Keep whisper models in memory (#233)

Co-authored-by: renovate[bot] <29139614+renovate[bot]@users.noreply.github.com>

.github/workflows/bump_deps.yaml

@@ -27,6 +27,9 @@ jobs:
           - repository: "go-skynet/go-bert.cpp"
             variable: "BERT_VERSION"
             branch: "master"
+          - repository: "go-skynet/bloomz.cpp"
+            variable: "BLOOMZ_VERSION"
+            branch: "main"
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v3

Makefile

@@ -5,11 +5,12 @@ BINARY_NAME=local-ai
 
 GOLLAMA_VERSION?=c03e8adbc45c866e0f6d876af1887d6b01d57eb4
 GOGPT4ALLJ_VERSION?=1f7bff57f66cb7062e40d0ac3abd2217815e5109
-GOGPT2_VERSION?=245a5bfe6708ab80dc5c733dcdbfbe3cfd2acdaa
+GOGPT2_VERSION?=abf038a7d8efa4eefdc7c891f05ad33d4e59e49d
 RWKV_REPO?=https://github.com/donomii/go-rwkv.cpp
-RWKV_VERSION?=af62fcc432be2847acb6e0688b2c2491d6588d58
+RWKV_VERSION?=07166da10cb2a9e8854395a4f210464dcea76e47
 WHISPER_CPP_VERSION?=bf2449dfae35a46b2cd92ab22661ce81a48d4993
 BERT_VERSION?=ec771ec715576ac050263bb7bb74bfd616a5ba13
+BLOOMZ_VERSION?=e9366e82abdfe70565644fbfae9651976714efd1
 
 
 GREEN  := $(shell tput -Txterm setaf 2)
@@ -18,8 +19,8 @@ WHITE  := $(shell tput -Txterm setaf 7)
 CYAN   := $(shell tput -Txterm setaf 6)
 RESET  := $(shell tput -Txterm sgr0)
 
-C_INCLUDE_PATH=$(shell pwd)/go-llama:$(shell pwd)/go-gpt4all-j:$(shell pwd)/go-gpt2:$(shell pwd)/go-rwkv:$(shell pwd)/whisper.cpp:$(shell pwd)/go-bert
-LIBRARY_PATH=$(shell pwd)/go-llama:$(shell pwd)/go-gpt4all-j:$(shell pwd)/go-gpt2:$(shell pwd)/go-rwkv:$(shell pwd)/whisper.cpp:$(shell pwd)/go-bert
+C_INCLUDE_PATH=$(shell pwd)/go-llama:$(shell pwd)/go-gpt4all-j:$(shell pwd)/go-gpt2:$(shell pwd)/go-rwkv:$(shell pwd)/whisper.cpp:$(shell pwd)/go-bert:$(shell pwd)/bloomz
+LIBRARY_PATH=$(shell pwd)/go-llama:$(shell pwd)/go-gpt4all-j:$(shell pwd)/go-gpt2:$(shell pwd)/go-rwkv:$(shell pwd)/whisper.cpp:$(shell pwd)/go-bert:$(shell pwd)/bloomz
 
 # Use this if you want to set the default behavior
 ifndef BUILD_TYPE
@@ -69,6 +70,18 @@ go-rwkv:
 go-rwkv/librwkv.a: go-rwkv
 	cd go-rwkv && cd rwkv.cpp &&	cmake . -DRWKV_BUILD_SHARED_LIBRARY=OFF &&	cmake --build . && 	cp librwkv.a .. && cp ggml/src/libggml.a ..
 
+## bloomz
+bloomz:
+	git clone --recurse-submodules https://github.com/go-skynet/bloomz.cpp bloomz
+	@find ./bloomz -type f -name "*.c" -exec sed -i'' -e 's/ggml_/ggml_bloomz_/g' {} +
+	@find ./bloomz -type f -name "*.cpp" -exec sed -i'' -e 's/ggml_/ggml_bloomz_/g' {} +
+	@find ./bloomz -type f -name "*.h" -exec sed -i'' -e 's/ggml_/ggml_bloomz_/g' {} +
+	@find ./bloomz -type f -name "*.cpp" -exec sed -i'' -e 's/gpt_/gpt_bloomz_/g' {} +
+	@find ./bloomz -type f -name "*.h" -exec sed -i'' -e 's/gpt_/gpt_bloomz_/g' {} +
+
+bloomz/libbloomz.a: bloomz
+	cd bloomz && make libbloomz.a
+
 go-bert/libgobert.a: go-bert
 	$(MAKE) -C go-bert libgobert.a
 
@@ -111,8 +124,9 @@ replace:
 	$(GOCMD) mod edit -replace github.com/donomii/go-rwkv.cpp=$(shell pwd)/go-rwkv
 	$(GOCMD) mod edit -replace github.com/ggerganov/whisper.cpp=$(shell pwd)/whisper.cpp
 	$(GOCMD) mod edit -replace github.com/go-skynet/go-bert.cpp=$(shell pwd)/go-bert
+	$(GOCMD) mod edit -replace github.com/go-skynet/bloomz.cpp=$(shell pwd)/bloomz
 
-prepare-sources: go-llama go-gpt2 go-gpt4all-j go-rwkv whisper.cpp go-bert
+prepare-sources: go-llama go-gpt2 go-gpt4all-j go-rwkv whisper.cpp go-bert bloomz
 	$(GOCMD) mod download
 
 ## GENERIC
@@ -123,9 +137,10 @@ rebuild: ## Rebuilds the project
 	$(MAKE) -C go-rwkv clean
 	$(MAKE) -C whisper.cpp clean
 	$(MAKE) -C go-bert clean
+	$(MAKE) -C bloomz clean
 	$(MAKE) build
 
-prepare: prepare-sources go-llama/libbinding.a go-gpt4all-j/libgptj.a go-bert/libgobert.a go-gpt2/libgpt2.a go-rwkv/librwkv.a whisper.cpp/libwhisper.a replace ## Prepares for building
+prepare: prepare-sources go-llama/libbinding.a go-gpt4all-j/libgptj.a go-bert/libgobert.a go-gpt2/libgpt2.a go-rwkv/librwkv.a whisper.cpp/libwhisper.a bloomz/libbloomz.a replace ## Prepares for building
 
 clean: ## Remove build related file
 	rm -fr ./go-llama
@@ -133,6 +148,7 @@ clean: ## Remove build related file
 	rm -rf ./go-gpt2
 	rm -rf ./go-rwkv
 	rm -rf ./go-bert
+	rm -rf ./bloomz
 	rm -rf $(BINARY_NAME)
 
 ## Build:

README.md

@@ -25,7 +25,7 @@ See [examples on how to integrate LocalAI](https://github.com/go-skynet/LocalAI/
 
 ## News
 
-- 10-05-2023: Added support for fast and accurate embeddings with `bert.cpp` ( https://github.com/go-skynet/LocalAI/pull/222 )
+- 10-05-2023: __1.8.0__ released! 🔥 Added support for fast and accurate embeddings with `bert.cpp` ( https://github.com/go-skynet/LocalAI/pull/222 )
 - 09-05-2023: Added experimental support for transcriptions endpoint ( https://github.com/go-skynet/LocalAI/pull/211 )
 - 08-05-2023: Support for embeddings with models using the `llama.cpp` backend ( https://github.com/go-skynet/LocalAI/pull/207 )
 - 02-05-2023: Support for `rwkv.cpp` models ( https://github.com/go-skynet/LocalAI/pull/158 ) and for `/edits` endpoint

api/api_test.go

@@ -79,7 +79,7 @@ var _ = Describe("API test", func() {
 		It("returns errors", func() {
 			_, 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: 6 errors occurred:"))
+			Expect(err.Error()).To(ContainSubstring("error, status code: 500, message: could not load model - all backends returned error: 9 errors occurred:"))
 		})
 
 	})

api/openai.go

@@ -5,6 +5,7 @@ import (
 	"bytes"
 	"encoding/json"
 	"fmt"
+	"io"
 	"net/http"
 	"os"
 	"path"
@@ -14,7 +15,6 @@ import (
 	model "github.com/go-skynet/LocalAI/pkg/model"
 	"github.com/go-skynet/LocalAI/pkg/whisper"
 	"github.com/gofiber/fiber/v2"
-	"github.com/otiai10/copy"
 	"github.com/rs/zerolog/log"
 	"github.com/valyala/fasthttp"
 )
@@ -409,24 +409,41 @@ func transcriptEndpoint(cm ConfigMerger, debug bool, loader *model.ModelLoader,
 		if err != nil {
 			return c.Status(http.StatusBadRequest).JSON(fiber.Map{"error": err.Error()})
 		}
+		f, err := file.Open()
+		if err != nil {
+			return c.Status(http.StatusBadRequest).JSON(fiber.Map{"error": err.Error()})
+		}
+		defer f.Close()
 		log.Debug().Msgf("Audio file: %+v", file)
 
 		dir, err := os.MkdirTemp("", "whisper")
+
 		if err != nil {
 			return err
 		}
 		defer os.RemoveAll(dir)
 
 		dst := filepath.Join(dir, path.Base(file.Filename))
-		if err := copy.Copy(file.Filename, dst); err != nil {
+		dstFile, err := os.Create(dst)
+		if err != nil {
+			return c.Status(http.StatusBadRequest).JSON(fiber.Map{"error": err.Error()})
+		}
+
+		if _, err := io.Copy(dstFile, f); err != nil {
+			log.Debug().Msgf("Audio file %+v - %+v - err %+v", file.Filename, dst, err)
 			return err
 		}
 
 		log.Debug().Msgf("Audio file copied to: %+v", dst)
 
-		tr, err := whisper.Transcript(filepath.Join(loader.ModelPath, config.Model), dst, input.Language)
+		whisperModel, err := loader.WhisperLoader("whisper", config.Model)
 		if err != nil {
-			return err
+			return c.Status(http.StatusBadRequest).JSON(fiber.Map{"error": err.Error()})
+		}
+
+		tr, err := whisper.Transcript(whisperModel, dst, input.Language)
+		if err != nil {
+			return c.Status(http.StatusBadRequest).JSON(fiber.Map{"error": err.Error()})
 		}
 
 		log.Debug().Msgf("Trascribed: %+v", tr)

api/prediction.go

@@ -8,6 +8,7 @@ import (
 
 	"github.com/donomii/go-rwkv.cpp"
 	model "github.com/go-skynet/LocalAI/pkg/model"
+	"github.com/go-skynet/bloomz.cpp"
 	bert "github.com/go-skynet/go-bert.cpp"
 	gpt2 "github.com/go-skynet/go-gpt2.cpp"
 	gptj "github.com/go-skynet/go-gpt4all-j.cpp"
@@ -198,6 +199,50 @@ func ModelInference(s string, loader *model.ModelLoader, c Config, tokenCallback
 
 			return response, nil
 		}
+	case *gpt2.RedPajama:
+		fn = func() (string, error) {
+			// Generate the prediction using the language model
+			predictOptions := []gpt2.PredictOption{
+				gpt2.SetTemperature(c.Temperature),
+				gpt2.SetTopP(c.TopP),
+				gpt2.SetTopK(c.TopK),
+				gpt2.SetTokens(c.Maxtokens),
+				gpt2.SetThreads(c.Threads),
+			}
+
+			if c.Batch != 0 {
+				predictOptions = append(predictOptions, gpt2.SetBatch(c.Batch))
+			}
+
+			if c.Seed != 0 {
+				predictOptions = append(predictOptions, gpt2.SetSeed(c.Seed))
+			}
+
+			return model.Predict(
+				s,
+				predictOptions...,
+			)
+		}
+	case *bloomz.Bloomz:
+		fn = func() (string, error) {
+			// Generate the prediction using the language model
+			predictOptions := []bloomz.PredictOption{
+				bloomz.SetTemperature(c.Temperature),
+				bloomz.SetTopP(c.TopP),
+				bloomz.SetTopK(c.TopK),
+				bloomz.SetTokens(c.Maxtokens),
+				bloomz.SetThreads(c.Threads),
+			}
+
+			if c.Seed != 0 {
+				predictOptions = append(predictOptions, bloomz.SetSeed(c.Seed))
+			}
+
+			return model.Predict(
+				s,
+				predictOptions...,
+			)
+		}
 	case *gpt2.StableLM:
 		fn = func() (string, error) {
 			// Generate the prediction using the language model
@@ -222,6 +267,30 @@ func ModelInference(s string, loader *model.ModelLoader, c Config, tokenCallback
 				predictOptions...,
 			)
 		}
+	case *gpt2.Dolly:
+		fn = func() (string, error) {
+			// Generate the prediction using the language model
+			predictOptions := []gpt2.PredictOption{
+				gpt2.SetTemperature(c.Temperature),
+				gpt2.SetTopP(c.TopP),
+				gpt2.SetTopK(c.TopK),
+				gpt2.SetTokens(c.Maxtokens),
+				gpt2.SetThreads(c.Threads),
+			}
+
+			if c.Batch != 0 {
+				predictOptions = append(predictOptions, gpt2.SetBatch(c.Batch))
+			}
+
+			if c.Seed != 0 {
+				predictOptions = append(predictOptions, gpt2.SetSeed(c.Seed))
+			}
+
+			return model.Predict(
+				s,
+				predictOptions...,
+			)
+		}
 	case *gpt2.GPT2:
 		fn = func() (string, error) {
 			// Generate the prediction using the language model

go.mod

@@ -4,12 +4,13 @@ go 1.19
 
 require (
 	github.com/donomii/go-rwkv.cpp v0.0.0-20230503112711-af62fcc432be
-	github.com/ggerganov/whisper.cpp/bindings/go v0.0.0-20230508180809-bf2449dfae35
+	github.com/ggerganov/whisper.cpp/bindings/go v0.0.0-20230509153812-1d17cd5bb37a
 	github.com/go-audio/wav v1.1.0
+	github.com/go-skynet/bloomz.cpp v0.0.0-20230510195113-ad7e89a0885f
 	github.com/go-skynet/go-bert.cpp v0.0.0-20230510101404-7bb183b147ea
-	github.com/go-skynet/go-gpt2.cpp v0.0.0-20230422085954-245a5bfe6708
+	github.com/go-skynet/go-gpt2.cpp v0.0.0-20230509180201-d49823284cc6
 	github.com/go-skynet/go-gpt4all-j.cpp v0.0.0-20230422090028-1f7bff57f66c
-	github.com/go-skynet/go-llama.cpp v0.0.0-20230509080828-f4d26f43f1d3
+	github.com/go-skynet/go-llama.cpp v0.0.0-20230510072905-70593fccbe4b
 	github.com/gofiber/fiber/v2 v2.45.0
 	github.com/hashicorp/go-multierror v1.1.1
 	github.com/onsi/ginkgo/v2 v2.9.4

go.sum

@@ -16,8 +16,12 @@ github.com/creack/pty v1.1.9/go.mod h1:oKZEueFk5CKHvIhNR5MUki03XCEU+Q6VDXinZuGJ3
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/donomii/go-rwkv.cpp v0.0.0-20230503112711-af62fcc432be h1:3Hic97PY6hcw/SY44RuR7kyONkxd744RFeRrqckzwNQ=
+github.com/donomii/go-rwkv.cpp v0.0.0-20230503112711-af62fcc432be/go.mod h1:gWy7FIWioqYmYxkaoFyBnaKApeZVrUkHhv9EV9pz4dM=
 github.com/ggerganov/whisper.cpp/bindings/go v0.0.0-20230508180809-bf2449dfae35 h1:sMg/SgnMPS/HNUO/2kGm72vl8R9TmNIwgLFr2TNwR3g=
 github.com/ggerganov/whisper.cpp/bindings/go v0.0.0-20230508180809-bf2449dfae35/go.mod h1:QIjZ9OktHFG7p+/m3sMvrAJKKdWrr1fZIK0rM6HZlyo=
+github.com/ggerganov/whisper.cpp/bindings/go v0.0.0-20230509153812-1d17cd5bb37a h1:MlyiDLNCM/wjbv8U5Elj18NvaAgl61SGiRUpqQz5dfs=
+github.com/ggerganov/whisper.cpp/bindings/go v0.0.0-20230509153812-1d17cd5bb37a/go.mod h1:QIjZ9OktHFG7p+/m3sMvrAJKKdWrr1fZIK0rM6HZlyo=
 github.com/go-audio/audio v1.0.0 h1:zS9vebldgbQqktK4H0lUqWrG8P0NxCJVqcj7ZpNnwd4=
 github.com/go-audio/audio v1.0.0/go.mod h1:6uAu0+H2lHkwdGsAY+j2wHPNPpPoeg5AaEFh9FlA+Zs=
 github.com/go-audio/riff v1.0.0 h1:d8iCGbDvox9BfLagY94fBynxSPHO80LmZCaOsmKxokA=
@@ -36,6 +40,16 @@ github.com/go-openapi/spec v0.20.4/go.mod h1:faYFR1CvsJZ0mNsmsphTMSoRrNV3TEDoAM7
 github.com/go-openapi/swag v0.19.5/go.mod h1:POnQmlKehdgb5mhVOsnJFsivZCEZ/vjK9gh66Z9tfKk=
 github.com/go-openapi/swag v0.19.15 h1:D2NRCBzS9/pEY3gP9Nl8aDqGUcPFrwG2p+CNFrLyrCM=
 github.com/go-openapi/swag v0.19.15/go.mod h1:QYRuS/SOXUCsnplDa677K7+DxSOj6IPNl/eQntq43wQ=
+github.com/go-skynet/bloomz.cpp v0.0.0-20230510195113-ad7e89a0885f h1:GW8RQa1RVeDF1dOuAP/y6xWVC+BRtf9tJOuEza6Asbg=
+github.com/go-skynet/bloomz.cpp v0.0.0-20230510195113-ad7e89a0885f/go.mod h1:wc0fJ9V04yiYTfgKvE5RUUSRQ5Kzi0Bo4I+U3nNOUuA=
+github.com/go-skynet/go-bert.cpp v0.0.0-20230510101404-7bb183b147ea h1:8Isk9D+Auth5OuXVAQPC3MO+5zF/2S7mvs2JZLw6a+8=
+github.com/go-skynet/go-bert.cpp v0.0.0-20230510101404-7bb183b147ea/go.mod h1:NHwIVvsg7Jh6p0M4uBLVmSMEaPUia6O6yjXUpLWVJmQ=
+github.com/go-skynet/go-gpt2.cpp v0.0.0-20230509180201-d49823284cc6 h1:XshpypO6ekU09CI19vuzke2a1Es1lV5ZaxA7CUehu0E=
+github.com/go-skynet/go-gpt2.cpp v0.0.0-20230509180201-d49823284cc6/go.mod h1:1Wj/xbkMfwQSOrhNYK178IzqQHstZbRfhx4s8p1M5VM=
+github.com/go-skynet/go-gpt4all-j.cpp v0.0.0-20230422090028-1f7bff57f66c h1:48I7jpLNGiQeBmF0SFVVbREh8vlG0zN13v9LH5ctXis=
+github.com/go-skynet/go-gpt4all-j.cpp v0.0.0-20230422090028-1f7bff57f66c/go.mod h1:5VZ9XbcINI0XcHhkcX8GPK8TplFGAzu1Hrg4tNiMCtI=
+github.com/go-skynet/go-llama.cpp v0.0.0-20230510072905-70593fccbe4b h1:qqxrjY8fYDXQahmCMTCACahm1tbiqHLPUHALkFLyBfo=
+github.com/go-skynet/go-llama.cpp v0.0.0-20230510072905-70593fccbe4b/go.mod h1:DLfsPD7tYYnpksERH83HSf7qVNW3FIwmz7/zfYO0/6I=
 github.com/go-task/slim-sprig v0.0.0-20230315185526-52ccab3ef572 h1:tfuBGBXKqDEevZMzYi5KSi8KkcZtzBcTgAUUtapy0OI=
 github.com/go-task/slim-sprig v0.0.0-20230315185526-52ccab3ef572/go.mod h1:9Pwr4B2jHnOSGXyyzV8ROjYa2ojvAY6HCGYYfMoC3Ls=
 github.com/godbus/dbus/v5 v5.0.4/go.mod h1:xhWf0FNVPg57R7Z0UbKHbJfkEywrmjJnf7w5xrFpKfA=
@@ -187,9 +201,8 @@ golang.org/x/xerrors v0.0.0-20200804184101-5ec99f83aff1/go.mod h1:I/5z698sn9Ka8T
 google.golang.org/protobuf v1.28.0 h1:w43yiav+6bVFTBQFZX0r7ipe9JQ1QsbMgHwbBziscLw=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/check.v1 v1.0.0-20180628173108-788fd7840127/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/check.v1 v1.0.0-20200227125254-8fa46927fb4f h1:BLraFXnmrev5lT+xlilqcH8XK9/i0At2xKjWk4p6zsU=
 gopkg.in/check.v1 v1.0.0-20200227125254-8fa46927fb4f/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
-gopkg.in/check.v1 v1.0.0-20200902074654-038fdea0a05b h1:QRR6H1YWRnHb4Y/HeNFCTJLFVxaq6wH4YuVdsUOr75U=
-gopkg.in/check.v1 v1.0.0-20200902074654-038fdea0a05b/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/yaml.v2 v2.2.2/go.mod h1:hI93XBmqTisBFMUTm0b8Fm+jr3Dg1NNxqwp+5A1VGuI=
 gopkg.in/yaml.v2 v2.4.0 h1:D8xgwECY7CYvx+Y2n4sBz93Jn9JRvxdiyyo8CTfuKaY=
 gopkg.in/yaml.v2 v2.4.0/go.mod h1:RDklbk79AGWmwhnvt/jBztapEOGDOx6ZbXqjP6csGnQ=

pkg/model/loader.go

@@ -10,14 +10,15 @@ import (
 	"sync"
 	"text/template"
 
-	"github.com/hashicorp/go-multierror"
-	"github.com/rs/zerolog/log"
-
 	rwkv "github.com/donomii/go-rwkv.cpp"
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
+	bloomz "github.com/go-skynet/bloomz.cpp"
 	bert "github.com/go-skynet/go-bert.cpp"
 	gpt2 "github.com/go-skynet/go-gpt2.cpp"
 	gptj "github.com/go-skynet/go-gpt4all-j.cpp"
 	llama "github.com/go-skynet/go-llama.cpp"
+	"github.com/hashicorp/go-multierror"
+	"github.com/rs/zerolog/log"
 )
 
 type ModelLoader struct {
@@ -28,10 +29,13 @@ type ModelLoader struct {
 	gptmodels         map[string]*gptj.GPTJ
 	gpt2models        map[string]*gpt2.GPT2
 	gptstablelmmodels map[string]*gpt2.StableLM
+	dollymodels       map[string]*gpt2.Dolly
+	redpajama         map[string]*gpt2.RedPajama
 	rwkv              map[string]*rwkv.RwkvState
+	bloomz            map[string]*bloomz.Bloomz
 	bert              map[string]*bert.Bert
-
-	promptsTemplates map[string]*template.Template
+	promptsTemplates  map[string]*template.Template
+	whisperModels     map[string]whisper.Model
 }
 
 func NewModelLoader(modelPath string) *ModelLoader {
@@ -40,10 +44,14 @@ func NewModelLoader(modelPath string) *ModelLoader {
 		gpt2models:        make(map[string]*gpt2.GPT2),
 		gptmodels:         make(map[string]*gptj.GPTJ),
 		gptstablelmmodels: make(map[string]*gpt2.StableLM),
+		dollymodels:       make(map[string]*gpt2.Dolly),
+		redpajama:         make(map[string]*gpt2.RedPajama),
 		models:            make(map[string]*llama.LLama),
 		rwkv:              make(map[string]*rwkv.RwkvState),
+		bloomz:            make(map[string]*bloomz.Bloomz),
 		bert:              make(map[string]*bert.Bert),
 		promptsTemplates:  make(map[string]*template.Template),
+		whisperModels:     make(map[string]whisper.Model),
 	}
 }
 
@@ -128,6 +136,70 @@ func (ml *ModelLoader) loadTemplateIfExists(modelName, modelFile string) error {
 	return nil
 }
 
+func (ml *ModelLoader) LoadRedPajama(modelName string) (*gpt2.RedPajama, error) {
+	ml.mu.Lock()
+	defer ml.mu.Unlock()
+
+	// Check if we already have a loaded model
+	if !ml.ExistsInModelPath(modelName) {
+		return nil, fmt.Errorf("model does not exist")
+	}
+
+	if m, ok := ml.redpajama[modelName]; ok {
+		log.Debug().Msgf("Model already loaded in memory: %s", modelName)
+		return m, nil
+	}
+
+	// Load the model and keep it in memory for later use
+	modelFile := filepath.Join(ml.ModelPath, modelName)
+	log.Debug().Msgf("Loading model in memory from file: %s", modelFile)
+
+	model, err := gpt2.NewRedPajama(modelFile)
+	if err != nil {
+		return nil, err
+	}
+
+	// If there is a prompt template, load it
+	if err := ml.loadTemplateIfExists(modelName, modelFile); err != nil {
+		return nil, err
+	}
+
+	ml.redpajama[modelName] = model
+	return model, err
+}
+
+func (ml *ModelLoader) LoadDollyModel(modelName string) (*gpt2.Dolly, error) {
+	ml.mu.Lock()
+	defer ml.mu.Unlock()
+
+	// Check if we already have a loaded model
+	if !ml.ExistsInModelPath(modelName) {
+		return nil, fmt.Errorf("model does not exist")
+	}
+
+	if m, ok := ml.dollymodels[modelName]; ok {
+		log.Debug().Msgf("Model already loaded in memory: %s", modelName)
+		return m, nil
+	}
+
+	// Load the model and keep it in memory for later use
+	modelFile := filepath.Join(ml.ModelPath, modelName)
+	log.Debug().Msgf("Loading model in memory from file: %s", modelFile)
+
+	model, err := gpt2.NewDolly(modelFile)
+	if err != nil {
+		return nil, err
+	}
+
+	// If there is a prompt template, load it
+	if err := ml.loadTemplateIfExists(modelName, modelFile); err != nil {
+		return nil, err
+	}
+
+	ml.dollymodels[modelName] = model
+	return model, err
+}
+
 func (ml *ModelLoader) LoadStableLMModel(modelName string) (*gpt2.StableLM, error) {
 	ml.mu.Lock()
 	defer ml.mu.Unlock()
@@ -192,6 +264,38 @@ func (ml *ModelLoader) LoadBERT(modelName string) (*bert.Bert, error) {
 	return model, err
 }
 
+func (ml *ModelLoader) LoadBloomz(modelName string) (*bloomz.Bloomz, error) {
+	ml.mu.Lock()
+	defer ml.mu.Unlock()
+
+	// Check if we already have a loaded model
+	if !ml.ExistsInModelPath(modelName) {
+		return nil, fmt.Errorf("model does not exist")
+	}
+
+	if m, ok := ml.bloomz[modelName]; ok {
+		log.Debug().Msgf("Model already loaded in memory: %s", modelName)
+		return m, nil
+	}
+
+	// Load the model and keep it in memory for later use
+	modelFile := filepath.Join(ml.ModelPath, modelName)
+	log.Debug().Msgf("Loading model in memory from file: %s", modelFile)
+
+	model, err := bloomz.New(modelFile)
+	if err != nil {
+		return nil, err
+	}
+
+	// If there is a prompt template, load it
+	if err := ml.loadTemplateIfExists(modelName, modelFile); err != nil {
+		return nil, err
+	}
+
+	ml.bloomz[modelName] = model
+	return model, err
+}
+
 func (ml *ModelLoader) LoadGPT2Model(modelName string) (*gpt2.GPT2, error) {
 	ml.mu.Lock()
 	defer ml.mu.Unlock()
@@ -320,6 +424,33 @@ func (ml *ModelLoader) LoadLLaMAModel(modelName string, opts ...llama.ModelOptio
 	return model, err
 }
 
+func (ml *ModelLoader) LoadWhisperModel(modelName string) (whisper.Model, error) {
+	ml.mu.Lock()
+	defer ml.mu.Unlock()
+
+	// Check if we already have a loaded model
+	if !ml.ExistsInModelPath(modelName) {
+		return nil, fmt.Errorf("model does not exist -- %s", modelName)
+	}
+
+	if m, ok := ml.whisperModels[modelName]; ok {
+		log.Debug().Msgf("Model already loaded in memory: %s", modelName)
+		return m, nil
+	}
+
+	// Load the model and keep it in memory for later use
+	modelFile := filepath.Join(ml.ModelPath, modelName)
+	log.Debug().Msgf("Loading model in memory from file: %s", modelFile)
+
+	model, err := whisper.New(modelFile)
+	if err != nil {
+		return nil, err
+	}
+
+	ml.whisperModels[modelName] = model
+	return model, err
+}
+
 const tokenizerSuffix = ".tokenizer.json"
 
 var loadedModels map[string]interface{} = map[string]interface{}{}
@@ -329,8 +460,14 @@ func (ml *ModelLoader) BackendLoader(backendString string, modelFile string, lla
 	switch strings.ToLower(backendString) {
 	case "llama":
 		return ml.LoadLLaMAModel(modelFile, llamaOpts...)
+	case "bloomz":
+		return ml.LoadBloomz(modelFile)
 	case "stablelm":
 		return ml.LoadStableLMModel(modelFile)
+	case "dolly":
+		return ml.LoadDollyModel(modelFile)
+	case "redpajama":
+		return ml.LoadRedPajama(modelFile)
 	case "gpt2":
 		return ml.LoadGPT2Model(modelFile)
 	case "gptj":
@@ -344,6 +481,16 @@ func (ml *ModelLoader) BackendLoader(backendString string, modelFile string, lla
 	}
 }
 
+func (ml *ModelLoader) WhisperLoader(backendString string, modelFile string) (model whisper.Model, err error) {
+	//TODO expose more whisper options in next PR
+	switch strings.ToLower(backendString) {
+	case "whisper":
+		return ml.LoadWhisperModel(modelFile)
+	default:
+		return nil, fmt.Errorf("whisper backend unsupported: %s", backendString)
+	}
+}
+
 func (ml *ModelLoader) GreedyLoader(modelFile string, llamaOpts []llama.ModelOption, threads uint32) (model interface{}, err error) {
 	updateModels := func(model interface{}) {
 		muModels.Lock()
@@ -391,6 +538,30 @@ func (ml *ModelLoader) GreedyLoader(modelFile string, llamaOpts []llama.ModelOpt
 		err = multierror.Append(err, modelerr)
 	}
 
+	model, modelerr = ml.LoadDollyModel(modelFile)
+	if modelerr == nil {
+		updateModels(model)
+		return model, nil
+	} else {
+		err = multierror.Append(err, modelerr)
+	}
+
+	model, modelerr = ml.LoadRedPajama(modelFile)
+	if modelerr == nil {
+		updateModels(model)
+		return model, nil
+	} else {
+		err = multierror.Append(err, modelerr)
+	}
+
+	model, modelerr = ml.LoadBloomz(modelFile)
+	if modelerr == nil {
+		updateModels(model)
+		return model, nil
+	} else {
+		err = multierror.Append(err, modelerr)
+	}
+
 	model, modelerr = ml.LoadRWKV(modelFile, modelFile+tokenizerSuffix, threads)
 	if modelerr == nil {
 		updateModels(model)

pkg/whisper/whisper.go

@@ -28,7 +28,7 @@ func audioToWav(src, dst string) error {
 	return nil
 }
 
-func Transcript(modelpath, audiopath, language string) (string, error) {
+func Transcript(model whisper.Model, audiopath, language string) (string, error) {
 
 	dir, err := os.MkdirTemp("", "whisper")
 	if err != nil {
@@ -58,13 +58,6 @@ func Transcript(modelpath, audiopath, language string) (string, error) {
 
 	data := buf.AsFloat32Buffer().Data
 
-	// Load the model
-	model, err := whisper.New(modelpath)
-	if err != nil {
-		return "", err
-	}
-	defer model.Close()
-
 	// Process samples
 	context, err := model.NewContext()
 	if err != nil {