dearwolf/LocalAI
1
0
Fork 0
mirror of https://github.com/mudler/LocalAI.git synced 2025-06-27 05:04:59 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 21

chore(llama-ggml): drop deprecated backend

Browse source 
The GGML format is now dead, since in the next version of LocalAI we
already bring many breaking compatibility changes, taking the occasion
also to drop ggml support (pre-gguf).

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
This commit is contained in:
Ettore Di Giacinto 2025-02-06 16:44:31 +01:00
parent 8d45670e41
commit 695935c184
6 changed files with 7 additions and 348 deletions
Download patch file Download diff file Expand all files Collapse all files

38
Makefile
View file

@ -6,8 +6,6 @@ BINARY_NAME=local-ai
DETECT_LIBS?=true
# llama.cpp versions
GOLLAMA_REPO?=https://github.com/go-skynet/go-llama.cpp
GOLLAMA_VERSION?=2b57a8ae43e4699d3dc5d1496a1ccd42922993be
CPPLLAMA_VERSION?=d774ab3acc4fee41fbed6dbfc192b57d5f79f34b
# whisper.cpp version
@ -151,7 +149,6 @@ ifeq ($(BUILD_TYPE),hipblas)
LD_LIBRARY_PATH ?= /opt/rocm/lib:/opt/rocm/llvm/lib
export CXX=$(ROCM_HOME)/llvm/bin/clang++
export CC=$(ROCM_HOME)/llvm/bin/clang
# llama-ggml has no hipblas support, so override it here.
export STABLE_BUILD_TYPE=
export GGML_HIP=1
GPU_TARGETS ?= gfx900,gfx906,gfx908,gfx940,gfx941,gfx942,gfx90a,gfx1030,gfx1031,gfx1100,gfx1101
@ -188,7 +185,6 @@ ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-avx
ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-avx2
ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-avx512
ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-fallback
ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-ggml
ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-grpc
ALL_GRPC_BACKENDS+=backend-assets/util/llama-cpp-rpc-server
ALL_GRPC_BACKENDS+=backend-assets/grpc/whisper
@ -222,19 +218,6 @@ endif
all: help
## go-llama.cpp
sources/go-llama.cpp:
mkdir -p sources/go-llama.cpp
cd sources/go-llama.cpp && \
git init && \
git remote add origin $(GOLLAMA_REPO) && \
git fetch origin && \
git checkout $(GOLLAMA_VERSION) && \
git submodule update --init --recursive --depth 1 --single-branch
sources/go-llama.cpp/libbinding.a: sources/go-llama.cpp
$(MAKE) -C sources/go-llama.cpp BUILD_TYPE=$(STABLE_BUILD_TYPE) libbinding.a
## bark.cpp
sources/bark.cpp:
git clone --recursive $(BARKCPP_REPO) sources/bark.cpp && \
@ -310,19 +293,17 @@ sources/whisper.cpp:
sources/whisper.cpp/libwhisper.a: sources/whisper.cpp
cd sources/whisper.cpp && $(MAKE) libwhisper.a libggml.a
get-sources: sources/go-llama.cpp sources/go-piper sources/stablediffusion-ggml.cpp sources/bark.cpp sources/whisper.cpp backend/cpp/llama/llama.cpp
get-sources: sources/go-piper sources/stablediffusion-ggml.cpp sources/bark.cpp sources/whisper.cpp backend/cpp/llama/llama.cpp
replace:
$(GOCMD) mod edit -replace github.com/ggerganov/whisper.cpp=$(CURDIR)/sources/whisper.cpp
$(GOCMD) mod edit -replace github.com/ggerganov/whisper.cpp/bindings/go=$(CURDIR)/sources/whisper.cpp/bindings/go
$(GOCMD) mod edit -replace github.com/mudler/go-piper=$(CURDIR)/sources/go-piper
$(GOCMD) mod edit -replace github.com/go-skynet/go-llama.cpp=$(CURDIR)/sources/go-llama.cpp
dropreplace:
$(GOCMD) mod edit -dropreplace github.com/ggerganov/whisper.cpp
$(GOCMD) mod edit -dropreplace github.com/ggerganov/whisper.cpp/bindings/go
$(GOCMD) mod edit -dropreplace github.com/mudler/go-piper
$(GOCMD) mod edit -dropreplace github.com/go-skynet/go-llama.cpp
prepare-sources: get-sources replace
$(GOCMD) mod download
@ -330,7 +311,6 @@ prepare-sources: get-sources replace
## GENERIC
rebuild: ## Rebuilds the project
$(GOCMD) clean -cache
$(MAKE) -C sources/go-llama.cpp clean
$(MAKE) -C sources/whisper.cpp clean
$(MAKE) -C sources/go-piper clean
$(MAKE) build
@ -434,7 +414,7 @@ run: prepare ## run local-ai
test-models/testmodel.ggml:
mkdir test-models
mkdir test-dir
wget -q https://huggingface.co/TheBloke/orca_mini_3B-GGML/resolve/main/orca-mini-3b.ggmlv3.q4_0.bin -O test-models/testmodel.ggml
wget -q https://huggingface.co/RichardErkhov/Qwen_-_Qwen2-1.5B-Instruct-gguf/resolve/main/Qwen2-1.5B-Instruct.Q2_K.gguf -O test-models/testmodel.ggml
wget -q https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin -O test-models/whisper-en
wget -q https://huggingface.co/mudler/all-MiniLM-L6-v2/resolve/main/ggml-model-q4_0.bin -O test-models/bert
wget -q https://cdn.openai.com/whisper/draft-20220913a/micro-machines.wav -O test-dir/audio.wav
@ -449,8 +429,7 @@ test: prepare test-models/testmodel.ggml grpcs
export GO_TAGS="tts debug"
$(MAKE) prepare-test
HUGGINGFACE_GRPC=$(abspath ./)/backend/python/transformers/run.sh TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="!llama && !llama-gguf" --flake-attempts $(TEST_FLAKES) --fail-fast -v -r $(TEST_PATHS)
$(MAKE) test-llama
$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="!llama-gguf" --flake-attempts $(TEST_FLAKES) --fail-fast -v -r $(TEST_PATHS)
$(MAKE) test-llama-gguf
$(MAKE) test-tts
$(MAKE) test-stablediffusion
@ -479,10 +458,6 @@ teardown-e2e:
rm -rf $(TEST_DIR) || true
docker stop $$(docker ps -q --filter ancestor=localai-tests)
test-llama: prepare-test
TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="llama" --flake-attempts $(TEST_FLAKES) -v -r $(TEST_PATHS)
test-llama-gguf: prepare-test
TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="llama-gguf" --flake-attempts $(TEST_FLAKES) -v -r $(TEST_PATHS)
@ -760,13 +735,6 @@ backend-assets/util/llama-cpp-rpc-server: backend-assets/grpc/llama-cpp-grpc
mkdir -p backend-assets/util/
cp -rf backend/cpp/llama-grpc/llama.cpp/build/bin/rpc-server backend-assets/util/llama-cpp-rpc-server
backend-assets/grpc/llama-ggml: sources/go-llama.cpp sources/go-llama.cpp/libbinding.a backend-assets/grpc
CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/sources/go-llama.cpp LIBRARY_PATH=$(CURDIR)/sources/go-llama.cpp \
$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/llama-ggml ./backend/go/llm/llama-ggml/
ifneq ($(UPX),)
$(UPX) backend-assets/grpc/llama-ggml
endif
backend-assets/grpc/bark-cpp: backend/go/bark/libbark.a backend-assets/grpc
CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/backend/go/bark/ LIBRARY_PATH=$(CURDIR)/backend/go/bark/ \
$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/bark-cpp ./backend/go/bark/

204
backend/go/llm/llama-ggml/llama.go
View file

@ -1,204 +0,0 @@
package main
// 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/go-llama.cpp"
"github.com/mudler/LocalAI/pkg/grpc/base"
pb "github.com/mudler/LocalAI/pkg/grpc/proto"
)
type LLM struct {
base.SingleThread
llama *llama.LLama
}
func (llm *LLM) Load(opts *pb.ModelOptions) error {
ropeFreqBase := float32(10000)
ropeFreqScale := float32(1)
if opts.RopeFreqBase != 0 {
ropeFreqBase = opts.RopeFreqBase
}
if opts.RopeFreqScale != 0 {
ropeFreqScale = opts.RopeFreqScale
}
llamaOpts := []llama.ModelOption{
llama.WithRopeFreqBase(ropeFreqBase),
llama.WithRopeFreqScale(ropeFreqScale),
}
if opts.NGQA != 0 {
llamaOpts = append(llamaOpts, llama.WithGQA(int(opts.NGQA)))
}
if opts.RMSNormEps != 0 {
llamaOpts = append(llamaOpts, llama.WithRMSNormEPS(opts.RMSNormEps))
}
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.ModelFile, llamaOpts...)
llm.llama = model
return err
}
func buildPredictOptions(opts *pb.PredictOptions) []llama.PredictOption {
ropeFreqBase := float32(10000)
ropeFreqScale := float32(1)
if opts.RopeFreqBase != 0 {
ropeFreqBase = opts.RopeFreqBase
}
if opts.RopeFreqScale != 0 {
ropeFreqScale = opts.RopeFreqScale
}
predictOptions := []llama.PredictOption{
llama.SetTemperature(opts.Temperature),
llama.SetTopP(opts.TopP),
llama.SetTopK(int(opts.TopK)),
llama.SetTokens(int(opts.Tokens)),
llama.SetThreads(int(opts.Threads)),
llama.WithGrammar(opts.Grammar),
llama.SetRopeFreqBase(ropeFreqBase),
llama.SetRopeFreqScale(ropeFreqScale),
llama.SetNegativePromptScale(opts.NegativePromptScale),
llama.SetNegativePrompt(opts.NegativePrompt),
}
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(opts.MirostatETA))
}
if opts.MirostatTAU != 0 {
predictOptions = append(predictOptions, llama.SetMirostatTAU(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(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(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(opts.TailFreeSamplingZ))
predictOptions = append(predictOptions, llama.SetTypicalP(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...)
}

19
backend/go/llm/llama-ggml/main.go
View file

@ -1,19 +0,0 @@
package main
import (
"flag"
grpc "github.com/mudler/LocalAI/pkg/grpc"
)
var (
addr = flag.String("addr", "localhost:50051", "the address to connect to")
)
func main() {
flag.Parse()
if err := grpc.StartServer(*addr, &LLM{}); err != nil {
panic(err)
}
}

71
core/http/app_test.go
View file

@ -526,77 +526,6 @@ var _ = Describe("API test", func() {
Expect(content["usage"]).To(ContainSubstring("You can test this model with curl like this"))
})
It("runs openllama(llama-ggml backend)", Label("llama"), func() {
if runtime.GOOS != "linux" {
Skip("test supported only on linux")
}
response := postModelApplyRequest("http://127.0.0.1:9090/models/apply", modelApplyRequest{
URL: "github:go-skynet/model-gallery/openllama_3b.yaml",
Name: "openllama_3b",
Overrides: map[string]interface{}{"backend": "llama-ggml", "mmap": true, "f16": true, "context_size": 128},
})
Expect(response["uuid"]).ToNot(BeEmpty(), fmt.Sprint(response))
uuid := response["uuid"].(string)
Eventually(func() bool {
response := getModelStatus("http://127.0.0.1:9090/models/jobs/" + uuid)
return response["processed"].(bool)
}, "360s", "10s").Should(Equal(true))
By("testing completion")
resp, err := client.CreateCompletion(context.TODO(), openai.CompletionRequest{Model: "openllama_3b", Prompt: "Count up to five: one, two, three, four, "})
Expect(err).ToNot(HaveOccurred())
Expect(len(resp.Choices)).To(Equal(1))
Expect(resp.Choices[0].Text).To(ContainSubstring("five"))
By("testing functions")
resp2, err := client.CreateChatCompletion(
context.TODO(),
openai.ChatCompletionRequest{
Model: "openllama_3b",
Messages: []openai.ChatCompletionMessage{
{
Role: "user",
Content: "What is the weather like in San Francisco (celsius)?",
},
},
Functions: []openai.FunctionDefinition{
openai.FunctionDefinition{
Name: "get_current_weather",
Description: "Get the current weather",
Parameters: jsonschema.Definition{
Type: jsonschema.Object,
Properties: map[string]jsonschema.Definition{
"location": {
Type: jsonschema.String,
Description: "The city and state, e.g. San Francisco, CA",
},
"unit": {
Type: jsonschema.String,
Enum: []string{"celcius", "fahrenheit"},
},
},
Required: []string{"location"},
},
},
},
})
Expect(err).ToNot(HaveOccurred())
Expect(len(resp2.Choices)).To(Equal(1))
Expect(resp2.Choices[0].Message.FunctionCall).ToNot(BeNil())
Expect(resp2.Choices[0].Message.FunctionCall.Name).To(Equal("get_current_weather"), resp2.Choices[0].Message.FunctionCall.Name)
var res map[string]string
err = json.Unmarshal([]byte(resp2.Choices[0].Message.FunctionCall.Arguments), &res)
Expect(err).ToNot(HaveOccurred())
Expect(res["location"]).To(ContainSubstring("San Francisco"), fmt.Sprint(res))
Expect(res["unit"]).To(Equal("celcius"), fmt.Sprint(res))
Expect(string(resp2.Choices[0].FinishReason)).To(Equal("function_call"), fmt.Sprint(resp2.Choices[0].FinishReason))
})
It("runs openllama gguf(llama-cpp)", Label("llama-gguf"), func() {
if runtime.GOOS != "linux" {
Skip("test supported only on linux")

17
docs/content/docs/features/text-generation.md
View file

@ -124,7 +124,7 @@ Note: rwkv models needs to specify the backend `rwkv` in the YAML config files a
{{% alert note %}}
The `ggml` file format has been deprecated. If you are using `ggml` models and you are configuring your model with a YAML file, specify, use the `llama-ggml` backend instead. If you are relying in automatic detection of the model, you should be fine. For `gguf` models, use the `llama` backend. The go backend is deprecated as well but still available as `go-llama`. The go backend supports still features not available in the mainline: speculative sampling and embeddings.
The `ggml` file format has been deprecated. If you are using `ggml` models and you are configuring your model with a YAML file, specify, use a LocalAI version older than v2.25.0. For `gguf` models, use the `llama` backend. The go backend is deprecated as well but still available as `go-llama`.
{{% /alert %}}
@ -175,25 +175,12 @@ name: llama
backend: llama
parameters:
# Relative to the models path
model: file.gguf.bin
```
In the example above we specify `llama` as the backend to restrict loading `gguf` models only.
For instance, to use the `llama-ggml` backend for `ggml` models:
```yaml
name: llama
backend: llama-ggml
parameters:
# Relative to the models path
model: file.ggml.bin
model: file.gguf
```
#### Reference
- [llama](https://github.com/ggerganov/llama.cpp)
- [binding](https://github.com/go-skynet/go-llama.cpp)
### exllama/2

6
pkg/model/initializers.go
View file

@ -43,8 +43,6 @@ var TypeAlias map[string]string = map[string]string{
var AutoDetect = os.Getenv("DISABLE_AUTODETECT") != "true"
const (
LlamaGGML = "llama-ggml"
LLamaCPP = "llama-cpp"
LLamaCPPAVX2 = "llama-cpp-avx2"
@ -143,10 +141,10 @@ func orderBackends(backends map[string][]string) ([]string, error) {
// sets a priority list - first has more priority
priorityList := []string{
// First llama.cpp(variants) and llama-ggml to follow.
// First llama.cpp(variants)
// We keep the fallback to prevent that if the llama.cpp variants
// that depends on shared libs if breaks have still a safety net.
LLamaCPP, LlamaGGML, LLamaCPPFallback,
LLamaCPP, LLamaCPPFallback,
}
toTheEnd := []string{