feat(backend): Add Sherpa ONNX backend and Omnilingual ASR
Adds a new Go backend wrapping sherpa-onnx via purego (no cgo). Same
approach as opus/stablediffusion-ggml/whisper — a thin C shim
(csrc/shim.c + shim.h → libsherpa-shim.so) wraps the bits purego
can't reach directly: nested struct config writes, result-struct field
reads, and the streaming TTS callback trampoline. The Go side uses
opaque uintptr handles and purego.NewCallback for the TTS callback.
Supports:
- VAD via sherpa-onnx's Silero VAD
- Offline ASR: Whisper, Paraformer, SenseVoice, Omnilingual CTC
- Online/streaming ASR: zipformer transducer with endpoint detection
(AudioTranscriptionStream emits delta events during decode)
- Offline TTS: VITS (LJS, etc.)
- Streaming TTS: sherpa-onnx's callback API → PCM chunks on a channel,
prefixed by a streaming WAV header
Gallery entries: omnilingual-0.3b-ctc-q8-sherpa (1600-language offline
ASR), streaming-zipformer-en-sherpa (low-latency streaming ASR),
silero-vad-sherpa, vits-ljs-sherpa.
E2E coverage: tests/e2e-backends for offline + streaming ASR,
tests/e2e for the full realtime pipeline (VAD + STT + TTS).
Assisted-by: claude-opus-4-7-1M [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
Bumps ik_llama.cpp pin to 16996aeab7. Upstream 286ce32...16996ae adds a
trailing `const struct quantize_user_data *` parameter to
`ggml_quantize_chunk` (PR ikawrakow/ik_llama.cpp#1677) but leaves
`examples/llava/clip.cpp` unchanged because their build has moved to
`examples/mtmd/`. LocalAI's prepare.sh still copies from
`examples/llava/`, so the dead 7-arg call reaches the grpc-server
compile and fails. Patch the call site to pass `nullptr` for the new
param.
Assisted-by: Claude:Opus-4.7 [Read] [Edit] [Bash]
* feat(react-ui): add Face & Voice Recognition pages
Expose the face and voice biometrics endpoints
(/v1/face/*, /v1/voice/*) through the React UI. Each page has four
tabs driving the six endpoints per modality: Analyze (demographics
with bounding boxes / waveform segments), Compare (verify with a
match gauge and live threshold slider), Enrollment (register /
identify / forget with a top-K matches view), Embedding (raw
vector inspector with sparkline + copy).
MediaInput supports file upload plus live capture: webcam
snap-to-canvas for face, MediaRecorder -> AudioContext ->
16-bit PCM mono WAV transcode for voice (libsndfile on the
backend only handles WAV/FLAC/OGG natively).
Sidebar gets a new Biometrics section feature-gated on
face_recognition / voice_recognition; routes are wrapped in
<RequireFeature>. No new dependencies -- Font Awesome icons
picked from the Free set.
Assisted-by: Claude:Opus 4.7
* fix(localai): accept data URI prefixes with codec/charset params
Browser MediaRecorder produces data URIs like
data:audio/webm;codecs=opus;base64,...
so the pre-';base64,' section can carry multiple parameter
segments. The `^data:([^;]+);base64,` regex in pkg/utils/base64.go
and core/http/endpoints/localai/audio.go only matched exactly one
segment, so recordings straight from the React UI's live-capture
tab failed the strip and then tripped the base64 decoder on the
leading 'data:' literal, surfacing as
"invalid audio base64: illegal base64 data at input byte 4"
Widened both regexes to `^data:[^,]+?;base64,` so any number of
';param=value' segments between the mime type and ';base64,' are
tolerated. Added a regression test covering the MediaRecorder
shape.
Assisted-by: Claude:Opus 4.7
* fix(insightface): scope pack ONNX loading to known manifests
LocalAI's gallery extracts buffalo_* zips flat into the models
directory, which inevitably mixes with ONNX files from other
backends (opencv face engine, MiniFASNet antispoof, WeSpeaker
voice embedding) and older buffalo pack installs. Feeding those
foreign files into insightface's model_zoo.get_model() blows up
inside the router -- it assumes a 4-D NCHW input and indexes
`input_shape[2]` on tensors that aren't shaped like a face model,
raising IndexError mid-load and leaving the backend unusable.
The router's dispatch isn't amenable to per-file try/except alone
(first-file-wins picks det_10g.onnx from buffalo_l even when the
user asked for buffalo_sc -- alphabetical order happens to favour
the wrong pack). Instead, ship an explicit manifest of the
upstream v0.7 pack contents and scope the glob to that when the
requested pack is known. The manifest is small and stable; future
packs can be added alongside or fall through to the tolerance
loop, which also swallows any remaining IndexError / ValueError
from foreign files with a clear `[insightface] skipped` stderr
line for diagnostics.
Assisted-by: Claude:Opus 4.7
* fix(speaker-recognition): extract FBank features for rank-3 ONNX encoders
Pre-exported speaker-encoder ONNX graphs come in two shapes:
rank-2 [batch, samples] -- some 3D-Speaker exports,
take raw waveform directly.
rank-3 [batch, frames, n_mels] -- WeSpeaker and most Kaldi-
lineage encoders, expect
pre-computed Kaldi FBank.
OnnxDirectEngine unconditionally fed `audio.reshape(1, -1)` --
correct for rank-2, IndexError-on-input_shape[3] on rank-3, which
surfaced to the UI as
"Invalid rank for input: feats Got: 2 Expected: 3"
Detect the input rank at session init and run Kaldi FBank
(80-dim, 25ms/10ms frames, dither=0.0, per-utterance CMN) before
the forward pass when rank>=3. All knobs are configurable via
backend options for encoders that deviate from defaults.
torchaudio.compliance.kaldi is already in the backend's
requirements (SpeechBrain pulls torchaudio in), so no new
dependency.
Assisted-by: Claude:Opus 4.7
* fix(biometrics): isolate face and voice vector stores
Face (ArcFace, 512-D) and voice (ECAPA-TDNN 192-D / WeSpeaker
256-D) biometric embeddings were colliding inside a single
in-memory local-store instance. Enrolling one after the other
failed with
"Try to add key with length N when existing length is M"
because local-store correctly refuses to mix dimensions in one
keyspace.
The registries were constructed with `storeName=""`, which in
StoreBackend() is just a WithModel() call. But ModelLoader's
cache is keyed on `modelID`, not `model` -- so both registries
collapsed to the same `modelID=""` slot and reused the same
backend process despite looking isolated on paper.
Three complementary fixes:
1. application.go -- give each registry a distinct default
namespace ("localai-face-biometrics" /
"localai-voice-biometrics"). The comment claimed
isolation, now it's actually enforced.
2. stores.go -- pass the storeName as both WithModelID and
WithModel so the ModelLoader cache key separates
namespaces and the loader spawns distinct processes.
3. local-store/store.go -- drop the Load() `opts.Model != ""`
guard. It was there to prevent generic model-loading loops
from picking up local-store by accident, but that auto-load
path is being retired; the guard now just blocks legitimate
namespace isolation. opts.Model is treated as a tag; the
per-tuple process isolation upstream handles discrimination.
Assisted-by: Claude:Opus 4.7
* fix(gallery): stale-file cleanup and upgrade-tmp directory safety
Two related robustness fixes for backend install/upgrade:
pkg/downloader/uri.go
OCI downloads passed through
if filepath.Ext(filePath) != "" ...
filePath = filepath.Dir(filePath)
which was intended to redirect file-shaped download targets
into their parent directory for OCI extraction. The heuristic
misfires on directory-shaped paths with a dot-suffix --
gallery.UpgradeBackend uses
tmpPath = "<backendsPath>/<name>.upgrade-tmp"
and Go's filepath.Ext treats ".upgrade-tmp" as an extension.
The rewrite landed the extraction at "<backendsPath>/", which
then **overwrote the real install** (backends/<name>/) with a
flat-layout file and left a stray run.sh at the top level. The
tmp dir itself stayed empty, so the validation step that
checked "<tmpPath>/run.sh" predictably failed with
"upgrade validation failed: run.sh not found in new backend"
Every manual upgrade silently corrupted the backends tree this
way. Guard the rewrite behind "target isn't already an existing
directory" -- InstallBackend / UpgradeBackend both pre-create
the target as a directory, so they get the correct behaviour;
existing file-path callers with a genuine dot-extension still
get the parent redirect.
core/gallery/backends.go
InstallBackend's MkdirAll returned ENOTDIR when something at
the target path was already a file (legacy dev builds dropped
golang backend binaries directly at `<backendsPath>/<name>`
instead of nesting them under their own subdir). That
permanently blocked reinstall and upgrade for anyone carrying
that state, since every retry hit the same error. Detect a
pre-existing non-directory, warn, and remove it before the
MkdirAll so the fresh install can write the correct nested
layout with metadata.json + run.sh.
Assisted-by: Claude:Opus 4.7
* fix(galleryop): refresh upgrade cache after backend ops
UpgradeChecker caches the last upgrade-check result and only
refreshes on the 6-hour tick or after an auto-upgrade cycle.
Manual upgrades (POST /api/backends/upgrade/:name) go through
the async galleryop worker, which completes the upgrade
correctly but never tells UpgradeChecker to re-check -- so
/api/backends/upgrades continued to list a just-upgraded backend
as upgradeable, indistinguishable from a failed upgrade, for up
to six hours.
Add an optional `OnBackendOpCompleted func()` hook on
GalleryService that fires after every successful install /
upgrade / delete on the backend channel (async, so a slow
callback doesn't stall the queue). startup.go wires it to
UpgradeChecker.TriggerCheck after both services exist. Result:
the upgrade banner clears within milliseconds of the worker
finishing.
Assisted-by: Claude:Opus 4.7
* build: prepend GOPATH/bin to PATH for protogen-go
install-go-tools runs `go install` for protoc-gen-go and
protoc-gen-go-grpc, which writes them into `go env GOPATH`/bin.
That directory isn't on every dev's PATH, and protoc resolves
its code-gen plugins via PATH, so the immediately-following
protoc invocation fails with
"protoc-gen-go: program not found"
which in turn blocks `make build` and any
`make backends/%` target that depends on build.
Prepend `go env GOPATH`/bin to PATH for the protoc invocation
so the freshly-installed plugins are found without requiring a
shell-profile change.
Assisted-by: Claude:Opus 4.7
* refactor(ui-api): non-blocking backend upgrade handler with opcache
POST /api/backends/upgrade/:name used to send the ManagementOp
directly onto the unbuffered BackendGalleryChannel, which blocked
the HTTP request whenever the galleryop worker was busy with a
prior operation. The op also didn't show up in /api/operations,
so the Backends UI couldn't reflect upgrade progress on the
affected row.
Register the op in opcache immediately, wrap it in a cancellable
context, store the cancellation function on the GalleryService,
and push onto the channel from a goroutine so the handler
returns right away. Response gains a `jobID` field and a
`message` string so clients have a consistent handle regardless
of whether the op is queued or running.
Pairs with the OnBackendOpCompleted hook added in the galleryop
commit — together the UI sees the upgrade start, watches
progress via /api/operations, and drops the "upgradeable" flag
the moment the worker finishes.
Assisted-by: Claude:Opus 4.7
* feat(insightface): add antispoofing (liveness) detection
Light up the anti_spoofing flag that was parked during the first pass.
Both FaceVerify and FaceAnalyze now run the Silent-Face MiniFASNetV2 +
MiniFASNetV1SE ensemble (~4 MB, Apache 2.0, CPU <10ms) when the flag is
set. Failed liveness on either image vetoes FaceVerify regardless of
embedding similarity. Every insightface* gallery entry now ships the
MiniFASNet ONNX weights so existing packs light up after reinstall.
Setting the flag against a model without the MiniFASNet files returns
FAILED_PRECONDITION (HTTP 412) with a clear install message — no
silent is_real=false.
FaceVerifyResponse gained per-image img{1,2}_is_real and
img{1,2}_antispoof_score (proto 9-12); FaceAnalysis's existing
is_real/antispoof_score fields are now populated. Schema fields are
pointers so they are fully absent from the JSON response when
anti_spoofing was not requested — avoids collapsing "not checked" with
"checked and fake" under Go's omitempty on bool.
Validated end-to-end over HTTP against a local install:
- verify + anti_spoofing, both real -> verified=true, score ~0.76
- verify + anti_spoofing, img2 spoof -> verified=false, img2_is_real=false
- analyze + anti_spoofing -> is_real and score per face
- flag against model without MiniFASNet -> HTTP 412 fail-loud
Assisted-by: Claude:claude-opus-4-7 go vet
* test(insightface): wire test target into test-extra
The root Makefile's `test-extra` already runs
`$(MAKE) -C backend/python/insightface test`, but the backend's
Makefile never defined the target — so the command silently errored
and the suite was never executed in CI. Adding the two-line target
(matching ace-step/Makefile) hooks `test.sh` → `runUnittests` →
`python -m unittest test.py`, which discovers both the pre-existing
engine classes (InsightFaceEngineTest, OnnxDirectEngineTest) and the
new AntispoofingTest. Each class skips gracefully when its weights
can't be downloaded from a network-restricted runner.
Assisted-by: Claude:claude-opus-4-7
* test(insightface): exercise antispoofing in e2e-backends (both paths)
Add a `face_antispoof` capability to the Ginkgo e2e suite and extend
the existing FaceVerify + FaceAnalyze specs with liveness assertions
covering BOTH paths:
real fixture -> is_real=true, score>0, verified stays true
spoof fixture -> is_real=false, verified vetoed to false
The spoof fixture is upstream's own `image_F2.jpg` (via the yakhyo
mirror) — verified locally against the MiniFASNetV2+V1SE ensemble to
classify as is_real=false with score ~0.013. That makes the assertion
deterministic across CI runs; synthetic/derived spoofs fool the model
unpredictably and would be flaky.
Makefile wires it up end-to-end:
- New INSIGHTFACE_ANTISPOOF_* cache dir + two ONNX downloads with
pinned SHAs, matching the gallery entries.
- insightface-antispoof-models target shared by both backend configs.
- FACE_SPOOF_IMAGE_URL passed via BACKEND_TEST_FACE_SPOOF_IMAGE_URL.
- Both e2e targets (buffalo-sc + opencv) now:
* depend on insightface-antispoof-models
* pass antispoof_v2_onnx / antispoof_v1se_onnx in BACKEND_TEST_OPTIONS
* include face_antispoof in BACKEND_TEST_CAPS
backend_test.go adds the new capability constant and a faceSpoofFile
fixture resolved the same way as faceFile1/2/3. Spoof assertions are
gated on both capFaceAntispoof AND faceSpoofFile being set, so a test
config that omits the spoof fixture degrades gracefully to "real path
only" instead of failing.
Assisted-by: Claude:claude-opus-4-7 go vet
* fix(llama-cpp): include server-chat.cpp in grpc-server translation unit
Upstream llama.cpp refactor (ggml-org/llama.cpp#20690) moved the
OAI/Anthropic/Responses and transcription conversion helpers out of
server-common.cpp into a new server-chat.cpp, and server-task.cpp and
server-context.cpp now call those symbols (convert_transcriptions_to_chatcmpl,
server_chat_convert_responses_to_chatcmpl, server_chat_convert_anthropic_to_oai,
server_chat_msg_diff_to_json_oaicompat) via server-chat.h.
grpc-server.cpp builds as a single translation unit by #include-ing the
upstream .cpp files directly. Without including server-chat.cpp, the
declarations are satisfied at compile time via server-chat.h but the
link step fails with undefined references once LLAMA_VERSION crosses
the refactor commit (134d6e54).
Guard the include with __has_include so the same source stays buildable
on older LLAMA_VERSION pins that predate the refactor (where prepare.sh
won't copy server-chat.cpp into tools/grpc-server/).
Assisted-by: Claude:claude-opus-4-7 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* chore(llama-cpp): bump LLAMA_VERSION to 0d0764dfd
Bump to ggml-org/llama.cpp@0d0764dfd2.
Paired with the preceding grpc-server server-chat.cpp include so the
refactor at 134d6e54 links cleanly. Supersedes PR #9494.
Assisted-by: Claude:claude-opus-4-7 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Upstream ik_llama.cpp commit e0596bf6 ("Autoparser") changed
common_params_sampling::grammar from std::string to a common_grammar
struct (type + grammar), which broke our two direct accesses:
- JSON ingest fed the field through json_value<common_grammar>(...),
for which nlohmann has no from_json adapter.
- JSON export emitted the struct directly, for which nlohmann has no
to_json adapter.
Wrap the incoming JSON string in common_grammar{COMMON_GRAMMAR_TYPE_USER, ...}
and serialize via the inner .grammar member, mirroring upstream's
examples/server/server-context.cpp.
Also bump IK_LLAMA_VERSION to 286ce324baed17c95faec77792eaa6bdb1c7a5f5
so the local-ai side lines up with the dependency bump in #9496.
Assisted-by: Claude-Code:claude-opus-4-7
* feat(voice-recognition): add /v1/voice/{verify,analyze,embed} + speaker-recognition backend
Audio analog to face recognition. Adds three gRPC RPCs
(VoiceVerify / VoiceAnalyze / VoiceEmbed), their Go service and HTTP
layers, a new FLAG_SPEAKER_RECOGNITION capability flag, and a Python
backend scaffold under backend/python/speaker-recognition/ wrapping
SpeechBrain ECAPA-TDNN with a parallel OnnxDirectEngine for
WeSpeaker / 3D-Speaker ONNX exports.
The kokoros Rust backend gets matching unimplemented trait stubs —
tonic's async_trait has no defaults, so adding an RPC without Rust
stubs breaks the build (same regression fixed by eb01c772 for face).
Swagger, /api/instructions, and the auth RouteFeatureRegistry /
APIFeatures list are updated so the endpoints surface everywhere a
client or admin UI looks.
Assisted-by: Claude:claude-opus-4-7
* feat(voice-recognition): add 1:N identify + register/forget endpoints
Mirrors the face-recognition register/identify/forget surface. New
package core/services/voicerecognition/ carries a Registry interface
and a local-store-backed implementation (same in-memory vector-store
plumbing facerecognition uses, separate instance so the embedding
spaces stay isolated).
Handlers under /v1/voice/{register,identify,forget} reuse
backend.VoiceEmbed to compute the probe vector, then delegate the
nearest-neighbour search to the registry. Default cosine-distance
threshold is tuned for ECAPA-TDNN on VoxCeleb (0.25, EER ~1.9%).
As with the face registry, the current backing is in-memory only — a
pgvector implementation is a future constructor-level swap.
Assisted-by: Claude:claude-opus-4-7
* feat(voice-recognition): gallery, docs, CI and e2e coverage
- backend/index.yaml: speaker-recognition backend entry + CPU and
CUDA-12 image variants (plus matching development variants).
- gallery/index.yaml: speechbrain-ecapa-tdnn (default) and
wespeaker-resnet34 model entries. The WeSpeaker SHA-256 is a
deliberate placeholder — the HF URI must be curl'd and its hash
filled in before the entry installs.
- docs/content/features/voice-recognition.md: API reference + quickstart,
mirrors the face-recognition docs.
- React UI: CAP_SPEAKER_RECOGNITION flag export (consumers follow face's
precedent — no dedicated tab yet).
- tests/e2e-backends: voice_embed / voice_verify / voice_analyze specs.
Helper resolveFaceFixture is reused as-is — the only thing face/voice
share is "download a file into workDir", so no need for a new helper.
- Makefile: docker-build-speaker-recognition + test-extra-backend-
speaker-recognition-{ecapa,all} targets. Audio fixtures default to
VCTK p225/p226 samples from HuggingFace.
- CI: test-extra.yml grows a tests-speaker-recognition-grpc job
mirroring insightface. backend.yml matrix gains CPU + CUDA-12 image
build entries — scripts/changed-backends.js auto-picks these up.
Assisted-by: Claude:claude-opus-4-7
* feat(voice-recognition): wire a working /v1/voice/analyze head
Adds AnalysisHead: a lazy-loading age / gender / emotion inference
wrapper that plugs into both SpeechBrainEngine and OnnxDirectEngine.
Defaults to two open-licence HuggingFace checkpoints:
- audeering/wav2vec2-large-robust-24-ft-age-gender (Apache 2.0) —
age regression + 3-way gender (female / male / child).
- superb/wav2vec2-base-superb-er (Apache 2.0) — 4-way emotion.
Both are optional and degrade gracefully when transformers or the
model can't be loaded — the engine raises NotImplementedError so the
gRPC layer returns 501 instead of a generic 500.
Emotion classes pass through from the model (neutral/happy/angry/sad
on the default checkpoint); the e2e test now accepts any non-empty
dominant gender so custom age_gender_model overrides don't fail it.
Adds transformers to the backend's CPU and CUDA-12 requirements.
Assisted-by: Claude:claude-opus-4-7
* fix(voice-recognition): pin real WeSpeaker ResNet34 ONNX SHA-256
Replaces the placeholder hash in gallery/index.yaml with the actual
SHA-256 (7bb2f06e…) of the upstream
Wespeaker/wespeaker-voxceleb-resnet34-LM ONNX at ~25MB. `local-ai
models install wespeaker-resnet34` now succeeds.
Assisted-by: Claude:claude-opus-4-7
* fix(voice-recognition): soundfile loader + honest analyze default
Two issues surfaced on first end-to-end smoke with the actual backend
image:
1. torchaudio.load in torchaudio 2.8+ requires the torchcodec package
for audio decoding. Switch SpeechBrainEngine._load_waveform to the
already-present soundfile (listed in requirements.txt) plus a numpy
linear resample to 16kHz. Drops a heavy ffmpeg-linked dep and the
codepath we never exercise (torchaudio's ffmpeg backend).
2. The AnalysisHead was defaulting to audeering/wav2vec2-large-robust-
24-ft-age-gender, but AutoModelForAudioClassification silently
mangles that checkpoint — it reports the age head weights as
UNEXPECTED and re-initialises the classifier head with random
values, so the "gender" output is noise and there is no age output
at all. Make age/gender opt-in instead (empty default; users wire
a cleanly-loadable Wav2Vec2ForSequenceClassification checkpoint via
age_gender_model: option). Emotion keeps its working Superb default.
Also broaden _infer_age_gender's tensor-shape handling and catch
runtime exceptions so a dodgy age/gender head never takes down the
whole analyze call.
Docs and README updated to match the new policy.
Verified with the branch-scoped gallery on localhost:
- voice/embed → 192-d ECAPA-TDNN vector
- voice/verify → same-clip dist≈6e-08 verified=true; cross-speaker
dist 0.76–0.99 verified=false (as expected)
- voice/register/identify/forget → round-trip works, 404 on unknown id
- voice/analyze → emotion populated, age/gender omitted (opt-in)
Assisted-by: Claude:claude-opus-4-7
* fix(voice-recognition): real CI audio fixtures + fixture-agnostic verify spec
Two issues surfaced after CI actually ran the speaker-recognition e2e
target (I'd curl-tested against a running server but hadn't run the
make target locally):
1. The default BACKEND_TEST_VOICE_AUDIO_* URLs pointed at
huggingface.co/datasets/CSTR-Edinburgh/vctk paths that return 404
(the dataset is gated). Swap them for the speechbrain test samples
served from github.com/speechbrain/speechbrain/raw/develop/ —
public, no auth, correct 16kHz mono format.
2. The VoiceVerify spec required d(file1,file2) < 0.4, assuming
file1/file2 were same-speaker. The speechbrain samples are three
different speakers (example1/2/5), and there is no easy un-gated
source of true same-speaker audio pairs (VoxCeleb/VCTK/LibriSpeech
are all license- or size-gated for CI use). Replace the ceiling
check with a relative-ordering assertion: d(pair) > d(same-clip)
for both file2 and file3 — that's enough to prove the embeddings
encode speaker info, and it works with any three non-identical
clips. Actual speaker ordering d(1,2) vs d(1,3) is logged but not
asserted.
Local run: 4/4 voice specs pass (Health, LoadModel, VoiceEmbed,
VoiceVerify) on the built backend image. 12 non-voice specs skipped
as expected.
Assisted-by: Claude:claude-opus-4-7
* fix(ci): checkout with submodules in the reusable backend_build workflow
The kokoros Rust backend build fails with
failed to read .../sources/Kokoros/kokoros/Cargo.toml: No such file
because the reusable backend_build.yml workflow's actions/checkout
step was missing `submodules: true`. Dockerfile.rust does `COPY .
/LocalAI`, and without the submodule files the subsequent `cargo
build` can't find the vendored Kokoros crate.
The bug pre-dates this PR — scripts/changed-backends.js only triggers
the kokoros image job when something under backend/rust/kokoros or
the shared proto changes, so master had been coasting past it. The
voice-recognition proto addition re-broke it.
Other checkouts in backend.yml (llama-cpp-darwin) and test-extra.yml
(insightface, kokoros, speaker-recognition) already pass
`submodules: true`; this brings the shared backend image builder in
line.
Assisted-by: Claude:claude-opus-4-7
The backend.proto was updated to add FaceVerify and FaceAnalyze RPCs
(face detection support), but the Rust KokorosService was never updated
to match the regenerated tonic trait, breaking compilation with E0046:
not all trait items implemented, missing: `face_verify`, `face_analyze`
Stubs both methods as unimplemented, matching the pattern used for the
other RPCs Kokoros does not support.
Assisted-by: Claude:claude-opus-4-7 [Claude Code]
* feat(face-recognition): add insightface backend for 1:1 verify, 1:N identify, embedding, detection, analysis
Adds face recognition as a new first-class capability in LocalAI via the
`insightface` Python backend, with a pluggable two-engine design so
non-commercial (insightface model packs) and commercial-safe
(OpenCV Zoo YuNet + SFace) models share the same gRPC/HTTP surface.
New gRPC RPCs (backend/backend.proto):
* FaceVerify(FaceVerifyRequest) returns FaceVerifyResponse
* FaceAnalyze(FaceAnalyzeRequest) returns FaceAnalyzeResponse
Existing Embedding and Detect RPCs are reused (face image in
PredictOptions.Images / DetectOptions.src) for face embedding and
face detection respectively.
New HTTP endpoints under /v1/face/:
* verify — 1:1 image pair same-person decision
* analyze — per-face age + gender (emotion/race reserved)
* register — 1:N enrollment; stores embedding in vector store
* identify — 1:N recognition; detect → embed → StoresFind
* forget — remove a registered face by opaque ID
Service layer (core/services/facerecognition/) introduces a
`Registry` interface with one in-memory `storeRegistry` impl backed
by LocalAI's existing local-store gRPC vector backend. HTTP handlers
depend on the interface, not on StoresSet/StoresFind directly, so a
persistent PostgreSQL/pgvector implementation can be slotted in via a
single constructor change in core/application (TODO marker in the
package doc).
New usecase flag FLAG_FACE_RECOGNITION; insightface is also wired
into FLAG_DETECTION so /v1/detection works for face bounding boxes.
Gallery (backend/index.yaml) ships three entries:
* insightface-buffalo-l — SCRFD-10GF + ArcFace R50 + genderage
(~326MB pre-baked; non-commercial research use only)
* insightface-opencv — YuNet + SFace (~40MB pre-baked; Apache 2.0)
* insightface-buffalo-s — SCRFD-500MF + MBF (runtime download; non-commercial)
Python backend (backend/python/insightface/):
* engines.py — FaceEngine protocol with InsightFaceEngine and
OnnxDirectEngine; resolves model paths relative to the backend
directory so the same gallery config works in docker-scratch and
in the e2e-backends rootfs-extraction harness.
* backend.py — gRPC servicer implementing Health, LoadModel, Status,
Embedding, Detect, FaceVerify, FaceAnalyze.
* install.sh — pre-bakes buffalo_l + OpenCV YuNet/SFace inside the
backend directory so first-run is offline-clean (the final scratch
image only preserves files under /<backend>/).
* test.py — parametrized unit tests over both engines.
Tests:
* Registry unit tests (go test -race ./core/services/facerecognition/...)
— in-memory fake grpc.Backend, table-driven, covers register/
identify/forget/error paths + concurrent access.
* tests/e2e-backends/backend_test.go extended with face caps
(face_detect, face_embed, face_verify, face_analyze); relative
ordering + configurable verifyCeiling per engine.
* Makefile targets: test-extra-backend-insightface-buffalo-l,
-opencv, and the -all aggregate.
* CI: .github/workflows/test-extra.yml gains tests-insightface-grpc,
auto-triggered by changes under backend/python/insightface/.
Docs:
* docs/content/features/face-recognition.md — feature page with
license table, quickstart (defaults to the commercial-safe model),
models matrix, API reference, 1:N workflow, storage caveats.
* Cross-refs in object-detection.md, stores.md, embeddings.md, and
whats-new.md.
* Contributor README at backend/python/insightface/README.md.
Verified end-to-end:
* buffalo_l: 6/6 specs (health, load, face_detect, face_embed,
face_verify, face_analyze).
* opencv: 5/5 specs (same minus face_analyze — SFace has no
demographic head; correctly skipped via BACKEND_TEST_CAPS).
Assisted-by: Claude:claude-opus-4-7
* fix(face-recognition): move engine selection to model gallery, collapse backend entries
The previous commit put engine/model_pack options on backend gallery
entries (`backend/index.yaml`). That was wrong — `GalleryBackend`
(core/gallery/backend_types.go:32) has no `options` field, so the
YAML decoder silently dropped those keys and all three "different
insightface-*" backend entries resolved to the same container image
with no distinguishing configuration.
Correct split:
* `backend/index.yaml` now has ONE `insightface` backend entry
shipping the CPU + CUDA 12 container images. The Python backend
bundles both the non-commercial insightface model packs
(buffalo_l / buffalo_s) and the commercial-safe OpenCV Zoo
weights (YuNet + SFace); the active engine is selected at
LoadModel time via `options: ["engine:..."]`.
* `gallery/index.yaml` gains three model entries —
`insightface-buffalo-l`, `insightface-opencv`,
`insightface-buffalo-s` — each setting the appropriate
`overrides.backend` + `overrides.options` so installing one
actually gives the user the intended engine. This matches how
`rfdetr-base` lives in the model gallery against the `rfdetr`
backend.
The earlier e2e tests passed despite this bug because the Makefile
targets pass `BACKEND_TEST_OPTIONS` directly to LoadModel via gRPC,
bypassing any gallery resolution entirely. No code changes needed.
Assisted-by: Claude:claude-opus-4-7
* feat(face-recognition): cover all supported models in the gallery + drop weight baking
Follows up on the model-gallery split: adds entries for every model
configuration either engine actually supports, and switches weight
delivery from image-baked to LocalAI's standard gallery mechanism.
Gallery now has seven `insightface-*` model entries (gallery/index.yaml):
insightface (family) — non-commercial research use
• buffalo-l (326MB) — SCRFD-10GF + ResNet50 + genderage, default
• buffalo-m (313MB) — SCRFD-2.5GF + ResNet50 + genderage
• buffalo-s (159MB) — SCRFD-500MF + MBF + genderage
• buffalo-sc (16MB) — SCRFD-500MF + MBF, recognition only
(no landmarks, no demographics — analyze
returns empty attributes)
• antelopev2 (407MB) — SCRFD-10GF + ResNet100@Glint360K + genderage
OpenCV Zoo family — Apache 2.0 commercial-safe
• opencv — YuNet + SFace fp32 (~40MB)
• opencv-int8 — YuNet + SFace int8 (~12MB, ~3x smaller, faster on CPU)
Model weights are no longer baked into the backend image. The image
now ships only the Python runtime + libraries (~275MB content size,
~1.18GB disk vs ~1.21GB when weights were baked). Weights flow through
LocalAI's gallery mechanism:
* OpenCV variants list `files:` with ONNX URIs + SHA-256, so
`local-ai models install insightface-opencv` pulls them into the
models directory exactly like any other gallery-managed model.
* insightface packs (upstream distributes .zip archives only, not
individual ONNX files) auto-download on first LoadModel via
FaceAnalysis' built-in machinery, rooted at the LocalAI models
directory so they live alongside everything else — same pattern
`rfdetr` uses with `inference.get_model()`.
Backend changes (backend/python/insightface/):
* backend.py — LoadModel propagates `ModelOptions.ModelPath` (the
LocalAI models directory) to engines via a `_model_dir` hint.
This replaces the earlier ModelFile-dirname approach; ModelPath
is the canonical "models directory" variable set by the Go loader
(pkg/model/initializers.go:144) and is always populated.
* engines.py::_resolve_model_path — picks up `model_dir` and searches
it (plus basename-in-model-dir) before falling back to the dev
script-dir. This is how OnnxDirectEngine finds gallery-downloaded
YuNet/SFace files by filename only.
* engines.py::_flatten_insightface_pack — new helper that works
around an upstream packaging inconsistency: buffalo_l/s/sc zips
expand flat, but buffalo_m and antelopev2 zips wrap their ONNX
files in a redundant `<name>/` directory. insightface's own
loader looks one level too shallow and fails. We call
`ensure_available()` explicitly, flatten if nested, then hand to
FaceAnalysis.
* engines.py::InsightFaceEngine.prepare — root-resolution order now
includes the `_model_dir` hint so packs download into the LocalAI
models directory by default.
* install.sh — no longer pre-downloads any weights. Everything is
gallery-managed now.
* smoke.py (new) — parametrized smoke test that iterates over every
gallery configuration, simulating the LocalAI install flow
(creates a models dir, fetches OpenCV files with checksum
verification, lets insightface auto-download its packs), then
runs detect + embed + verify (+ analyze where supported) through
the in-process BackendServicer.
* test.py — OnnxDirectEngineTest no longer hardcodes `/models/opencv/`
paths; downloads ONNX files to a temp dir at setUpClass time and
passes ModelPath accordingly.
Registry change (core/services/facerecognition/store_registry.go):
* `dim=0` in NewStoreRegistry now means "accept whatever dimension
arrives" — needed because the backend supports 512-d ArcFace/MBF
and 128-d SFace via the same Registry. A non-zero dim still fails
fast with ErrDimensionMismatch.
* core/application plumbs `faceEmbeddingDim = 0`, explaining the
rationale in the comment.
Backend gallery description updated to reflect that the image carries
no weights — it's just Python + engines.
Smoke-tested all 7 configurations against the rebuilt image (with the
flatten fix applied), exit 0:
PASS: insightface-buffalo-l faces=6 dim=512 same-dist=0.000
PASS: insightface-buffalo-sc faces=6 dim=512 same-dist=0.000
PASS: insightface-buffalo-s faces=6 dim=512 same-dist=0.000
PASS: insightface-buffalo-m faces=6 dim=512 same-dist=0.000
PASS: insightface-antelopev2 faces=6 dim=512 same-dist=0.000
PASS: insightface-opencv faces=6 dim=128 same-dist=0.000
PASS: insightface-opencv-int8 faces=6 dim=128 same-dist=0.000
7/7 passed
Assisted-by: Claude:claude-opus-4-7
* fix(face-recognition): pre-fetch OpenCV ONNX for e2e target; drop stale pre-baked claim
CI regression from the previous commit: I moved OpenCV Zoo weight
delivery to LocalAI's gallery `files:` mechanism, but the
test-extra-backend-insightface-opencv target was still passing
relative paths `detector_onnx:models/opencv/yunet.onnx` in
BACKEND_TEST_OPTIONS. The e2e suite drives LoadModel directly over
gRPC without going through the gallery, so those relative paths
resolved to nothing and OpenCV's ONNXImporter failed:
LoadModel failed: Failed to load face engine:
OpenCV(4.13.0) ... Can't read ONNX file: models/opencv/yunet.onnx
Fix: add an `insightface-opencv-models` prerequisite target that
fetches the two ONNX files (YuNet + SFace) to a deterministic host
cache at /tmp/localai-insightface-opencv-cache/, verifies SHA-256,
and skips the download on re-runs. The opencv test target depends on
it and passes absolute paths in BACKEND_TEST_OPTIONS, so the backend
finds the files via its normal absolute-path resolution branch.
Also refresh the buffalo_l comment: it no longer says "pre-baked"
(nothing is — the pack auto-downloads from upstream's GitHub release
on first LoadModel, same as in CI).
Locally verified: `make test-extra-backend-insightface-opencv` passes
5/5 specs (health, load, face_detect, face_embed, face_verify).
Assisted-by: Claude:claude-opus-4-7
* feat(face-recognition): add POST /v1/face/embed + correct /v1/embeddings docs
The docs promised that /v1/embeddings returns face vectors when you
send an image data-URI. That was never true: /v1/embeddings is
OpenAI-compatible and text-only by contract — its handler goes
through `core/backend/embeddings.go::ModelEmbedding`, which sets
`predictOptions.Embeddings = s` (a string of TEXT to embed) and never
populates `predictOptions.Images[]`. The Python backend's Embedding
gRPC method does handle Images[] (that's how /v1/face/register reaches
it internally via `backend.FaceEmbed`), but the HTTP embeddings
endpoint wasn't wired to populate it.
Rather than overload /v1/embeddings with image-vs-text detection —
messy, and the endpoint is OpenAI-compatible by design — add a
dedicated /v1/face/embed endpoint that wraps `backend.FaceEmbed`
(already used internally by /v1/face/register and /v1/face/identify).
Matches LocalAI's convention of a dedicated path per non-standard flow
(/v1/rerank, /v1/detection, /v1/face/verify etc.).
Response:
{
"embedding": [<dim> floats, L2-normed],
"dim": int, // 512 for ArcFace R50 / MBF, 128 for SFace
"model": "<name>"
}
Live-tested on the opencv engine: returns a 128-d L2-normalized vector
(sum(x^2) = 1.0000). Sentinel in docs updated to note /v1/embeddings
is text-only and point image users at /v1/face/embed instead.
Assisted-by: Claude:claude-opus-4-7
* fix(http): map malformed image input + gRPC status codes to proper 4xx
Image-input failures on LocalAI's single-image endpoints (/v1/detection,
/v1/face/{verify,analyze,embed,register,identify}) have historically
returned 500 — even when the client was the one who sent garbage.
Classic example: you POST an "image" that isn't a URL, isn't a
data-URI, and isn't a valid JPEG/PNG — the server shouldn't claim
that's its fault.
Two helpers land in core/http/endpoints/localai/images.go and every
single-image handler is switched over:
* decodeImageInput(s)
Wraps utils.GetContentURIAsBase64 and turns any failure
(invalid URL, not a data-URI, download error, etc.) into
echo.NewHTTPError(400, "invalid image input: ...").
* mapBackendError(err)
Inspects the gRPC status on a backend call error and maps:
INVALID_ARGUMENT → 400 Bad Request
NOT_FOUND → 404 Not Found
FAILED_PRECONDITION → 412 Precondition Failed
Unimplemented → 501 Not Implemented
All other codes fall through unchanged (still 500).
Before, my 1×1 PNG error-path test returned:
HTTP 500 "rpc error: code = InvalidArgument desc = failed to decode one or both images"
After:
HTTP 400 "failed to decode one or both images"
Scope-limited to the LocalAI single-image endpoints. The multi-modal
paths (middleware/request.go, openresponses/responses.go,
openai/realtime.go) intentionally log-and-skip individual media parts
when decoding fails — different design intent (graceful degradation
of a multi-part message), not a 400-worthy failure. Left untouched.
Live-verified: every error case in /tmp/face_errors.py now returns
4xx with a meaningful message; the "image with no face (1x1 PNG)"
case specifically went from 500 → 400.
Assisted-by: Claude:claude-opus-4-7
* refactor(face-recognition): insightface packs go through gallery files:, drop FaceAnalysis
Follows up on the discovery that LocalAI's gallery `files:` mechanism
handles archives (zip, tar.gz, …) via mholt/archiver/v3 — the rhasspy
piper voices use exactly this pattern. Insightface packs are zip
archives, so we can now deliver them the same way every other
gallery-managed model gets delivered: declaratively, checksum-verified,
through LocalAI's standard download+extract pipeline.
Two changes:
1. Gallery (gallery/index.yaml) — every insightface-* entry gains a
`files:` list with the pack zip's URI + SHA-256. `local-ai models
install insightface-buffalo-l` now fetches the zip, verifies the
hash, and extracts it into the models directory. No more reliance
on insightface's library-internal `ensure_available()` auto-download
or its hardcoded `BASE_REPO_URL`.
2. InsightFaceEngine (backend/python/insightface/engines.py) — drops
the FaceAnalysis wrapper and drives insightface's `model_zoo`
directly. The ~50 lines FaceAnalysis provides — glob ONNX files,
route each through `model_zoo.get_model()`, build a
`{taskname: model}` dict, loop per-face at inference — are
reimplemented in `InsightFaceEngine`. The actual inference classes
(RetinaFace, ArcFaceONNX, Attribute, Landmark) are still
insightface's — we only replicate the glue, so drift risk against
upstream is minimal.
Why drop FaceAnalysis: it hard-codes a `<root>/models/<name>/*.onnx`
layout that doesn't match what LocalAI's zip extraction produces.
LocalAI unpacks archives flat into `<models_dir>`. Upstream packs
are inconsistent — buffalo_l/s/sc ship ONNX at the zip root (lands
at `<models_dir>/*.onnx`), buffalo_m/antelopev2 wrap in a redundant
`<name>/` dir (lands at `<models_dir>/<name>/*.onnx`). The new
`_locate_insightface_pack` helper searches both locations plus
legacy paths and returns whichever has ONNX files. Replaces the
earlier `_flatten_insightface_pack` helper (which tried to fight
FaceAnalysis's layout expectations; now we just find the files
wherever they are).
Net effect for users: install once via LocalAI's managed flow,
weights live alongside every other model, progress shows in the
jobs endpoint, no first-load network call. Same API surface,
cleaner plumbing.
Assisted-by: Claude:claude-opus-4-7
* fix(face-recognition): CI's insightface e2e path needs the pack pre-fetched
The e2e suite drives LoadModel over gRPC without going through LocalAI's
gallery flow, so the engine's `_model_dir` option (normally populated
from ModelPath) is empty. Previously the insightface target relied on
FaceAnalysis auto-download to paper over this, but we dropped
FaceAnalysis in favor of direct model_zoo calls — so the buffalo_l
target started failing at LoadModel with "no insightface pack found".
Mirror the opencv target's pre-fetch pattern: download buffalo_sc.zip
(same SHA as the gallery entry), extract it on the host, and pass
`root:<dir>` so the engine locates the pack without needing
ModelPath. Switched to buffalo_sc (smallest pack, ~16MB) to keep CI
fast; it covers the same insightface engine code path as buffalo_l.
Face analyze cap dropped since buffalo_sc has no age/gender head.
Assisted-by: Claude:claude-opus-4-7[1m]
* feat(face-recognition): surface face-recognition in advertised feature maps
The six /v1/face/* endpoints were missing from every place LocalAI
advertises its feature surface to clients:
* api_instructions — the machine-readable capability index at
GET /api/instructions. Added `face-recognition` as a dedicated
instruction area with an intro that calls out the in-memory
registry caveat and the /v1/face/embed vs /v1/embeddings split.
* auth/permissions — added FeatureFaceRecognition constant, routed
all six face endpoints through it so admins can gate them per-user
like any other API feature. Default ON (matches the other API
features).
* React UI capabilities — CAP_FACE_RECOGNITION symbol mapped to
FLAG_FACE_RECOGNITION. Declared only for now; the Face page is a
follow-up (noted in the plan).
Instruction count bumped 9 → 10; test updated.
Assisted-by: Claude:claude-opus-4-7[1m]
* docs(agents): capture advertising-surface steps in the endpoint guide
Before this change, adding a new /v1/* endpoint reliably missed one or
more of: the swagger @Tags annotation, the /api/instructions registry,
the auth RouteFeatureRegistry, and the React UI CAP_* symbol. The
endpoint would work but be invisible to API consumers, admins, and the
UI — and nothing in the existing docs said to look in those places.
Extend .agents/api-endpoints-and-auth.md with a new "Advertising
surfaces" section covering all four surfaces (swagger tags, /api/
instructions, capabilities.js, docs/), and expand the closing checklist
so it's impossible to ship a feature without visiting each one. Hoist a
one-liner reminder into AGENTS.md's Quick Reference so agents skim it
before diving in.
Assisted-by: Claude:claude-opus-4-7[1m]
whisperx has no upstream AMD GPU support and its core transcription path
(faster-whisper -> ctranslate2) falls back to CPU on AMD since the PyPI
ctranslate2 is CUDA-only. The torch rocm wheels would accelerate only the
alignment/diarization stages, producing a misleadingly half-working image.
Drop the hipblas variant rather than shipping a partially accelerated build
users can't distinguish from the real thing. AMD hosts now fall through
the capability map to cpu-whisperx / cpu-whisperx-development.
Also removes the now-dangling rocm-whisperx assertion from
pkg/system/capabilities_test.go and the ROCm mention from the whisperx
row in docs/content/reference/compatibility-table.md.
Assisted-by: Claude Code:claude-opus-4-7
gen_video's ffmpeg subprocess was relying on the filename extension to
choose the output container. Distributed LocalAI hands the backend a
staging path (e.g. /staging/localai-output-NNN.tmp) that is renamed to
.mp4 only after the backend returns, so ffmpeg saw a .tmp extension and
bailed with "Unable to choose an output format". Inference had already
completed and the frames were piped in, producing the cryptic
"video inference failed (code 1)" at the API layer.
Pass -f mp4 explicitly so the container is selected by flag instead of
by filename suffix.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(turboquant): drop ignore-eos patch, bump fork to b8967-627ebbc
The upstream PR #21203 (server: respect the ignore_eos flag) has been
merged into the TheTom/llama-cpp-turboquant feature/turboquant-kv-cache
branch. With the fix now in-tree, 0001-server-respect-the-ignore-eos-flag.patch
no longer applies (git apply sees its additions already present) and the
nightly turboquant bump fails.
Retire the patch and bump the pin to the first fork revision that carries
the merged fix (tag feature-turboquant-kv-cache-b8967-627ebbc). This matches
the contract in apply-patches.sh: drop patches once the fork catches up.
* fix(turboquant): patch out get_media_marker() call in grpc-server copy
CI turboquant docker build was failing with:
grpc-server.cpp:2825:40: error: use of undeclared identifier
'get_media_marker'
The call was added by 7809c5f5 (PR #9412) to propagate the mtmd random
per-server media marker upstream landed in ggml-org/llama.cpp#21962. The
TheTom/llama-cpp-turboquant fork branched before that PR, so its
server-common.cpp has no such symbol.
Extend patch-grpc-server.sh to substitute get_media_marker() with the
legacy "<__media__>" literal in the build-time grpc-server.cpp copy
under turboquant-<flavor>-build/. The fork's mtmd_default_marker()
returns exactly that string, and the Go layer falls back to the same
sentinel when media_marker is empty, so behavior on the turboquant path
is unchanged. Patched copy only — the shared source under
backend/cpp/llama-cpp/ keeps compiling against vanilla upstream.
Verified by running `make docker-build-turboquant` locally end-to-end:
all five flavors (avx, avx2, avx512, fallback, grpc+rpc-server) now
compile past the previous failure and the image tags successfully.
The backend.proto was updated to add AudioTranscriptionStream RPC, but
the Rust KokorosService was never updated to match the regenerated
tonic trait, breaking compilation with E0046.
Stubs the new streaming method as unimplemented, matching the pattern
used for the other streaming RPCs Kokoros does not support.
Add gfx1151 (AMD Strix Halo / Ryzen AI MAX) to the default AMDGPU_TARGETS
list in the llama-cpp backend Makefile. ROCm 7.2.1 ships with gfx1151
Tensile libraries, so this architecture should be included in default builds.
Also expose AMDGPU_TARGETS as an ARG/ENV in Dockerfile.llama-cpp so that
users building for non-default GPU architectures can override the target
list via --build-arg AMDGPU_TARGETS=<arch>. Previously, passing
-DAMDGPU_TARGETS=<arch> through CMAKE_ARGS was silently overridden by
the Makefile's own append of the default target list.
Fixes#9374
Signed-off-by: Keith Mattix <keithmattix2@gmail.com>
Co-authored-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
The shared grpc-server CMakeLists hardcoded `llama-common`, the post-rename
target name in upstream llama.cpp. The turboquant fork branched before that
rename and still exposes the helpers library as `common`, so the name
silently degraded to a plain `-llama-common` link flag, the PUBLIC include
directory was never propagated, and tools/server/server-task.h failed to
find common.h during turboquant-<flavor> builds.
Upstream llama.cpp (PR #21962) switched the server-side mtmd media
marker to a random per-server string and removed the legacy
"<__media__>" backward-compat replacement in mtmd_tokenizer. The
Go layer still emitted the hardcoded "<__media__>", so on the
non-tokenizer-template path the prompt arrived with a marker mtmd
did not recognize and tokenization failed with "number of bitmaps
(1) does not match number of markers (0)".
Report the active media marker via ModelMetadataResponse.media_marker
and substitute the sentinel "<__media__>" with it right before the
gRPC call, after the backend has been loaded and probed. Also skip
the Go-side multimodal templating entirely when UseTokenizerTemplate
is true — llama.cpp's oaicompat_chat_params_parse already injects its
own marker and StringContent is unused in that path. Backends that do
not expose the field keep the legacy "<__media__>" behavior.
Upstream llama.cpp (45cac7ca) renamed the CMake library target
`common` to `llama-common`. Linking the old name caused
`target_include_directories(... PUBLIC .)` from the common/ dir
to not propagate, so `#include "common.h"` failed when building
grpc-server.
chore: ⬆️ Update TheTom/llama-cpp-turboquant to `45f8a066ed5f5bb38c695cec532f6cef9f4efa9d`
Drop 0002-ggml-rpc-bump-op-count-to-97.patch; the fork now has
GGML_OP_COUNT == 97 and RPC_PROTO_PATCH_VERSION 2 upstream.
Fetch all tags in backend/cpp/llama-cpp/Makefile so tag-only commits
(the new turboquant pin is reachable only through the tag
feature-turboquant-kv-cache-b8821-45f8a06) can be checked out.
Drop the 295-line vendor/llama.py fork in favor of `tinygrad.apps.llm`,
which now provides the Transformer blocks, GGUF loader (incl. Q4/Q6/Q8
quantization), KV-cache and generate loop we were maintaining ourselves.
What changed:
- New vendor/appsllm_adapter.py (~90 LOC) — HF -> GGUF-native state-dict
keymap, Transformer kwargs builder, `_embed_hidden` helper, and a hard
rejection of qkv_bias models (Qwen2 / 2.5 are no longer supported; the
apps.llm Transformer ties `bias=False` on Q/K/V projections).
- backend.py routes both safetensors and GGUF paths through
apps.llm.Transformer. Generation now delegates to its (greedy-only)
`generate()`; Temperature / TopK / TopP / RepetitionPenalty are still
accepted on the wire but ignored — documented in the module docstring.
- Jinja chat render now passes `enable_thinking=False` so Qwen3's
reasoning preamble doesn't eat the tool-call token budget on small
models.
- Embedding path uses `_embed_hidden` (block stack + output_norm) rather
than the custom `embed()` method we were carrying on the vendored
Transformer.
- test.py gains TestAppsLLMAdapter covering the keymap rename, tied
embedding fallback, unknown-key skipping, and qkv_bias rejection.
- Makefile fixtures move from Qwen/Qwen2.5-0.5B-Instruct to Qwen/Qwen3-0.6B
(apps.llm-compatible) and tool_parser from qwen3_xml to hermes (the
HF chat template emits hermes-style JSON tool calls).
Verified with the docker-backed targets:
test-extra-backend-tinygrad 5/5 PASS
test-extra-backend-tinygrad-embeddings 3/3 PASS
test-extra-backend-tinygrad-whisper 4/4 PASS
test-extra-backend-tinygrad-sd 3/3 PASS
* feat(backends): add sglang
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* fix(sglang): force AVX-512 CXXFLAGS and disable CI e2e job
sgl-kernel's shm.cpp uses __m512 AVX-512 intrinsics unconditionally;
-march=native fails on CI runners without AVX-512 in /proc/cpuinfo.
Force -march=sapphirerapids so the build always succeeds, matching
sglang upstream's docker/xeon.Dockerfile recipe.
The resulting binary still requires an AVX-512 capable CPU at runtime,
so disable tests-sglang-grpc in test-extra.yml for the same reason
tests-vllm-grpc is disabled. Local runs with make test-extra-backend-sglang
still work on hosts with the right SIMD baseline.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* fix(sglang): patch CMakeLists.txt instead of CXXFLAGS for AVX-512
CXXFLAGS with -march=sapphirerapids was being overridden by
add_compile_options(-march=native) in sglang's CPU CMakeLists.txt,
since CMake appends those flags after CXXFLAGS. Sed-patch the
CMakeLists.txt directly after cloning to replace -march=native.
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat(backend): add tinygrad multimodal backend
Wire tinygrad as a new Python backend covering LLM text generation with
native tool-call extraction, embeddings, Stable Diffusion 1.x image
generation, and Whisper speech-to-text from a single self-contained
container.
Backend (`backend/python/tinygrad/`):
- `backend.py` gRPC servicer with LLM Predict/PredictStream (auto-detects
Llama / Qwen2 / Mistral architecture from `config.json`, supports
safetensors and GGUF), Embedding via mean-pooled last hidden state,
GenerateImage via the vendored SD1.x pipeline, AudioTranscription +
AudioTranscriptionStream via the vendored Whisper inference loop, plus
Tokenize / ModelMetadata / Status / Free.
- Vendored upstream model code under `vendor/` (MIT, headers preserved):
llama.py with an added `qkv_bias` flag for Qwen2-family bias support
and an `embed()` method that returns the last hidden state, plus
clip.py, unet.py, stable_diffusion.py (trimmed to drop the MLPerf
training branch that pulls `mlperf.initializers`), audio_helpers.py
and whisper.py (trimmed to drop the pyaudio listener).
- Pluggable tool-call parsers under `tool_parsers/`: hermes (Qwen2.5 /
Hermes), llama3_json (Llama 3.1+), qwen3_xml (Qwen 3), mistral
(Mistral / Mixtral). Auto-selected from model architecture or `Options`.
- `install.sh` pins Python 3.11.14 (tinygrad >=0.12 needs >=3.11; the
default portable python is 3.10).
- `package.sh` bundles libLLVM.so.1 + libedit/libtinfo/libgomp/libsndfile
into the scratch image. `run.sh` sets `CPU_LLVM=1` and `LLVM_PATH` so
tinygrad's CPU device uses the in-process libLLVM JIT instead of
shelling out to the missing `clang` binary.
- Local unit tests for Health and the four parsers in `test.py`.
Build wiring:
- Root `Makefile`: `.NOTPARALLEL`, `prepare-test-extra`, `test-extra`,
`BACKEND_TINYGRAD = tinygrad|python|.|false|true`,
docker-build-target eval, and `docker-build-backends` aggregator.
- `.github/workflows/backend.yml`: cpu / cuda12 / cuda13 build matrix
entries (mirrors the transformers backend placement).
- `backend/index.yaml`: `&tinygrad` meta + cpu/cuda12/cuda13 image
entries (latest + development).
E2E test wiring:
- `tests/e2e-backends/backend_test.go` gains an `image` capability that
exercises GenerateImage and asserts a non-empty PNG is written to
`dst`. New `BACKEND_TEST_IMAGE_PROMPT` / `BACKEND_TEST_IMAGE_STEPS`
knobs.
- Five new make targets next to `test-extra-backend-vllm`:
- `test-extra-backend-tinygrad` — Qwen2.5-0.5B-Instruct + hermes,
mirrors the vllm target 1:1 (5/9 specs in ~57s).
- `test-extra-backend-tinygrad-embeddings` — same model, embeddings
via LLM hidden state (3/9 in ~10s).
- `test-extra-backend-tinygrad-sd` — stable-diffusion-v1-5 mirror,
health/load/image (3/9 in ~10min, 4 diffusion steps on CPU).
- `test-extra-backend-tinygrad-whisper` — openai/whisper-tiny.en
against jfk.wav from whisper.cpp samples (4/9 in ~49s).
- `test-extra-backend-tinygrad-all` aggregate.
All four targets land green on the first MVP pass: 15 specs total, 0
failures across LLM+tools, embeddings, image generation, and speech
transcription.
* refactor(tinygrad): collapse to a single backend image
tinygrad generates its own GPU kernels (PTX renderer for CUDA, the
autogen ctypes wrappers for HIP / Metal / WebGPU) and never links
against cuDNN, cuBLAS, or any toolkit-version-tied library. The only
runtime dependency that varies across hosts is the driver's libcuda.so.1
/ libamdhip64.so, which are injected into the container at run time by
the nvidia-container / rocm runtimes. So unlike torch- or vLLM-based
backends, there is no reason to ship per-CUDA-version images.
- Drop the cuda12-tinygrad and cuda13-tinygrad build-matrix entries
from .github/workflows/backend.yml. The sole remaining entry is
renamed to -tinygrad (from -cpu-tinygrad) since it is no longer
CPU-only.
- Collapse backend/index.yaml to a single meta + development pair.
The meta anchor carries the latest uri directly; the development
entry points at the master tag.
- run.sh picks the tinygrad device at launch time by probing
/usr/lib/... for libcuda.so.1 / libamdhip64.so. When libcuda is
visible we set CUDA=1 + CUDA_PTX=1 so tinygrad uses its own PTX
renderer (avoids any nvrtc/toolkit dependency); otherwise we fall
back to HIP or CLANG. CPU_LLVM=1 + LLVM_PATH keep the in-process
libLLVM JIT for the CLANG path.
- backend.py's _select_tinygrad_device() is trimmed to a CLANG-only
fallback since production device selection happens in run.sh.
Re-ran test-extra-backend-tinygrad after the change:
Ran 5 of 9 Specs in 56.541 seconds — 5 Passed, 0 Failed
* feat(backend): add turboquant llama.cpp-fork backend
turboquant is a llama.cpp fork (TheTom/llama-cpp-turboquant, branch
feature/turboquant-kv-cache) that adds a TurboQuant KV-cache scheme.
It ships as a first-class backend reusing backend/cpp/llama-cpp sources
via a thin wrapper Makefile: each variant target copies ../llama-cpp
into a sibling build dir and invokes llama-cpp's build-llama-cpp-grpc-server
with LLAMA_REPO/LLAMA_VERSION overridden to point at the fork. No
duplication of grpc-server.cpp — upstream fixes flow through automatically.
Wires up the full matrix (CPU, CUDA 12/13, L4T, L4T-CUDA13, ROCm, SYCL
f32/f16, Vulkan) in backend.yml and the gallery entries in index.yaml,
adds a tests-turboquant-grpc e2e job driven by BACKEND_TEST_CACHE_TYPE_K/V=q8_0
to exercise the KV-cache config path (backend_test.go gains dedicated env
vars wired into ModelOptions.CacheTypeKey/Value — a generic improvement
usable by any llama.cpp-family backend), and registers a nightly auto-bump
PR in bump_deps.yaml tracking feature/turboquant-kv-cache.
scripts/changed-backends.js gets a special-case so edits to
backend/cpp/llama-cpp/ also retrigger the turboquant CI pipeline, since
the wrapper reuses those sources.
* feat(turboquant): carry upstream patches against fork API drift
turboquant branched from llama.cpp before upstream commit 66060008
("server: respect the ignore eos flag", #21203) which added the
`logit_bias_eog` field to `server_context_meta` and a matching
parameter to `server_task::params_from_json_cmpl`. The shared
backend/cpp/llama-cpp/grpc-server.cpp depends on that field, so
building it against the fork unmodified fails.
Cherry-pick that commit as a patch file under
backend/cpp/turboquant/patches/ and apply it to the cloned fork
sources via a new apply-patches.sh hook called from the wrapper
Makefile. Simplifies the build flow too: instead of hopping through
llama-cpp's build-llama-cpp-grpc-server indirection, the wrapper now
drives the copied Makefile directly (clone -> patch -> build).
Drop the corresponding patch whenever the fork catches up with
upstream — the build fails fast if a patch stops applying, which
is the signal to retire it.
* docs: add turboquant backend section + clarify cache_type_k/v
Document the new turboquant (llama.cpp fork with TurboQuant KV-cache)
backend alongside the existing llama-cpp / ik-llama-cpp sections in
features/text-generation.md: when to pick it, how to install it from
the gallery, and a YAML example showing backend: turboquant together
with cache_type_k / cache_type_v.
Also expand the cache_type_k / cache_type_v table rows in
advanced/model-configuration.md to spell out the accepted llama.cpp
quantization values and note that these fields apply to all
llama.cpp-family backends, not just vLLM.
* feat(turboquant): patch ggml-rpc GGML_OP_COUNT assertion
The fork adds new GGML ops bringing GGML_OP_COUNT to 97, but
ggml/include/ggml-rpc.h static-asserts it equals 96, breaking
the GGML_RPC=ON build paths (turboquant-grpc / turboquant-rpc-server).
Carry a one-line patch that updates the expected count so the
assertion holds. Drop this patch whenever the fork fixes it upstream.
* feat(turboquant): allow turbo* KV-cache types and exercise them in e2e
The shared backend/cpp/llama-cpp/grpc-server.cpp carries its own
allow-list of accepted KV-cache types (kv_cache_types[]) and rejects
anything outside it before the value reaches llama.cpp's parser. That
list only contains the standard llama.cpp types — turbo2/turbo3/turbo4
would throw "Unsupported cache type" at LoadModel time, meaning
nothing the LocalAI gRPC layer accepted was actually fork-specific.
Add a build-time augmentation step (patch-grpc-server.sh, called from
the turboquant wrapper Makefile) that inserts GGML_TYPE_TURBO2_0/3_0/4_0
into the allow-list of the *copied* grpc-server.cpp under
turboquant-<flavor>-build/. The original file under backend/cpp/llama-cpp/
is never touched, so the stock llama-cpp build keeps compiling against
vanilla upstream which has no notion of those enum values.
Switch test-extra-backend-turboquant to set
BACKEND_TEST_CACHE_TYPE_K=turbo3 / _V=turbo3 so the e2e gRPC suite
actually runs the fork's TurboQuant KV-cache code paths (turbo3 also
auto-enables flash_attention in the fork). Picking q8_0 here would
only re-test the standard llama.cpp path that the upstream llama-cpp
backend already covers.
Refresh the docs (text-generation.md + model-configuration.md) to
list turbo2/turbo3/turbo4 explicitly and call out that you only get
the TurboQuant code path with this backend + a turbo* cache type.
* fix(turboquant): rewrite patch-grpc-server.sh in awk, not python3
The builder image (ubuntu:24.04 stage-2 in Dockerfile.turboquant)
does not install python3, so the python-based augmentation step
errored with `python3: command not found` at make time. Switch to
awk, which ships in coreutils and is already available everywhere
the rest of the wrapper Makefile runs.
* Apply suggestion from @mudler
Signed-off-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
---------
Signed-off-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
