Files
LocalAI/core/services/nodes/inflight.go
LocalAI [bot] 294170d3ed feat(backend): add depth-anything (Depth Anything 3) C++/ggml backend + gallery (#10352)
* feat(backend): add depth-anything (Depth Anything 3) C++/ggml backend + gallery

Mirrors the locate-anything-cpp backend to register a new depth-anything
backend that wraps the Depth Anything 3 ggml port (depth-anything.cpp) via
purego (cgo-less, no Python at inference).

- backend/go/depth-anything-cpp/: gRPC backend (Load + Predict + GenerateImage),
  purego binding to the da_capi_* C ABI, CMake/Makefile/run/package/test scripts
  building depth-anything.cpp's DA_SHARED static .so per CPU variant.
- backend/index.yaml: depth-anything backend meta + all hardware-variant
  capability entries (cpu/cuda12/cuda13/intel-sycl-f32+f16/vulkan/nvidia-l4t).
- gallery/index.yaml: 8 Depth Anything 3 GGUF models (base q4_k/q8_0/f16/f32,
  small, large, giant, mono-large).
- .github/backend-matrix.yml: one build entry per hardware variant.

Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* feat(depth): typed Depth RPC + REST endpoint exposing full DA3 data

Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* fix(depth): pin depth-anything.cpp to e0b6814 (ABI 3 dense C-API)

The Depth RPC handler calls da_capi_depth_dense / da_capi_points (C-API ABI 3);
pin the native build to the commit that exports them.

Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* fix(depth): pin depth-anything.cpp to v0.1.0 release (b515c31)

Repoint the native version from the now-orphaned e0b6814 to the
b515c31 release commit, kept alive by the upstream v0.1.0 tag.
C-API is unchanged (da_capi_abi_version == 3).

Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* fix(depth): wire depth-anything-cpp into build, CI bump, and importer

The backend dir, gallery index, and CI build-matrix were present but the
backend was never wired into the integration points that adding-backends.md
requires:

- root Makefile: add to .NOTPARALLEL, the test-extra chain, a BACKEND_*
  definition, the docker-build target eval, and docker-build-backends
  (mirrors parakeet-cpp; the backend's own Makefile already documented that
  its `test` target is driven by test-extra).
- bump_deps.yaml: register the DEPTHANYTHING_VERSION pin so the daily
  auto-bump bot tracks mudler/depth-anything.cpp master (it cannot see an
  unregistered Makefile pin).
- import form: add a preference-only KnownBackend entry so depth-anything is
  selectable at /import-model (mirrors sam3-cpp; no reliable GGUF auto-detect
  signal, so pref-only per the doc's default).

changed-backends.js needs no entry: the generic golang suffix branch already
resolves backend/go/depth-anything-cpp/.

Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* feat(depth): auto-detect importer for depth-anything GGUFs

Replace the preference-only entry with a real auto-detect importer
(mirrors parakeet-cpp / locate-anything):

- DepthAnythingImporter matches a .gguf whose name carries a
  depth-anything token (depth-anything-<size>-<quant>.gguf), so
  /import-model recognises mudler/depth-anything.cpp-gguf repos and direct
  GGUF URLs without an explicit backend preference. preferences.backend=
  "depth-anything" still forces it.
- Registered before LlamaCPPImporter so its GGUF bundles aren't claimed by
  the generic .gguf importer; the narrow name match means it cannot claim
  arbitrary llama GGUFs or the upstream safetensors PyTorch repos.
- Multi-quant repos pick the smallest quant by default (q4_k -> ... -> f32,
  depth stays >0.998 corr even at q4_k); quantizations preference overrides.
- Drops the now-redundant knownPrefOnlyBackends entry (importer-backed
  backends are not listed there, matching parakeet-cpp).
- Table-driven Ginkgo test covers detection, negative cases (llama GGUF,
  upstream safetensors), default/override/fallback quant pick, and direct
  URL import. 10/10 specs pass.

Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

* fix(depth): check conn.Close error in grpc Depth client (errcheck)

The new Depth() client method used a bare `defer conn.Close()`. golangci-lint
runs with new-from-merge-base, so although the 39 sibling methods use the same
bare form (grandfathered), the newly added line trips errcheck. Drop the result
explicitly to satisfy the linter.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8

* fix(depth): bump depth-anything.cpp to v0.1.1 (embeddable CMake)

v0.1.0 (b515c31) used ${CMAKE_SOURCE_DIR} for its include dirs, which
points at the parent project when built via add_subdirectory() as this
backend does, so the container build failed with missing stb_image.h /
da_gguf_keys.h. v0.1.1 (2d42897) switches to project-relative paths.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8

* fix(depth): resolve gosec findings in the backend wrapper

The code-scanning gate flagged three new failure-level alerts in
godepthanythingcpp.go (gosec runs with -no-fail; GitHub gates on new alerts):

- G301: export dirs were created with 0o755. Tighten to 0o750 (no world
  access needed for backend-written export output).
- G304: writeDepthPNG creates req.GetDst(). That path is chosen by the
  LocalAI core as the intended output destination (same pattern every
  image backend uses), not attacker input, so annotate with #nosec G304
  and document why.

The remaining G103 "audit unsafe" notes on the unsafe.Slice C-buffer copies
are warning-level (the same purego interop whisper/parakeet use) and do not
gate the check, per the supertonic exclusion precedent in secscan.yaml.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8

* fix(depth): bump depth-anything.cpp to v0.1.2 (CUDA cross-build arch)

v0.1.1 forced CMAKE_CUDA_ARCHITECTURES=native, which breaks the GPU-less
l4t/cublas CI builds (nvcc "Unsupported gpu architecture 'compute_'" on
CMake 3.22). v0.1.2 (442eea4) drops the override and lets ggml pick its
default cross-build arch list.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8

---------

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
2026-06-16 16:28:28 +02:00

245 lines
10 KiB
Go

package nodes
import (
"context"
"sync"
"time"
"github.com/mudler/LocalAI/pkg/grpc"
"github.com/mudler/LocalAI/pkg/grpc/grpcerrors"
pb "github.com/mudler/LocalAI/pkg/grpc/proto"
"github.com/mudler/xlog"
ggrpc "google.golang.org/grpc"
)
// InFlightTrackingClient wraps a grpc.Backend and tracks active inference requests
// in the NodeRegistry. This allows the router's eviction logic to know which models
// are actively serving and should not be unloaded.
//
// Per-replica: a single tracker instance is bound to (nodeID, modelName, replicaIndex).
// The router constructs one tracker per Route() result, so each in-flight tick lands
// on the correct row even when multiple replicas of the same model live on the same node.
type InFlightTrackingClient struct {
grpc.Backend // embed for passthrough of untracked methods
registry InFlightTracker
nodeID string
modelName string
replicaIndex int
firstOnce sync.Once // guards onFirstComplete
onFirstComplete func() // called once after the first tracked inference call completes
}
// NewInFlightTrackingClient wraps a gRPC backend client with in-flight tracking.
func NewInFlightTrackingClient(inner grpc.Backend, registry InFlightTracker, nodeID, modelName string, replicaIndex int) *InFlightTrackingClient {
return &InFlightTrackingClient{
Backend: inner,
registry: registry,
nodeID: nodeID,
modelName: modelName,
replicaIndex: replicaIndex,
}
}
// OnFirstComplete registers a callback that fires once after the first tracked
// inference call completes. This is used to release the initial in-flight
// reservation (set during model load) after the triggering request finishes,
// so that in-flight returns to 0 when the model is idle.
func (c *InFlightTrackingClient) OnFirstComplete(fn func()) {
c.onFirstComplete = fn
}
func (c *InFlightTrackingClient) track(ctx context.Context) func() {
if err := c.registry.IncrementInFlight(ctx, c.nodeID, c.modelName, c.replicaIndex); err != nil {
xlog.Warn("Failed to increment in-flight counter", "node", c.nodeID, "model", c.modelName, "replica", c.replicaIndex, "error", err)
return func() {}
}
return func() {
decCtx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
defer cancel()
c.registry.DecrementInFlight(decCtx, c.nodeID, c.modelName, c.replicaIndex)
// Release the initial reservation after the first inference call completes
if c.onFirstComplete != nil {
c.firstOnce.Do(c.onFirstComplete)
}
}
}
// reconcile self-heals stale routing: when a backend reports that the model is
// no longer loaded (the process survived but the model was evicted, while the
// registry still lists it as loaded), it drops the replica row so the next
// request triggers a fresh load instead of routing back here. Without this the
// model stays unreachable until the controller restarts. The original error is
// returned unchanged.
func (c *InFlightTrackingClient) reconcile(err error) error {
if !grpcerrors.IsModelNotLoaded(err) {
return err
}
rmCtx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
defer cancel()
if rmErr := c.registry.RemoveNodeModel(rmCtx, c.nodeID, c.modelName, c.replicaIndex); rmErr != nil {
xlog.Warn("Failed to drop stale replica after model-not-loaded",
"node", c.nodeID, "model", c.modelName, "replica", c.replicaIndex, "error", rmErr)
} else {
xlog.Warn("Backend reports model not loaded; dropped stale replica so the next request reloads",
"node", c.nodeID, "model", c.modelName, "replica", c.replicaIndex)
}
return err
}
// --- Tracked inference methods ---
func (c *InFlightTrackingClient) Predict(ctx context.Context, in *pb.PredictOptions, opts ...ggrpc.CallOption) (*pb.Reply, error) {
defer c.track(ctx)()
reply, err := c.Backend.Predict(ctx, in, opts...)
return reply, c.reconcile(err)
}
func (c *InFlightTrackingClient) PredictStream(ctx context.Context, in *pb.PredictOptions, f func(reply *pb.Reply), opts ...ggrpc.CallOption) error {
defer c.track(ctx)()
return c.reconcile(c.Backend.PredictStream(ctx, in, f, opts...))
}
func (c *InFlightTrackingClient) Embeddings(ctx context.Context, in *pb.PredictOptions, opts ...ggrpc.CallOption) (*pb.EmbeddingResult, error) {
defer c.track(ctx)()
res, err := c.Backend.Embeddings(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) GenerateImage(ctx context.Context, in *pb.GenerateImageRequest, opts ...ggrpc.CallOption) (*pb.Result, error) {
defer c.track(ctx)()
res, err := c.Backend.GenerateImage(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) GenerateVideo(ctx context.Context, in *pb.GenerateVideoRequest, opts ...ggrpc.CallOption) (*pb.Result, error) {
defer c.track(ctx)()
res, err := c.Backend.GenerateVideo(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) TTS(ctx context.Context, in *pb.TTSRequest, opts ...ggrpc.CallOption) (*pb.Result, error) {
defer c.track(ctx)()
res, err := c.Backend.TTS(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) TTSStream(ctx context.Context, in *pb.TTSRequest, f func(reply *pb.Reply), opts ...ggrpc.CallOption) error {
defer c.track(ctx)()
return c.reconcile(c.Backend.TTSStream(ctx, in, f, opts...))
}
func (c *InFlightTrackingClient) SoundGeneration(ctx context.Context, in *pb.SoundGenerationRequest, opts ...ggrpc.CallOption) (*pb.Result, error) {
defer c.track(ctx)()
res, err := c.Backend.SoundGeneration(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) AudioTranscription(ctx context.Context, in *pb.TranscriptRequest, opts ...ggrpc.CallOption) (*pb.TranscriptResult, error) {
defer c.track(ctx)()
res, err := c.Backend.AudioTranscription(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) AudioTranscriptionStream(ctx context.Context, in *pb.TranscriptRequest, f func(chunk *pb.TranscriptStreamResponse), opts ...ggrpc.CallOption) error {
defer c.track(ctx)()
return c.reconcile(c.Backend.AudioTranscriptionStream(ctx, in, f, opts...))
}
func (c *InFlightTrackingClient) Detect(ctx context.Context, in *pb.DetectOptions, opts ...ggrpc.CallOption) (*pb.DetectResponse, error) {
defer c.track(ctx)()
res, err := c.Backend.Detect(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) Depth(ctx context.Context, in *pb.DepthRequest, opts ...ggrpc.CallOption) (*pb.DepthResponse, error) {
defer c.track(ctx)()
res, err := c.Backend.Depth(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) Rerank(ctx context.Context, in *pb.RerankRequest, opts ...ggrpc.CallOption) (*pb.RerankResult, error) {
defer c.track(ctx)()
res, err := c.Backend.Rerank(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) VAD(ctx context.Context, in *pb.VADRequest, opts ...ggrpc.CallOption) (*pb.VADResponse, error) {
defer c.track(ctx)()
res, err := c.Backend.VAD(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) Diarize(ctx context.Context, in *pb.DiarizeRequest, opts ...ggrpc.CallOption) (*pb.DiarizeResponse, error) {
defer c.track(ctx)()
res, err := c.Backend.Diarize(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) FaceVerify(ctx context.Context, in *pb.FaceVerifyRequest, opts ...ggrpc.CallOption) (*pb.FaceVerifyResponse, error) {
defer c.track(ctx)()
res, err := c.Backend.FaceVerify(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) FaceAnalyze(ctx context.Context, in *pb.FaceAnalyzeRequest, opts ...ggrpc.CallOption) (*pb.FaceAnalyzeResponse, error) {
defer c.track(ctx)()
res, err := c.Backend.FaceAnalyze(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) VoiceVerify(ctx context.Context, in *pb.VoiceVerifyRequest, opts ...ggrpc.CallOption) (*pb.VoiceVerifyResponse, error) {
defer c.track(ctx)()
res, err := c.Backend.VoiceVerify(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) VoiceAnalyze(ctx context.Context, in *pb.VoiceAnalyzeRequest, opts ...ggrpc.CallOption) (*pb.VoiceAnalyzeResponse, error) {
defer c.track(ctx)()
res, err := c.Backend.VoiceAnalyze(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) VoiceEmbed(ctx context.Context, in *pb.VoiceEmbedRequest, opts ...ggrpc.CallOption) (*pb.VoiceEmbedResponse, error) {
defer c.track(ctx)()
res, err := c.Backend.VoiceEmbed(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) TokenClassify(ctx context.Context, in *pb.TokenClassifyRequest, opts ...ggrpc.CallOption) (*pb.TokenClassifyResponse, error) {
defer c.track(ctx)()
res, err := c.Backend.TokenClassify(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) Score(ctx context.Context, in *pb.ScoreRequest, opts ...ggrpc.CallOption) (*pb.ScoreResponse, error) {
defer c.track(ctx)()
res, err := c.Backend.Score(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) AudioEncode(ctx context.Context, in *pb.AudioEncodeRequest, opts ...ggrpc.CallOption) (*pb.AudioEncodeResult, error) {
defer c.track(ctx)()
res, err := c.Backend.AudioEncode(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) AudioDecode(ctx context.Context, in *pb.AudioDecodeRequest, opts ...ggrpc.CallOption) (*pb.AudioDecodeResult, error) {
defer c.track(ctx)()
res, err := c.Backend.AudioDecode(ctx, in, opts...)
return res, c.reconcile(err)
}
func (c *InFlightTrackingClient) AudioTransform(ctx context.Context, in *pb.AudioTransformRequest, opts ...ggrpc.CallOption) (*pb.AudioTransformResult, error) {
defer c.track(ctx)()
res, err := c.Backend.AudioTransform(ctx, in, opts...)
return res, c.reconcile(err)
}
// AudioTransformStream, AudioToAudioStream and Forward are deliberately left as
// embedded passthrough: they return a stream client and the inference spans the
// stream's lifetime, not the constructor call. Wrapping the constructor with
// track() would increment and immediately decrement (and fire onFirstComplete)
// before any audio flows. Tracking those correctly needs the done() func tied to
// stream close, which the current Backend interface doesn't surface here.