feat(llama.cpp): do not specify backends to autoload and add llama.cpp variants (#2232)

* feat(initializer): do not specify backends to autoload We can simply try to autoload the backends extracted in the asset dir. This will allow to build variants of the same backend (for e.g. with different instructions sets), so to have a single binary for all the variants. Signed-off-by: mudler <mudler@localai.io> * refactor(prepare): refactor out llama.cpp prepare steps Make it so are idempotent and that we can re-build Signed-off-by: mudler <mudler@localai.io> * [TEST] feat(build): build noavx version along Signed-off-by: mudler <mudler@localai.io> * build: make build parallel Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * build: do not override CMAKE_ARGS Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * build: add fallback variant Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * Fixups Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * fix(huggingface-langchain): fail if no token is set Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * fix(huggingface-langchain): rename Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * fix: do not autoload local-store Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * fix: give priority between the listed backends Signed-off-by: Ettore Di Giacinto <mudler@localai.io> --------- Signed-off-by: mudler <mudler@localai.io> Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
2024-12-18 20:27:57 +00:00 · 2024-05-04 17:56:12 +02:00 · 2024-05-04 17:56:12 +02:00 · 530bec9c64
commit 530bec9c64
parent fa10302dd2
7 changed files with 161 additions and 54 deletions
--- a/41
+++ b/41
@ -152,9 +152,11 @@ ifeq ($(findstring tts,$(GO_TAGS)),tts)
 	OPTIONAL_GRPC+=backend-assets/grpc/piper
 endif

-ALL_GRPC_BACKENDS=backend-assets/grpc/langchain-huggingface
+ALL_GRPC_BACKENDS=backend-assets/grpc/huggingface
 ALL_GRPC_BACKENDS+=backend-assets/grpc/bert-embeddings
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp
+ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-noavx
+ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-fallback
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-ggml
 ALL_GRPC_BACKENDS+=backend-assets/grpc/gpt4all
 ALL_GRPC_BACKENDS+=backend-assets/grpc/rwkv
@ -293,6 +295,7 @@ clean: ## Remove build related file
 	rm -rf backend-assets/*
 	$(MAKE) -C backend/cpp/grpc clean
 	$(MAKE) -C backend/cpp/llama clean
+	rm -rf backend/cpp/llama-* || true
 	$(MAKE) dropreplace
 	$(MAKE) protogen-clean
 	rmdir pkg/grpc/proto || true
@ -311,7 +314,7 @@ build: prepare backend-assets grpcs ## Build the project
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o $(BINARY_NAME) ./

 build-minimal:
-	BUILD_GRPC_FOR_BACKEND_LLAMA=true GRPC_BACKENDS=backend-assets/grpc/llama-cpp GO_TAGS=none $(MAKE) build
+	BUILD_GRPC_FOR_BACKEND_LLAMA=true GRPC_BACKENDS="backend-assets/grpc/llama-cpp" GO_TAGS=none $(MAKE) build

 build-api:
 	BUILD_GRPC_FOR_BACKEND_LLAMA=true BUILD_API_ONLY=true GO_TAGS=none $(MAKE) build
@ -616,8 +619,8 @@ backend-assets/grpc/gpt4all: sources/gpt4all sources/gpt4all/gpt4all-bindings/go
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/sources/gpt4all/gpt4all-bindings/golang/ LIBRARY_PATH=$(CURDIR)/sources/gpt4all/gpt4all-bindings/golang/ \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/gpt4all ./backend/go/llm/gpt4all/

-backend-assets/grpc/langchain-huggingface: backend-assets/grpc
-	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/langchain-huggingface ./backend/go/llm/langchain/
+backend-assets/grpc/huggingface: backend-assets/grpc
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/huggingface ./backend/go/llm/langchain/

 backend/cpp/llama/llama.cpp:
 	LLAMA_VERSION=$(CPPLLAMA_VERSION) $(MAKE) -C backend/cpp/llama llama.cpp
@ -629,7 +632,7 @@ ADDED_CMAKE_ARGS=-Dabsl_DIR=${INSTALLED_LIB_CMAKE}/absl \
 				 -Dutf8_range_DIR=${INSTALLED_LIB_CMAKE}/utf8_range \
 				 -DgRPC_DIR=${INSTALLED_LIB_CMAKE}/grpc \
 				 -DCMAKE_CXX_STANDARD_INCLUDE_DIRECTORIES=${INSTALLED_PACKAGES}/include
-backend/cpp/llama/grpc-server:
+build-llama-cpp-grpc-server:
 # Conditionally build grpc for the llama backend to use if needed
 ifdef BUILD_GRPC_FOR_BACKEND_LLAMA
 	$(MAKE) -C backend/cpp/grpc build
@ -638,19 +641,37 @@ ifdef BUILD_GRPC_FOR_BACKEND_LLAMA
 	PATH="${INSTALLED_PACKAGES}/bin:${PATH}" \
 	CMAKE_ARGS="${CMAKE_ARGS} ${ADDED_CMAKE_ARGS}" \
 	LLAMA_VERSION=$(CPPLLAMA_VERSION) \
-	$(MAKE) -C backend/cpp/llama grpc-server
+	$(MAKE) -C backend/cpp/${VARIANT} grpc-server
 else
 	echo "BUILD_GRPC_FOR_BACKEND_LLAMA is not defined."
-	LLAMA_VERSION=$(CPPLLAMA_VERSION) $(MAKE) -C backend/cpp/llama grpc-server
+	LLAMA_VERSION=$(CPPLLAMA_VERSION) $(MAKE) -C backend/cpp/${VARIANT} grpc-server
 endif

-backend-assets/grpc/llama-cpp: backend-assets/grpc backend/cpp/llama/grpc-server
-	cp -rfv backend/cpp/llama/grpc-server backend-assets/grpc/llama-cpp
+backend-assets/grpc/llama-cpp: backend-assets/grpc
+	$(info ${GREEN}I llama-cpp build info:standard${RESET})
+	cp -rf backend/cpp/llama backend/cpp/llama-default
+	$(MAKE) -C backend/cpp/llama-default purge
+	$(MAKE) VARIANT="llama-default" build-llama-cpp-grpc-server
+	cp -rfv backend/cpp/llama-default/grpc-server backend-assets/grpc/llama-cpp
 # TODO: every binary should have its own folder instead, so can have different metal implementations
 ifeq ($(BUILD_TYPE),metal)
-	cp backend/cpp/llama/llama.cpp/build/bin/default.metallib backend-assets/grpc/
+	cp backend/cpp/llama-default/llama.cpp/build/bin/default.metallib backend-assets/grpc/
 endif

+backend-assets/grpc/llama-cpp-noavx: backend-assets/grpc
+	cp -rf backend/cpp/llama backend/cpp/llama-noavx
+	$(MAKE) -C backend/cpp/llama-noavx purge
+	$(info ${GREEN}I llama-cpp build info:noavx${RESET})
+	CMAKE_ARGS="$(CMAKE_ARGS) -DLLAMA_AVX512=OFF -DLLAMA_AVX2=OFF" $(MAKE) VARIANT="llama-noavx" build-llama-cpp-grpc-server
+	cp -rfv backend/cpp/llama-noavx/grpc-server backend-assets/grpc/llama-cpp-noavx
+
+backend-assets/grpc/llama-cpp-fallback: backend-assets/grpc
+	cp -rf backend/cpp/llama backend/cpp/llama-fallback
+	$(MAKE) -C backend/cpp/llama-fallback purge
+	$(info ${GREEN}I llama-cpp build info:fallback${RESET})
+	CMAKE_ARGS="$(CMAKE_ARGS) -DLLAMA_F16C=OFF -DLLAMA_AVX512=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF" $(MAKE) VARIANT="llama-fallback" build-llama-cpp-grpc-server
+	cp -rfv backend/cpp/llama-fallback/grpc-server backend-assets/grpc/llama-cpp-fallback
+
 backend-assets/grpc/llama-ggml: sources/go-llama.cpp sources/go-llama.cpp/libbinding.a backend-assets/grpc
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/sources/go-llama.cpp LIBRARY_PATH=$(CURDIR)/sources/go-llama.cpp \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/llama-ggml ./backend/go/llm/llama-ggml/
--- a/backend/cpp/llama/Makefile
+++ b/backend/cpp/llama/Makefile
@ -43,31 +43,23 @@ llama.cpp:

 llama.cpp/examples/grpc-server: llama.cpp
 	mkdir -p llama.cpp/examples/grpc-server
-	cp -r $(abspath ./)/CMakeLists.txt llama.cpp/examples/grpc-server/
-	cp -r $(abspath ./)/grpc-server.cpp llama.cpp/examples/grpc-server/
-	cp -rfv $(abspath ./)/json.hpp llama.cpp/examples/grpc-server/
-	cp -rfv $(abspath ./)/utils.hpp llama.cpp/examples/grpc-server/
-	echo "add_subdirectory(grpc-server)" >> llama.cpp/examples/CMakeLists.txt
-## XXX: In some versions of CMake clip wasn't being built before llama.
-## This is an hack for now, but it should be fixed in the future.
-	cp -rfv llama.cpp/examples/llava/clip.h llama.cpp/examples/grpc-server/clip.h
-	cp -rfv llama.cpp/examples/llava/llava.cpp llama.cpp/examples/grpc-server/llava.cpp
-	echo '#include "llama.h"' > llama.cpp/examples/grpc-server/llava.h
-	cat llama.cpp/examples/llava/llava.h >> llama.cpp/examples/grpc-server/llava.h
-	cp -rfv llama.cpp/examples/llava/clip.cpp llama.cpp/examples/grpc-server/clip.cpp
+	bash prepare.sh

 rebuild:
-	cp -rfv $(abspath ./)/CMakeLists.txt llama.cpp/examples/grpc-server/
-	cp -rfv $(abspath ./)/grpc-server.cpp llama.cpp/examples/grpc-server/
-	cp -rfv $(abspath ./)/json.hpp llama.cpp/examples/grpc-server/
+	bash prepare.sh
 	rm -rf grpc-server
 	$(MAKE) grpc-server

-clean:
-	rm -rf llama.cpp
+purge:
+	rm -rf llama.cpp/build
+	rm -rf llama.cpp/examples/grpc-server
 	rm -rf grpc-server

+clean: purge
+	rm -rf llama.cpp
+
 grpc-server: llama.cpp llama.cpp/examples/grpc-server
+	@echo "Building grpc-server with $(BUILD_TYPE) build type and $(CMAKE_ARGS)"
 ifneq (,$(findstring sycl,$(BUILD_TYPE)))
 	bash -c "source $(ONEAPI_VARS); \
 	cd llama.cpp && mkdir -p build && cd build && cmake .. $(CMAKE_ARGS) && cmake --build . --config Release"	
--- a/backend/cpp/llama/prepare.sh
+++ b/backend/cpp/llama/prepare.sh
@ -0,0 +1,20 @@
+#!/bin/bash
+
+cp -r CMakeLists.txt llama.cpp/examples/grpc-server/
+cp -r grpc-server.cpp llama.cpp/examples/grpc-server/
+cp -rfv json.hpp llama.cpp/examples/grpc-server/
+cp -rfv utils.hpp llama.cpp/examples/grpc-server/
+    
+if grep -q "grpc-server" llama.cpp/examples/CMakeLists.txt; then
+    echo "grpc-server already added"
+else
+    echo "add_subdirectory(grpc-server)" >> llama.cpp/examples/CMakeLists.txt
+fi
+
+## XXX: In some versions of CMake clip wasn't being built before llama.
+## This is an hack for now, but it should be fixed in the future.
+cp -rfv llama.cpp/examples/llava/clip.h llama.cpp/examples/grpc-server/clip.h
+cp -rfv llama.cpp/examples/llava/llava.cpp llama.cpp/examples/grpc-server/llava.cpp
+echo '#include "llama.h"' > llama.cpp/examples/grpc-server/llava.h
+cat llama.cpp/examples/llava/llava.h >> llama.cpp/examples/grpc-server/llava.h
+cp -rfv llama.cpp/examples/llava/clip.cpp llama.cpp/examples/grpc-server/clip.cpp
--- a/backend/go/llm/langchain/langchain.go
+++ b/backend/go/llm/langchain/langchain.go
@ -4,6 +4,7 @@ package main
 // It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
 import (
 	"fmt"
+	"os"

 	"github.com/go-skynet/LocalAI/pkg/grpc/base"
 	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
@ -18,9 +19,14 @@ type LLM struct {
 }

 func (llm *LLM) Load(opts *pb.ModelOptions) error {
-	llm.langchain, _ = langchain.NewHuggingFace(opts.Model)
+	var err error
+	hfToken := os.Getenv("HUGGINGFACEHUB_API_TOKEN")
+	if hfToken == "" {
+		return fmt.Errorf("no huggingface token provided")
+	}
+	llm.langchain, err = langchain.NewHuggingFace(opts.Model, hfToken)
 	llm.model = opts.Model
-	return nil
+	return err
 }

 func (llm *LLM) Predict(opts *pb.PredictOptions) (string, error) {
--- a/core/http/app_test.go
+++ b/core/http/app_test.go
@ -787,11 +787,11 @@ var _ = Describe("API test", func() {
 		})

 		It("returns errors", func() {
-			backends := len(model.AutoLoadBackends) + 1 // +1 for huggingface
 			_, err := client.CreateCompletion(context.TODO(), openai.CompletionRequest{Model: "foomodel", Prompt: testPrompt})
 			Expect(err).To(HaveOccurred())
-			Expect(err.Error()).To(ContainSubstring(fmt.Sprintf("error, status code: 500, message: could not load model - all backends returned error: %d errors occurred:", backends)))
+			Expect(err.Error()).To(ContainSubstring("error, status code: 500, message: could not load model - all backends returned error:"))
 		})
+
 		It("transcribes audio", func() {
 			if runtime.GOOS != "linux" {
 				Skip("test supported only on linux")
--- a/pkg/langchain/huggingface.go
+++ b/pkg/langchain/huggingface.go
@ -2,6 +2,7 @@ package langchain

 import (
 	"context"
+	"fmt"

 	"github.com/tmc/langchaingo/llms"
 	"github.com/tmc/langchaingo/llms/huggingface"
@ -9,11 +10,16 @@ import (

 type HuggingFace struct {
 	modelPath string
+	token     string
 }

-func NewHuggingFace(repoId string) (*HuggingFace, error) {
+func NewHuggingFace(repoId, token string) (*HuggingFace, error) {
+	if token == "" {
+		return nil, fmt.Errorf("no huggingface token provided")
+	}
 	return &HuggingFace{
 		modelPath: repoId,
+		token:     token,
 	}, nil
 }

@ -21,7 +27,7 @@ func (s *HuggingFace) PredictHuggingFace(text string, opts ...PredictOption) (*P
 	po := NewPredictOptions(opts...)

 	// Init client
-	llm, err := huggingface.New()
+	llm, err := huggingface.New(huggingface.WithToken(s.token))
 	if err != nil {
 		return nil, err
 	}
--- a/pkg/model/initializers.go
+++ b/pkg/model/initializers.go
@ -2,14 +2,15 @@ package model

 import (
 	"context"
+	"errors"
 	"fmt"
 	"os"
 	"path/filepath"
+	"slices"
 	"strings"
 	"time"

 	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
-	"github.com/hashicorp/go-multierror"
 	"github.com/phayes/freeport"
 	"github.com/rs/zerolog/log"
 )
@ -18,11 +19,15 @@ var Aliases map[string]string = map[string]string{
 	"go-llama":              LLamaCPP,
 	"llama":                 LLamaCPP,
 	"embedded-store":        LocalStoreBackend,
+	"langchain-huggingface": LCHuggingFaceBackend,
 }

 const (
 	LlamaGGML = "llama-ggml"
 	LLamaCPP  = "llama-cpp"
+
+	LLamaCPPFallback = "llama-cpp-fallback"
+
 	Gpt4AllLlamaBackend = "gpt4all-llama"
 	Gpt4AllMptBackend   = "gpt4all-mpt"
 	Gpt4AllJBackend     = "gpt4all-j"
@ -34,21 +39,73 @@ const (
 	StableDiffusionBackend = "stablediffusion"
 	TinyDreamBackend       = "tinydream"
 	PiperBackend           = "piper"
-	LCHuggingFaceBackend   = "langchain-huggingface"
+	LCHuggingFaceBackend   = "huggingface"

 	LocalStoreBackend = "local-store"
 )

-var AutoLoadBackends []string = []string{
-	LLamaCPP,
-	LlamaGGML,
-	Gpt4All,
+func backendPath(assetDir, backend string) string {
+	return filepath.Join(assetDir, "backend-assets", "grpc", backend)
+}
+
+func backendsInAssetDir(assetDir string) ([]string, error) {
+	excludeBackends := []string{"local-store"}
+	entry, err := os.ReadDir(backendPath(assetDir, ""))
+	if err != nil {
+		return nil, err
+	}
+	var backends []string
+ENTRY:
+	for _, e := range entry {
+		for _, exclude := range excludeBackends {
+			if e.Name() == exclude {
+				continue ENTRY
+			}
+		}
+		if !e.IsDir() {
+			backends = append(backends, e.Name())
+		}
+	}
+
+	// order backends from the asset directory.
+	// as we scan for backends, we want to keep some order which backends are tried of.
+	// for example, llama.cpp should be tried first, and we want to keep the huggingface backend at the last.
+
+	// sets a priority list
+	// First has more priority
+	priorityList := []string{
+		// First llama.cpp and llama-ggml
+		LLamaCPP, LLamaCPPFallback, LlamaGGML, Gpt4All,
+	}
+	toTheEnd := []string{
+		// last has to be huggingface
+		LCHuggingFaceBackend,
+		// then bert embeddings
 		BertEmbeddingsBackend,
-	RwkvBackend,
-	WhisperBackend,
-	StableDiffusionBackend,
-	TinyDreamBackend,
-	PiperBackend,
+	}
+	slices.Reverse(priorityList)
+	slices.Reverse(toTheEnd)
+
+	// order certain backends first
+	for _, b := range priorityList {
+		for i, be := range backends {
+			if be == b {
+				backends = append([]string{be}, append(backends[:i], backends[i+1:]...)...)
+				break
+			}
+		}
+	}
+	// make sure that some others are pushed at the end
+	for _, b := range toTheEnd {
+		for i, be := range backends {
+			if be == b {
+				backends = append(append(backends[:i], backends[i+1:]...), be)
+				break
+			}
+		}
+	}
+
+	return backends, nil
 }

 // starts the grpcModelProcess for the backend, and returns a grpc client
@ -99,7 +156,7 @@ func (ml *ModelLoader) grpcModel(backend string, o *Options) func(string, string
 				client = ModelAddress(uri)
 			}
 		} else {
-			grpcProcess := filepath.Join(o.assetDir, "backend-assets", "grpc", backend)
+			grpcProcess := backendPath(o.assetDir, backend)
 			// Check if the file exists
 			if _, err := os.Stat(grpcProcess); os.IsNotExist(err) {
 				return "", fmt.Errorf("grpc process not found: %s. some backends(stablediffusion, tts) require LocalAI compiled with GO_TAGS", grpcProcess)
@ -243,7 +300,12 @@ func (ml *ModelLoader) GreedyLoader(opts ...Option) (grpc.Backend, error) {

 	// autoload also external backends
 	allBackendsToAutoLoad := []string{}
-	allBackendsToAutoLoad = append(allBackendsToAutoLoad, AutoLoadBackends...)
+	autoLoadBackends, err := backendsInAssetDir(o.assetDir)
+	if err != nil {
+		return nil, err
+	}
+	log.Debug().Msgf("Loading from the following backends (in order): %+v", autoLoadBackends)
+	allBackendsToAutoLoad = append(allBackendsToAutoLoad, autoLoadBackends...)
 	for _, b := range o.externalBackends {
 		allBackendsToAutoLoad = append(allBackendsToAutoLoad, b)
 	}
@ -271,10 +333,10 @@ func (ml *ModelLoader) GreedyLoader(opts ...Option) (grpc.Backend, error) {
 			log.Info().Msgf("[%s] Loads OK", b)
 			return model, nil
 		} else if modelerr != nil {
-			err = multierror.Append(err, modelerr)
+			err = errors.Join(err, modelerr)
 			log.Info().Msgf("[%s] Fails: %s", b, modelerr.Error())
 		} else if model == nil {
-			err = multierror.Append(err, fmt.Errorf("backend returned no usable model"))
+			err = errors.Join(err, fmt.Errorf("backend returned no usable model"))
 			log.Info().Msgf("[%s] Fails: %s", b, "backend returned no usable model")
 		}
 	}