grpc-server.cpp forced params.cache_ram_mib = -1 (no limit) since #7009.
Since v4.3 kv_unified and cache_idle_slots are on by default, so every
distinct prompt now leaves its slot KV state in the host-side prompt
cache, and without a limit the backend grows until the host runs out of
memory.
Measured on gfx1151 (Strix Halo, 128 GB), llama-cpp backend, one request
at a time, 100 distinct prompts of ~2000 characters plus a fixed system
prompt, max_tokens 200:
model cache_ram RSS loaded -> after 100
gemma-4-26B-A4B (q8_0 KV) -1 (default) 1.4 GB -> 25.3 GB
Qwen3.6-35B-A3B (q8_0 KV) -1 (default) 1.1 GB -> 19.5 GB
gemma-4-26B-A4B -1, same prompt 100x 1.4 GB -> 1.6 GB
gemma-4-26B-A4B 4096 1.4 GB -> 5.4 GB (flat from
request 20 on, same latency)
Qwen3.6-35B-A3B 4096 1.1 GB -> 5.1 GB (flat)
The memory is not released when idle. In production a document
classification pass pushed the daily chat model to 34 GB RSS overnight.
Drop the override so llama.cpp's own default (8192 MiB) applies; the
cache_ram option still accepts -1 for users who want no limit. Update
both docs tables (the option reference and the prompt-cache table) and
note what -1 does.
Assisted-by: Claude:claude-opus-5-5
Assisted-by: Codex:GPT-6
Signed-off-by: Stefan Walcz <stefan.walcz@walcz.de>
utils.VerifyPath joins its argument onto the base path, so a path that
the caller already joined always passes. Several callers gave it joined
paths, and their checks could not fail:
- modeladmin (config view, patch, edit, pin and state): the config file
path from the loader. A config loaded from outside the models
directory (--models-config-file) could be pinned, and the pin wrote
the outside file. The patch and state paths stopped later, in the
mutation snapshot, with a different error.
- core/backend/tts.go: the model path joined onto the models path.
- The trellis2cpp and stablediffusion-ggml backends: option paths
(*_path) joined onto the model path. A "../" value outside the model
directory was accepted.
Add utils.VerifyResolvedPath for a full path. modeladmin and tts use
it. The backends now check the relative option value before they join
it. A rename in modeladmin checks the new relative name.
For models from a config file outside the models directory, the admin
API and web UI now return ErrPathNotTrusted for view, edit, pin, and
enable or disable. The docs describe this.
Assisted-by: Claude:claude-opus-5-5 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
* fix(distributed): stage every shard of a split GGUF
A split GGUF is configured by its first shard only. llama.cpp opens the
other "-0000N-of-0000M.gguf" files from the same directory by name. The
router staged only the configured path, so the worker received shard 1
and the load failed with "failed to load GGUF split".
The router now stages the remaining shards next to the first one. A
missing shard fails the load and names the file. The file count for
progress and the payload size also include all shards. The payload size
feeds the load deadline and the disk-headroom check. For a 111 GB model
whose first shard is 10 MB, both were sized for less than 1 GB.
Assisted-by: Claude:claude-opus-5-5 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* fix(distributed): stage the files a model install declares
Replace the split GGUF file name matching with the model's own file
list. A gallery install or an import records every file of the model in
._gallery_<name>.yaml (files:), and a config can list more under
download_files:. The router now stages all of these files, not only the
files that the config's path fields name. This includes the other
shards of a split GGUF, which llama.cpp opens by name.
The application gives the router a resolver that reads the two lists.
The resolver looks up the files by model name when it stages them, so a
replica that the reconciler loads from saved load options gets the same
files. backend.proto does not change.
A declared file that is missing on the frontend is skipped with a
warning. The load deadline and the disk headroom check include the
declared files.
Assisted-by: Claude:claude-opus-5-5 [Claude Code]
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>
- compatibility table: voxtral.c, OuteTTS and VoxCPM repos live under
antirez, edwko and OpenBMB
- distributed inferencing: llama.cpp RPC README moved to tools/rpc under ggml-org
- customize-model: the phi-2 example config moved to LocalAI-examples, and
embedded/models was replaced by the gallery
- model-gallery: malformed URL; link the gallery index
- GPU acceleration: ROCm install guide moved
- integrations: Wave Terminal docs page moved to ai-presets
Assisted-by: Claude:claude-fable-5-1
Signed-off-by: Pratik Gandhi <travpreneur@gmail.com>
* sglang backend: pass through thinking_budget + require_reasoning
sglang's raw Engine.async_generate() API (which this backend calls
directly, bypassing sglang's own OpenAI server) supports a precise,
tokenizer-derived reasoning-length budget via
sampling_params["custom_params"]["thinking_budget"] plus
require_reasoning=True, gated behind --enable-strict-thinking. Neither
was reachable through LocalAI: this backend built sampling_params only
from a fixed field mapping (temperature, top_p, ...) with no custom_params
key, and never passed require_reasoning to async_generate at all.
- LoadModel now reads a model-level "thinking_budget" option (same
mechanism as the existing tool_parser/reasoning_parser options), and
_build_sampling_params adds it as custom_params.thinking_budget on
every request when configured.
- _new_reasoning_parser already derives, from the rendered prompt, whether
the model's chat template pre-opened a reasoning block (Qwen3-style
templates append <think> to the prompt instead of letting the model
emit it) -- the same signal sglang's own OpenAI server computes from
per-template config to decide require_reasoning. This backend has no
template manager, so it now returns that signal too and _predict
forwards it to async_generate(require_reasoning=...).
Verified against production (NVFP4, sm_121, Qwen3.6-35B-A3B) via a raw
Engine.async_generate() call bypassing this backend: 301 reasoning
tokens against a 300-token budget, clean completion, ~27s. Not yet
verified through this backend's own gRPC path end-to-end (no local
CUDA/sglang environment available here) -- existing + new unit tests in
test.py cover the pure-Python merge/passthrough logic only.
Scope note: require_reasoning is derived only from the existing
prompt-suffix heuristic, not sglang's full per-template
_get_reasoning_from_request decision tree (minimax-m3/hunyuan special
cases etc.) -- this backend has no template manager to evaluate that
tree against, and the prompt-suffix check is the one heuristic already
validated in this file (test_reasoning_parser_forced_when_template_prefills_think_tag).
Signed-off-by: pos-ei-don <1822533+pos-ei-don@users.noreply.github.com>
* sglang backend: honour a model-level reasoning_default
A model YAML can already carry "parameters: reasoning_effort:", but that
value only reaches this backend when a *caller* sets it per request (the Go
side turns it into Metadata["enable_thinking"]). As a model-level default it
is silently dropped: a config reading "reasoning_effort: none" still produces
full reasoning on every request, so the config says one thing and the model
does another.
That gap is expensive in practice. On a self-hosted Qwen3.6-35B-A3B the
reasoning phase consumed the entire max_tokens budget before any content was
produced - 90% of code completions came back empty at max_tokens=768, and the
server log filled with "backend produced only reasoning, retrying". The
config looked like reasoning was off the whole time.
This adds "reasoning_default:off" (or ":on") on the same model-level
options: mechanism as thinking_budget. A per-request value always wins; the
default only fills in when the request is silent.
Measured on the stack above (sglang 0.5.20, NVFP4, GB10/sm_121) after
applying it:
default (nothing set) -> 0 chars reasoning, 27 tokens
"reasoning_effort": "none" -> 0 chars reasoning, 27 tokens
metadata enable_thinking=true -> capped at the 512-token thinking_budget,
541 tokens total, finish_reason stop
Tests: three cases added to backend/python/sglang/test.py covering the
default, per-request override in both directions, and the unconfigured case
(which must leave the template untouched).
Signed-off-by: pos-ei-don <1822533+pos-ei-don@users.noreply.github.com>
* sglang backend: validate thinking_budget instead of crashing LoadModel
Addresses the review on this PR:
- `int(thinking_budget)` raised on values like "5000.0" or "abc" and took
LoadModel down. The option is now parsed by _parse_thinking_budget():
integral numbers in any spelling are accepted, anything else is ignored
with a warning on stderr.
- Zero and negative budgets are ignored with a warning instead of being
passed to sglang, where they have no defined meaning. Turning reasoning
off is what reasoning_default:off is for.
- A load-time warning when thinking_budget is set but enable_strict_thinking
is not in engine_args, since sglang then ignores the budget silently.
- Tests for integral spellings, unset, zero, negative, non-integer and the
strict-thinking warning.
Signed-off-by: pos-ei-don <1822533+pos-ei-don@users.noreply.github.com>
* docs(sglang): explain reasoning options
Document the reasoning budget, strict-thinking requirement, and
precedence of request metadata over the model-level default.
Also note that the budget has to stay well below max_tokens (otherwise
it never triggers and the reply can end up empty), and that
POST /models/reload or a backend-only restart does not pick up changed
options; LocalAI itself has to be restarted.
Assisted-by: Codex:GPT-6
Signed-off-by: pos-ei-don <1822533+pos-ei-don@users.noreply.github.com>
* docs(sglang): clarify configuration reloads
Distinguish rereading model configuration from updating a running backend. Keep the full LocalAI restart recommendation for changed reasoning options.
Assisted-by: Codex:GPT-6
* sglang backend: only pass require_reasoning when sglang supports it
Engine.async_generate() gained the require_reasoning keyword in sglang
0.5.13 and takes no **kwargs. The CPU profile builds v0.5.11 from source
and the other profiles only set a >=0.5.11 floor, so passing the keyword
unconditionally made every request fail with TypeError. Detect support
once at import time, as the file already does for sampling_seed.
enable_strict_thinking first appears in sglang 0.5.12; fix the comment.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-5-5 [Claude Code]
---------
Signed-off-by: pos-ei-don <1822533+pos-ei-don@users.noreply.github.com>
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: localai-org-maint-bot <localai-org-maint-bot@users.noreply.github.com>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
The quickstart compose file still requested phi-2, which is no longer in
the gallery. Use phi-2-chat instead and fix model preload error wrapping
so discover/install failures report the real error instead of %!w(<nil>).
Keep earlier model failures when discovery fails for another model.
Document the Compose gallery default.
Fixes#11974
Signed-off-by: lei_lei <imleilei123@gmail.com>
* fix(models): fallback to application config default context size (#12202)
Honor appConfig.ContextSize in /v1/models/capabilities when model context_size is unset.
* docs(models): explain context size fallback
Describe the application default used by capability discovery and
preserve the distinction between total context and per-request limits.
Assisted-by: Codex:GPT-6
* fix(models): apply the default context size only when context_size is unset
The request path applies the application default context size only
when a model leaves context_size unset. An explicit 0 or -1 falls
through to the backend fallback. The capabilities endpoint now does
the same, so it reports the value the backend uses.
Assisted-by: Claude:claude-opus-5-5 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
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>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
fix(router): re-seed the knn corpus index when the vector store comes back empty
The corpus manager records a store as synced by file fingerprint and
embedding fingerprint. The local-store backend behind it is an in-memory
gRPC process the model loader may evict (active-backend cap, memory
pressure) or the idle watchdog may kill, and relaunch on the next
request — empty. The file is unchanged, so EnsureLoaded returned early
and the router went blind: every probe fell back with similarity 0 while
corpus/stats kept reporting the full count.
Measured on a production router (LOCALAI_MAX_ACTIVE_BACKENDS=6, four
resident models + two router stores): loading any further backend
evicted a store, and the idle watchdog killed both after 15 minutes;
/stores/find returned 0 hits against a 100-line corpus file whose stored
vectors matched fresh embeddings with cosine 1.000.
Two parts, because the knn classifier is built once and cached
(GetOrBuildClassifier), so the sync at build time is otherwise the only
one for the process lifetime:
- corpus.Manager remembers one vector it inserted (probe) and, on the
synced path, asks the live index for it. A miss means the index was
relaunched — fall through and re-seed from the file (no re-embedding).
- The router middleware wraps the knn classifier's store so every
lookup runs EnsureLoaded first; the loader gets the raw store, so its
probe never re-enters the wrapper. A sync error fails the lookup
closed, like the build-time load.
Specs: corpus package (relaunched empty store is re-seeded under an
unchanged file), middleware (relaunched index behind the cached
classifier is re-seeded instead of falling back; the spec is red without
the wrapper). The test fake now answers Search for inserted vectors.
Folds in the maintainer's follow-up (router-corpus-reseed-after-store-relaunch): reviewed and accepted.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Stefan Walcz <stefan.walcz@walcz.de>
* feat: return 429 when backends are saturated
When backends are at capacity (per-model max_concurrent or the
process-wide --max-concurrent-backend-requests ceiling), the response
was 503. The OpenAI SDK, litellm, and most agent harnesses key on 429
for rate-limit backoff and treat 503 as a hard error.
Both saturation paths now return 429 with the existing Retry-After
header and type: "rate_limit_error" in the JSON body. The per-model
admission middleware keeps admission_rejected as the code field so
existing alerts that match on it still fire.
Non-saturation 503s are unchanged: model cold-loading (with progress
body), model-load failure cooldown, PII detector fail-closed, and
classifier unavailable. These mean "not ready" rather than "busy".
Assisted-by: AGENT:regolo/glm5.2 [TOOL]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* fix: return 503 when scheduler has no available nodes
When the scheduler cannot find any healthy node to serve a model —
all nodes are full and eviction cannot free a slot, or a node_selector
excludes every candidate — the error fell through to 500. A 500 tells
clients something is broken when the condition is transient and
retryable.
The router now wraps these errors with a new ErrNoAvailableNodes
sentinel. The HTTP error handler maps it to 503 via applyNoAvailableNodes,
following the same pattern as applyBackendAdmission (429). Unrelated
scheduler errors (DB timeouts, registry lookups) still return 500.
Three return sites are wrapped:
- resolveSelectorCandidates: selector matches zero healthy nodes
- scheduleNewModel eviction-busy: all models have in-flight requests
- scheduleNewModel eviction-failed: eviction itself errored
The existing scheduleAndLoad wrapper ("no available nodes: %w") preserves
the sentinel through the chain via errors.Is, as does ModelRouterAdapter.
Assisted-by: AGENT:regolo/glm5.2 [TOOL]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* test(http): use Ginkgo for admission tests
Replace forbidden testing.T calls with Ginkgo and Gomega so the lint
check accepts the admission handler tests.
Assisted-by: Codex:GPT-6 forbidigo
* fix(middleware): show the recorded status for admission rejections
The admission audit row now records 429, but the Middleware page still
printed a hard-coded 503. Read the status from the event, and update
the two package comments that still said 503.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-5-5 [Claude Code]
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com>
Hardcoding size/size_vram as 0 made Ollama clients treat loaded models as
free. Prefer ModelFileName+ModelPath Stat when available, and omit size_vram
(and size) when the value is unknown instead of emitting literal zeros.
Resolve each listed model by its stored ID so tagged variants use their
own weights.
Fixes#11969
Signed-off-by: lei_lei <imleilei123@gmail.com>
LocalAGI 7e0947d added allowed_tools/excluded_tools and the
required_tool_before_finish gate. Single-node agents get them through
LocalAGI's runtime, but the distributed executor drives cogito directly
and its static config meta did not list the fields, so the agent form
hid them and the worker ignored them.
The distributed config now parses the tool lists from a JSON array or a
comma/newline separated string, and the meta entries match LocalAGI's.
The executor filters the knowledge base, skill and MCP tools (MCP via
cogito.WithMCPToolFilter) before the model sees them, and re-prompts the
model when it answers before the required tool returned "ok": true, up
to the configured number of reminders.
LocalAGI keeps its filter and gate helpers unexported, so a minimal copy
lives in core/services/agents/toolpolicy.go. A spec compares the meta
entries with LocalAGI's to catch drift.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-5-5 [Claude Code]
Pick up the LocalAGI PRs merged after f2a2af4:
- per-collection embedding and reranker models, locked per collection
so one agent's upload or rerank no longer stalls the others (#499)
- required_tool_before_finish: a tool the agent must call successfully
before it may answer (#495)
- allowed_tools / excluded_tools per agent, applied to MCP tools too
(#480)
Document the new agent settings. They show up in the single-node agent
form, which reads LocalAGI's config metadata; distributed mode keeps its
own field list and does not offer them yet.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-5-5 [Claude Code]
Design specs and implementation plans are working notes and are not
kept in the tree.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-5-5 [Claude Code]
* feat(gallery): publish signed OCI fallbacks
Publish both official gallery indexes with their local base configs so
an outage of the HTTP and GitHub sources can fall back to Quay.
Keep artifact signing policies separate from backend image policies,
and expose each moving gallery tag only after its digest is signed.
Assisted-by: Codex:gpt-6
* fix(gallery): confine packaged files to selected roots
Use directory-scoped file access to reject symlink escapes during gallery packaging. Create private bundle files for the publishing runner.
Assisted-by: Codex:GPT-6
---------
Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com>
* feat(system): report per-model DRM VRAM
Expose optional resident device memory for local backend process trees.
Deduplicate DRM clients and omit unsupported or incomplete readings.
Document accounting limits and preserve a measured zero in JSON.
Closes#11970.
Assisted-by: Codex:gpt-6
* fix(system): document trusted procfs reads
Scope G304 annotations to paths built from the fixed procfs root,
integer process IDs, and kernel directory entries. These reads accept
no user-controlled path components.
Assisted-by: Codex:GPT-6 gosec
---------
Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com>
Keep each message and reasoning item at its announced output index.
Include the answer in completed responses with reasoning or fallback
function calls, and retain reasoning supplied through backend deltas.
Add regression coverage for stream indices, final output, plain text,
and automatic tool parsing.
Assisted-by: Codex:GPT-6
Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com>
Partial JSON parsing heals a name-only chunk into a tool call. The
stream emits that call with empty arguments and skips later chunks.
Require complete JSON before emitting terminal tool-call events.
Preserve complete calls before an unfinished trailing call, and count
only actual tool calls. Add split-chunk regression tests and docs.
Refs #11635. The non-streaming report remains unconfirmed.
Assisted-by: Codex:GPT-6
Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com>
The legacy NVIDIA device reservation requests utility without compute.
Docker derives driver capabilities from that list, leaving CUDA libraries
unavailable even when monitoring works.
Include compute in the legacy example and clarify the matching docs.
Assisted-by: Codex:GPT-6
Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com>
Older Go linkers stamp pure-Go hosts with SDK metadata that disables
modern Metal APIs. Select Go 1.27 for Darwin builds and document the
backend rebuild requirement.
Assisted-by: Codex:gpt-6
Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com>
The link text still said chatbot-ui, but it now pointed at the examples
repository root. Link the configurations directory, which holds the
example model config files, and describe it as such.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-5-5 [Claude Code]
Add IQ4_XS and Q4_K_M GGUF builds with a BF16 vision projector.
Pin verified artifacts and document installation and variant selection.
Assisted-by: Codex:gpt-6
Add the localai-proxy known limits (no grammar or media forwarding,
TTS streams that end cleanly after an upstream failure, the /v1 path in
upstream_url), state that the Unimplemented skip covers the APIs that
answer HTTP 501, and describe a NATS-partitioned leader and pin
re-sync in distributed mode.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The failover chain editor leaves the warm toggle enabled on every row
(the model list has no backend field to gate on) and relies on the
server warning instead, so the docs describing it as disabled for
remote targets were wrong. Separately, syncstate's hydrate() returns
early with no Store or Loader, so a Reconcile tick is a no-op rather
than one that empties the map — correct that claim everywhere it was
repeated (contributor guide, distsync comment, design spec).
No behavior change.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Add the distributed-aware state contributor rule: any feature that
keeps runtime state must choose shared (syncstate), single-runner
(advisorylock), stateless, or documented per-instance behaviour, so it
behaves correctly across multiple frontends instead of diverging
silently. Also sweeps the failover/localai-proxy docs for gaps found
along the way: the UI (chain editor field, health strip, overview
page, chain badge), the 429->ResourceExhausted trip and 501->skip
mappings, and a spec correction for the live-transcription bridge's
actual close behavior.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The e2e suite now registers the localai-proxy binary and points proxy
models back at the test server itself, so a request leaves LocalAI
through the backend, returns over REST and is answered by a mock model.
Chat, embeddings, TTS and transcription through the proxy return the
upstream model's answer; a chain whose proxy target's upstream model
fails to load serves from the local target; and a realtime pipeline
whose LLM stage is a chain on a remote target completes a turn, then
switches to the local target with a localai.model.failover trip event
when a gate in front of the upstream starts answering 503.
The docs describe the localai-proxy backend next to cloud-proxy and add
a per-stage remote LocalAI example to the failover page.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
A leader whose host died without closing its connection kept the
advisory lock for about two hours of OS keepalive defaults, and no other
frontend could probe. The lock session now sets short TCP keepalives and
tcp_user_timeout, so the server drops it within about 30 seconds.
Shutdown now closes the lock for good, so a tick that runs after it
cannot take the lock back.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
A lock taken per tick passed between frontends on almost every tick, so
several frontends probed at once and each change of leader re-sent the
warm set and all state. The leader now holds a dedicated PostgreSQL
session with the advisory lock and keeps it until it shuts down or the
session dies.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Failover state is per frontend, remote chain targets only cover chat,
and chains have no UI. Design shared pins, health and chain state for
distributed mode, a localai-proxy backend for every API including live
transcription, a chain editor and health view, and a contributor rule
for distributed-aware state.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
The spec promised a load-time warning when a chain marks a remote
target warm, where the flag does nothing; the loader now logs it. The
remote-backend test moves into ModelConfig.IsRemoteProxy so the loader
and the failover manager agree on what is remote.
The spec now says what ships: a load blocked by pinned warm targets
proceeds over the limit after eviction retries, without an error that
names them.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Transcription-only and sound-detection-only realtime sessions passed a
chain config straight to the model loader. It has no backend, so the
loader fell back to greedy backend auto-detection: slow, and ending in
an unhelpful error. Sound-only sessions are a main use of chains.
The stage routing of the full pipeline moves into a stageRouter that
both realtime model kinds embed. Every stage resolves to the chain's
active target at build time and goes through the failover plan per
call. The session sends failover events for any model with chain
stages, and restarts them when a transcription session.update swaps
the model.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
A chain request reached a cloud-proxy target with the client's model,
the chain name, whenever the target set no upstream_model: passthrough
forwards the body's model and translate falls back to it. The upstream
answered 404, which neither retries nor trips, while the liveness
probe, which checks the target's own name, kept passing.
PrepareTarget now sets the upstream model of a remote target to
proxy.upstream_model or the target name, the same name the probe uses.
The request pipeline and realtime chain stages both call it.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
A warm target's liveness probe called ModelLoader.Load, which blocked
until the model finished loading (while the warm preload loaded it
too). Tick waited for every probe, so all probing froze, and the probe
then ran HealthCheck on an expired context and tripped the target at
every startup.
The prober now takes a function that returns the running backend
without loading it. A target that is not loaded passes liveness; its
recovery is neither confirmed nor failed and it returns to healthy
after min_dwell, like a cold target. Tick no longer waits for probes:
each probe applies its own result and a target whose probe is running
is skipped.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
A missing file is not a reliable signal: models download on first use,
some backends need no file, and dotted names like Phi-3.5-mini look like
paths. Marking such a fallback down removed the retry a chain exists
for. Cold targets are now judged only by real requests.
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
