Files
LocalAI/pkg/grpc/server.go
LocalAI [bot] 600dafd20b feat(ced): sound-event classification backend (CED audio tagger) (#10425)
* feat(ced): sketch sound-classification backend (CED audio tagger)

Wires ced.cpp (CED, 527-class AudioSet sound-event tagger; baby cry,
footsteps, glass, alarms, dog bark) into LocalAI as a Go/purego backend.

SKETCH (backend skeleton real; core REST wiring + CI/gallery is a checklist
in DESIGN.md):
- backend/backend.proto: new SoundDetection rpc + SoundClass messages
  (run `make protogen-go` to regenerate pkg/grpc/proto).
- backend/go/ced: main.go (purego dlopen libced.so + ced_capi.h),
  goced.go (Ced gRPC backend: Load + SoundDetection), Makefile
  (clone-at-pin CED_VERSION, ggml static-PIC shared build), run.sh,
  package.sh, .gitignore.
- DESIGN.md: REST /v1/audio/classification wiring (handler/route/capability
  registration checklist), gallery/index + CI registration, and a scoping
  note for the realtime/websocket live-recognition path (sliding-window
  classify over the existing ws transport + voicegate; the ced C-API
  per-PCM entry point is already window-friendly).

Backend code does not compile until protogen-go regenerates the pb types
and a libced.so is built (Makefile clones+builds it).

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* feat(ced): REST /v1/audio/classification endpoint + capability registration

Wires the ced sound-event classification backend (AudioSet audio tagger)
end to end through the REST surface, mirroring the transcription path.

- Handler: core/http/endpoints/openai/sound_classification.go parses the
  multipart audio upload, temp-files it, resolves the model config and
  calls the SoundDetection RPC; returns {model, detections[]} JSON.
- Backend wrapper: core/backend/sound_classification.go (ModelSoundDetection)
  loads the model and normalizes the proto response into schema types.
- Schema: core/schema/sound_classification.go (SoundClassificationResult).
- gRPC layer: SoundDetection wired through the LocalAI wrapper (interface,
  Backend client, Client, embed, server, base default) so the loader-typed
  client exposes the RPC; proto regenerated via make protogen-go.
- Route: POST /v1/audio/classification (+ /audio/classification alias) with
  the audio/multipart default-model middleware in routes/openai.go.
- Capability surfaces: swagger @Tags/@Router on the handler; FLAG_SOUND_
  CLASSIFICATION usecase flag + UsecaseSoundClassification + UsecaseInfoMap +
  GuessUsecases + ModalityGroups + GetAllModelConfigUsecases; meta usecase
  option; /api/instructions audio area updated; auth RouteFeatureRegistry +
  FeatureAudioClassification (APIFeatures, default ON) + FeatureMetas; UI
  usecaseFilters, capabilities.js CAP_SOUND_CLASSIFICATION, Models.jsx filter
  + i18n; docs page features/audio-classification.md + whats-new + crosslink.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* feat(ced): realtime sound-event detection over the websocket API

When a realtime pipeline configures a sound-classification model, each
VAD-committed utterance (the same window the transcription path produces)
is also run through the CED sound-event classifier and the scored AudioSet
tags are emitted as a new server event. No new backend rpc is needed: the
SoundDetection gRPC method already exists on this branch.

- config: add Pipeline.SoundDetection (yaml/json sound_detection,omitempty)
  beside Transcription/VAD.
- realtime: add Model.SoundDetection(ctx, audio, topK, threshold) to the
  ModelInterface; implement it on wrappedModel and transcriptOnlyModel by
  calling backend.ModelSoundDetection with the session's sound-classification
  model config (mirrors how Transcribe dispatches). Load the optional config
  in newModel / newTranscriptionOnlyModel; nil config keeps it additive.
- types: add ConversationItemSoundDetectionEvent (item_id, content_index,
  detections[]{label,score,index}) with type conversation.item.sound_detection,
  its ServerEventType constant and MarshalJSON, mirroring the transcription
  completed event.
- realtime: add emitSoundDetection (unary path: classify the committed window,
  build the event, t.SendEvent) and wire it at the utterance-commit hook right
  after emitTranscription; gated on session.SoundDetectionEnabled (resolved
  from Pipeline.SoundDetection at session setup, defaults top_k=5, threshold=0).
  Its error is logged via xlog but never aborts the turn.
- test: Ginkgo specs for emitSoundDetection (tags emitted, empty detections,
  classifier error) plus a SoundDetection method on the fakeModel double.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* fix(ced): implement SoundDetection in nodes backend test doubles

The SoundDetection method added to the grpc backend interface left two
test doubles (fakeBackendClient, fakeGRPCBackend) incomplete, so
core/services/nodes failed to compile under `go vet`/`go test` (go build
missed it: the doubles live in _test.go). Add the method to both,
mirroring their existing Detect mock. Repairs CI for the nodes package.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* feat(ced): decouple realtime sound detection from VAD (sound-only sessions)

Sound-event detection must activate on sounds, not speech, so it no longer
runs through the voice VAD/transcription path. A sound-detection-only
pipeline (sound_detection set, no transcription/LLM) now:

- is accepted by prepareRealtimeConfig (sound_detection counts as a pipeline
  stage),
- builds a lightweight model via newSoundDetectionOnlyModel (no VAD/STT/LLM/TTS
  loaded), and
- defaults the session to turn_detection none (no VAD) with no transcription
  stage, so the client drives windowing via input_audio_buffer.commit
  (option A: client-side sliding window). The per-PCM C-API already supports
  arbitrary windows.

commitUtterance gains a sound-only branch: it emits the
conversation.item.sound_detection event (scored AudioSet tags) and stops -
no transcription, no LLM response. generateResponse is now guarded on a
transcription stage being present, so a sound-only turn never invokes the LLM.

Existing transcription/VAD sessions are unchanged (additive). Added a
commitUtterance sound-only Ginkgo spec asserting it emits the sound event and
neither transcribes nor generates a response. go vet + golangci-lint
(new-from-merge-base) clean; openai suite green.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* feat(ced): register sound-classification backend in gallery + CI

Mechanical backend-image registration for the ced sound-event classifier,
mirroring the parakeet-cpp Go/purego backend everywhere it is wired up.

- .github/backend-matrix.yml: add the ced build matrix, field-for-field copies
  of the parakeet-cpp entries (cpu amd64/arm64, cublas cuda 12/13 amd64,
  l4t cuda-13 arm64, l4t-jetpack cuda-12 arm64, sycl f32/f16, vulkan
  amd64/arm64, rocm hipblas, and the metal darwin entry), changing only
  backend and tag-suffix. dockerfile stays ./backend/Dockerfile.golang.
- backend/index.yaml: add the &ced meta anchor (capabilities map per platform)
  plus ced-development and the per-arch image entries, each uri/mirror
  tag-suffix matching the matrix exactly. The model gallery (GGUF) entry is
  intentionally deferred pending the HuggingFace publish (TODO note inline).
- scripts/changed-backends.js: add an explicit item.backend === "ced" branch in
  inferBackendPath mapping to backend/go/ced/, same mechanism and ordering as
  the parakeet-cpp branch (before the generic golang fallthrough).
- .github/workflows/bump_deps.yaml: register mudler/ced.cpp -> CED_VERSION in
  backend/go/ced/Makefile so the daily bot bumps the pin.
- swagger/{docs.go,swagger.json,swagger.yaml}: regenerated via make swagger so
  the existing /v1/audio/classification annotations land in the generated spec.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* feat(ced): server-side windowing for realtime sound detection (option B)

Adds an optional server-driven sliding-window classifier so a sound-only
realtime client only has to stream audio (no input_audio_buffer.commit):

- Pipeline.sound_detection_window_ms / sound_detection_hop_ms config knobs.
  When both > 0 on a sound-only session, the server classifies the last
  window of streamed audio every hop and emits a conversation.item.sound_
  detection event; the input buffer is trimmed to one window so a long
  stream stays bounded. When unset, the session stays client-driven
  (option A). Runs independent of VAD (sound events are not speech).
- handleSoundWindow (ticker) + classifySoundWindow (one tick, extracted so
  it is unit-testable) + writeWindowWAV, which declares the true
  InputSampleRate (NewWAVHeaderWithRate) so the classifier resamples
  correctly. Goroutine is started after toggleVAD and torn down with the
  session (close + wg.Wait).
- Register pipeline.sound_detection (+window_ms/hop_ms) in the config meta
  registry; the earlier realtime commit added pipeline.sound_detection
  without a registry entry, failing TestAllFieldsHaveRegistryEntries. This
  fixes that and covers the two new knobs.

Tests: classifySoundWindow emits an event + trims the buffer to one window,
no-ops on too-little audio; writeWindowWAV declares the given sample rate.
go build/vet + golangci-lint (new-from-merge-base) clean; config + openai
suites green.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* feat(ced): add ced-base GGUF model gallery entries (f16 + q8_0)

The ced-base weights are now published at mudler/ced-base-gguf (Apache-2.0,
converted from mispeech/ced-base). Adds gallery/ced.yaml (backend: ced +
known_usecases: sound_classification) and two gallery/index.yaml entries
(ced-base-f16 default, ced-base-q8 smallest) with sha256-pinned files, and
removes the now-resolved TODO from backend/index.yaml.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* feat(ced): add tiny/mini/small GGUF model gallery entries

Publishes the rest of the CED family (same architecture, metadata-driven port
verified end-to-end on ced-tiny) to mudler/ced-{tiny,mini,small}-gguf and adds
their f16 + q8_0 gallery entries:

  ced-tiny  (5.5M, edge/Pi-class)  f16 11MB / q8_0 6MB
  ced-mini  (9.6M)                 f16 19MB / q8_0 11MB
  ced-small (22M)                  f16 42MB / q8_0 23MB

All sha256-pinned. ced-base remains the accuracy default.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* chore(ced): point gallery entries at the consolidated mudler/ced-gguf repo

All CED quantizations (tiny/mini/small/base, f16/q8_0) now live in a single
HuggingFace repo, mudler/ced-gguf, instead of per-model repos. Repoint the 8
gallery model entries' urls + file uris accordingly. sha256 and filenames are
unchanged.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* chore(ced): bump CED_VERSION to the short-clip fix

Pin the ced backend to ced.cpp 99c6ed3, which fixes a crash on any clip
shorter than target_length (~10.11s): time_pos_embed was added at its full
63-frame grid instead of being sliced to the clip's actual time grid, tripping
ggml_can_repeat in ggml_add. Surfaced by the live realtime e2e (sub-10s
windows) and gated with a short-clip parity test upstream.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* docs(ced): list ced.cpp as a LocalAI-team engine + backend-guide directive

- README.md: add ced.cpp to the "native C/C++/GGML engines developed and
  maintained by the LocalAI project" table.
- docs/content/features/backends.md: add a Sound Classification backend
  category (sound-event classification / audio tagging) listing ced.cpp.
- .agents/adding-backends.md: add a "Documenting the backend" section and two
  verification-checklist items requiring new backends to be documented in the
  backends.md category list, and in-house native engines to be added to the
  README maintained-engines table. This directive was missing.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* chore(ced): repin CED_VERSION to the v0.1.0 release commit

ced.cpp history was squashed into a single release commit (tagged v0.1.0), so
the previous pin (99c6ed3) no longer exists upstream. Pin to c04ac14, the
v0.1.0 release commit, so the backend builds against a commit that exists.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* fix(ced): silence gosec G304/G103 + govet unsafeptr on audited paths

- sound_classification.go: os.Create(dst) where dst = temp dir + path.Base of
  the upload (no traversal). #nosec G304, matching the depth-anything-cpp handler.
- goced.go: reading a NUL-terminated C string from a libced-owned buffer.
  #nosec G103 (gosec) + //nolint:govet (golangci-lint's unsafeptr check), since
  the uintptr is a C-owned malloc'd buffer, not Go-GC memory.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

---------

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
2026-06-22 01:00:28 +02:00

898 lines
22 KiB
Go

package grpc
import (
"context"
"crypto/subtle"
"errors"
"fmt"
"io"
"log"
"net"
"os"
"strings"
pb "github.com/mudler/LocalAI/pkg/grpc/proto"
"google.golang.org/grpc"
"google.golang.org/grpc/codes"
"google.golang.org/grpc/metadata"
"google.golang.org/grpc/status"
)
// A GRPC Server that allows to run LLM inference.
// It is used by the LLMServices to expose the LLM functionalities that are called by the client.
// The GRPC Service is general, trying to encompass all the possible LLM options models.
// It depends on the real implementer then what can be done or not.
//
// The server is implemented as a GRPC service, with the following methods:
// - Predict: to run the inference with options
// - PredictStream: to run the inference with options and stream the results
// server is used to implement helloworld.GreeterServer.
type server struct {
pb.UnimplementedBackendServer
llm AIModel
}
func (s *server) Health(ctx context.Context, in *pb.HealthMessage) (*pb.Reply, error) {
return newReply("OK"), nil
}
func (s *server) Embedding(ctx context.Context, in *pb.PredictOptions) (*pb.EmbeddingResult, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
embeds, err := s.llm.Embeddings(in)
if err != nil {
return nil, err
}
return &pb.EmbeddingResult{Embeddings: embeds}, nil
}
func (s *server) LoadModel(ctx context.Context, in *pb.ModelOptions) (*pb.Result, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
err := s.llm.Load(in)
if err != nil {
return &pb.Result{Message: fmt.Sprintf("Error loading model: %s", err.Error()), Success: false}, err
}
return &pb.Result{Message: "Loading succeeded", Success: true}, nil
}
func (s *server) Predict(ctx context.Context, in *pb.PredictOptions) (*pb.Reply, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
if rich, ok := s.llm.(AIModelRich); ok {
return rich.PredictRich(in)
}
result, err := s.llm.Predict(in)
return newReply(result), err
}
func (s *server) GenerateImage(ctx context.Context, in *pb.GenerateImageRequest) (*pb.Result, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
err := s.llm.GenerateImage(in)
if err != nil {
return &pb.Result{Message: fmt.Sprintf("Error generating image: %s", err.Error()), Success: false}, err
}
return &pb.Result{Message: "Image generated", Success: true}, nil
}
func (s *server) GenerateVideo(ctx context.Context, in *pb.GenerateVideoRequest) (*pb.Result, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
err := s.llm.GenerateVideo(in)
if err != nil {
return &pb.Result{Message: fmt.Sprintf("Error generating video: %s", err.Error()), Success: false}, err
}
return &pb.Result{Message: "Video generated", Success: true}, nil
}
func (s *server) TTS(ctx context.Context, in *pb.TTSRequest) (*pb.Result, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
err := s.llm.TTS(in)
if err != nil {
return &pb.Result{Message: fmt.Sprintf("Error generating audio: %s", err.Error()), Success: false}, err
}
return &pb.Result{Message: "TTS audio generated", Success: true}, nil
}
func (s *server) TTSStream(in *pb.TTSRequest, stream pb.Backend_TTSStreamServer) error {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
audioChan := make(chan []byte)
done := make(chan bool)
go func() {
for audioChunk := range audioChan {
stream.Send(&pb.Reply{Audio: audioChunk})
}
done <- true
}()
err := s.llm.TTSStream(in, audioChan)
<-done
return err
}
func (s *server) SoundGeneration(ctx context.Context, in *pb.SoundGenerationRequest) (*pb.Result, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
err := s.llm.SoundGeneration(in)
if err != nil {
return &pb.Result{Message: fmt.Sprintf("Error generating audio: %s", err.Error()), Success: false}, err
}
return &pb.Result{Message: "Sound Generation audio generated", Success: true}, nil
}
func (s *server) Detect(ctx context.Context, in *pb.DetectOptions) (*pb.DetectResponse, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.Detect(in)
if err != nil {
return nil, err
}
return &res, nil
}
func (s *server) Depth(ctx context.Context, in *pb.DepthRequest) (*pb.DepthResponse, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.Depth(in)
if err != nil {
return nil, err
}
return &res, nil
}
func (s *server) FaceVerify(ctx context.Context, in *pb.FaceVerifyRequest) (*pb.FaceVerifyResponse, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.FaceVerify(in)
if err != nil {
return nil, err
}
return &res, nil
}
func (s *server) FaceAnalyze(ctx context.Context, in *pb.FaceAnalyzeRequest) (*pb.FaceAnalyzeResponse, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.FaceAnalyze(in)
if err != nil {
return nil, err
}
return &res, nil
}
func (s *server) VoiceVerify(ctx context.Context, in *pb.VoiceVerifyRequest) (*pb.VoiceVerifyResponse, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.VoiceVerify(in)
if err != nil {
return nil, err
}
return &res, nil
}
func (s *server) VoiceAnalyze(ctx context.Context, in *pb.VoiceAnalyzeRequest) (*pb.VoiceAnalyzeResponse, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.VoiceAnalyze(in)
if err != nil {
return nil, err
}
return &res, nil
}
func (s *server) VoiceEmbed(ctx context.Context, in *pb.VoiceEmbedRequest) (*pb.VoiceEmbedResponse, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.VoiceEmbed(in)
if err != nil {
return nil, err
}
return &res, nil
}
func (s *server) AudioTranscription(ctx context.Context, in *pb.TranscriptRequest) (*pb.TranscriptResult, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
result, err := s.llm.AudioTranscription(ctx, in)
if err != nil {
return nil, err
}
tresult := &pb.TranscriptResult{}
for _, s := range result.Segments {
tks := []int32{}
for _, t := range s.Tokens {
tks = append(tks, int32(t))
}
words := make([]*pb.TranscriptWord, 0, len(s.Words))
for _, w := range s.Words {
words = append(words, &pb.TranscriptWord{
Start: int64(w.Start),
End: int64(w.End),
Text: w.Text,
})
}
tresult.Segments = append(tresult.Segments,
&pb.TranscriptSegment{
Text: s.Text,
Id: int32(s.Id),
Start: int64(s.Start),
End: int64(s.End),
Tokens: tks,
Speaker: s.Speaker,
Words: words,
})
}
tresult.Text = result.Text
tresult.Language = result.Language
tresult.Duration = result.Duration
return tresult, nil
}
func (s *server) AudioTranscriptionStream(in *pb.TranscriptRequest, stream pb.Backend_AudioTranscriptionStreamServer) error {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
resultChan := make(chan *pb.TranscriptStreamResponse)
done := make(chan bool)
go func() {
for chunk := range resultChan {
stream.Send(chunk)
}
done <- true
}()
err := s.llm.AudioTranscriptionStream(stream.Context(), in, resultChan)
<-done
return err
}
func (s *server) PredictStream(in *pb.PredictOptions, stream pb.Backend_PredictStreamServer) error {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
if rich, ok := s.llm.(AIModelRich); ok {
replyChan := make(chan *pb.Reply)
done := make(chan bool)
go func() {
for reply := range replyChan {
// Send errors here mean the client disconnected;
// drain the rest of the channel so the producer
// (PredictStreamRich) doesn't block on the next
// reply forever.
_ = stream.Send(reply)
}
done <- true
}()
// Server-side close: PredictStreamRich implementations send into
// the channel and return when finished; closing is the host's
// concern so impls don't have to remember `defer close(...)`.
err := rich.PredictStreamRich(in, replyChan)
close(replyChan)
<-done
return err
}
resultChan := make(chan string)
done := make(chan bool)
go func() {
for result := range resultChan {
stream.Send(newReply(result))
}
done <- true
}()
err := s.llm.PredictStream(in, resultChan)
<-done
return err
}
func (s *server) TokenizeString(ctx context.Context, in *pb.PredictOptions) (*pb.TokenizationResponse, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.TokenizeString(in)
if err != nil {
return nil, err
}
castTokens := make([]int32, len(res.Tokens))
for i, v := range res.Tokens {
castTokens[i] = int32(v)
}
return &pb.TokenizationResponse{
Length: int32(res.Length),
Tokens: castTokens,
}, err
}
func (s *server) Status(ctx context.Context, in *pb.HealthMessage) (*pb.StatusResponse, error) {
res, err := s.llm.Status()
if err != nil {
return nil, err
}
return &res, nil
}
func (s *server) StoresSet(ctx context.Context, in *pb.StoresSetOptions) (*pb.Result, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
err := s.llm.StoresSet(in)
if err != nil {
return &pb.Result{Message: fmt.Sprintf("Error setting entry: %s", err.Error()), Success: false}, err
}
return &pb.Result{Message: "Set key", Success: true}, nil
}
func (s *server) StoresDelete(ctx context.Context, in *pb.StoresDeleteOptions) (*pb.Result, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
err := s.llm.StoresDelete(in)
if err != nil {
return &pb.Result{Message: fmt.Sprintf("Error deleting entry: %s", err.Error()), Success: false}, err
}
return &pb.Result{Message: "Deleted key", Success: true}, nil
}
func (s *server) StoresGet(ctx context.Context, in *pb.StoresGetOptions) (*pb.StoresGetResult, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.StoresGet(in)
if err != nil {
return nil, err
}
return &res, nil
}
func (s *server) StoresFind(ctx context.Context, in *pb.StoresFindOptions) (*pb.StoresFindResult, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.StoresFind(in)
if err != nil {
return nil, err
}
return &res, nil
}
func (s *server) VAD(ctx context.Context, in *pb.VADRequest) (*pb.VADResponse, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.VAD(in)
if err != nil {
return nil, err
}
return &res, nil
}
func (s *server) Diarize(ctx context.Context, in *pb.DiarizeRequest) (*pb.DiarizeResponse, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.Diarize(in)
if err != nil {
return nil, err
}
return &res, nil
}
func (s *server) SoundDetection(ctx context.Context, in *pb.SoundDetectionRequest) (*pb.SoundDetectionResponse, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
return s.llm.SoundDetection(ctx, in)
}
func (s *server) AudioEncode(ctx context.Context, in *pb.AudioEncodeRequest) (*pb.AudioEncodeResult, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.AudioEncode(in)
if err != nil {
return nil, err
}
return res, nil
}
func (s *server) AudioDecode(ctx context.Context, in *pb.AudioDecodeRequest) (*pb.AudioDecodeResult, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.AudioDecode(in)
if err != nil {
return nil, err
}
return res, nil
}
func (s *server) AudioTransform(ctx context.Context, in *pb.AudioTransformRequest) (*pb.AudioTransformResult, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.AudioTransform(in)
if err != nil {
return nil, err
}
return res, nil
}
func (s *server) AudioTransformStream(stream pb.Backend_AudioTransformStreamServer) error {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
in := make(chan *pb.AudioTransformFrameRequest, 4)
out := make(chan *pb.AudioTransformFrameResponse, 4)
// Pump incoming frames from the gRPC stream into `in`. EOF closes the
// channel, which signals the backend that the client is done sending.
recvErrCh := make(chan error, 1)
go func() {
defer close(in)
for {
req, err := stream.Recv()
if err != nil {
if errors.Is(err, io.EOF) {
recvErrCh <- nil
return
}
recvErrCh <- err
return
}
select {
case in <- req:
case <-stream.Context().Done():
recvErrCh <- stream.Context().Err()
return
}
}
}()
// Pump outgoing frames from `out` to the gRPC stream. The backend closes
// `out` on completion.
sendDone := make(chan error, 1)
go func() {
for resp := range out {
if err := stream.Send(resp); err != nil {
sendDone <- err
// Drain `out` so the backend can finish.
for range out {
}
return
}
}
sendDone <- nil
}()
backendErr := s.llm.AudioTransformStream(in, out)
sendErr := <-sendDone
recvErr := <-recvErrCh
if backendErr != nil {
return backendErr
}
if sendErr != nil {
return sendErr
}
return recvErr
}
// AudioToAudioStream is the bidirectional any-to-any S2S handler. The
// shape mirrors AudioTransformStream exactly (recv → in chan, out chan →
// send) so backends can implement either via the same goroutine idiom.
func (s *server) AudioToAudioStream(stream pb.Backend_AudioToAudioStreamServer) error {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
in := make(chan *pb.AudioToAudioRequest, 8)
out := make(chan *pb.AudioToAudioResponse, 8)
recvErrCh := make(chan error, 1)
go func() {
defer close(in)
for {
req, err := stream.Recv()
if err != nil {
if errors.Is(err, io.EOF) {
recvErrCh <- nil
return
}
recvErrCh <- err
return
}
select {
case in <- req:
case <-stream.Context().Done():
recvErrCh <- stream.Context().Err()
return
}
}
}()
sendDone := make(chan error, 1)
go func() {
for resp := range out {
if err := stream.Send(resp); err != nil {
sendDone <- err
for range out {
}
return
}
}
sendDone <- nil
}()
backendErr := s.llm.AudioToAudioStream(in, out)
sendErr := <-sendDone
recvErr := <-recvErrCh
if backendErr != nil {
return backendErr
}
if sendErr != nil {
return sendErr
}
return recvErr
}
// Forward is the bidi-stream handler for the cloud-proxy backend's
// passthrough mode. Same recv→in / out→send goroutine idiom as
// AudioTransformStream / AudioToAudioStream above. Buffer size 8 to
// keep SSE token streams flowing — at 4, a half-RTT slow gRPC client
// makes the body-read goroutine in the backend block on out<- after
// every few token frames.
func (s *server) Forward(stream pb.Backend_ForwardServer) error {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
in := make(chan *pb.ForwardRequest, 8)
out := make(chan *pb.ForwardReply, 8)
recvErrCh := make(chan error, 1)
go func() {
defer close(in)
for {
req, err := stream.Recv()
if err != nil {
if errors.Is(err, io.EOF) {
recvErrCh <- nil
return
}
recvErrCh <- err
return
}
select {
case in <- req:
case <-stream.Context().Done():
recvErrCh <- stream.Context().Err()
return
}
}
}()
sendDone := make(chan error, 1)
go func() {
for resp := range out {
if err := stream.Send(resp); err != nil {
sendDone <- err
for range out {
}
return
}
}
sendDone <- nil
}()
backendErr := s.llm.Forward(stream.Context(), in, out)
sendErr := <-sendDone
recvErr := <-recvErrCh
if backendErr != nil {
return backendErr
}
if sendErr != nil {
return sendErr
}
return recvErr
}
func (s *server) StartFineTune(ctx context.Context, in *pb.FineTuneRequest) (*pb.FineTuneJobResult, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.StartFineTune(in)
if err != nil {
return &pb.FineTuneJobResult{Success: false, Message: fmt.Sprintf("Error starting fine-tune: %s", err.Error())}, err
}
return res, nil
}
func (s *server) FineTuneProgress(in *pb.FineTuneProgressRequest, stream pb.Backend_FineTuneProgressServer) error {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
updateChan := make(chan *pb.FineTuneProgressUpdate)
done := make(chan bool)
go func() {
for update := range updateChan {
stream.Send(update)
}
done <- true
}()
err := s.llm.FineTuneProgress(in, updateChan)
<-done
return err
}
func (s *server) StopFineTune(ctx context.Context, in *pb.FineTuneStopRequest) (*pb.Result, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
err := s.llm.StopFineTune(in)
if err != nil {
return &pb.Result{Message: fmt.Sprintf("Error stopping fine-tune: %s", err.Error()), Success: false}, err
}
return &pb.Result{Message: "Fine-tune stopped", Success: true}, nil
}
func (s *server) ListCheckpoints(ctx context.Context, in *pb.ListCheckpointsRequest) (*pb.ListCheckpointsResponse, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.ListCheckpoints(in)
if err != nil {
return nil, err
}
return res, nil
}
func (s *server) ExportModel(ctx context.Context, in *pb.ExportModelRequest) (*pb.Result, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
err := s.llm.ExportModel(in)
if err != nil {
return &pb.Result{Message: fmt.Sprintf("Error exporting model: %s", err.Error()), Success: false}, err
}
return &pb.Result{Message: "Model exported", Success: true}, nil
}
func (s *server) StartQuantization(ctx context.Context, in *pb.QuantizationRequest) (*pb.QuantizationJobResult, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.StartQuantization(in)
if err != nil {
return &pb.QuantizationJobResult{Success: false, Message: fmt.Sprintf("Error starting quantization: %s", err.Error())}, err
}
return res, nil
}
func (s *server) QuantizationProgress(in *pb.QuantizationProgressRequest, stream pb.Backend_QuantizationProgressServer) error {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
updateChan := make(chan *pb.QuantizationProgressUpdate)
done := make(chan bool)
go func() {
for update := range updateChan {
stream.Send(update)
}
done <- true
}()
err := s.llm.QuantizationProgress(in, updateChan)
<-done
return err
}
func (s *server) StopQuantization(ctx context.Context, in *pb.QuantizationStopRequest) (*pb.Result, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
err := s.llm.StopQuantization(in)
if err != nil {
return &pb.Result{Message: fmt.Sprintf("Error stopping quantization: %s", err.Error()), Success: false}, err
}
return &pb.Result{Message: "Quantization stopped", Success: true}, nil
}
func (s *server) ModelMetadata(ctx context.Context, in *pb.ModelOptions) (*pb.ModelMetadataResponse, error) {
if s.llm.Locking() {
s.llm.Lock()
defer s.llm.Unlock()
}
res, err := s.llm.ModelMetadata(in)
if err != nil {
return nil, err
}
return res, nil
}
func (s *server) Free(ctx context.Context, in *pb.HealthMessage) (*pb.Result, error) {
if err := s.llm.Free(); err != nil {
return &pb.Result{Success: false, Message: err.Error()}, nil
}
return &pb.Result{Success: true}, nil
}
// NewBackendServer creates a pb.BackendServer.
func NewBackendServer(model AIModel) pb.BackendServer {
return &server{llm: model}
}
// AuthTokenEnvVar is the environment variable used to configure gRPC bearer token auth.
const AuthTokenEnvVar = "LOCALAI_GRPC_AUTH_TOKEN"
// validateToken extracts the bearer token from gRPC metadata and validates it.
func validateToken(ctx context.Context, expected string) error {
md, ok := metadata.FromIncomingContext(ctx)
if !ok {
return status.Error(codes.Unauthenticated, "missing metadata")
}
values := md.Get("authorization")
if len(values) == 0 {
return status.Error(codes.Unauthenticated, "missing authorization header")
}
raw := values[0]
if !strings.HasPrefix(raw, "Bearer ") {
return status.Error(codes.Unauthenticated, "authorization must use Bearer scheme")
}
token := strings.TrimPrefix(raw, "Bearer ")
if subtle.ConstantTimeCompare([]byte(token), []byte(expected)) != 1 {
return status.Error(codes.Unauthenticated, "invalid token")
}
return nil
}
func tokenUnaryInterceptor(token string) grpc.UnaryServerInterceptor {
return func(ctx context.Context, req any, info *grpc.UnaryServerInfo, handler grpc.UnaryHandler) (any, error) {
if err := validateToken(ctx, token); err != nil {
return nil, err
}
return handler(ctx, req)
}
}
func tokenStreamInterceptor(token string) grpc.StreamServerInterceptor {
return func(srv any, ss grpc.ServerStream, info *grpc.StreamServerInfo, handler grpc.StreamHandler) error {
if err := validateToken(ss.Context(), token); err != nil {
return err
}
return handler(srv, ss)
}
}
// serverOpts returns the common gRPC server options, including auth interceptors
// when LOCALAI_GRPC_AUTH_TOKEN is set.
func serverOpts() []grpc.ServerOption {
opts := []grpc.ServerOption{
grpc.MaxRecvMsgSize(maxGRPCMessageSize),
grpc.MaxSendMsgSize(maxGRPCMessageSize),
}
if token := os.Getenv(AuthTokenEnvVar); token != "" {
opts = append(opts,
grpc.UnaryInterceptor(tokenUnaryInterceptor(token)),
grpc.StreamInterceptor(tokenStreamInterceptor(token)),
)
log.Printf("gRPC auth enabled via %s", AuthTokenEnvVar)
}
return opts
}
func StartServer(address string, model AIModel) error {
lis, err := net.Listen("tcp", address)
if err != nil {
return err
}
s := grpc.NewServer(serverOpts()...)
pb.RegisterBackendServer(s, &server{llm: model})
log.Printf("gRPC Server listening at %v", lis.Addr())
if err := s.Serve(lis); err != nil {
return err
}
return nil
}
func RunServer(address string, model AIModel) (func() error, error) {
lis, err := net.Listen("tcp", address)
if err != nil {
return nil, err
}
s := grpc.NewServer(serverOpts()...)
pb.RegisterBackendServer(s, &server{llm: model})
log.Printf("gRPC Server listening at %v", lis.Addr())
if err = s.Serve(lis); err != nil {
return func() error {
return lis.Close()
}, err
}
return func() error {
s.GracefulStop()
return nil
}, nil
}