Bump llama.cpp, downgrade gpt4all-j (#149)

Makefile

@@ -3,9 +3,9 @@ GOTEST=$(GOCMD) test
 GOVET=$(GOCMD) vet
 BINARY_NAME=local-ai
 # renovate: datasource=github-tags depName=go-skynet/go-llama.cpp
-GOLLAMA_VERSION?=llama.cpp-7f15c5c
+GOLLAMA_VERSION?=llama.cpp-f4cef87
 # renovate: datasource=git-refs packageNameTemplate=https://github.com/go-skynet/go-gpt4all-j.cpp currentValueTemplate=master depNameTemplate=go-gpt4all-j.cpp
-GOGPT4ALLJ_VERSION?=d57834a2d24e8be64c78b9496a870d18393066fd
+GOGPT4ALLJ_VERSION?=1f7bff57f66cb7062e40d0ac3abd2217815e5109
 # renovate: datasource=git-refs packageNameTemplate=https://github.com/go-skynet/go-gpt2.cpp currentValueTemplate=master depNameTemplate=go-gpt2.cpp
 GOGPT2_VERSION?=245a5bfe6708ab80dc5c733dcdbfbe3cfd2acdaa
 
@@ -46,7 +46,7 @@ generic-build: ## Build the project using generic
 ## GPT4ALL-J
 go-gpt4all-j:
 	git clone --recurse-submodules https://github.com/go-skynet/go-gpt4all-j.cpp go-gpt4all-j
-	cd go-gpt4all-j && git checkout -b build $(GOGPT4ALLJ_VERSION)
+	cd go-gpt4all-j && git checkout -b build $(GOGPT4ALLJ_VERSION) && git submodule update --init --recursive --depth 1
 	# This is hackish, but needed as both go-llama and go-gpt4allj have their own version of ggml..
 	@find ./go-gpt4all-j -type f -name "*.c" -exec sed -i'' -e 's/ggml_/ggml_gptj_/g' {} +
 	@find ./go-gpt4all-j -type f -name "*.cpp" -exec sed -i'' -e 's/ggml_/ggml_gptj_/g' {} +
@@ -58,12 +58,12 @@ go-gpt4all-j:
 	@find ./go-gpt4all-j -type f -name "*.cpp" -exec sed -i'' -e 's/::replace/::json_gptj_replace/g' {} +
 
 go-gpt4all-j/libgptj.a: go-gpt4all-j
-	$(MAKE) -C go-gpt4all-j $(GENERIC_PREFIX)libgptj.a example
+	$(MAKE) -C go-gpt4all-j $(GENERIC_PREFIX)libgptj.a
 
 # CEREBRAS GPT
 go-gpt2:
 	git clone --recurse-submodules https://github.com/go-skynet/go-gpt2.cpp go-gpt2
-	cd go-gpt2 && git checkout -b build $(GOGPT2_VERSION)
+	cd go-gpt2 && git checkout -b build $(GOGPT2_VERSION) && git submodule update --init --recursive --depth 1
 	# This is hackish, but needed as both go-llama and go-gpt4allj have their own version of ggml..
 	@find ./go-gpt2 -type f -name "*.c" -exec sed -i'' -e 's/ggml_/ggml_gpt2_/g' {} +
 	@find ./go-gpt2 -type f -name "*.cpp" -exec sed -i'' -e 's/ggml_/ggml_gpt2_/g' {} +

api/prediction.go

@@ -37,7 +37,7 @@ func ModelInference(s string, loader *model.ModelLoader, c Config) (func() (stri
 	// TODO: this is ugly, better identifying the model somehow! however, it is a good stab for a first implementation..
 	model, llamaerr = loader.LoadLLaMAModel(modelFile, llamaOpts...)
 	if llamaerr != nil {
-		gptModel, gptjerr = loader.LoadGPTJModel(modelFile, gptj.SetThreads(c.Threads))
+		gptModel, gptjerr = loader.LoadGPTJModel(modelFile)
 		if gptjerr != nil {
 			gpt2Model, gpt2err = loader.LoadGPT2Model(modelFile)
 			if gpt2err != nil {
@@ -108,12 +108,17 @@ func ModelInference(s string, loader *model.ModelLoader, c Config) (func() (stri
 				gptj.SetTopP(c.TopP),
 				gptj.SetTopK(c.TopK),
 				gptj.SetTokens(c.Maxtokens),
+				gptj.SetThreads(c.Threads),
 			}
 
 			if c.Batch != 0 {
 				predictOptions = append(predictOptions, gptj.SetBatch(c.Batch))
 			}
 
+			if c.Seed != 0 {
+				predictOptions = append(predictOptions, gptj.SetSeed(c.Seed))
+			}
+
 			return gptModel.Predict(
 				s,
 				predictOptions...,

go.mod

@@ -4,7 +4,7 @@ go 1.19
 
 require (
 	github.com/go-skynet/go-gpt2.cpp v0.0.0-20230422085954-245a5bfe6708
-	github.com/go-skynet/go-gpt4all-j.cpp v0.0.0-20230501181604-77bf8c1bccf4
+	github.com/go-skynet/go-gpt4all-j.cpp v0.0.0-20230422090028-1f7bff57f66c
 	github.com/go-skynet/go-llama.cpp v0.0.0-20230430075552-377fd245eae2
 	github.com/gofiber/fiber/v2 v2.44.0
 	github.com/jaypipes/ghw v0.10.0

go.sum

@@ -25,8 +25,6 @@ github.com/go-skynet/go-gpt4all-j.cpp v0.0.0-20230422090028-1f7bff57f66c h1:48I7
 github.com/go-skynet/go-gpt4all-j.cpp v0.0.0-20230422090028-1f7bff57f66c/go.mod h1:5VZ9XbcINI0XcHhkcX8GPK8TplFGAzu1Hrg4tNiMCtI=
 github.com/go-skynet/go-gpt4all-j.cpp v0.0.0-20230501160437-8417608f0e94 h1:klkEyXTg7bpchNNpIQH1f2wX/C17lFLti8isCCC3mYo=
 github.com/go-skynet/go-gpt4all-j.cpp v0.0.0-20230501160437-8417608f0e94/go.mod h1:5VZ9XbcINI0XcHhkcX8GPK8TplFGAzu1Hrg4tNiMCtI=
-github.com/go-skynet/go-gpt4all-j.cpp v0.0.0-20230501181604-77bf8c1bccf4 h1:nN3SXI7zUlxj4dWqbYG8A+yKp6bWmDNItcLYP9gUFbs=
-github.com/go-skynet/go-gpt4all-j.cpp v0.0.0-20230501181604-77bf8c1bccf4/go.mod h1:5VZ9XbcINI0XcHhkcX8GPK8TplFGAzu1Hrg4tNiMCtI=
 github.com/go-skynet/go-llama.cpp v0.0.0-20230428071219-3d084e4299e9 h1:N/0SBefkMFao6GiGhIF7+5EdYOMHn4KnCG2AFcIXPt0=
 github.com/go-skynet/go-llama.cpp v0.0.0-20230428071219-3d084e4299e9/go.mod h1:35AKIEMY+YTKCBJIa/8GZcNGJ2J+nQk1hQiWo/OnEWw=
 github.com/go-skynet/go-llama.cpp v0.0.0-20230429125915-9bf702fe56b9 h1:20/tdOA4+b7Y7lCob+q2sczfOSz0pp+14L32adYJ+uQ=

pkg/model/loader.go

@@ -193,7 +193,7 @@ func (ml *ModelLoader) LoadGPT2Model(modelName string) (*gpt2.GPT2, error) {
 	return model, err
 }
 
-func (ml *ModelLoader) LoadGPTJModel(modelName string, opts ...gptj.ModelOption) (*gptj.GPTJ, error) {
+func (ml *ModelLoader) LoadGPTJModel(modelName string) (*gptj.GPTJ, error) {
 	ml.mu.Lock()
 	defer ml.mu.Unlock()
 
@@ -222,7 +222,7 @@ func (ml *ModelLoader) LoadGPTJModel(modelName string, opts ...gptj.ModelOption)
 	modelFile := filepath.Join(ml.ModelPath, modelName)
 	log.Debug().Msgf("Loading model in memory from file: %s", modelFile)
 
-	model, err := gptj.New(modelFile, opts...)
+	model, err := gptj.New(modelFile)
 	if err != nil {
 		return nil, err
 	}