fix(distributed): don't let a dead worker pin the model-load advisory lock (#10600)

* fix(distributed): don't let a dead worker pin the model-load advisory lock

In distributed mode a chat request could fail with:

  failed to route model with internal loader: routing model ...:
  loading model ...: advisorylock: acquiring lock <id>:
  ERROR: canceling statement due to lock timeout (SQLSTATE 55P03)

Root cause is two independent defects in the cross-replica model-load path:

1. SmartRouter.Route holds a per-model PostgreSQL advisory lock for the whole
   cold-load sequence, which includes installBackendOnNode -> InstallBackend,
   a NATS request-reply with a 15m deadline (DefaultBackendInstallTimeout) that
   ignored ctx. When the chosen worker died mid-install, the holder sat on the
   lock for up to 15m. The detached loadCtx (WithoutCancel) had no deadline, so
   nothing capped the hold.

2. The acquiring statement, pg_advisory_lock(), is subject to any deployment
   global lock_timeout. A common operator setting (e.g. 10s) aborts the wait
   with SQLSTATE 55P03, so every other replica's request for that model hard
   -errored instead of waiting for the in-progress load and reusing it. For the
   ~15m window the model was effectively unroutable.

Fixes:

- advisorylock.WithLockCtx (postgres): SET lock_timeout = 0 on its dedicated
  connection (RESET before it returns to the pool) so the Go context, not a
  deployment-wide GUC, governs how long we wait. Waiters now block and then
  re-check, reusing the model another replica just loaded.

- SmartRouter: bound the detached loadCtx with a single ModelLoadCeiling so the
  lock is always released in bounded time even if a sub-step wedges. Default is
  the configured backend.install deadline + 10m (staging + LoadModel margin),
  so a legitimately slow load is never cut.

- installBackendOnNode: use singleflight.DoChan + select on ctx.Done() so the
  install wait honors cancellation; the ceiling can then actually free a caller
  pinned behind a dead worker. The shared install still coalesces via
  singleflight.

Reproduced both defects as failing tests first (a real 55P03 against a
testcontainer with a short lock_timeout; a wedged install that blocks Route)
and confirmed green.

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

* fix(distributed): bound advisory-lock wait instead of disabling lock_timeout

Setting lock_timeout = 0 to override a deployment's short global lock_timeout
meant "wait forever" server-side. Safe for SmartRouter.Route (its loadCtx now
carries the model-load ceiling) but unsafe for the schema-migration callers
that pass context.Background(): a holder whose session never releases would
hang them indefinitely.

Derive the server-side lock_timeout from the caller's context instead: its
remaining budget plus a margin (so the Go context's cancellation still wins
with a clean error and the server bound is only a backstop), or a finite
30m backstop when the context has no deadline. Never zero - "wait forever"
is no longer possible, while a deployment's hostile short lock_timeout is
still overridden so legitimate cross-replica waits don't fail with 55P03.

Added a spec proving a deadline-less waiter gives up at the (shrunk) backstop
rather than hanging.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
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>
Co-authored-by: Ettore Di Giacinto <mudler@users.noreply.github.com>

Dockerfile

@@ -171,6 +171,17 @@ RUN if [ "${BUILD_TYPE}" = "hipblas" ]; then \
     ln -s /opt/rocm-**/lib/llvm/lib/libomp.so /usr/lib/libomp.so \
     ; fi
 
+# ROCm's bundled libdrm_amdgpu is built with a hardcoded fallback lookup path
+# for the ASIC ID table (/opt/amdgpu/share/libdrm/amdgpu.ids), which only exists
+# if AMD's full amdgpu graphics/DKMS stack is installed. This compute-only image
+# doesn't have it, so hipblas/rocBLAS log "No such file or directory" on every
+# model load and can fail to identify the GPU. Point it at the equivalent file
+# Ubuntu's libdrm-common package already ships.
+RUN if [ "${BUILD_TYPE}" = "hipblas" ] && [ -f /usr/share/libdrm/amdgpu.ids ] && [ ! -e /opt/amdgpu/share/libdrm/amdgpu.ids ]; then \
+    mkdir -p /opt/amdgpu/share/libdrm && \
+    ln -s /usr/share/libdrm/amdgpu.ids /opt/amdgpu/share/libdrm/amdgpu.ids \
+    ; fi
+
 RUN expr "${BUILD_TYPE}" = intel && echo "intel" > /run/localai/capability || echo "not intel"
 
 # Cuda

backend/cpp/ik-llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-IK_LLAMA_VERSION?=f74a6fb87b315b2c3154166e075360e15021a61d
+IK_LLAMA_VERSION?=068b173649f2fd8dc96b35ada5a0b76d8985105d
 LLAMA_REPO?=https://github.com/ikawrakow/ik_llama.cpp
 
 CMAKE_ARGS?=

backend/cpp/llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-LLAMA_VERSION?=6f4f53f2b7da54fcdbbecaaa734337c337ad6176
+LLAMA_VERSION?=4fc4ec5541b243957ae5099edb67372f8f3b550e
 LLAMA_REPO?=https://github.com/ggerganov/llama.cpp
 
 CMAKE_ARGS?=

backend/go/crispasr/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # CrispASR version (release tag)
 CRISPASR_REPO?=https://github.com/CrispStrobe/CrispASR
-CRISPASR_VERSION?=3b93758f9725d400eca82976f895e4cec3f31260
+CRISPASR_VERSION?=fcbc8718e654995e3bd2d0c98bcb8e55e297d23c
 SO_TARGET?=libgocrispasr.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF

backend/go/parakeet-cpp/Makefile

@@ -1,6 +1,6 @@
 # parakeet-cpp backend Makefile.
 #
-# Upstream pin lives below as PARAKEET_VERSION?=f469a57270a1cc4554acb15febf60e56619673b9
+# Upstream pin lives below as PARAKEET_VERSION?=e8acc6172a94e20a952cf1843decace5d771a94b
 # (.github/bump_deps.sh) can find and update it - matches the
 # whisper.cpp / ds4 / vibevoice-cpp convention.
 #
@@ -15,7 +15,7 @@
 # That's what the L0 smoke test uses. The default target below does the
 # proper clone-at-pin + cmake build so CI doesn't need a side-checkout.
 
-PARAKEET_VERSION?=f469a57270a1cc4554acb15febf60e56619673b9
+PARAKEET_VERSION?=e8acc6172a94e20a952cf1843decace5d771a94b
 PARAKEET_REPO?=https://github.com/mudler/parakeet.cpp
 
 GOCMD?=go

backend/go/stablediffusion-ggml/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # stablediffusion.cpp (ggml)
 STABLEDIFFUSION_GGML_REPO?=https://github.com/leejet/stable-diffusion.cpp
-STABLEDIFFUSION_GGML_VERSION?=3b6c9ca97cfcda8e68e719e6670d06379fcbe943
+STABLEDIFFUSION_GGML_VERSION?=3590aa8d626e671a1b1dc84506ea2932a243a480
 
 CMAKE_ARGS+=-DGGML_MAX_NAME=128
 

backend/go/stablediffusion-ggml/cpp/gosd.cpp

@@ -798,6 +798,7 @@ void sd_img_gen_params_set_seed(sd_img_gen_params_t *params, int64_t seed) {
 int gen_image(sd_img_gen_params_t *p, int steps, char *dst, float cfg_scale, char *src_image, float strength, char *mask_image, char* ref_images[], int ref_images_count) {
 
     sd_image_t* results;
+    int num_results_out = 0;
 
     std::vector<int> skip_layers = {7, 8, 9};
 
@@ -994,10 +995,14 @@ int gen_image(sd_img_gen_params_t *p, int steps, char *dst, float cfg_scale, cha
             sd_ctx_params_to_str(&ctx_params),
             sd_img_gen_params_to_str(p));
 
-    results = generate_image(sd_c, p);
+    bool gen_ok = generate_image(sd_c, p, &results, &num_results_out);
 
     std::free(p);
 
+    if (!gen_ok || num_results_out == 0) {
+        results = NULL;
+    }
+
     if (results == NULL) {
         fprintf (stderr, "NO results\n");
         if (input_image_buffer) free(input_image_buffer);

backend/go/whisper/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggml-org/whisper.cpp
-WHISPER_CPP_VERSION?=0ae02cdb2c7317b50991367c165736ce42ed96ac
+WHISPER_CPP_VERSION?=6fc7c33b4c3a2cec83e4b65abd5e96a890480375
 SO_TARGET?=libgowhisper.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF

backend/python/vllm/backend.py

@@ -748,7 +748,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         # When (A) native streaming ran cleanly, per-delta yields above already
         # delivered everything — do NOT extract again on the full text or we'd
         # duplicate content/tool_calls into the final chunk.
-        if has_tool_parser and not (native_streaming and not native_streaming_error):
+        # NOTE: `native_streaming` is a capability flag ("streaming parser is
+        # available"), not a state flag ("streaming actually ran"). For
+        # non-streaming requests it is still True but the per-delta loop was
+        # never entered, so we MUST still run extract_tool_calls here. Hence
+        # the explicit `streaming and …` guard on both branches.
+        if has_tool_parser and not (streaming and native_streaming and not native_streaming_error):
             try:
                 tp = tp_instance
                 if tp is None:
@@ -770,7 +775,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
                         ))
             except Exception as e:
                 print(f"Tool parser error: {e}", file=sys.stderr)
-        elif native_streaming and not native_streaming_error:
+        elif streaming and native_streaming and not native_streaming_error:
             # Per-delta path already emitted content + tool_calls; the final
             # chat_delta should carry only metadata (token counts, logprobs).
             content = ""

backend/python/vllm/install.sh

@@ -104,7 +104,7 @@ if [ "$(uname -s)" = "Darwin" ]; then
     # can rewrite it. Darwin therefore follows vllm-metal and can lag the Linux
     # vllm pin (requirements-cublas13-after.txt, bumped independently against
     # vllm/vllm) until vllm-metal supports a newer vLLM.
-    VLLM_METAL_VERSION="v0.3.0.dev20260628073537"
+    VLLM_METAL_VERSION="v0.3.0.dev20260701132215"
 
     # The coupled vLLM source version is whatever this vllm-metal release builds
     # against -- it declares it in its own installer as `vllm_v=`. Derive it from

backend/python/vllm/requirements-cublas13-after.txt

@@ -3,8 +3,8 @@
 # on a cu130 host. Pull the cu130-flavoured wheel from vLLM's per-tag index
 # instead — the cublas13 case in install.sh adds --index-strategy=unsafe-best-match
 # so uv consults this index alongside PyPI.
---extra-index-url https://wheels.vllm.ai/0.23.0/cu130
+--extra-index-url https://wheels.vllm.ai/0.24.0/cu130
 # VERSION COUPLING: darwin/Apple-Silicon builds use vllm-metal (see install.sh),
 # which pins this exact vLLM version. Bumping vllm here means coordinating with a
 # vllm-metal release that supports the new version, or macOS/Metal builds break.
-vllm==0.23.0
+vllm==0.24.0

backend/rust/kokoros/src/service.rs

@@ -351,6 +351,16 @@ impl Backend for KokorosService {
         Err(Status::unimplemented("Not supported"))
     }
 
+    type AudioTranscriptionLiveStream =
+        ReceiverStream<Result<backend::TranscriptLiveResponse, Status>>;
+
+    async fn audio_transcription_live(
+        &self,
+        _: Request<tonic::Streaming<backend::TranscriptLiveRequest>>,
+    ) -> Result<Response<Self::AudioTranscriptionLiveStream>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
     async fn diarize(
         &self,
         _: Request<backend::DiarizeRequest>,

cmd/launcher/internal/launcher.go

@@ -207,12 +207,20 @@ func (l *Launcher) StartLocalAI() error {
 	}
 
 	// Build command arguments
+	dataPath := l.GetDataPath()
 	args := []string{
 		"run",
 		"--models-path", l.config.ModelsPath,
 		"--backends-path", l.config.BackendsPath,
 		"--address", l.config.Address,
 		"--log-level", l.config.LogLevel,
+		// Keep persistent data and dynamic config under the launcher's data
+		// directory (~/.localai) rather than letting the server resolve them
+		// to ${basepath}/{data,configuration}. ${basepath} expands to the
+		// launcher process's CWD (often the user's home root), which puts
+		// ~/data and ~/configuration outside ~/.localai. See #10610.
+		"--data-path", filepath.Join(dataPath, "data"),
+		"--localai-config-dir", filepath.Join(dataPath, "configuration"),
 	}
 
 	l.localaiCmd = exec.CommandContext(l.ctx, binaryPath, args...)

core/application/distributed.go

@@ -356,6 +356,12 @@ func initDistributed(cfg *config.ApplicationConfig, authDB *gorm.DB, configLoade
 		PrefixConfig:     prefixCfg,
 		Pressure:         pressure,
 		SharedModels:     cfg.Distributed.SharedModels,
+		// Cap how long a cold load may hold the per-model advisory lock: the
+		// configured backend.install deadline plus a margin for file staging and
+		// the remote LoadModel. Derived from the install timeout so raising it
+		// (for slow links pulling multi-GB images) widens the ceiling too,
+		// instead of letting the static default cut a legitimately slow load.
+		ModelLoadCeiling: cfg.Distributed.BackendInstallTimeoutOrDefault() + 10*time.Minute,
 	})
 
 	// Wire staging-progress broadcasting so file-staging shows up on every

core/services/advisorylock/advisorylock.go

@@ -6,10 +6,39 @@ import (
 	"hash/fnv"
 	"strings"
 	"sync"
+	"time"
 
 	"gorm.io/gorm"
 )
 
+// advisoryLockWaitBackstop bounds, server-side, how long we will wait to
+// acquire a blocking advisory lock when the caller's context carries no
+// deadline (e.g. a startup schema migration using context.Background()). It
+// only exists so such a caller cannot hang forever behind a holder whose
+// session never releases the lock; it is far longer than any legitimate
+// guarded section. A var (not const) so tests can shrink it.
+var advisoryLockWaitBackstop = 30 * time.Minute
+
+// advisoryLockTimeoutMargin is added to a context's remaining budget when
+// deriving the server-side lock_timeout, so the Go context's own (cleaner)
+// cancellation fires first and the server bound is only ever a backstop.
+const advisoryLockTimeoutMargin = 30 * time.Second
+
+// advisoryLockWaitBudget returns the server-side lock_timeout to use for a
+// blocking acquire: the caller context's remaining time plus a margin (so the
+// Go context still governs), or the backstop when the context has no deadline.
+// Never returns zero - "wait forever" must not be possible.
+func advisoryLockWaitBudget(ctx context.Context) time.Duration {
+	if dl, ok := ctx.Deadline(); ok {
+		budget := time.Until(dl) + advisoryLockTimeoutMargin
+		if budget < time.Second {
+			budget = time.Second
+		}
+		return budget
+	}
+	return advisoryLockWaitBackstop
+}
+
 // localLocks holds one buffered channel (capacity 1) per lock key, used as an
 // in-process mutex for non-PostgreSQL dialects (SQLite). A SQLite auth DB is
 // effectively single-process, so serializing guarded sections within this
@@ -130,6 +159,27 @@ func WithLockCtx(ctx context.Context, db *gorm.DB, key int64, fn func() error) e
 	}
 	defer conn.Close()
 
+	// Override any deployment-wide lock_timeout on this dedicated connection.
+	// Operators commonly set a short global lock_timeout (on the role or
+	// database) to bound ordinary row-lock waits. Applied to the blocking
+	// pg_advisory_lock below, it aborts the wait with SQLSTATE 55P03 and turns
+	// LocalAI's intentional cross-replica "wait your turn, then re-check"
+	// coordination into a hard error for the caller (e.g. a chat request that
+	// just wanted to reuse a model another replica is loading).
+	//
+	// We do NOT disable it outright (lock_timeout = 0 would wait forever, which
+	// is unsafe for the schema-migration callers that pass context.Background()).
+	// Instead we set a bound derived from the caller's context: its remaining
+	// budget plus a margin so the Go context's cancellation wins with a clean
+	// error, or a finite backstop when the context has no deadline.
+	waitBudget := advisoryLockWaitBudget(ctx)
+	if _, err := conn.ExecContext(ctx,
+		fmt.Sprintf("SET lock_timeout = %d", waitBudget.Milliseconds())); err != nil {
+		return fmt.Errorf("advisorylock: setting lock_timeout: %w", err)
+	}
+	// Restore the session default before this pooled connection is reused.
+	defer func() { _, _ = conn.ExecContext(context.Background(), "RESET lock_timeout") }()
+
 	if _, err := conn.ExecContext(ctx, "SELECT pg_advisory_lock($1)", key); err != nil {
 		return fmt.Errorf("advisorylock: acquiring lock %d: %w", key, err)
 	}

core/services/advisorylock/advisorylock_test.go

@@ -158,6 +158,87 @@ var _ = Describe("AdvisoryLock", func() {
 			Expect(err).To(HaveOccurred())
 		})
 
+		It("waits out a short server-side lock_timeout instead of failing with 55P03", func() {
+			const lockKey int64 = 703
+
+			// Reproduce the production deployment that triggered this: a short
+			// global lock_timeout set on the database. Without the fix, a waiter
+			// blocked on pg_advisory_lock() is aborted by the server after this
+			// window and surfaces SQLSTATE 55P03 ("canceling statement due to
+			// lock timeout") to the caller instead of waiting for its turn.
+			Expect(db.Exec("ALTER DATABASE testdb SET lock_timeout = '300ms'").Error).ToNot(HaveOccurred())
+			sqlDB, err := db.DB()
+			Expect(err).ToNot(HaveOccurred())
+			// Drop pooled connections so subsequent ones reconnect and inherit
+			// the new database-level lock_timeout default.
+			sqlDB.SetMaxIdleConns(0)
+
+			holding := make(chan struct{})
+			released := make(chan struct{})
+			go func() {
+				defer GinkgoRecover()
+				herr := WithLockCtx(context.Background(), db, lockKey, func() error {
+					close(holding)
+					// Hold well past the 300ms server lock_timeout.
+					time.Sleep(1 * time.Second)
+					return nil
+				})
+				Expect(herr).ToNot(HaveOccurred())
+				close(released)
+			}()
+
+			<-holding // ensure the holder owns the lock before we contend
+
+			ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+			defer cancel()
+			executed := false
+			start := time.Now()
+			werr := WithLockCtx(ctx, db, lockKey, func() error {
+				executed = true
+				return nil
+			})
+			Expect(werr).ToNot(HaveOccurred(),
+				"waiter should wait out the in-progress hold, not fail with lock_timeout (55P03)")
+			Expect(executed).To(BeTrue())
+			Expect(time.Since(start)).To(BeNumerically(">=", 400*time.Millisecond),
+				"waiter should have actually waited for the holder to release")
+			<-released
+		})
+
+		It("bounds a deadline-less waiter with the backstop instead of waiting forever", func() {
+			const lockKey int64 = 704
+
+			// A caller with no context deadline (e.g. startup schema migration
+			// passing context.Background()) must not hang forever if the holder
+			// never releases. Shrink the backstop so the test is fast.
+			origBackstop := advisoryLockWaitBackstop
+			advisoryLockWaitBackstop = 500 * time.Millisecond
+			DeferCleanup(func() { advisoryLockWaitBackstop = origBackstop })
+
+			holding := make(chan struct{})
+			release := make(chan struct{})
+			go func() {
+				defer GinkgoRecover()
+				_ = WithLockCtx(context.Background(), db, lockKey, func() error {
+					close(holding)
+					<-release // hold until the test releases us
+					return nil
+				})
+			}()
+			defer close(release)
+
+			<-holding
+
+			start := time.Now()
+			err := WithLockCtx(context.Background(), db, lockKey, func() error {
+				Fail("waiter should not have acquired the still-held lock")
+				return nil
+			})
+			Expect(err).To(HaveOccurred(), "deadline-less waiter should give up at the backstop, not hang")
+			Expect(time.Since(start)).To(BeNumerically("<", 5*time.Second),
+				"backstop must cap the wait well under the test timeout")
+		})
+
 		It("serializes concurrent WithLockCtx on same key", func() {
 			const lockKey int64 = 702
 

core/services/nodes/router.go

@@ -68,6 +68,13 @@ type SmartRouterOptions struct {
 	// the absolute model paths untouched so the worker loads them directly from
 	// the shared volume (#10556). See config.DistributedConfig.SharedModels.
 	SharedModels bool
+	// ModelLoadCeiling is the hard upper bound on how long a single cold-load
+	// attempt (node selection -> backend install -> file staging -> LoadModel)
+	// may run while holding the per-model advisory lock. It backstops every
+	// sub-step's own timeout so a wedged worker can never pin the lock - and
+	// every other replica's request for that model - indefinitely. Zero selects
+	// defaultModelLoadCeiling.
+	ModelLoadCeiling time.Duration
 }
 
 // SmartRouter routes inference requests to the best available backend node.
@@ -101,8 +108,18 @@ type SmartRouter struct {
 	// sharedModels skips file staging when all nodes mount the same models
 	// directory at the same path (see SmartRouterOptions.SharedModels).
 	sharedModels bool
+	// modelLoadCeiling bounds how long a cold load may hold the per-model
+	// advisory lock (see SmartRouterOptions.ModelLoadCeiling).
+	modelLoadCeiling time.Duration
 }
 
+// defaultModelLoadCeiling is the fallback hold ceiling for a cold model load.
+// It must comfortably exceed the slowest legitimate load - a multi-GB backend
+// install (DefaultBackendInstallTimeout, 15m) plus staging and the remote
+// LoadModel (5m) - so it never cuts a real load short; it only ever fires when
+// a step is genuinely wedged (e.g. a worker that died mid-install).
+const defaultModelLoadCeiling = 25 * time.Minute
+
 // probeCacheTTL is how long a successful gRPC HealthCheck on a backend is
 // trusted before the next request re-probes. Matches healthCheckTTL in
 // pkg/model/model.go so the single-process and distributed paths share a
@@ -117,6 +134,10 @@ func NewSmartRouter(registry ModelRouter, opts SmartRouterOptions) *SmartRouter
 	if factory == nil {
 		factory = &tokenClientFactory{token: opts.AuthToken}
 	}
+	ceiling := opts.ModelLoadCeiling
+	if ceiling <= 0 {
+		ceiling = defaultModelLoadCeiling
+	}
 	return &SmartRouter{
 		registry:         registry,
 		unloader:         opts.Unloader,
@@ -131,6 +152,7 @@ func NewSmartRouter(registry ModelRouter, opts SmartRouterOptions) *SmartRouter
 		prefixConfig:     opts.PrefixConfig,
 		pressure:         opts.Pressure,
 		sharedModels:     opts.SharedModels,
+		modelLoadCeiling: ceiling,
 	}
 }
 
@@ -383,11 +405,19 @@ func (r *SmartRouter) Route(ctx context.Context, modelID, modelName, backendType
 	// the request context. If staging were bound to it, the multi-GB upload
 	// aborts with "context canceled" mid-transfer and large models can never
 	// finish staging (the model-load outage). WithoutCancel keeps the request's
-	// values (prefix chain, etc.) but drops its cancellation/deadline. Each
-	// long step still has its own bound (the file stager's resume budget,
-	// LoadModel's 5m timeout), and the per-model advisory lock below de-dupes
-	// concurrent loaders across replicas.
-	loadCtx := context.WithoutCancel(ctx)
+	// values (prefix chain, etc.) but drops its cancellation/deadline.
+	//
+	// Detaching from the caller is necessary, but it must not be unbounded: the
+	// load runs while holding the per-model advisory lock, and a worker that
+	// dies mid-install (its backend.install never replies) would otherwise pin
+	// that lock (and every other replica's request for the same model) until
+	// the NATS install deadline alone expires. Re-impose a single hard ceiling
+	// over the whole sequence so the lock is always released in bounded time,
+	// even if a sub-step wedges. Each long step still has its own (tighter)
+	// bound; this only backstops them. The per-model advisory lock below
+	// de-dupes concurrent loaders across replicas.
+	loadCtx, cancelLoad := context.WithTimeout(context.WithoutCancel(ctx), r.modelLoadCeiling)
+	defer cancelLoad()
 	loadModel := func(ctx context.Context) (*RouteResult, error) {
 		// Re-check after acquiring lock — another request may have loaded it
 		node, nm, err := r.registry.FindAndLockNodeWithModel(ctx, trackingKey, candidateNodeIDs, pref)
@@ -916,7 +946,14 @@ func (r *SmartRouter) installBackendOnNode(ctx context.Context, node *BackendNod
 	}
 
 	key := fmt.Sprintf("%s|%s|%s|%d", node.ID, backendType, modelID, replicaIndex)
-	v, err, _ := r.installFlight.Do(key, func() (any, error) {
+	// DoChan rather than Do so this wait honors ctx cancellation. InstallBackend
+	// blocks for its full NATS deadline (15m by default) when a worker accepts
+	// the request but never replies (e.g. it died mid-install). Without ctx
+	// awareness the caller (holding the per-model advisory lock) would sit there
+	// the whole time; here a cancelled ctx (typically the model-load ceiling)
+	// frees the caller promptly. The shared install keeps running in the
+	// background and still coalesces other callers via singleflight.
+	resCh := r.installFlight.DoChan(key, func() (any, error) {
 		reply, err := r.unloader.InstallBackend(node.ID, backendType, modelID, r.galleriesJSON, "", "", "", replicaIndex, "", nil)
 		if err != nil {
 			return "", err
@@ -931,10 +968,15 @@ func (r *SmartRouter) installBackendOnNode(ctx context.Context, node *BackendNod
 		}
 		return addr, nil
 	})
-	if err != nil {
-		return "", err
+	select {
+	case <-ctx.Done():
+		return "", ctx.Err()
+	case res := <-resCh:
+		if res.Err != nil {
+			return "", res.Err
+		}
+		return res.Val.(string), nil
 	}
-	return v.(string), nil
 }
 
 func (r *SmartRouter) buildClientForAddr(node *BackendNode, addr string, parallel bool) grpc.Backend {

core/services/nodes/router_test.go

@@ -493,6 +493,44 @@ var _ = Describe("SmartRouter", func() {
 				Expect(result.Node.ID).To(Equal("n3"))
 			})
 		})
+
+		Context("worker wedges mid-install (dead node holding the lock)", func() {
+			It("aborts the load at the ModelLoadCeiling instead of blocking forever", func() {
+				// Simulate the production incident: the chosen worker accepts the
+				// backend.install but never replies (it died), so InstallBackend
+				// would otherwise block for its full NATS deadline (15m by
+				// default) while pinning the per-model advisory lock. Route must
+				// give up at the ceiling so the lock is released promptly.
+				reg.findAndLockErr = errors.New("not found")
+				reg.findIdleNode = &BackendNode{ID: "n4", Name: "dead-node", Address: "10.0.0.4:50051"}
+
+				block := make(chan struct{})
+				defer close(block) // let the background install goroutine drain at test end
+				unloader.installHook = func() { <-block }
+
+				router := NewSmartRouter(reg, SmartRouterOptions{
+					Unloader:         unloader,
+					ClientFactory:    factory,
+					ModelLoadCeiling: 200 * time.Millisecond,
+				})
+
+				done := make(chan error, 1)
+				start := time.Now()
+				go func() {
+					defer GinkgoRecover()
+					_, err := router.Route(context.Background(), "wedged-model",
+						"models/wedged.gguf", "llama-cpp",
+						&pb.ModelOptions{Model: "models/wedged.gguf"}, false)
+					done <- err
+				}()
+
+				var routeErr error
+				Eventually(done, 5*time.Second).Should(Receive(&routeErr),
+					"Route must not block on a wedged install past the ceiling")
+				Expect(routeErr).To(HaveOccurred())
+				Expect(time.Since(start)).To(BeNumerically("<", 5*time.Second))
+			})
+		})
 	})
 
 	Describe("scheduleNewModel (mock-based, via Route)", func() {

docs/data/version.json

@@ -1,3 +1,3 @@
 {
-  "version": "v4.5.5"
+  "version": "v4.5.6"
 }

gallery/index.yaml

@@ -1716,7 +1716,7 @@
       - use_jinja:true
     parameters:
       min_p: 0.15
-      model: llama-cpp/models/LFM2.5-8B-A1B-GGUF/LFM2.5-8B-A1B-Q4_K_M.gguf
+      model: llama-cpp/models/LFM2.5-8B-A1B-GGUF/LFM2.5-8B-A1B-Q8_0.gguf
       repeat_penalty: 1.05
       temperature: 0.1
       top_k: 50
@@ -1724,9 +1724,9 @@
     template:
       use_tokenizer_template: true
   files:
-    - filename: llama-cpp/models/LFM2.5-8B-A1B-GGUF/LFM2.5-8B-A1B-Q4_K_M.gguf
-      uri: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-GGUF/resolve/main/LFM2.5-8B-A1B-Q4_K_M.gguf
-      sha256: 4923ec14f06b968b74d663e5949867d2d9c3bf13a20b8be1a9f9af39989b2bb0
+    - filename: llama-cpp/models/LFM2.5-8B-A1B-GGUF/LFM2.5-8B-A1B-Q8_0.gguf
+      uri: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-GGUF/resolve/main/LFM2.5-8B-A1B-Q8_0.gguf
+      sha256: 33ab3b8ce6a964fb8ebac89360c9b3cf72c4fa418d5e4c0a94d46883124d5c02
 - name: "qwopus3.5-9b-coder-mtp"
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:
@@ -1758,8 +1758,8 @@
       use_tokenizer_template: true
   files:
     - filename: llama-cpp/models/Qwopus3.5-9B-Coder-MTP-GGUF/Qwopus3.5-9B-Coder-MTP-Q4_K_M.gguf
-      sha256: f6fc5d193045796d9e1870cbc40f827fe55f53f70593c3f5c1968b82b9331991
       uri: https://huggingface.co/Jackrong/Qwopus3.5-9B-Coder-MTP-GGUF/resolve/main/Qwopus3.5-9B-Coder-MTP-Q4_K_M.gguf
+      sha256: 9ea3ecd122a5165b8b81655f29eaf09d71daf841503e4c4212bdfadb36ab3712
     - filename: llama-cpp/mmproj/Qwopus3.5-9B-Coder-MTP-GGUF/Qwopus3.5-9B-Coder-MTP-mmproj.gguf
       sha256: f48daca405a1c768a9514e392c3955dcc4a9d66a5cf64cf45e064092b5f20ee4
       uri: https://huggingface.co/Jackrong/Qwopus3.5-9B-Coder-MTP-GGUF/resolve/main/Qwopus3.5-9B-Coder-MTP-mmproj.gguf

pkg/grpc/grpcerrors/errors.go

@@ -58,6 +58,23 @@ func IsLiveTranscriptionUnsupported(err error) bool {
 	return strings.Contains(strings.ToLower(err.Error()), "unimplemented")
 }
 
+// IsUnimplemented reports whether err is a gRPC Unimplemented status — the
+// signal a backend gives for an RPC it does not implement. The generated
+// UnimplementedBackendServer stub returns exactly this for any RPC a backend
+// (e.g. a Python or external backend) has not overridden, so callers can treat
+// an optional RPC as a no-op rather than a failure. Prefers the typed status
+// code and falls back to the message for paths that lose the status (e.g. errors
+// wrapped across non-gRPC boundaries).
+func IsUnimplemented(err error) bool {
+	if err == nil {
+		return false
+	}
+	if status.Code(err) == codes.Unimplemented {
+		return true
+	}
+	return strings.Contains(strings.ToLower(err.Error()), "unimplemented")
+}
+
 // StreamTranscriptionUnsupported returns the canonical error a backend returns
 // when it (or the loaded model) cannot serve the server-streaming
 // AudioTranscriptionStream RPC. It carries codes.Unimplemented like the live

pkg/grpc/grpcerrors/errors_test.go

@@ -55,6 +55,18 @@ var _ = Describe("grpcerrors", func() {
 		Expect(grpcerrors.IsModelNotLoaded(err)).To(BeFalse())
 	})
 
+	DescribeTable("IsUnimplemented",
+		func(err error, want bool) {
+			Expect(grpcerrors.IsUnimplemented(err)).To(Equal(want))
+		},
+		Entry("nil", nil, false),
+		Entry("typed code", status.Error(codes.Unimplemented, "method Free not implemented"), true),
+		Entry("stale stub message (Unknown code)", errors.New("rpc error: code = Unimplemented desc = "), true),
+		Entry("unrelated error", errors.New("context deadline exceeded"), false),
+		Entry("unrelated grpc code", status.Error(codes.Unavailable, "connection refused"), false),
+		Entry("model not loaded is NOT unimplemented", grpcerrors.ModelNotLoaded("parakeet-cpp"), false),
+	)
+
 	It("StreamTranscriptionUnsupported carries Unimplemented and is not ModelNotLoaded", func() {
 		err := grpcerrors.StreamTranscriptionUnsupported("parakeet-cpp", "not a streaming model")
 		Expect(status.Code(err)).To(Equal(codes.Unimplemented))

pkg/model/process.go

@@ -11,6 +11,7 @@ import (
 	"time"
 
 	"github.com/hpcloud/tail"
+	"github.com/mudler/LocalAI/pkg/grpc/grpcerrors"
 	"github.com/mudler/LocalAI/pkg/signals"
 	process "github.com/mudler/go-processmanager"
 	"github.com/mudler/xlog"
@@ -52,10 +53,21 @@ func (ml *ModelLoader) deleteProcess(s string) error {
 		hook(s)
 	}
 
-	// Free GPU resources before stopping the process to ensure VRAM is released
+	// Free GPU resources before stopping the process to ensure VRAM is released.
+	// Free is optional: backends that don't override it (the generated stub, many
+	// Python/external backends, or a federation proxy in distributed mode) return
+	// gRPC Unimplemented. That is expected, not a failure — VRAM is reclaimed when
+	// the process is stopped below, or by the remote unloader for remote backends —
+	// so don't surface it as an error.
 	xlog.Debug("Calling Free() to release GPU resources", "model", s)
 	if err := model.GRPC(false, ml.wd).Free(context.Background()); err != nil {
-		xlog.Warn("Error freeing GPU resources", "error", err, "model", s)
+		if grpcerrors.IsUnimplemented(err) {
+			xlog.Debug("Backend does not implement Free(); GPU release handled on process stop", "model", s)
+		} else {
+			// Now that the expected Unimplemented case is filtered out above, a
+			// remaining error is a genuine failure to release VRAM — surface it.
+			xlog.Error("Error freeing GPU resources", "error", err, "model", s)
+		}
 	}
 
 	process := model.Process()