The nvidia-l4t-cuda-13-arm64 vLLM backend left `vllm` unpinned, so the
prebuilt image drifted onto whatever aarch64 wheel was latest at build
time (0.23.x). On GB10 / DGX Spark (Grace Blackwell, unified memory),
0.23 crashes deterministically during cold model loads with an empty
"Engine core initialization failed" set and pins GPU memory until a host
reboot.
vLLM 0.24.0 carries vllm-project/vllm#45179 ("release cached device
memory under pressure on UMA GPUs during weight loading"), which the
reporter verified fixes the crash on GB10. Pin the L4T requirements to
0.24.0 to match the already-pinned cublas13 build
(requirements-cublas13-after.txt) and keep the image deterministic.
Editing this file also re-triggers the single-arch L4T image build via
the path filter, republishing the gallery image with 0.24.0 (the
single-arch matrix builds again after #10703).
Closes#10722
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
Allow llama.cpp model configs to select the backend devices used for offload, matching upstream --device behavior so users can exclude a display or debug GPU.
Signed-off-by: rvmzes <rvmzes@rvmzess-MacBook-Pro.local>
Co-authored-by: rvmzes <rvmzes@rvmzess-MacBook-Pro.local>
* ⬆️ Update leejet/stable-diffusion.cpp
Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* fix(stablediffusion-ggml): pass chroma knobs via model_args after upstream API change
Upstream stable-diffusion.cpp bb849711 removed the dedicated
chroma_use_dit_mask / chroma_use_t5_mask / chroma_t5_mask_pad fields from
sd_ctx_params_t and now reads them from the generic model_args key=value
spec (parse_key_value_args). Assigning the old struct members broke the
gosd.cpp build. Emit the three options into model_args instead so the
existing chroma controls keep working. Verified by building
libgosd-fallback.so against the pinned upstream commit.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
---------
Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: mudler <2420543+mudler@users.noreply.github.com>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
GitHub Actions refuses to instantiate a matrix that would generate more
than 256 jobs. It does so silently: the job hangs forever at "Waiting for
pending jobs" and the whole run is marked `failure` while every other job
stays green. This is exactly what happened on the v4.6.1 tag build
(run 28786533892): the single-arch build matrix had grown to 268 entries,
so `backend-jobs-singlearch` (and its downstream merge) never produced a
single job, and the release build "failed" with no failing job to point at.
The single-arch list is the one that grows unbounded as backends are added,
so shard it across a fixed number of matrix jobs (SINGLEARCH_SHARDS=4,
~67 entries each today, headroom to ~1020 backends). Each merge shard
`needs:` only its matching build shard, preserving the "merge waits only on
its own build" property that keeps slow CUDA/ROCm builds from gating
multi-arch manifest assembly.
changed-backends.js now emits per-shard matrix/has-* outputs and throws
loudly if a shard ever reaches the 256 limit (telling the maintainer to
bump SINGLEARCH_SHARDS and add matching job blocks) instead of letting
GitHub drop the overflow silently. backend.yml and backend_pr.yml define
the four build + four merge shard jobs; multi-arch and darwin groups are
untouched.
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
* fix(llama-cpp): cap single-pass embedding batch to fit VRAM
Embedding/score/rerank all decode or pool the whole input in one physical
batch, so EffectiveBatchSize sized the batch to the full context window. For
a large context that makes n_ubatch huge, and the per-device CUDA compute
buffer (forward-graph scratch, ~n_ubatch * n_ctx, NOT split across GPUs)
balloons into multi-GiB: a large-context embedding model then aborts on load
(exitCode=-1) even with plenty of free VRAM. Reproduced with qwen3-embedding-4b
(context 40960 -> n_batch 40960 -> abort) and qwen3-embedding-0.6b
(n_batch 8192); pinning batch:512 avoided it.
This is the same root cause as issue #10485 (a large context turns the batch
into multi-GiB of scratch that must fit on a SINGLE card), but the single-pass
path bypassed the VRAM headroom guard the config layer already had — it
returned the unbounded context as the batch with no GPU awareness.
Make the single-pass batch VRAM-aware: cap it to the largest batch whose
compute buffer fits the per-device VRAM headroom, clamped to
[DefaultPhysicalBatch, ctx], reusing the existing computeBufferBytesPerCell and
headroom-divisor math (no duplication). Unknown per-device VRAM (0) stays
conservative (DefaultPhysicalBatch, not the context) so a detection gap can't
OOM. The GPU is resolved through an injectable package var (config.LocalGPU,
backed by sync.Once-cached xsysinfo detection) so the per-request router call
stays cheap and tests inject a deterministic device. Explicit batch: still
wins. An input longer than the cap can no longer be pooled in one pass — the
accepted tradeoff, since a batch that OOMs the device processes nothing.
Assisted-by: Claude:claude-opus-4-8 golangci-lint go-test
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* fix(config): single-pass batch follows context on unknown VRAM
The single-pass (embedding/score/rerank) batch cap must only shrink the batch
when the per-device VRAM ceiling is KNOWN. On unknown VRAM (CPU-only or a GPU
detection gap) SinglePassBatchForContext returned DefaultPhysicalBatch, which
under-sized the batch below the context — over-trimming score/embed/rerank
inputs (the modelTokenTrim middleware regression) with no OOM benefit on CPU
where the compute buffer lives in system RAM. Return the full context instead,
preserving the original single-pass behavior; the VRAM cap stays a downward
safety that only engages when VRAM is known.
Assisted-by: Claude:claude-opus-4-8 [go-test go-vet]
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>
When a model is imported without an explicit context_size, the GGUF
importer defaulted the model's context to its full trained window
(n_ctx_train). For long-context models (128k / 256k / 1M) that KV cache
cannot fit a consumer GPU, so the backend aborts on load (exitCode=-1)
even though the model file is perfectly fine. Reproduced live:
gemma-4-26b-a4b-it-qat-q4_0 defaulted to context=262144 and
qwythos-9b-claude-mythos-5-1m to 1048576, both aborting on a 20 GB card.
Instead of chasing the trained max, auto-derive a conservative default:
min(trainedMax, DefaultAutoContextSize=8192). A small model keeps its
trained window; a long-context model caps at 8k and users opt into more
via context_size. This cap applies always, including CPU / unknown-VRAM
hosts, so it never regresses those paths.
Per-device VRAM is used only as a DOWNWARD safety: when a per-device
ceiling is detected (xsysinfo.MinPerGPUVRAM) and even the 8k cap would
not fit it with headroom, step down through candidate contexts to the
largest that fits, floored at DefaultContextSize. When VRAM is unknown
(0) or no GPU is detected we do NOT clamp — the bug is GPU OOM and the
8k cap is already safe, so detection gaps must not shrink the window.
The footprint estimate reuses gpustack/gguf-parser-go's
EstimateLLaMACppRun at a given context with all layers offloaded, taking
the per-device NonUMA VRAM figure. The estimate and VRAM detection are
package vars so tests inject deterministic values. Explicit context_size
always wins (guessGGUFFromFile only acts when it is nil).
Assisted-by: Claude:claude-opus-4-8 [golangci-lint go-test]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
The `--generated-content-path` and `--upload-path` defaults were the fixed
shared locations `/tmp/generated/content` and `/tmp/localai/upload`. On any
multi-user host these collide across accounts: macOS routes `/tmp` to the
shared `/private/tmp` for every user, so whichever account starts LocalAI
first creates the parent with 0750 perms and every other account then fails
startup with:
unable to create ImageDir: "mkdir /tmp/generated/content: permission denied"
unable to create UploadDir: "mkdir /tmp/localai/upload: permission denied"
The same happens on Linux once a stale root-owned `/tmp/generated` (e.g. from
a prior `sudo` run) is left behind. This bites the desktop launcher and any
app embedding the raw binary (Wingman, nib-desktop), which start `local-ai
run` with no path flags.
Default both paths under the OS temp dir (`os.TempDir()`, honoring `$TMPDIR`;
already per-user on macOS) namespaced by the current UID
(`TMPDIR/localai-<uid>/...`), so accounts never collide while the paths stay
ephemeral. Wired via new kong vars in main.go so every consumer of the raw
binary inherits the fix. All content subdirs (audio, images) derive from
`GeneratedContentDir`, so they are fixed transitively.
As defense in depth, the launcher also anchors these two paths under its own
per-user data directory (mirroring the #10610 fix for data/config), extracted
into a testable `BuildRunArgs`.
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
* fix(reasoning): don't persist request-scoped reasoning_effort into model config
When a model sets `reasoning_effort: none` (or any default) in its YAML
without an explicit `reasoning.disable`, ApplyReasoningEffort resolves that
default at request time and sets ReasoningConfig.DisableReasoning on the
request-scoped config copy. The post-load thinking/marker probe then wrote
that request-scoped value back into the loader's persistent config via
UpdateModelConfig, making it look as though the operator had explicitly set
reasoning.disable=true. From then on, per-request `reasoning_effort` overrides
were silently ignored (an explicit operator disable wins over a request
asking to think).
DetectThinkingSupportFromBackend only fills reasoning slots that are still
nil, so a slot already set here came from ApplyReasoningEffort, not the probe.
Snapshot which slots were nil before the probe and only persist those, so the
probe's genuine backend detection is still saved while request-time reasoning
effort never leaks into the persistent config.
Fixes#10622
Signed-off-by: Tai An <antai12232931@outlook.com>
* test(reasoning): cover persist-guard added in this PR, extract for testability
ModelInference's post-probe persistence of ReasoningConfig.DisableReasoning /
DisableReasoningTagPrefill had no test: the guard logic lived inline in a
closure only reachable through a live gRPC backend. Extract it into
persistProbedReasoning (pure refactor, no behavior change) so it can be
exercised directly against a ModelConfigLoader, then add specs covering:
- a probe-filled slot (nil beforehand) gets persisted
- a slot that already carried a request-scoped value (e.g. from
reasoning_effort: none) is left alone, i.e. the #10622 regression stays
fixed
- an operator's explicit persisted disable is preserved when the guard is
false
- the media marker still persists unconditionally
Verified red/green: reverting persistProbedReasoning to the old unconditional
copy fails exactly the two guard specs.
Assisted-by: Claude:claude-sonnet-5 go vet
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* test(reasoning): ignore os.Remove error in temp file cleanup (errcheck)
Signed-off-by: Tai An <antai12232931@outlook.com>
* chore: empty commit to re-trigger flaky Agent Jobs CI test
Signed-off-by: Tai An <antai12232931@outlook.com>
---------
Signed-off-by: Tai An <antai12232931@outlook.com>
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
Operators need a scrape-friendly signal for agent-turn health (completing,
erroring, cancelled, duration) — log-derived counters proved brittle (ANSI/
timezone parsing, restart gaps). Adds localai_agent_runs_total{agent,outcome}
and localai_agent_run_seconds histogram, recorded at the Chat() response
handoff (single choke point of the local execution path). Lazy meter init,
same pattern as the PII events counter (#10641).
Signed-off-by: Stefan Walcz <stefan.walcz@walcz.de>
The no-models getting-started wizard (`.home-wizard`) rendered
left-aligned instead of centered. `.home-page` is a column flexbox with
the default `align-items: stretch`; a child with `max-width: 48rem`
cannot be stretched past its max-width, so it falls back to the
cross-start (left) edge. The populated home branch never exposed this
because its children are full-width.
Add `margin: 0 auto` to `.home-wizard` so the max-width block centers
horizontally, for both the admin getting-started wizard and the
non-admin no-models hero.
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
Additive superset of /v1/models that enriches each model entry with the
capabilities it supports plus its input/output modalities
(text / image / audio / video). Clients that only understand /v1/models
are unaffected -- they simply never call the new route.
Audio and video *input* are derived from the model's multimodal limits
(vLLM limit_mm_per_prompt), which no single usecase FLAG expresses. That
gap is exactly why a plain capability list is insufficient and this
enriched endpoint exists: an attachment router can now decide whether an
image/audio/video file can go to the active model directly, or must be
converted/transcribed first.
Capability derivation lives in core/config as the single source of truth
(ModelConfig.Capabilities / InputModalities / OutputModalities /
VisionSupported / ...); the Ollama capability surface now delegates to
it instead of keeping a parallel copy. Vision is gated on
chat/completion capability so a MediaMarker hydrated onto a non-chat
model (e.g. a pure ASR/TTS backend) no longer reports a false vision
capability.
Read-only listing: no new FLAG_* flag, reuses the existing `models`
swagger tag, and intentionally exposes no MCP admin tool (there is
nothing to manage conversationally).
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
The OAuth callback discarded the error returned by user-info resolution
before sending the generic 500, so real failures were completely opaque
in the logs: ID-token verification errors (e.g. issuer/audience mismatch
behind a reverse proxy), a missing id_token, claim-parse errors, or a
rejecting GitHub userinfo endpoint all collapsed into
"failed to fetch user info" with nothing logged.
Log the wrapped cause with xlog.Error (provider + error), matching the
code-exchange step just above it. The client-facing message is unchanged,
so no internal detail leaks to the browser.
Refs #10677
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
qwen3-tts-cpp, omnivoice-cpp, acestep-cpp and vibevoice-cpp shipped
rocm-* variants that silently ran on CPU ([Load] backend: CPU). Two
coupled defects:
- The Makefiles passed -DGGML_HIPBLAS=ON, but the vendored ggml only
understands -DGGML_HIP=ON (GGML_HIPBLAS was removed upstream), so the
ggml-hip backend target was never created and no GPU code was built.
- The CMake foreach that links the ggml GPU backends into the module
listed blas/cuda/metal/vulkan but not hip, so even a built ggml-hip
would not have been linked and its static backend registration would
never run.
CUDA users were unaffected because cublas passes the correct GGML_CUDA=ON
and the foreach already links cuda. Mirror the proven llama-cpp hipblas
block (ROCm clang CC/CXX + AMDGPU_TARGETS) and add hip to each foreach.
Upstream picks the best device via ggml_backend_init_best(), so no
runtime flag is needed once HIP is compiled and linked.
Assisted-by: Claude:claude-opus-4-8[1m] [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
The ROCm packager copied rocBLAS kernel data (rocblas/library/*.dat) into the
bundled lib/ dir and run.sh pointed ROCBLAS_TENSILE_LIBPATH at it, but the
parallel hipBLASLt data dir (hipblaslt/library/TensileLibrary_lazy_gfx*.dat)
was never packaged and no HIPBLASLT_TENSILE_LIBPATH was set. The bundled
libhipblaslt.so therefore resolved its per-arch kernel data relative to itself,
found nothing, and silently fell back to slow generic kernels, logging:
rocblaslt error: Cannot read "TensileLibrary_lazy_gfx1201.dat": No such file or directory
rocblaslt error: Could not load "TensileLibrary_lazy_gfx1201.dat"
Fix, mirroring the existing rocBLAS handling:
- package-gpu-libs.sh: extract the rocblas data-dir copy into a reusable
copy_rocm_data_dir helper and call it for both rocblas and hipblaslt.
- llama-cpp/turboquant run.sh: export HIPBLASLT_TENSILE_LIBPATH when the
bundled hipblaslt/library dir exists.
The helper takes an optional ROCM_BASE_DIRS override so the copy is unit
testable without a real ROCm install; add a regression test that runs
package_rocm_libs against a fabricated ROCm tree and asserts both data dirs
are bundled.
Note: this bundles whatever gfx*.dat the build image's ROCm provides. If a
given arch's tensile data is absent from the shipped ROCm, that arch still
needs a ROCm bump; the packaging gap itself is fixed for every supported arch.
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
POST /models/apply with an empty "id" fetches the attacker-supplied
"url" gallery config directly via http.Client, with no check that the
URL resolves to a public IP. In the default Docker deployment no API key
is configured, so any network-reachable client can coerce LocalAI into
issuing requests to internal services or cloud-metadata endpoints (and
exfiltrate a small slice of the response through the job error message).
Guard the config fetch chokepoints (GetGalleryConfigFromURL and
GetGalleryConfigFromURLWithContext, which back both the /models/apply
worker and gallery installs) with utils.ValidateExternalURL, matching
the protection already applied to the CORS proxy and image/video/audio
download paths. Only plain http(s) URLs are validated; non-network
schemes (huggingface://, github:, oci://, ollama://, file://) resolve to
fixed public services or local files and are left untouched.
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
The PII EventStore ring buffer is capacity-bound and meant for
recent-audit browsing via /api/pii/events; operators also want a
monotonic, scrape-friendly signal on /metrics — how many
detections/masks/blocks per hour, per origin, and whether the filter
stopped firing after a deploy (silent-failure class).
EventStore.Record is the single choke point every producer already goes
through (request middleware, response scrubbing, MITM proxy
connects/intercepts), so one lazily-initialised counter there covers all
paths without touching any producer:
localai_pii_events_total{kind, origin, action, direction}
Same lazy otel.Meter pattern as core/services/routing/billing, so the
counter lands on the Prometheus-backed global MeterProvider installed by
the monitoring service. No behaviour change; label cardinality is
bounded (enum-like fields only, no pattern IDs or user IDs).
Assisted-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Signed-off-by: stefanwalcz <stefan.walcz@walcz.de>
Realtime sessions previously lazy-loaded each pipeline sub-model (VAD,
transcription, LLM, TTS) on first use, so every cold session paid a
per-request model-load stall and load errors only surfaced mid-stream.
Warm the whole pipeline eagerly and blockingly at session start
(including the voice-gate speaker-recognition model, which an enforced
gate blocks each utterance on; compaction's summary_model stays lazy
since it only runs off the response path):
- Add backend.PreloadModel / PreloadModelByName as the single load path
for every modality (no transcription special-case; backend-omitted
configs are deprecated).
- The realtime session blocks on Model.Warmup and returns a
model_load_error to the client if any stage fails to load;
updateSession warms in the background. Opt out per pipeline with
pipeline.disable_warmup, exposed as a UI toggle via the
config-metadata registry.
Add a LocalAI-native POST /backend/load (and /v1/backend/load) that
pre-loads a model -- expanding realtime pipelines into their sub-models
-- as the inverse of /backend/shutdown. There is one preload engine
(backend.PreloadStages): the realtime Warmup methods, /backend/load and
the --load-to-memory startup flag all use it, so --load-to-memory now
also expands pipeline models and records load-failure traces. Pipeline
sub-model alias resolution is likewise shared
(ModelConfigLoader.LoadResolvedModelConfig). Surface the endpoint
everywhere an admin manages models:
- MCP admin tool load_model (httpapi + inproc clients, safety/catalog
prompts, catalog/dispatch tests).
- "Load into memory" action in the React models UI.
- Swagger regenerated; docs moved to the general backend-monitor page
since it is not realtime-specific.
Fix a Traces UI crash ("json: unsupported value: -Inf"): audio-snippet
RMS/peak now floor at a finite dBFS, and backend-trace data is sanitized
to drop non-finite floats before marshaling. The sanitizer is
copy-on-write -- it runs on every RecordBackendTrace, so containers are
only re-allocated on the paths that actually changed.
Migrate core/http/openresponses_test.go onto the prebuilt mock-backend
the rest of the http suite already uses -- it was the last spec still
pointing at a real HuggingFace model, so it 404'd wherever no vision
backend was built -- and fix its item_reference specs to send the
spec's "id" field instead of "item_id", which the handler never
accepted.
Assisted-by: Claude:claude-opus-4-8 Claude Code
Signed-off-by: Richard Palethorpe <io@richiejp.com>
Two bugs broke OpenAI-style tool calling on the MLX backend (and any
Python backend sharing backend/python/common), reproduced end-to-end on
LocalAI v4.5.5 with the metal-mlx backend and
mlx-community/Qwen3.5-2B-MLX-8bit.
messages_to_dicts left each tool call's function.arguments as the raw
OpenAI-wire JSON string. HuggingFace chat templates (e.g. Qwen3.5)
iterate arguments as a mapping (.items()), so any request whose history
contained a prior assistant tool_calls message failed with HTTP 500
"Generation failed: Can only get item pairs from a mapping." — breaking
every agent loop on its second turn. Decode the string back into a dict
so the template sees a mapping.
split_reasoning returned ("", text) whenever the opening think tag was
absent. Models like Qwen3.5 open the assistant turn already inside
thinking, so the generated text carries only the closing </think>; the
whole chain-of-thought leaked into content. When the opener is missing
but the closer is present, treat everything before the closer as
reasoning.
Adds platform-independent unit tests under backend/python/common
(stdlib-only, no MLX/venv required, following parent_watch_test.py).
Assisted-by: Claude Code:claude-opus-4-8
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
