mirror of
https://github.com/mudler/LocalAI.git
synced 2025-05-20 02:24:59 +00:00
feat(llama.cpp): add distributed llama.cpp inferencing (#2324)
* feat(llama.cpp): support distributed llama.cpp Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * feat: let tweak how chat messages are merged together Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * refactor Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * Makefile: register to ALL_GRPC_BACKENDS Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * refactoring, allow disable auto-detection of backends Signed-off-by: Ettore Di Giacinto <mudler@localai.io> * minor fixups Signed-off-by: mudler <mudler@localai.io> * feat: add cmd to start rpc-server from llama.cpp Signed-off-by: mudler <mudler@localai.io> * ci: add ccache Signed-off-by: mudler <mudler@localai.io> --------- Signed-off-by: Ettore Di Giacinto <mudler@localai.io> Signed-off-by: mudler <mudler@localai.io>
This commit is contained in:
Ettore Di Giacinto
GitHub
parent
29909666c3
commit
c89271b2e4
11 changed files with 222 additions and 82 deletions
|
|
@ -12,9 +12,9 @@ import (
|
|||
|
||||
grpc "github.com/go-skynet/LocalAI/pkg/grpc"
|
||||
"github.com/go-skynet/LocalAI/pkg/xsysinfo"
|
||||
"github.com/klauspost/cpuid/v2"
|
||||
"github.com/phayes/freeport"
|
||||
"github.com/rs/zerolog/log"
|
||||
"golang.org/x/sys/cpu"
|
||||
|
||||
"github.com/elliotchance/orderedmap/v2"
|
||||
)
|
||||
|
|
@ -26,16 +26,18 @@ var Aliases map[string]string = map[string]string{
|
|||
"langchain-huggingface": LCHuggingFaceBackend,
|
||||
}
|
||||
|
||||
var autoDetect = os.Getenv("DISABLE_AUTODETECT") != "true"
|
||||
|
||||
const (
|
||||
LlamaGGML = "llama-ggml"
|
||||
|
||||
LLamaCPP = "llama-cpp"
|
||||
LLamaCPP = "llama-cpp"
|
||||
|
||||
LLamaCPPCUDA12 = "llama-cpp-cuda12"
|
||||
LLamaCPPAVX2 = "llama-cpp-avx2"
|
||||
LLamaCPPAVX = "llama-cpp-avx"
|
||||
LLamaCPPFallback = "llama-cpp-fallback"
|
||||
LLamaCPPCUDA = "llama-cpp-cuda"
|
||||
LLamaCPPGRPC = "llama-cpp-grpc"
|
||||
|
||||
Gpt4AllLlamaBackend = "gpt4all-llama"
|
||||
Gpt4AllMptBackend = "gpt4all-mpt"
|
||||
|
|
@ -59,7 +61,7 @@ func backendPath(assetDir, backend string) string {
|
|||
|
||||
// backendsInAssetDir returns the list of backends in the asset directory
|
||||
// that should be loaded
|
||||
func backendsInAssetDir(assetDir string) (*orderedmap.OrderedMap[string, any], error) {
|
||||
func backendsInAssetDir(assetDir string) ([]string, error) {
|
||||
// Exclude backends from automatic loading
|
||||
excludeBackends := []string{LocalStoreBackend}
|
||||
entry, err := os.ReadDir(backendPath(assetDir, ""))
|
||||
|
|
@ -74,27 +76,46 @@ ENTRY:
|
|||
continue ENTRY
|
||||
}
|
||||
}
|
||||
if !e.IsDir() {
|
||||
if !strings.Contains(e.Name(), LLamaCPP) || strings.Contains(e.Name(), LLamaCPPFallback) {
|
||||
backends[e.Name()] = []string{}
|
||||
}
|
||||
if e.IsDir() {
|
||||
continue
|
||||
}
|
||||
|
||||
// Skip the llama.cpp variants if we are autoDetecting
|
||||
// But we always load the fallback variant if it exists
|
||||
if strings.Contains(e.Name(), LLamaCPP) && !strings.Contains(e.Name(), LLamaCPPFallback) && autoDetect {
|
||||
continue
|
||||
}
|
||||
|
||||
backends[e.Name()] = []string{}
|
||||
}
|
||||
|
||||
foundLCPPAVX, foundLCPPAVX2, foundLCPPFallback := false, false, false
|
||||
if _, ok := backends[LLamaCPP]; !ok {
|
||||
for _, e := range entry {
|
||||
if strings.Contains(e.Name(), LLamaCPPAVX2) && !foundLCPPAVX2 {
|
||||
backends[LLamaCPP] = append(backends[LLamaCPP], LLamaCPPAVX2)
|
||||
foundLCPPAVX2 = true
|
||||
}
|
||||
if strings.Contains(e.Name(), LLamaCPPAVX) && !foundLCPPAVX {
|
||||
backends[LLamaCPP] = append(backends[LLamaCPP], LLamaCPPAVX)
|
||||
foundLCPPAVX = true
|
||||
}
|
||||
if strings.Contains(e.Name(), LLamaCPPFallback) && !foundLCPPFallback {
|
||||
backends[LLamaCPP] = append(backends[LLamaCPP], LLamaCPPFallback)
|
||||
foundLCPPFallback = true
|
||||
// if we are autoDetecting, we want to show the llama.cpp variants as a single backend
|
||||
if autoDetect {
|
||||
// if we find the llama.cpp variants, show them of as a single backend (llama-cpp) as later we are going to pick that up
|
||||
// when starting the service
|
||||
foundLCPPAVX, foundLCPPAVX2, foundLCPPFallback, foundLCPPGRPC, foundLCPPCuda := false, false, false, false, false
|
||||
if _, ok := backends[LLamaCPP]; !ok {
|
||||
for _, e := range entry {
|
||||
if strings.Contains(e.Name(), LLamaCPPAVX2) && !foundLCPPAVX2 {
|
||||
backends[LLamaCPP] = append(backends[LLamaCPP], LLamaCPPAVX2)
|
||||
foundLCPPAVX2 = true
|
||||
}
|
||||
if strings.Contains(e.Name(), LLamaCPPAVX) && !foundLCPPAVX {
|
||||
backends[LLamaCPP] = append(backends[LLamaCPP], LLamaCPPAVX)
|
||||
foundLCPPAVX = true
|
||||
}
|
||||
if strings.Contains(e.Name(), LLamaCPPFallback) && !foundLCPPFallback {
|
||||
backends[LLamaCPP] = append(backends[LLamaCPP], LLamaCPPFallback)
|
||||
foundLCPPFallback = true
|
||||
}
|
||||
if strings.Contains(e.Name(), LLamaCPPGRPC) && !foundLCPPGRPC {
|
||||
backends[LLamaCPP] = append(backends[LLamaCPP], LLamaCPPGRPC)
|
||||
foundLCPPGRPC = true
|
||||
}
|
||||
if strings.Contains(e.Name(), LLamaCPPCUDA) && !foundLCPPCuda {
|
||||
backends[LLamaCPP] = append(backends[LLamaCPP], LLamaCPPCUDA)
|
||||
foundLCPPCuda = true
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -102,10 +123,13 @@ ENTRY:
|
|||
// 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
|
||||
|
||||
// sets a priority list - first has more priority
|
||||
priorityList := []string{
|
||||
// First llama.cpp and llama-ggml
|
||||
|
||||
// First llama.cpp(variants) and llama-ggml to follow.
|
||||
// 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, Gpt4All, LLamaCPPFallback,
|
||||
}
|
||||
|
||||
|
|
@ -139,7 +163,57 @@ ENTRY:
|
|||
}
|
||||
}
|
||||
|
||||
return orderedBackends, nil
|
||||
return orderedBackends.Keys(), nil
|
||||
}
|
||||
|
||||
// selectGRPCProcess selects the GRPC process to start based on system capabilities
|
||||
func selectGRPCProcess(backend, assetDir string) string {
|
||||
foundCUDA := false
|
||||
var grpcProcess string
|
||||
|
||||
// Select backend now just for llama.cpp
|
||||
if backend != LLamaCPP {
|
||||
return ""
|
||||
}
|
||||
|
||||
// Note: This environment variable is read by the LocalAI's llama.cpp grpc-server
|
||||
if os.Getenv("LLAMACPP_GRPC_SERVERS") != "" {
|
||||
log.Info().Msgf("[%s] attempting to load with GRPC variant", LLamaCPPGRPC)
|
||||
return backendPath(assetDir, LLamaCPPGRPC)
|
||||
}
|
||||
|
||||
gpus, err := xsysinfo.GPUs()
|
||||
if err == nil {
|
||||
for _, gpu := range gpus {
|
||||
if strings.Contains(gpu.String(), "nvidia") {
|
||||
p := backendPath(assetDir, LLamaCPPCUDA)
|
||||
if _, err := os.Stat(p); err == nil {
|
||||
log.Info().Msgf("[%s] attempting to load with CUDA variant", backend)
|
||||
grpcProcess = p
|
||||
foundCUDA = true
|
||||
} else {
|
||||
log.Info().Msgf("GPU device found but no CUDA backend present")
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if foundCUDA {
|
||||
return grpcProcess
|
||||
}
|
||||
|
||||
if xsysinfo.HasCPUCaps(cpuid.AVX2) {
|
||||
log.Info().Msgf("[%s] attempting to load with AVX2 variant", backend)
|
||||
grpcProcess = backendPath(assetDir, LLamaCPPAVX2)
|
||||
} else if xsysinfo.HasCPUCaps(cpuid.AVX) {
|
||||
log.Info().Msgf("[%s] attempting to load with AVX variant", backend)
|
||||
grpcProcess = backendPath(assetDir, LLamaCPPAVX)
|
||||
} else {
|
||||
log.Info().Msgf("[%s] attempting to load with fallback variant", backend)
|
||||
grpcProcess = backendPath(assetDir, LLamaCPPFallback)
|
||||
}
|
||||
|
||||
return grpcProcess
|
||||
}
|
||||
|
||||
// starts the grpcModelProcess for the backend, and returns a grpc client
|
||||
|
|
@ -192,33 +266,10 @@ func (ml *ModelLoader) grpcModel(backend string, o *Options) func(string, string
|
|||
} else {
|
||||
grpcProcess := backendPath(o.assetDir, backend)
|
||||
|
||||
foundCUDA := false
|
||||
// for llama-cpp, check CPU capabilities and load the appropriate variant
|
||||
if backend == LLamaCPP {
|
||||
gpus, err := xsysinfo.GPUs()
|
||||
if err == nil {
|
||||
for _, gpu := range gpus {
|
||||
if strings.Contains(gpu.String(), "nvidia") {
|
||||
log.Info().Msgf("[%s] attempting to load with CUDA variant", backend)
|
||||
grpcProcess = backendPath(o.assetDir, LLamaCPPCUDA)
|
||||
if _, err := os.Stat(grpcProcess); err == nil {
|
||||
foundCUDA = true
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if !foundCUDA {
|
||||
if cpu.X86.HasAVX2 {
|
||||
log.Info().Msgf("[%s] attempting to load with AVX2 variant", backend)
|
||||
grpcProcess = backendPath(o.assetDir, LLamaCPPAVX2)
|
||||
} else if cpu.X86.HasAVX {
|
||||
log.Info().Msgf("[%s] attempting to load with AVX variant", backend)
|
||||
grpcProcess = backendPath(o.assetDir, LLamaCPPAVX)
|
||||
} else {
|
||||
log.Info().Msgf("[%s] attempting to load with fallback variant", backend)
|
||||
grpcProcess = backendPath(o.assetDir, LLamaCPPFallback)
|
||||
}
|
||||
if autoDetect {
|
||||
// autoDetect GRPC process to start based on system capabilities
|
||||
if selectedProcess := selectGRPCProcess(backend, o.assetDir); selectedProcess != "" {
|
||||
grpcProcess = selectedProcess
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -363,28 +414,24 @@ func (ml *ModelLoader) GreedyLoader(opts ...Option) (grpc.Backend, error) {
|
|||
|
||||
var err error
|
||||
|
||||
// autoload also external backends
|
||||
allBackendsToAutoLoad := orderedmap.NewOrderedMap[string, any]()
|
||||
// get backends embedded in the binary
|
||||
autoLoadBackends, err := backendsInAssetDir(o.assetDir)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
|
||||
// append externalBackends supplied by the user via the CLI
|
||||
for _, b := range o.externalBackends {
|
||||
autoLoadBackends = append(autoLoadBackends, b)
|
||||
}
|
||||
|
||||
log.Debug().Msgf("Loading from the following backends (in order): %+v", autoLoadBackends)
|
||||
|
||||
for _, k := range autoLoadBackends.Keys() {
|
||||
v, _ := autoLoadBackends.Get(k)
|
||||
allBackendsToAutoLoad.Set(k, v)
|
||||
}
|
||||
|
||||
for _, b := range o.externalBackends {
|
||||
allBackendsToAutoLoad.Set(b, []string{})
|
||||
}
|
||||
|
||||
if o.model != "" {
|
||||
log.Info().Msgf("Trying to load the model '%s' with the backend '%s'", o.model, allBackendsToAutoLoad.Keys())
|
||||
log.Info().Msgf("Trying to load the model '%s' with the backend '%s'", o.model, autoLoadBackends)
|
||||
}
|
||||
|
||||
for _, key := range allBackendsToAutoLoad.Keys() {
|
||||
for _, key := range autoLoadBackends {
|
||||
log.Info().Msgf("[%s] Attempting to load", key)
|
||||
options := []Option{
|
||||
WithBackendString(key),
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue