The worker answered model.unload by calling Free on the first backend
process in its map, whatever model the request named. Every idle
scale-down or LRU eviction of one model could therefore empty another
model's backend on the same node. LocalAI still counted that model as
loaded, so its next request failed. A parakeet diarization model then
returned 501 "speaker profiles require a loaded speaker encoder" until
someone reloaded it by hand.
Resolve the target from the model name (all replicas), prefer an
address when the request carries one, and free nothing for an unknown
model.
Also let parakeet-cpp Diarize check the diarization model before the
speaker-profile capability. A backend with nothing loaded now answers
FailedPrecondition, which LocalAI treats as a stale replica and
reloads, instead of a final Unimplemented.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
parakeet_capi_speaker_embed_pcm and parakeet_capi_free_floats, which
VoiceEmbed binds, come from this commit.
Assisted-by: Claude Code:claude-sonnet-5-5
VoiceVerify used a fixed distance of 0.5 when the request had none. Read
voice_verify_threshold, a distance in (0, 2), from the model options and
keep 0.5 as the default. The real-library spec now cuts clips from a
two-voice recording and checks the bundle embedding size, the encoder
identity, determinism and the same-voice versus different-voice distance.
Assisted-by: Claude Code:claude-sonnet-5-5
Bind parakeet_capi_speaker_embed_pcm and parakeet_capi_free_floats with a
Dlsym probe and implement VoiceEmbed on the speaker context, so a model
with a speaker_model or a bundle speaker_component can serve the
realtime voice_recognition stage and /v1/voice/*. The response carries
the sha256 identity of the encoder weights.
VoiceVerify embeds both clips and compares them by cosine distance. It
refuses anti_spoofing because there is no such head.
A libparakeet.so without the symbols answers Unimplemented, and a model
without a speaker encoder answers FailedPrecondition.
Assisted-by: Claude Code:claude-sonnet-5-5
* chore(parakeet-cpp): bump parakeet.cpp to 2de154c
Brings in the speaker registry encoder fingerprint, the VAD segment trim
and the opt-in word filter, a fix for a per-call thread count that stayed
set on the process-wide backend after a Silero VAD pass, and bundle
components loaded from memory.
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
* feat(parakeet-cpp): encoder fingerprint for speaker naming, vad_trim and guard_* options
Speaker naming. A registered voice now carries the encoder that made it:
the embedding family (voicedetect:<arch>:<name>:<dim>) and the sha256 of
the weights. The backend reports the family of the loaded speaker model in
Status, voice enrollment from speaker_profiles stores it as encoder_family
(old entries load without it), and the registry sent to parakeet.cpp is
built with parakeet_capi_speaker_registry_add_embedding_fp. The library
then refuses a registry of another encoder family and the error names both
families; another quantization of the same family only warns. A voice with
only a weights hash gets the loaded family when the hashes are equal.
Voices without a fingerprint (registered from audio: libvoicedetect cannot
report one) keep the file-name rule and are used with a warning. The
library cannot mix them with fingerprinted voices in one registry, so a
request that has any uses the old registry for all. speaker_strict:true
drops them instead. A library without the symbols behaves as before.
Transcription. vad_trim (seconds, 0 keeps the whole cuts) goes through the
VAD options JSON, so it reaches /v1/vad and the segmenter. The guard_*
options guard_min_local_conf, guard_local_radius and guard_drop_punct_only
turn on the word filter through parakeet_capi_transcribe_path_json_with,
or through the segmenter with vad:true. They are off by default, bad
values fail the load, and a library without the symbol fails it with a
clear message. The dropped word count is logged at debug level.
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
---------
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
Saved voices send ref_text, but Fish Audio requires reference_text.
Derive the canonical parameter while preserving explicit overrides.
Both TTS modes use the shared builder.
Add regression cases and document the parameter alias.
Assisted-by: Codex:gpt-6
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com>
Newer ROCm releases split the rocBLAS and hipBLASLt TensileLibrary data
into one folder per GPU architecture (lib/hipblaslt/library/gfx1151/...).
The run.sh of the ROCm-capable backends exports ROCBLAS_TENSILE_LIBPATH and
HIPBLASLT_TENSILE_LIBPATH as the parent folder, and both libraries only look
directly in that folder, so they miss every kernel:
rocblaslt error: Cannot read ".../hipblaslt/library/TensileLibrary_lazy_gfx1151.dat"
hipModuleLoad failed: .../hipblaslt/library/Kernels.so-000-gfx1151.hsaco
and fall back to slower code paths.
rocm_tensile_dir keeps the folder when it has files at the top level (the
older flat layout) or several architecture folders, and descends into the
architecture folder when the bundle carries exactly one.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Stefan Walcz <stefan.walcz@walcz.de>
The speaker encoder metadata field makes the Rust status initializer
incomplete. Use protobuf defaults for absent metadata and memory fields.
Assisted-by: Codex:gpt-6
Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com>
* feat(parakeet-cpp): load bundle GGUF files and use their components by role
A bundle GGUF holds several models (ASR, VAD, diarization, sound events,
speaker encoder) in one file, each with its own licence. Detect a bundle
at load through parakeet_capi_bundle_components_json and open components
with parakeet_capi_load_component. The three symbols are probed together,
so an older libparakeet.so still loads plain files as before.
The only ASR component is the primary model; bundle_asr:<name> picks one
when there are several. A Silero VAD component of the primary bundle is
loaded without an option and serves /v1/vad and vad:true. The diar, ced
and voice components load on request: diar_component, sound_component and
speaker_component, or a companion option (diarization_model, sound_model,
speaker_model, vad_model) that names a bundle, even the model file itself.
vad_component picks a VAD component and implies vad:true.
A role the bundle cannot fill fails the load with the component list, and
a diarization or sound request on a model without that role names the
bundle components. Every existing option and single-file model behaves as
before.
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
* chore(parakeet-cpp): bump parakeet.cpp to 781a973
Brings in the bundle GGUF format and its C-API (parakeet_capi_load_component,
parakeet_capi_bundle_components_json, parakeet_capi_load_error).
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
* feat(parakeet-cpp): gallery entries for the bundle GGUF files, docs
Add parakeet-cpp-bundle-small (338 MB: Parakeet TDT+CTC 110M, Nemotron-3-
Diarization, CED-Small, WeSpeaker ResNet34-LM, Silero VAD), -standard
(1.1 GB, Parakeet TDT 0.6B v3 instead of the 110M model) and
-moondream-redux (215 MB: packed Redux and Silero VAD, CPU only). One
install serves transcription, VAD, diarization, sound events and speaker
naming through the component options. The existing single-purpose entries
stay.
A bundle has no single licence, so the entries use license: other and
state the licence and credit of each component in the description, with
the upstream inconsistency of the CED licence. The docs get a section on
bundles in audio-to-text with the entries, the roles, the options and the
licence notice, and pointers from the VAD, diarization and sound
classification pages. A gallery test checks the file names, checksums,
usecases and options.
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
---------
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
* feat(parakeet-cpp): add gallery entries for the VAD-only Moondream slices
Add parakeet-cpp-vad-moondream-redux and parakeet-cpp-vad-moondream-ultra.
They install the VAD head of Moondream Redux and Ultra (Q8_0) as small
files of 10 MB and 6 MB, cut out of the full models without retraining,
for the VAD endpoint. The files cannot transcribe, and a transcription
request fails with a clear error.
The files load only with a parakeet.cpp build that has VAD-only GGUF
support (parakeet.cpp pull request 87). The backend pin must move to a
commit that includes it before these entries work in a released image.
The parakeet-cpp-vad entry keeps installing Silero.
The docs list the files with the size, load time and memory compared
with loading a whole model. A gallery test checks the usecase, the file
name and the checksum of each entry.
Assisted-by: Claude Code:claude-sonnet-5-5 [golangci-lint]
* chore(parakeet-cpp): bump parakeet.cpp to e53a253
Brings in the VAD-only GGUF loader.
Assisted-by: Claude Code:claude-sonnet-5-5 [git] [gh]
* docs(gallery): link the parakeet.cpp VAD docs instead of the merged PR
Assisted-by: Claude Code:claude-sonnet-5-5 [git]
---------
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
* fix(schema): preserve SystemOne image inputs
Assisted-by: OpenAI
* test(schema): follow Ginkgo conventions for decision inputs
Assisted-by: OpenAI
* feat(llama-cpp): dispatch native decisions through Score
Upgrade the stock dependency and reconcile Score/TTS patches. Reuse native decision parsing, tasks, formatting and response-reader cleanup; preserve ordinary scoring admission and guard older dependencies.
Assisted-by: OpenAI
* refactor(systemone): share request and model validation
Assisted-by: OpenAI:gpt-5
* fix(systemone): preserve HTTP wire-byte validation limit
Keep structural validation separate from the serialized internal request bound so HTML escaping cannot reject valid HTTP payloads.
Assisted-by: OpenAI:gpt-5
* feat(systemone): bound images and account native decisions
Preserve public wire limits independently from router serialization. Reject unsupported NER images, map native request/capability errors, and stamp explicit usage once. Advertise decisions for stock llama-cpp.
Assisted-by: OpenAI
* fix(systemone): record usage on registered native route
Exercise real registration and billing with a mock native backend. Reject empty native responses, malformed image URLs, trailing JSON, and wire overflow including whitespace.
Assisted-by: OpenAI
* feat(router): add lazy native decision transport
Bind named models through internal ModelSystemOne calls with shared validation and bounded abandoned operations. Remove request and echoed-error contents from decision traces.
Assisted-by: OpenAI:gpt-5
* feat(router): classify overlapping policies with native decisions
Ask independent noul questions, validate probabilities and preserve first-superset routing. Wire the central factory with config-sensitive invalidation and cancellation-safe resolution. Document native framing and bounded operation limits.
Assisted-by: OpenAI:gpt-5
* feat(gallery): add pinned Julia-1 native decision model
Add a separate text-only llama-cpp Q8 entry with pinned Apache-2.0 source provenance and checksum. Installed using the gallery installer and exercised choice, score and noul on CPU.
Assisted-by: OpenAI
* test(router): verify native decisions through central factory
Add an opt-in real-model Ginkgo integration covering the native Go loader and C++ transport, token usage, independent overlapping labels, and candidate selection. Document owned-server execution and the intentionally non-quality threshold.
Assisted-by: Codex:gpt-5
* fix(llama-cpp): align upstream pin and preserve decision signatures
Advance to bed0a856 without losing the automated upstream bump. Detect full-request fill_task support at compile time and forward every question for Nimble framing while retaining the earlier native signature. Preserve reconciled SCORE/TTS patches; add standalone compatibility coverage.
Assisted-by: Codex:gpt-5
* feat(gallery): add native decision family defaults
Pin Laya, Kev-4B, lev, OpenJev and Nimble artifacts. Verify Laya/Kev/lev gallery installs and CPU contracts on both native pins; clearly mark OpenJev/Nimble runtime validation pending and their noncommercial licenses.
Assisted-by: OpenAI
* docs(decisions): clarify integrated Nimble prerequisite
Record the exact combined backend pin while retaining pending OpenJev and Nimble installation/runtime validation status.
Assisted-by: Codex:gpt-5
* fix(gallery): indent native decision model sequences
Match repository yamllint indentation for Laya, Kev, lev and OpenJev list fields. Parsed gallery data is unchanged; reproduce CI gallery lint failure before the whitespace-only fix and pass the same command afterward.
Assisted-by: Codex:gpt-5
* docs(decisions): record OpenJev and Nimble CPU validation
Record gallery installation, checksum/metadata verification and multiquestion native smoke results on bed0a856. Retain noncommercial and text-only limitations without accuracy or deterministic-output claims.
Assisted-by: OpenAI
* fix(ui): expose native Decisions router classifiers
Select classifier models using metadata-driven capability routing, retain tuned thresholds, and validate native decision selections before saving. Cover both native backends and create/save/reopen in the real React editor.
Assisted-by: Codex:gpt-5
* fix(router): exclude aliases from native decision discovery
Check the originally named config before advertising native Decisions eligibility. Retain target capability inheritance for ordinary generation aliases. Exercise the actual capabilities endpoint with native models on both backends, aliases, and disabled models.
Assisted-by: Codex:gpt-5
* feat(systemone): share bounded multimodal input validation
Preserve text wire limits while admitting bounded PNG/JPEG decision input. Share collection and header validation across internal and public callers and keep the native runner response budget independent.
Assisted-by: OpenAI:API-assistant
* fix(systemone): bound admission lifetimes and validate complete images
Retain shared admission leases through actual work completion, including abandoned internal operations. Decode bounded image pixels, cap public native responses before usage stamping, and preserve oversized malformed text status precedence.
Assisted-by: OpenAI:API-assistant
* fix(router): classify images before media fetching
Preserve ordered structured probes for native decisions. Defer OpenAI
media preparation until routing selects the served model, so rejected
decision URLs cannot trigger downloads before shared validation.
Guard direct image collection with context-aware shared admission.
Keep text classifiers and embedding caches from discarding image input.
Retain fail-closed classifier configuration and runtime fallback policy.
Add middleware, typed-content, admission, cancellation and cache tests.
Assisted-by: OpenAI:API-assistant
* fix(router): bound extraction before serialization
Check probe budgets before copying text or marshaling message state.
Count JSON escaping so oversized internal inputs fail before allocation.
Preserve typed Anthropic blocks through selected-model conversion and
fallback. Keep retry coverage in Ginkgo without global test registration.
Assisted-by: OpenAI
* fix(router): bound supported probe serialization
Arbitrary structs can bypass the probe budget through pointer marshalers,
string tags, and promoted fields. Accept concrete chat schema types and
plain JSON values instead of emulating arbitrary struct serialization.
Budget escaped direct prompts before marshaling so raw length cannot hide
serialized expansion. Preserve runtime fallback and reject oversized
input before invoking the decision runner.
Add Ginkgo allocation, boundary, and marshaler invocation regressions.
Six-package tests, three-package race tests, and full-T2 delta lint pass.
Assisted-by: OpenAI:GPT-5 golangci-lint
* feat(decisions): enable bounded OpenJev images
Validate native decision images before permissive media parsing and pixel
allocation. Require both decision image support and a vision projector;
missing or audio-only projectors cannot silently become text decisions.
Pin the OpenJev Q8 projector and document its license and disk footprint.
Add native safety tests, canonical limit parity, gallery and load-option
checks, and a reproducible CPU direct-RPC contrasting-image smoke.
Assisted-by: OpenAI:GPT-5
* fix(decisions): reject incomplete image streams
stb accepts corrupt PNG Adler checksums and truncated JPEG scans.
Use bounded zlib validation and strict libjpeg decoding before parsing.
Keep dimension and aggregate pixel checks ahead of decoder allocations.
Wire decoder dependencies into native builds and runtime packaging.
Add regressions for appended EOI and embedded marker bypasses.
Assisted-by: OpenAI:GPT-5
* fix(ci): gate native decision image validation
Run the decoder security tests outside the stdlib-only native suite.
Fetch vendor headers at the backend pin and provision decoder dependencies.
Gate Go limit parity and production CMake wiring without model downloads.
Assisted-by: OpenAI:GPT-5
* test(decisions): cover multimodal public API paths
Exercise shared image contracts through the registered HTTP routes and
external mock backend. Add opt-in cached gallery installation and real
OpenJev image decisions through SystemOne and both routing APIs.
Assisted-by: Codex:gpt-5
* test(decisions): assert isolation and cache bypass
Observe external RPC calls and compare complete classifier history.
Winner-only and cache-miss checks could hide dropped history or cache use.
Give real inference its own application and model directory so shared
backend mappings and loaded processes cannot affect mixed suite order.
Assisted-by: OpenAI:ChatGPT
* test(decisions): isolate fixture globals
Disable optional global services in the isolated HTTP fixture and register
cleanup before setup assertions. Verify meter provider identity survives
fixture creation and destruction.
Snapshot observed usage before assertions so failures cannot retain the
mutex. Require a successful usage stamp before checking error responses.
Assisted-by: Codex:gpt-5 golangci-lint
* fix(application): honor optional telemetry controls
Skip failover gauge registration when metrics are disabled. Register
against the application meter rather than looking up the global provider.
Allow embedders to retain the bounded routing log without billing stats.
Keep the existing default when stats are disabled. The isolated HTTP
fixture uses this option without losing its native router assertions.
Assisted-by: Codex:gpt-5 golangci-lint
---------
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
* feat(parakeet-cpp): implement the VAD call and add the vad_model option
The backend now serves the VAD gRPC call (POST /vad and /v1/vad) with the
standalone VAD of libparakeet. It accepts a Silero VAD GGUF as the model
file, or an ASR model with a VAD head (Moondream Ultra and Redux). The
request audio is float32 PCM at 16 kHz; the response lists the speech
segments in seconds, like the silero-vad backend. A model with neither
fails the request with the library message.
The vad_threshold, vad_min_pause, vad_min_speech, vad_speech_pad and
vad_max_segment options tune the segmenter. Unset values keep the
defaults of the detector in use, and a bad value fails the load.
The vad_model option names a Silero GGUF, resolved against the models
directory like the other companion files. It lets any ASR model cut long
audio at pauses through parakeet_capi_transcribe_path_json_vad_with, and
it implies vad. vad:true alone still uses the model's own head.
The new symbols are probed like the existing optional ones. A library
without them still loads; the feature that needs one fails with a clear
message only when it is used.
Assisted-by: Claude:claude-sonnet-5-5 [go test]
* feat(gallery): add parakeet-cpp VAD entries and a v3 plus Silero example
Add VAD-only entries for the parakeet-cpp backend: the VAD heads of
Moondream Redux (packed, CPU) and Ultra (Q8_0), which share their files
with the existing ASR entries, and Silero VAD v6.2.3 as a GGUF (MIT,
Silero Team). The parakeet-cpp-vad entry installs Silero; it has no variants,
because variant ranking prefers the larger build that fits and these are
different detectors.
Add parakeet-cpp-tdt-0.6b-v3-silero-vad, a v3 entry that sets vad_model
so long audio is cut at pauses by Silero.
The Silero GGUF entries point at the intended Hugging Face URL of the
file; the existing silero-vad entries are unchanged. A test checks the
usecases, the shared files and the default entry and the vad_model reference.
Assisted-by: Claude:claude-sonnet-5-5 [go test]
* docs: describe parakeet-cpp VAD and the vad_model option
Document the VAD endpoint on the parakeet-cpp backend (Silero GGUF and
the VAD heads of Moondream Ultra and Redux), the vad_* tuning options,
and the vad_model option that lets an ASR model without a VAD head cut
long audio with Silero.
Assisted-by: Claude:claude-sonnet-5-5
* chore(parakeet-cpp): bump parakeet.cpp to 6165e3d
Pin the release that adds the standalone VAD (Ultra/Redux head and
Silero) and the C API calls the backend now uses.
Assisted-by: Claude:claude-sonnet-5-5
---------
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
* chore(parakeet-cpp): bump parakeet.cpp to 11c1a0f
Picks up support for the Moondream Ultra and Redux models and the
VAD-segmented transcription entry point (C-API ABI is still 10).
Assisted-by: Claude:claude-sonnet-5-5
* feat(parakeet-cpp): add vad option for long-audio transcription
Models with a VAD head (Moondream Ultra and Redux) can cut long audio at
pauses. Setting vad:true in the model options routes offline
transcription through parakeet_capi_transcribe_path_json_vad. The symbol
is probed at startup like the other optional entry points, and vad:true
fails the load with a clear message when the library lacks it. A model
without a VAD head fails the request with the library's own message.
The option is off by default and does not affect streaming.
Also say in the load error that a packed ternary Redux model is CPU only,
because the library reports its refusal on a GPU backend through its log,
not through the C API.
Assisted-by: Claude:claude-sonnet-5-5
* feat(gallery): add Moondream Ultra and Redux for parakeet-cpp
Add five entries from the public parakeet-cpp GGUF repository: Ultra in
F16 and Q8_0, and Redux as packed ternary (CPU only, offline only) and
as dequantized F16 and Q8_0 (any backend). The entries enable vad:true so
long audio is cut at pauses. Checksums come from the repository's LFS
metadata. The weights are CC-BY-4.0.
Assisted-by: Claude:claude-sonnet-5-5
* docs(parakeet-cpp): document Moondream Ultra, Redux and the vad option
Assisted-by: Claude:claude-sonnet-5-5
---------
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
* feat(silero-vad): allow threshold/silence/pad via model options
Signed-off-by: anton ziderer <Antonziderer@mail.ru>
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(silero-vad): ignore NaN thresholds
NaN passes the detector validation but prevents speech comparisons from succeeding.
Ignore it like malformed input and document the option validation.
Convert the option tests to Ginkgo and cover invalid overrides.
Assisted-by: Codex:gpt-6
Signed-off-by: anton ziderer <Antonziderer@mail.ru>
---------
Signed-off-by: anton ziderer <Antonziderer@mail.ru>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com>
* feat(gallery): add nimble-9b-vllm-cpp decision model
Add Bespoke Nimble 9B, converted for vllm.cpp and pinned to the weights
commit 52eead25 of mudler/Bespoke-Nimble-9B-vllm-cpp (HEAD only adds the
model card). It is a redistribution of bespokelabs/Bespoke-Nimble-9B
with the LoRA merged into Qwen3.5-9B; config.json names NimbleModel, so
no hf_overrides are needed.
The artifact sits under overrides, where the installer reads it. The
entry sets an 8192-token context, Nimble's own prompt limit, and a KV
pool of 1024 blocks of 32 tokens for 4 sequences (about 1 GiB at 32 KiB
per token for the 8 full-attention layers).
Installed with local-ai models install and served on CPU through the
vllm-cpp backend: the model card's billing request gives billing
(0.986), refund 0.998 and urgency 0.33. Peak resident memory was
18.4 GB, so the description asks for about 20 GB of free RAM.
List the entry in the decisions gallery table. CLM stays out of the
gallery: the pinned engine cannot load the published head layout.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude Code:claude-sonnet-5-5
* feat(gallery): add clm-v0.1-8b-vllm-cpp and bump vllm.cpp for CLM
The CLM checkpoint on mudler/CLM-v0.1-8B-vllm-cpp stores the heads with
the reference's own tensor names (state_head.inp, hidden.N, norms.N,
out). The pinned vllm.cpp 96788348 still expects the old .0/.2/.4/.6
layout and refuses the load with "head.safetensors incomplete for
state_head". vllm.cpp a19294a9 matches the reference layout and adds the
converter that produced the upload, so move the pin there. The ABI stays
at v30.
Add the CLM entry, pinned to the weights commit 0d1903b1 (HEAD only adds
the model card), with a 4096-token context and a KV pool for 4 sequences
(about 2.25 GiB at 144 KiB per token for Qwen3-8B).
Installed with local-ai models install and served on CPU against a
libvllm built at a19294a9: the model card example (john works at google,
entity type) gives person 0.950, the same as the card. Peak resident
memory was 17.9 GB.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude Code:claude-sonnet-5-5
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
When the first result of a streamed request is an error (for example a
prompt that exceeds the context), PredictStream wrote the error message
as a Reply and only then returned the error status. LocalAI treated that
Reply as the first token: it sent the assistant role chunk and the error
text as `content` on an HTTP 200 stream. Because a chunk had already been
written, the pre-stream HTTP error path from #12204 never triggered, so
streaming clients still got a 200 with the error as model output, while
the same request without streaming correctly returns a 400.
Return the error only as the gRPC status. The e2e backend suite gets a
`context_overflow` capability (enabled for llama-cpp) that streams an
over-long prompt and asserts an error status with no content.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Stefan Walcz <stefan.walcz@walcz.de>
When an Anthropic model declines to answer, the Messages API returns
HTTP 200 with `stop_reason: "refusal"` and empty `content`. The
translate mode mapped that to a normal reply with no content, so the
OpenAI-compatible response looked like a successful completion
(`finish_reason: "stop"`, empty message). Routers, agents and UIs could
not tell "the model declined" from "the model had nothing to say", and
no fallback was triggered.
Return an explicit error for `stop_reason: "refusal"` in both the
non-streaming path and the streaming path (`message_delta`). Regular
replies, including empty `end_turn` replies, are unchanged.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Stefan Walcz <stefan.walcz@walcz.de>
AudioTranscription caught every exception and returned an empty
TranscriptResult. A failed diarization step therefore discarded a
transcript that was already finished: with an HF token that has not
accepted the terms of the gated pyannote pipeline, the download fails
with 403 and every transcription came back as an empty text with
HTTP 200.
Diarization now degrades: if it fails, the transcript is returned
without speaker labels and the reason is logged. Any other failure
aborts the call with INTERNAL instead of pretending success with an
empty text.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Stefan Walcz <stefan.walcz@walcz.de>
The environment fallback was applied whenever n_parallel was still 1
after option parsing. An explicit `parallel: 1` in the model YAML is
indistinguishable from the default that way, so it was replaced by
LLAMACPP_PARALLEL. The docs say options in the YAML take precedence
over environment variables; a single model could not be forced to one
slot while the global variable was set.
Track whether the options set the slot count and resolve it in a small
helper (parallel_params.h): option first, then LLAMACPP_PARALLEL, then
1. The helper gets a standalone unit test picked up by
`make test-backend-cpp`.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Stefan Walcz <stefan.walcz@walcz.de>
* feat(schema): validate portable speaker profiles
Add the versioned profile schema for explicit speaker enrollment.
Validate compatibility against separately supplied loaded-encoder metadata.
Reject unusable speakers, invalid vectors, and inconsistent clean spans.
This slice does not change HTTP routes, backend integration, or the UI.
Assisted-by: OpenAI:unknown
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat(parakeet): export profiles with transcripts
Export opt-in speaker profiles and trusted encoder metadata.
Replay registrations by ID so duplicate display names keep independent
vectors.
Use one profile-capable diarization for slots, names, and clean spans.
Assign timestamped ASR words to those slots without a second diarization.
Preserve legacy opt-out and no-ASR behavior, and propagate failures.
Assisted-by: OpenAI:unknown
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat(audio): enroll portable speaker profiles
Gate profile exports with voice-recognition permission and validate
registration against metadata from the loaded encoder. Preserve audio
enrollment and independent registrations with duplicate display names.
Exclude diarization and registration exchanges before API trace capture
so persisted traces cannot retain profile vectors or JSON audio.
Defer candidate dimensions to trusted loaded metadata. Sort candidates
by registration ID so incompatible profiles cannot suppress legacy voices
through registry iteration order. Keep portable identity checks closed
when trusted metadata is unavailable.
Test persisted traces, explicit slot zero, and selection through offline
and live transport. Document privacy and the ephemeral registry lifecycle.
Assisted-by: OpenAI:unknown
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat(ui): remember speakers from diarization
Add a Studio page for diarization and opt-in speaker profiles. Preview
clean intervals from the original recording before explicit registration.
Join profiles by raw speaker labels, preserve duplicate names, and relabel
turns only after a successful save. Discard stale results when the model
or recording changes. Share registration metadata with voice management
without storing vectors or recordings from this flow.
Document permissions and the global, ephemeral registry. Cover enrollment,
permissions, previews, and asynchronous races with mocked Playwright tests.
Assisted-by: OpenAI:unknown
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* docs: clarify HTTP speaker enrollment support
Replace the stale enrollment limitation with the current HTTP workflow.
Distinguish native transport from explicit registration and link its docs.
Assisted-by: OpenAI:unknown
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* chore(parakeet): pin merged speaker profile support
Use the merged commit from mudler/parakeet.cpp#80.
Its tree matches the previously accepted native pin.
Assisted-by: OpenAI:unknown
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* docs: add diarization enrollment setup example
Connect the existing gallery modes to the speaker enrollment workflow.
Show installation, private profile export, explicit raw-slot registration,
and later recognition without another export.
Assisted-by: OpenAI:unknown
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* docs(blog): explain diarization speaker profiles
Put the diarization walkthrough on the LocalAI website in the feature PR.
Cover the three gallery modes, explicit enrollment, and privacy limits.
Link setup instructions and keep availability conditional on feature support.
Assisted-by: OpenAI:unknown
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* docs(blog): focus diarization on everyday use
Explain what users can do with recordings before the setup steps.
Replace the technical walkthrough with a short Studio guide and link
readers to the existing reference for model names and developer use.
Assisted-by: OpenAI:unknown
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* docs(blog): lead with speaker capabilities
Present speaker recognition through everyday uses and a short UI flow.
Keep technical reference details in the existing documentation.
Assisted-by: OpenAI:unknown
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* fix(diarization): satisfy Go lint checks
Avoid copying protobuf message state when extending backend status, check the multipart reader close result, and document the focused testing.T lint exemptions.
Assisted-by: nib:gpt-5.6-sol
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>
FunASR left Transformers and Hugging Face Hub unconstrained, which let uv backtrack to tokenizers 0.10.3 without Python 3.12 wheels. Keep Transformers on the supported 4.x range, including the Intel upgrade profile.\n\nAssisted-by: nib:gpt-5.6-sol\nSigned-off-by: Ettore Di Giacinto <mudler@localai.io>
* ⬆️ Update ggml-org/llama.cpp
Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* fix(llama-cpp): migrate score batches to the new API
The pinned engine removes common_batch_add and the raw batch view.
Use common_batch entries and llama_process for score suffix decoding.
Read shared-prefix scores from the current common_batch view.
Validation: reproduce both compiler errors on the original patch.
The patched server context and complete grpc-server translation unit
pass g++ -std=c++17 -fsyntax-only with generated protobuf headers.
Assisted-by: Codex:gpt-6
---------
Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: mudler <2420543+mudler@users.noreply.github.com>
Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com>
* feat(voice): list registered voices and record which encoder made them
The voice registry could register, identify and forget but not list, and
it did not remember which speaker encoder produced an embedding. Add
Metadata.Model and Registry.List, answered from the index the store
registry already keeps for Forget. Needed so a backend can be given the
registered voices that match its own speaker encoder.
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
* feat(voice): store the encoder model with a registered voice
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
* feat(voice): pick the registered voices that match a speaker model
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
* feat(proto): carry known voices and speaker names on diarize and live messages
Assisted-by: Claude:claude-haiku-4-5 [Claude Code]
* feat(diarization): name speakers from the voice registry
When a diarization model has a speaker_model option, the endpoint sends
the registered voices made by that encoder to the backend. The backend's
name and name_score come back as extra fields next to the normalized
SPEAKER_NN speaker, and the speakers summary carries the first name seen
for each speaker. RTTM output and results without names are unchanged.
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
* feat(live): pass registered voices to a live session and surface speaker names
Live sessions now send the registered voices that match the model's
speaker_model to the backend, and each speaker segment carries the name
the backend matched. The realtime segment event gains an optional
speaker_name field.
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
* feat(parakeet-cpp): load a speaker model and build per-request voice registries
Adds the speaker bindings (ABI v9 and v10, probed separately), the
speaker_model, speaker_threshold and speaker_margin options, and a
per-request registry builder over the known voices.
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
* feat(parakeet-cpp): name the speakers in Diarize from the known voices
Diarize builds a per-request speaker registry from the known voices when a
speaker model is loaded, calls the named C functions, and puts each slot's
registered name and score on the segments. The registry is freed on every
path. A library without ABI 10 reports Unimplemented instead of dropping
the names.
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
* feat(parakeet-cpp): name speakers in the live scene stream
The live scene stream now begins with a known-voice registry when a
speaker model is loaded and the live config carries voices, and each
closed speaker segment takes its slot's current name from the feed's
names map. A segment that closes before its slot is identified has an
empty name. The registry is freed after the stream, on every path.
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
* feat(gallery): speaker naming entries and docs for parakeet-cpp
Add three gallery entries that load the WeSpeaker ResNet34 speaker model
next to the diarization or realtime scene models, and document speaker
names in the voice recognition, diarization, audio to text and realtime
pages.
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
* fix(parakeet-cpp): skip an unusable registered voice instead of failing the request
A registered voice with the wrong embedding size, or one the C side
refused, failed the whole diarization request, so one legacy voice broke
the model for every user. Skip such voices with a warning that does not
carry the voice name, and take the plain path when none is left.
Also map an exact 0 speaker threshold or margin to a tiny positive value,
since the C side reads 0 as "use the default", and fix a stale comment
about which contexts Free() walks.
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
* fix(diarization): warn once per model about voices from another encoder; document the privacy limit
The different-encoder warning fired on every request. Log it once per
feature and speaker model, then at debug level. Document that the global
voice registry lets any caller of a speaker_model model learn matching
names, and that skipped wrong-sized voices are logged.
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
* chore(parakeet-cpp): bump parakeet.cpp to 8c8cec0 (C-API v10) and check speaker naming against the real library
The pin moves from 623a968 to 8c8cec0, which brings in everything merged
in parakeet.cpp since: the voice identification change (C-API v9, #78) and
raw-embedding enroll plus diarize-only speaker naming (C-API v10, #79).
New real-library specs (gated on PARAKEET_BACKEND_TEST_SPEAKER_MODEL,
_DIAR_MODEL, _WAV and, for the live path, _STREAM_MODEL) name the two
speakers of two_speakers.wav from a committed pair of WeSpeaker embeddings,
with the voices passed in reversed order. They also check that the float32
threshold reaches C through purego. The shared test loader now registers
the v9/v10 and scene symbols as main.go does.
The rebase onto origin/master had no conflicts.
Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]
---------
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
G304 flags reading a path built from a variable. The directory is the model
directory from the operator's own model config, not a request input, so it is
annotated the way the other backends do it.
Assisted-by: Claude Code:claude-sonnet-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* chore(vllm-cpp): bump vllm.cpp to 967883486 (ABI v30)
Moves the pin from c3bebc357 to 967883486. On top of the Nimble decision
adapter and the Qwen3.5 vision-loader fix, this brings Tev1 on
/v1/systemone and vllm_decide (opt-in through a "Tev1Model" architecture
in config.json), a tokenizer/ subdirectory fallback so the Laya HF
snapshot loads as downloaded, a stop-token fix, a logprobs fix under async
scheduling and a pinned parakeet.cpp fetch for the diarization build.
ABI v30 only adds the diarization and speaker-attributed ASR entry
points; no existing struct or signature changed, so the purego mirrors
keep their layout and only abiVersion moves to 30. Between 4479dc99f and
967883486 vllm.h changed only in a comment.
v30 turns VLLM_CPP_WITH_DIARIZATION on by default. The fetch is pinned
now, but ON still downloads parakeet.cpp at configure time and links a
second ggml into libvllm for calls this backend never makes, so build
with the option off: the symbols stay present as refusing stubs.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude Code:claude-sonnet-5-5
* feat(vllm-cpp): add the hf_overrides engine arg
vLLM parity: engine_args.hf_overrides is a JSON object of top-level
config.json keys merged over the model directory's own config.json. The
main use is opting a published checkpoint into an engine adapter its
config does not name, such as {"architectures": ["Tev1Model"]} on the
Tev1 snapshots, which declare Qwen3_5ForConditionalGeneration.
The C ABI has no override input and the engine reads config.json from
the directory it is given, so Load builds a private overlay directory:
the merged config.json plus a symlink to every other entry of the model
directory, and passes that to the engine. The download is never written.
Free, a failed load and the next Load remove the overlay.
validModelPath and the DFlash draft resolution still see the real
directory.
A value that is not a JSON object, a .gguf model or a directory without
config.json fails the load instead of being skipped like an unknown
engine_args key, because loading the unmodified config would serve a
different architecture than the one configured.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude Code:claude-sonnet-5-5
* fix(gallery): nest vllm-cpp artifacts under overrides
artifacts: is a model-config key, and the installer reads model-config
keys only from overrides:. Five vllm-cpp entries (laya, gliner25-decide,
qwen3-vl-4b, cua-s1-forms and gliner2.5) declared it at the entry top
level, where it is silently dropped: the install reports success, writes
a config whose model is the bare HF repo id and downloads nothing, and
vllm-cpp (which does not infer artifacts) then fails the first load with
"model path not found".
Move each block under overrides:, and add a guard test that refuses a
top-level artifacts: key in the index.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude Code:claude-sonnet-5-5
* feat(gallery): add Tev1 4B and 0.8B on vllm-cpp
Two decisions entries for Together AI's Tev1 checkpoints, pinned to the
current HF revisions. Tev1 is autoregressive: vllm.cpp answers
/v1/systemone by scoring the option letters, and the same engine still
serves chat completions. The published config.json names
Qwen3_5ForConditionalGeneration, so each entry sets
hf_overrides: {architectures: [Tev1Model]} to enable the decision route
without editing the download. known_usecases is [decisions] only, since
a declared decisions list is authoritative for reservation.
The descriptions state what was checked: agreement with transformers on
CPU over seven questions (4B 7/7, max probability difference 0.0004;
0.8B 6/7 with one near tie), CPU-only for the decision route, and a
fine-tune license the model card says is still being finalized, so no
license key is set.
The Decisions API page lists both entries, drops the note that Tev1
does not serve /v1/systemone and documents the 24-option limit (Ollama
allows 26).
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude Code:claude-sonnet-5-5
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
