	github.com/golang/protobuf v1.5.3 // indirect
 	github.com/golang/snappy v0.0.2 // indirect
 	github.com/klauspost/pgzip v1.2.5 // indirect

pkg/grpc/llm/llama-master/llama.go

@@ -0,0 +1,168 @@
+package llama
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+
+	"github.com/go-skynet/LocalAI/pkg/grpc/base"
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+	"github.com/go-skynet/go-llama.cpp-master"
+)
+
+type LLM struct {
+	base.Base
+
+	llama *llama.LLama
+}
+
+func (llm *LLM) Load(opts *pb.ModelOptions) error {
+	llamaOpts := []llama.ModelOption{}
+
+	if opts.ContextSize != 0 {
+		llamaOpts = append(llamaOpts, llama.SetContext(int(opts.ContextSize)))
+	}
+	if opts.F16Memory {
+		llamaOpts = append(llamaOpts, llama.EnableF16Memory)
+	}
+	if opts.Embeddings {
+		llamaOpts = append(llamaOpts, llama.EnableEmbeddings)
+	}
+	if opts.NGPULayers != 0 {
+		llamaOpts = append(llamaOpts, llama.SetGPULayers(int(opts.NGPULayers)))
+	}
+
+	llamaOpts = append(llamaOpts, llama.SetMMap(opts.MMap))
+	llamaOpts = append(llamaOpts, llama.SetMainGPU(opts.MainGPU))
+	llamaOpts = append(llamaOpts, llama.SetTensorSplit(opts.TensorSplit))
+	if opts.NBatch != 0 {
+		llamaOpts = append(llamaOpts, llama.SetNBatch(int(opts.NBatch)))
+	} else {
+		llamaOpts = append(llamaOpts, llama.SetNBatch(512))
+	}
+
+	if opts.NUMA {
+		llamaOpts = append(llamaOpts, llama.EnableNUMA)
+	}
+
+	if opts.LowVRAM {
+		llamaOpts = append(llamaOpts, llama.EnabelLowVRAM)
+	}
+
+	model, err := llama.New(opts.Model, llamaOpts...)
+	llm.llama = model
+	return err
+}
+
+func buildPredictOptions(opts *pb.PredictOptions) []llama.PredictOption {
+	predictOptions := []llama.PredictOption{
+		llama.SetTemperature(float64(opts.Temperature)),
+		llama.SetTopP(float64(opts.TopP)),
+		llama.SetTopK(int(opts.TopK)),
+		llama.SetTokens(int(opts.Tokens)),
+		llama.SetThreads(int(opts.Threads)),
+	}
+
+	if opts.PromptCacheAll {
+		predictOptions = append(predictOptions, llama.EnablePromptCacheAll)
+	}
+
+	if opts.PromptCacheRO {
+		predictOptions = append(predictOptions, llama.EnablePromptCacheRO)
+	}
+
+	// Expected absolute path
+	if opts.PromptCachePath != "" {
+		predictOptions = append(predictOptions, llama.SetPathPromptCache(opts.PromptCachePath))
+	}
+
+	if opts.Mirostat != 0 {
+		predictOptions = append(predictOptions, llama.SetMirostat(int(opts.Mirostat)))
+	}
+
+	if opts.MirostatETA != 0 {
+		predictOptions = append(predictOptions, llama.SetMirostatETA(float64(opts.MirostatETA)))
+	}
+
+	if opts.MirostatTAU != 0 {
+		predictOptions = append(predictOptions, llama.SetMirostatTAU(float64(opts.MirostatTAU)))
+	}
+
+	if opts.Debug {
+		predictOptions = append(predictOptions, llama.Debug)
+	}
+
+	predictOptions = append(predictOptions, llama.SetStopWords(opts.StopPrompts...))
+
+	if opts.PresencePenalty != 0 {
+		predictOptions = append(predictOptions, llama.SetPenalty(float64(opts.PresencePenalty)))
+	}
+
+	if opts.NKeep != 0 {
+		predictOptions = append(predictOptions, llama.SetNKeep(int(opts.NKeep)))
+	}
+
+	if opts.Batch != 0 {
+		predictOptions = append(predictOptions, llama.SetBatch(int(opts.Batch)))
+	}
+
+	if opts.F16KV {
+		predictOptions = append(predictOptions, llama.EnableF16KV)
+	}
+
+	if opts.IgnoreEOS {
+		predictOptions = append(predictOptions, llama.IgnoreEOS)
+	}
+
+	if opts.Seed != 0 {
+		predictOptions = append(predictOptions, llama.SetSeed(int(opts.Seed)))
+	}
+
+	//predictOptions = append(predictOptions, llama.SetLogitBias(c.Seed))
+
+	predictOptions = append(predictOptions, llama.SetFrequencyPenalty(float64(opts.FrequencyPenalty)))
+	predictOptions = append(predictOptions, llama.SetMlock(opts.MLock))
+	predictOptions = append(predictOptions, llama.SetMemoryMap(opts.MMap))
+	predictOptions = append(predictOptions, llama.SetPredictionMainGPU(opts.MainGPU))
+	predictOptions = append(predictOptions, llama.SetPredictionTensorSplit(opts.TensorSplit))
+	predictOptions = append(predictOptions, llama.SetTailFreeSamplingZ(float64(opts.TailFreeSamplingZ)))
+	predictOptions = append(predictOptions, llama.SetTypicalP(float64(opts.TypicalP)))
+	return predictOptions
+}
+
+func (llm *LLM) Predict(opts *pb.PredictOptions) (string, error) {
+	return llm.llama.Predict(opts.Prompt, buildPredictOptions(opts)...)
+}
+
+func (llm *LLM) PredictStream(opts *pb.PredictOptions, results chan string) error {
+	predictOptions := buildPredictOptions(opts)
+
+	predictOptions = append(predictOptions, llama.SetTokenCallback(func(token string) bool {
+		results <- token
+		return true
+	}))
+
+	go func() {
+		_, err := llm.llama.Predict(opts.Prompt, predictOptions...)
+		if err != nil {
+			fmt.Println("err: ", err)
+		}
+		close(results)
+	}()
+
+	return nil
+}
+
+func (llm *LLM) Embeddings(opts *pb.PredictOptions) ([]float32, error) {
+	predictOptions := buildPredictOptions(opts)
+
+	if len(opts.EmbeddingTokens) > 0 {
+		tokens := []int{}
+		for _, t := range opts.EmbeddingTokens {
+			tokens = append(tokens, int(t))
+		}
+		return llm.llama.TokenEmbeddings(tokens, predictOptions...)
+	}
+
+	return llm.llama.Embeddings(opts.Embeddings, predictOptions...)
+}

pkg/model/initializers.go

@@ -37,6 +37,7 @@ const (
 	Gpt4All             = "gpt4all"
 	FalconBackend       = "falcon"
 	FalconGGMLBackend   = "falcon-ggml"
+	LlamaMasterBackend  = "llama-master"
 
 	BertEmbeddingsBackend  = "bert-embeddings"
 	RwkvBackend            = "rwkv"
@@ -47,22 +48,20 @@ const (
 	//GGLLMFalconBackend     = "falcon"
 )
 
-var autoLoadBackends []string = []string{
+var AutoLoadBackends []string = []string{
 	LlamaBackend,
 	Gpt4All,
-	RwkvBackend,
 	FalconBackend,
-	WhisperBackend,
 	GPTNeoXBackend,
 	BertEmbeddingsBackend,
+	LlamaMasterBackend,
 	FalconGGMLBackend,
 	GPTJBackend,
 	Gpt2Backend,
 	DollyBackend,
 	MPTBackend,
 	ReplitBackend,
 	StarcoderBackend,
-	BloomzBackend,
 }
 
 func (ml *ModelLoader) StopGRPC() {
@@ -186,7 +185,7 @@ func (ml *ModelLoader) BackendLoader(opts ...Option) (model *grpc.Client, err er
 
 	backend := strings.ToLower(o.backendString)
 	switch backend {
-	case LlamaBackend, GPTJBackend, DollyBackend,
+	case LlamaBackend, LlamaMasterBackend, GPTJBackend, DollyBackend,
 		MPTBackend, Gpt2Backend, FalconBackend,
 		GPTNeoXBackend, ReplitBackend, StarcoderBackend, BloomzBackend,
 		RwkvBackend, LCHuggingFaceBackend, BertEmbeddingsBackend, FalconGGMLBackend, StableDiffusionBackend, WhisperBackend:
@@ -217,10 +216,7 @@ func (ml *ModelLoader) GreedyLoader(opts ...Option) (*grpc.Client, error) {
 	ml.mu.Unlock()
 	var err error
 
-	for _, b := range autoLoadBackends {
-		if b == BloomzBackend || b == WhisperBackend || b == RwkvBackend { // do not autoload bloomz/whisper/rwkv
-			continue
-		}
+	for _, b := range AutoLoadBackends {
 		log.Debug().Msgf("[%s] Attempting to load", b)
 
 		model, modelerr := ml.BackendLoader(
@@ -236,6 +232,9 @@ func (ml *ModelLoader) GreedyLoader(opts ...Option) (*grpc.Client, error) {
 		} else if modelerr != nil {
 			err = multierror.Append(err, modelerr)
 			log.Debug().Msgf("[%s] Fails: %s", b, modelerr.Error())
+		} else if model == nil {
+			err = multierror.Append(err, modelerr)
+			log.Debug().Msgf("[%s] Fails: %s", b, "backend returned no usable model")
 		}
 	}