ci: push immutable sha- and moving main-cuda- tags alongside the version label

Push-triggered builds always defaulted to the v230 label, silently
overwriting it on every main merge — a stack pinned to any other
version label (v231) kept pulling stale images. Every build now also
pushes sha-<short> (immutable pin) and main-cuda-<llamacpp> (explicit
latest-main), so deployments can choose mutable-tracking or exact-pin
deliberately.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AXQxVhXBw8PwFAtsVrXSmj

.gitea/workflows/build-cuda-image.yml

@@ -0,0 +1,89 @@
+name: Build CUDA image (fork)
+
+# Builds this fork's llama-swap (serial scheduler + embedded UI) from source and
+# layers it on a pinned llama.cpp CUDA server base, then pushes to the Gitea
+# container registry, e.g. gitea.stevedudenhoeffer.com/steve/llama-swap:v230-cuda-b9821
+#
+# Requires repo secrets: REGISTRY_USER, REGISTRY_PASSWORD (push to the registry).
+
+on:
+  workflow_dispatch:
+    inputs:
+      llama_swap_version:
+        description: "llama-swap version label (image tag prefix)"
+        required: false
+        default: "v230"
+      llamacpp_build:
+        description: "llama.cpp CUDA server build (base image tag suffix)"
+        required: false
+        default: "b9821"
+  # Any change that lands in the shipped binary or image kicks off a fresh
+  # build (source, UI, deps, build definition). Docs/config-example edits
+  # don't.
+  push:
+    branches: [main]
+    paths:
+      - ".gitea/workflows/build-cuda-image.yml"
+      - "docker/fork-cuda.Containerfile"
+      - "**/*.go"
+      - "go.mod"
+      - "go.sum"
+      - "ui-svelte/**"
+      - "llama-swap.go"
+
+env:
+  REGISTRY: gitea.stevedudenhoeffer.com
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Compute image metadata
+        id: meta
+        run: |
+          LS_VER="${{ inputs.llama_swap_version || 'v230' }}"
+          LCPP="${{ inputs.llamacpp_build || 'b9821' }}"
+          SHORT_SHA="$(echo '${{ github.sha }}' | cut -c1-7)"
+          IMAGE="${REGISTRY}/${{ github.repository }}"
+          # Three tags per build: the version label (mutable, human-facing),
+          # an explicit moving main-cuda tag, and an immutable sha tag so a
+          # deployment can always pin the exact build.
+          {
+            echo "image=${IMAGE}"
+            echo "tag=${LS_VER}-cuda-${LCPP}"
+            echo "tags=${IMAGE}:${LS_VER}-cuda-${LCPP},${IMAGE}:main-cuda-${LCPP},${IMAGE}:sha-${SHORT_SHA}"
+            echo "base_tag=server-cuda-${LCPP}"
+            echo "ls_version=${LS_VER}"
+            echo "build_date=$(date -u +%Y-%m-%dT%H:%M:%SZ)"
+          } >> "$GITHUB_OUTPUT"
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Log in to Gitea registry
+        uses: docker/login-action@v3
+        with:
+          registry: ${{ env.REGISTRY }}
+          username: ${{ secrets.REGISTRY_USER }}
+          password: ${{ secrets.REGISTRY_PASSWORD }}
+
+      - name: Build and push
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: docker/fork-cuda.Containerfile
+          push: true
+          provenance: false
+          build-args: |
+            BASE_TAG=${{ steps.meta.outputs.base_tag }}
+            LS_VERSION=${{ steps.meta.outputs.ls_version }}
+            GIT_HASH=${{ github.sha }}
+            BUILD_DATE=${{ steps.meta.outputs.build_date }}
+          tags: ${{ steps.meta.outputs.tags }}
+
+      - name: Summary
+        run: |
+          echo "Pushed ${{ steps.meta.outputs.tags }}" >> "$GITHUB_STEP_SUMMARY"

AGENTS.md

@@ -5,16 +5,14 @@ llama-swap is a light weight, transparent proxy server that provides automatic m
 ## Tech stack
 
 - golang
-- typescript, vite and svelt5 for UI (located in ui/)
+- typescript, vite and svelte 5 for UI (located in ui-svelte/)
 
 ## Workflow Tasks
 
 - when summarizing changes only include details that require further action
-- just say "Done." when there is no further action
-- use the github CLI `gh` to create pull requests and work with github
 - Rules for creating pull requests:
-  - keep them short and focused on changes.
+  - keep them short and focused on changes
-  - never include a test plan
+  - skip the test plan
   - write the summary using the same style rules as commit message
 
 ## Testing
@@ -30,7 +28,7 @@ llama-swap is a light weight, transparent proxy server that provides automatic m
 ### Commit message example format:
 
 ```
-proxy: add new feature
+internal/server: add new feature
 
 Add new feature that implements functionality X and Y.
 

config-schema.json

@@ -601,10 +601,11 @@
                         "use": {
                             "type": "string",
                             "enum": [
+                                "serial",
                                 "fifo"
                             ],
-                            "default": "fifo",
+                            "default": "serial",
-                            "description": "Scheduler to use. Only 'fifo' is currently supported."
+                            "description": "Scheduler to use. 'serial' (default on this fork): strict one-model-at-a-time, requests run in exact arrival order, switching models evicts every other model first. 'fifo': throughput-oriented, batches same-model requests and allows parallel/co-resident models."
                         },
                         "settings": {
                             "type": "object",

config.example.yaml

@@ -556,11 +556,21 @@ routing:
           # expands to: [L]
           full: "L"
 
-  # scheduler: how queued requests are ordered.
+  # scheduler: how queued requests are ordered and run.
-  # The default and only valid scheduler is "fifo"
+  # - optional, default on this fork: "serial"
+  # - valid values:
+  #   - "serial": strict one-model-at-a-time. Requests run in exact arrival
+  #     order; only one request runs at a time; switching to a different model
+  #     evicts every other running model first so a single model occupies memory
+  #     at a time. This ignores group/matrix co-residency entirely. The "fifo"
+  #     settings below (priority) do not apply.
+  #   - "fifo": throughput-oriented. Same-model requests are batched to reduce
+  #     swaps and a model serves up to its concurrencyLimit in parallel; models
+  #     in non-exclusive groups can run concurrently. Requests may be reordered.
   scheduler:
-    use: fifo
+    use: serial
     settings:
+      # fifo settings only apply when use: fifo
       fifo:
         # priority: a dictionary of model ID -> priority
         # - optional, default: empty dictionary

docker/fork-cuda.Containerfile

@@ -0,0 +1,74 @@
+# Build a CUDA llama-swap image FROM THIS FORK's source (includes the serial
+# scheduler) and layer it on a pinned llama.cpp CUDA server base. Produces e.g.:
+#   gitea.stevedudenhoeffer.com/steve/llama-swap:v230-cuda-b9821
+#
+# BASE_TAG selects the llama.cpp CUDA runtime + llama-server build, e.g.
+# "server-cuda-b9821". The llama-swap binary (with the embedded Svelte UI) is
+# compiled from the repo at build time, so no GitHub release is required.
+#
+# Build context is the repo root:
+#   docker build -f docker/fork-cuda.Containerfile \
+#     --build-arg BASE_TAG=server-cuda-b9821 -t llama-swap:v230-cuda-b9821 .
+
+ARG BASE_IMAGE=ghcr.io/ggml-org/llama.cpp
+ARG BASE_TAG=server-cuda-b9821
+
+# ---- Stage 1: build the Svelte UI (embedded into the binary) ----
+FROM node:22-bookworm-slim AS ui
+WORKDIR /src/ui-svelte
+# Install deps first for layer caching. .npmrc carries legacy-peer-deps=true,
+# which the project relies on (tailwind/vite peer ranges), so copy it before
+# npm ci or the strict resolver fails with ERESOLVE.
+COPY ui-svelte/package.json ui-svelte/package-lock.json ui-svelte/.npmrc ./
+RUN npm ci
+COPY ui-svelte/ ./
+# `npm run build` is `vite build --emptyOutDir`; vite.config.ts writes to
+# ../internal/server/ui_dist, which //go:embed picks up in the next stage.
+RUN mkdir -p /src/internal/server && npm run build
+
+# ---- Stage 2: build the llama-swap binary with the embedded UI ----
+FROM golang:1.26-bookworm AS build
+WORKDIR /src
+# Cache modules independently of source churn.
+COPY go.mod go.sum ./
+RUN go mod download
+COPY . .
+# Overlay the freshly built UI so //go:embed ui_dist ships the real assets
+# instead of the committed placeholder.
+COPY --from=ui /src/internal/server/ui_dist/ ./internal/server/ui_dist/
+ARG LS_VERSION=v230
+ARG GIT_HASH=unknown
+ARG BUILD_DATE=unknown
+RUN CGO_ENABLED=0 GOOS=linux go build \
+    -ldflags="-X main.version=${LS_VERSION} -X main.commit=${GIT_HASH} -X main.date=${BUILD_DATE}" \
+    -o /out/llama-swap .
+
+# ---- Stage 3: runtime image on the pinned llama.cpp CUDA base ----
+FROM ${BASE_IMAGE}:${BASE_TAG}
+
+# Run as root by default to match the upstream `vNNN-cuda-bNNNN` (non-suffixed)
+# image that ragnaros pulls today: it needs root to reach the mounted docker
+# socket for container-backed models (sd-server). Override UID/GID at build time
+# for a non-root variant.
+ARG UID=0
+ARG GID=0
+ARG USER_HOME=/root
+ENV HOME=$USER_HOME
+
+RUN set -eux; \
+    if [ "$UID" -ne 0 ]; then \
+      if [ "$GID" -ne 0 ]; then groupadd --system --gid "$GID" app; fi; \
+      useradd --system --uid "$UID" --gid "$GID" --home "$USER_HOME" app; \
+    fi; \
+    mkdir --parents "$HOME" /app; \
+    chown --recursive "$UID:$GID" "$HOME" /app
+
+COPY --from=build --chown=$UID:$GID /out/llama-swap /app/llama-swap
+COPY --chown=$UID:$GID docker/config.example.yaml /app/config.yaml
+
+USER $UID:$GID
+WORKDIR /app
+ENV PATH="/app:${PATH}"
+
+HEALTHCHECK CMD curl -f http://localhost:8080/ || exit 1
+ENTRYPOINT [ "/app/llama-swap", "-config", "/app/config.yaml" ]

internal/config/config_posix_test.go

@@ -277,7 +277,7 @@ groups:
 				},
 			},
 			Scheduler: SchedulerConfig{
-				Use: "fifo",
+				Use: "serial",
 			},
 		},
 	}

internal/config/config_test.go

@@ -1572,7 +1572,7 @@ groups:
 	assert.Equal(t, "group", cfg.Routing.Router.Use)
 	// default group injected for orphaned models (none here) still leaves g1
 	assert.Contains(t, cfg.Routing.Router.Settings.Groups, "g1")
-	assert.Equal(t, "fifo", cfg.Routing.Scheduler.Use)
+	assert.Equal(t, "serial", cfg.Routing.Scheduler.Use)
 }
 
 func TestConfig_Routing_LegacyTopLevelMatrix(t *testing.T) {
@@ -1631,7 +1631,7 @@ func TestConfig_Routing_DefaultsToGroup(t *testing.T) {
 	cfg, err := LoadConfigFromReader(strings.NewReader(twoModels))
 	require.NoError(t, err)
 	assert.Equal(t, "group", cfg.Routing.Router.Use)
-	assert.Equal(t, "fifo", cfg.Routing.Scheduler.Use)
+	assert.Equal(t, "serial", cfg.Routing.Scheduler.Use)
 }
 
 func TestConfig_Routing_LegacyAndRoutingConflict(t *testing.T) {

internal/config/config_windows_test.go

@@ -266,7 +266,7 @@ groups:
 				},
 			},
 			Scheduler: SchedulerConfig{
-				Use: "fifo",
+				Use: "serial",
 			},
 		},
 	}

internal/config/load.go

@@ -358,11 +358,16 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 	config.Routing.Router.Settings.Matrix = config.Matrix
 	config.Routing.Router.Settings.Groups = config.Groups
 
+	// This fork defaults to the "serial" scheduler: one model loaded at a time,
+	// requests served in strict arrival order. Set use: fifo for the upstream
+	// throughput-oriented behavior that batches same-model requests.
 	if config.Routing.Scheduler.Use == "" {
-		config.Routing.Scheduler.Use = "fifo"
+		config.Routing.Scheduler.Use = "serial"
 	}
-	if config.Routing.Scheduler.Use != "fifo" {
+	switch config.Routing.Scheduler.Use {
-		return Config{}, fmt.Errorf("routing.scheduler.use: unknown scheduler %q (valid: fifo)", config.Routing.Scheduler.Use)
+	case "fifo", "serial":
+	default:
+		return Config{}, fmt.Errorf("routing.scheduler.use: unknown scheduler %q (valid: fifo, serial)", config.Routing.Scheduler.Use)
 	}
 	for modelID := range config.Routing.Scheduler.Settings.Fifo.Priority {
 		if _, found := config.RealModelName(modelID); !found {

internal/config/model_config.go

@@ -14,6 +14,7 @@ var validModalities = map[string]struct{}{
 	"text":  {},
 	"audio": {},
 	"image": {},
+	"video": {},
 }
 
 // ModelCapConfig defines what modalities and features a model supports.
@@ -37,12 +38,12 @@ func (c ModelCapConfig) Empty() bool {
 func (c ModelCapConfig) Validate() error {
 	for _, m := range c.In {
 		if _, ok := validModalities[m]; !ok {
-			return fmt.Errorf("capabilities.in: invalid modality %q, must be one of: text, audio, image", m)
+			return fmt.Errorf("capabilities.in: invalid modality %q, must be one of: text, audio, image, video", m)
 		}
 	}
 	for _, m := range c.Out {
 		if _, ok := validModalities[m]; !ok {
-			return fmt.Errorf("capabilities.out: invalid modality %q, must be one of: text, audio, image", m)
+			return fmt.Errorf("capabilities.out: invalid modality %q, must be one of: text, audio, image, video", m)
 		}
 	}
 	if c.Context < 0 {

internal/config/model_config_test.go

@@ -296,20 +296,25 @@ func TestConfig_ModelCapabilities_Validate(t *testing.T) {
 		assert.NoError(t, caps.Validate())
 	})
 
+	t.Run("valid_video_modality", func(t *testing.T) {
+		caps := ModelCapConfig{In: []string{"text", "image"}, Out: []string{"video"}}
+		assert.NoError(t, caps.Validate())
+	})
+
 	t.Run("invalid_in_modality", func(t *testing.T) {
-		caps := ModelCapConfig{In: []string{"video"}}
+		caps := ModelCapConfig{In: []string{"smell"}}
 		err := caps.Validate()
 		assert.Error(t, err)
 		assert.Contains(t, err.Error(), "capabilities.in")
-		assert.Contains(t, err.Error(), "video")
+		assert.Contains(t, err.Error(), "smell")
 	})
 
 	t.Run("invalid_out_modality", func(t *testing.T) {
-		caps := ModelCapConfig{Out: []string{"video"}}
+		caps := ModelCapConfig{Out: []string{"smell"}}
 		err := caps.Validate()
 		assert.Error(t, err)
 		assert.Contains(t, err.Error(), "capabilities.out")
-		assert.Contains(t, err.Error(), "video")
+		assert.Contains(t, err.Error(), "smell")
 	})
 
 	t.Run("negative_context", func(t *testing.T) {
@@ -327,10 +332,10 @@ models:
     capabilities:
       in:
         - text
-        - video
+        - smell
 `
 		_, err := LoadConfigFromReader(strings.NewReader(content))
 		assert.Error(t, err)
-		assert.Contains(t, err.Error(), "video")
+		assert.Contains(t, err.Error(), "smell")
 	})
 }

internal/router/scheduler/scheduler.go

@@ -92,9 +92,14 @@ type Effects interface {
 	StopProcesses(timeout time.Duration, ids []string)
 }
 
-// New returns a Scheduler selected by conf.Routing.Scheduler.Use, configured
+// New returns a Scheduler selected by conf.Routing.Scheduler.Use, configured from
-// from conf and bound to the given planner and effects. Currently only "fifo"
+// conf and bound to the given planner and effects. Supported values are "fifo"
-// (the default) is supported.
+// (throughput-oriented, batches same-model requests) and "serial" (strict
+// one-model-at-a-time, exact arrival order).
+//
+// The deployment default is applied by config loading (LoadConfig sets Use to
+// "serial" when unset). The "" fallback here is the library default and remains
+// "fifo" so callers that build a Config directly keep the original behavior.
 func New(conf config.Config, name string, logger *logmon.Monitor, planner Swapper, eff Effects) (Scheduler, error) {
 	use := conf.Routing.Scheduler.Use
 	if use == "" {
@@ -103,6 +108,9 @@ func New(conf config.Config, name string, logger *logmon.Monitor, planner Swappe
 	switch use {
 	case "fifo":
 		return NewFIFO(name, logger, planner, conf.Routing.Scheduler.Settings.Fifo, conf.Models, eff), nil
+	case "serial":
+		// Serial ignores the group planner: it always evicts every other model.
+		return NewSerial(name, logger, eff), nil
 	default:
 		return nil, fmt.Errorf("unsupported scheduler type: %q", use)
 	}

internal/router/scheduler/serial.go

@@ -0,0 +1,253 @@
+package scheduler
+
+import (
+	"fmt"
+	"sort"
+	"time"
+
+	"github.com/mostlygeek/llama-swap/internal/logmon"
+	"github.com/mostlygeek/llama-swap/internal/process"
+)
+
+// Serial is a strict one-model-at-a-time scheduler. Unlike FIFO it never reorders
+// or batches: requests run in exact arrival order and at most one request runs at
+// any instant. When the next request targets a model other than the one loaded,
+// every other running model is evicted and the target is loaded before it runs,
+// so a single model occupies memory at a time — at the cost of throughput.
+//
+// Example: A B C A is served as A B C A. The final A reloads its model even
+// though it ran first, because B and C displaced it in between. (FIFO, by
+// contrast, would batch the two A requests: A A B C.)
+//
+// Serial ignores group/eviction policy entirely: it always evicts every other
+// running model, regardless of how groups are configured. That is what makes the
+// single-model guarantee a property of the scheduler rather than of the config.
+//
+// Like FIFO, every method runs on the router's single run-loop goroutine, so no
+// internal locking is needed.
+type Serial struct {
+	name    string
+	logger  *logmon.Monitor
+	effects Effects
+
+	// queued holds requests in strict arrival order. It is never reordered.
+	queued []HandlerReq
+
+	// active is the one request currently being processed (loading or serving),
+	// or nil when idle. phase is meaningful only while active != nil.
+	active *HandlerReq
+	phase  serialPhase
+}
+
+// serialPhase is the lifecycle stage of the active request.
+type serialPhase int
+
+const (
+	phaseIdle     serialPhase = iota
+	phaseSwapping             // waiting for OnSwapDone for active.Model
+	phaseServing              // waiting for OnServeDone for active.Model
+)
+
+// NewSerial builds a Serial scheduler. It takes no Swapper: eviction is always
+// "stop every other running model", so the group planner is not consulted.
+func NewSerial(name string, logger *logmon.Monitor, eff Effects) *Serial {
+	return &Serial{
+		name:    name,
+		logger:  logger,
+		effects: eff,
+	}
+}
+
+// OnRequest validates the model and appends the request to the tail of the queue,
+// then tries to start the next job. Unknown models fail immediately.
+func (s *Serial) OnRequest(req HandlerReq) {
+	if _, ok := s.effects.ModelState(req.Model); !ok {
+		s.logger.Debugf("%s: model %s not handled by this router", s.name, req.Model)
+		s.effects.GrantError(req, ErrModelNotFound)
+		return
+	}
+	s.queued = append(s.queued, req)
+	broadcastQueuePositions(s.queued)
+	s.startNext()
+}
+
+// startNext begins processing the head of the queue when nothing is active. It
+// fast-paths a request whose model is already the sole loaded-and-ready process;
+// otherwise it launches a swap that evicts every other running model first. The
+// loop skips over requests for models that vanished (e.g. a config reload) and
+// requests whose caller disconnected before they could be served.
+func (s *Serial) startNext() {
+	if s.active != nil {
+		return // a job is already loading or serving
+	}
+	for len(s.queued) > 0 {
+		req := s.queued[0]
+		s.queued = s.queued[1:]
+		broadcastQueuePositions(s.queued)
+
+		state, ok := s.effects.ModelState(req.Model)
+		if !ok {
+			s.effects.GrantError(req, ErrModelNotFound)
+			continue
+		}
+
+		r := req
+		s.active = &r
+
+		evict := s.otherRunning(req.Model)
+		if state == process.StateReady && len(evict) == 0 {
+			// Already loaded and the only model running — serve immediately.
+			s.logger.Debugf("%s: serving model %s (already loaded)", s.name, req.Model)
+			if s.serve() {
+				return
+			}
+			continue // caller gone; pick the next request
+		}
+
+		s.logger.Debugf("%s: swapping to model %s, evicting %v", s.name, req.Model, evict)
+		s.phase = phaseSwapping
+		s.effects.StartSwap(req.Model, evict)
+		return
+	}
+}
+
+// serve hands the active request its tracked handler. It returns true when the
+// request is now serving (await OnServeDone); false when the caller had already
+// disconnected, in which case active is cleared so the next job can start.
+func (s *Serial) serve() bool {
+	if s.effects.GrantServe(*s.active, s.active.Model) {
+		s.phase = phaseServing
+		return true
+	}
+	s.logger.Debugf("%s: caller for model %s gone before serve", s.name, s.active.Model)
+	s.active = nil
+	s.phase = phaseIdle
+	return false
+}
+
+// OnSwapDone fires when the load for the active request completes. On success the
+// request is served; on failure its caller receives the error and the queue
+// advances. A SwapDone that does not match the active load (e.g. its request was
+// unloaded or cancelled mid-load) is ignored.
+func (s *Serial) OnSwapDone(ev SwapDone) {
+	if s.active == nil || s.phase != phaseSwapping || s.active.Model != ev.ModelID {
+		return
+	}
+	if ev.Err != nil {
+		s.logger.Debugf("%s: swap for model %s failed: %v", s.name, ev.ModelID, ev.Err)
+		s.effects.GrantError(*s.active, ev.Err)
+		s.active = nil
+		s.phase = phaseIdle
+		s.startNext()
+		return
+	}
+	if !s.serve() {
+		s.startNext() // caller vanished while the model loaded; move on
+	}
+}
+
+// OnServeDone fires when the active request's handler returns. The slot is freed
+// and the next queued request begins.
+func (s *Serial) OnServeDone(ev ServeDoneEvent) {
+	if s.active == nil || s.phase != phaseServing {
+		return
+	}
+	s.active = nil
+	s.phase = phaseIdle
+	s.startNext()
+}
+
+// OnCancel removes a disconnected client's request from the queue. A request that
+// is already active is left to finish: if it was loading, OnSwapDone's serve()
+// will find the caller gone (GrantServe false) and advance; if it was serving,
+// its handler returns normally and reaches OnServeDone.
+func (s *Serial) OnCancel(req HandlerReq) {
+	if len(s.queued) == 0 {
+		return
+	}
+	kept := s.queued[:0]
+	removed := false
+	for _, q := range s.queued {
+		if q.Respond == req.Respond {
+			removed = true
+			continue
+		}
+		kept = append(kept, q)
+	}
+	s.queued = kept
+	if removed {
+		s.logger.Debugf("%s: cancelled request for model %s pruned from queue", s.name, req.Model)
+		broadcastQueuePositions(s.queued)
+	}
+}
+
+// OnUnload reconciles state for an unload, stops the targeted processes, and
+// advances the queue. It mirrors the FIFO contract: queued requests for unloaded
+// models are failed; an active *loading* request for an unloaded model is failed
+// (its swap goroutine is left to finish and its SwapDone is then ignored); an
+// active *serving* request is left for its handler to end when StopProcesses
+// kills the upstream. The Stop is synchronous so callers of Unload can rely on
+// the processes being stopped on return.
+func (s *Serial) OnUnload(targets []string, timeout time.Duration) {
+	unloadErr := fmt.Errorf("%s: model unloaded", s.name)
+
+	targetSet := make(map[string]bool, len(targets))
+	for _, id := range targets {
+		targetSet[id] = true
+	}
+
+	if s.active != nil && s.phase == phaseSwapping && targetSet[s.active.Model] {
+		s.effects.GrantError(*s.active, unloadErr)
+		s.active = nil
+		s.phase = phaseIdle
+	}
+
+	if len(s.queued) > 0 {
+		kept := s.queued[:0]
+		for _, q := range s.queued {
+			if targetSet[q.Model] {
+				s.effects.GrantError(q, unloadErr)
+				continue
+			}
+			kept = append(kept, q)
+		}
+		s.queued = kept
+		broadcastQueuePositions(s.queued)
+	}
+
+	s.effects.StopProcesses(timeout, targets)
+
+	// A still-serving active request advances via OnServeDone when its killed
+	// handler returns; only start the next job when nothing is active now.
+	if s.active == nil {
+		s.startNext()
+	}
+}
+
+// OnShutdown grants err to every request the scheduler still holds: an active
+// loading request and all queued requests. A serving request is torn down with
+// its process by the baseRouter.
+func (s *Serial) OnShutdown(err error) {
+	if s.active != nil && s.phase == phaseSwapping {
+		s.effects.GrantError(*s.active, err)
+		s.active = nil
+		s.phase = phaseIdle
+	}
+	for _, q := range s.queued {
+		s.effects.GrantError(q, err)
+	}
+	s.queued = nil
+}
+
+// otherRunning returns every running model except target, sorted for
+// deterministic eviction.
+func (s *Serial) otherRunning(target string) []string {
+	var out []string
+	for id := range s.effects.RunningModels() {
+		if id != target {
+			out = append(out, id)
+		}
+	}
+	sort.Strings(out)
+	return out
+}

internal/router/scheduler/serial_test.go

@@ -0,0 +1,391 @@
+package scheduler
+
+import (
+	"errors"
+	"io"
+	"testing"
+	"time"
+
+	"github.com/mostlygeek/llama-swap/internal/logmon"
+	"github.com/mostlygeek/llama-swap/internal/process"
+)
+
+// Serial methods all run on the router's single run-loop goroutine, so these
+// tests drive them directly and synchronously, reusing fakeEffects and the
+// req/reqCh helpers from fifo_test.go. A load completes via OnSwapDone and a
+// served request finishes via OnServeDone — the events the run loop delivers.
+
+func newSerial(eff Effects) *Serial {
+	return NewSerial("test", logmon.NewWriter(io.Discard), eff)
+}
+
+// lastStart returns the most recent StartSwap record.
+func lastStart(t *testing.T, eff *fakeEffects) startRec {
+	t.Helper()
+	if len(eff.starts) == 0 {
+		t.Fatal("no StartSwap recorded")
+	}
+	return eff.starts[len(eff.starts)-1]
+}
+
+func sameSet(a, b []string) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	m := map[string]int{}
+	for _, x := range a {
+		m[x]++
+	}
+	for _, x := range b {
+		m[x]--
+	}
+	for _, v := range m {
+		if v != 0 {
+			return false
+		}
+	}
+	return true
+}
+
+// servedOrder returns the model IDs of every successful serve grant in order.
+func servedOrder(eff *fakeEffects) []string {
+	var out []string
+	for _, g := range eff.grants {
+		if g.err == nil && g.serve {
+			out = append(out, g.model)
+		}
+	}
+	return out
+}
+
+func TestSerial_FastPath_AlreadyLoaded(t *testing.T) {
+	eff := newFakeEffects()
+	eff.states["a"] = process.StateReady
+	s := newSerial(eff)
+
+	s.OnRequest(req("a"))
+
+	if got := len(eff.starts); got != 0 {
+		t.Errorf("StartSwap calls=%d want 0 (already loaded, no swap)", got)
+	}
+	if got := eff.served("a"); got != 1 {
+		t.Errorf("served(a)=%d want 1", got)
+	}
+}
+
+func TestSerial_ColdStart_LoadsThenServes(t *testing.T) {
+	eff := newFakeEffects()
+	eff.states["a"] = process.StateStopped
+	s := newSerial(eff)
+
+	s.OnRequest(req("a"))
+	if got := eff.startsFor("a"); got != 1 {
+		t.Fatalf("StartSwap(a)=%d want 1", got)
+	}
+	if got := eff.served("a"); got != 0 {
+		t.Errorf("served(a)=%d want 0 before load completes", got)
+	}
+
+	eff.states["a"] = process.StateReady
+	s.OnSwapDone(SwapDone{ModelID: "a"})
+	if got := eff.served("a"); got != 1 {
+		t.Errorf("served(a)=%d want 1 after load", got)
+	}
+}
+
+func TestSerial_UnknownModel(t *testing.T) {
+	eff := newFakeEffects() // no states => unknown
+	s := newSerial(eff)
+
+	s.OnRequest(req("ghost"))
+
+	if len(eff.starts) != 0 {
+		t.Errorf("StartSwap calls=%d want 0", len(eff.starts))
+	}
+	if eff.errored("ghost") != 1 {
+		t.Fatalf("errored(ghost)=%d want 1", eff.errored("ghost"))
+	}
+	if !errors.Is(eff.grants[0].err, ErrModelNotFound) {
+		t.Errorf("err=%v want ErrModelNotFound", eff.grants[0].err)
+	}
+}
+
+func TestSerial_EvictsEveryOtherModel(t *testing.T) {
+	eff := newFakeEffects()
+	eff.states["x"] = process.StateReady // already running
+	eff.states["y"] = process.StateReady // also running (e.g. left over)
+	eff.states["a"] = process.StateStopped
+	s := newSerial(eff)
+
+	s.OnRequest(req("a"))
+
+	st := lastStart(t, eff)
+	if st.model != "a" {
+		t.Fatalf("loading %s want a", st.model)
+	}
+	if !sameSet(st.evict, []string{"x", "y"}) {
+		t.Errorf("evict=%v want [x y] (serial evicts ALL other models)", st.evict)
+	}
+}
+
+// TestSerial_OneJobAtATime verifies a second request waits while the first is
+// serving, and only starts after the first finishes.
+func TestSerial_OneJobAtATime(t *testing.T) {
+	eff := newFakeEffects()
+	eff.states["a"] = process.StateReady
+	eff.states["b"] = process.StateStopped
+	s := newSerial(eff)
+
+	s.OnRequest(req("a")) // served immediately
+	s.OnRequest(req("b")) // must wait — a is serving
+
+	if got := eff.startsFor("b"); got != 0 {
+		t.Fatalf("StartSwap(b)=%d want 0 while a is serving", got)
+	}
+	if got := eff.served("a"); got != 1 {
+		t.Fatalf("served(a)=%d want 1", got)
+	}
+
+	// a finishes -> b may now load (evicting a).
+	s.OnServeDone(ServeDoneEvent{ModelID: "a"})
+	if got := eff.startsFor("b"); got != 1 {
+		t.Fatalf("StartSwap(b)=%d want 1 after a finished", got)
+	}
+	if st := lastStart(t, eff); !sameSet(st.evict, []string{"a"}) {
+		t.Errorf("b evict=%v want [a]", st.evict)
+	}
+}
+
+// TestSerial_SameModelConsecutive_NoReload verifies back-to-back requests for the
+// already-loaded model run without a reload, one after another.
+func TestSerial_SameModelConsecutive_NoReload(t *testing.T) {
+	eff := newFakeEffects()
+	eff.states["a"] = process.StateStopped
+	s := newSerial(eff)
+
+	s.OnRequest(req("a")) // cold load
+	s.OnRequest(req("a")) // queued behind the first
+
+	eff.states["a"] = process.StateReady
+	s.OnSwapDone(SwapDone{ModelID: "a"}) // first serves
+	if got := eff.served("a"); got != 1 {
+		t.Fatalf("served(a)=%d want 1 (one at a time)", got)
+	}
+
+	s.OnServeDone(ServeDoneEvent{ModelID: "a"}) // first done -> second serves
+	if got := eff.served("a"); got != 2 {
+		t.Fatalf("served(a)=%d want 2", got)
+	}
+	if got := eff.startsFor("a"); got != 1 {
+		t.Errorf("StartSwap(a)=%d want 1 (second request must not reload)", got)
+	}
+}
+
+// TestSerial_StrictArrivalOrder is the core guarantee: qwen36, qwen35, sdxl,
+// qwen36 execute in EXACTLY that order with evictions between each model switch,
+// including reloading qwen36 at the end even though it ran first.
+func TestSerial_StrictArrivalOrder(t *testing.T) {
+	eff := newFakeEffects()
+	for _, m := range []string{"qwen36", "qwen35", "sdxl"} {
+		eff.states[m] = process.StateStopped
+	}
+	s := newSerial(eff)
+
+	for _, m := range []string{"qwen36", "qwen35", "sdxl", "qwen36"} {
+		s.OnRequest(req(m))
+	}
+
+	// Only the first job starts loading; the rest wait their turn.
+	if len(eff.starts) != 1 || eff.starts[0].model != "qwen36" {
+		t.Fatalf("starts=%+v want only [qwen36] loading first", eff.starts)
+	}
+
+	// step completes the current model's load+serve and returns control to the
+	// scheduler, which must start the next queued model.
+	step := func(model string, wantEvict []string) {
+		t.Helper()
+		st := lastStart(t, eff)
+		if st.model != model {
+			t.Fatalf("loading %q want %q", st.model, model)
+		}
+		if !sameSet(st.evict, wantEvict) {
+			t.Fatalf("loading %q evict=%v want %v", model, st.evict, wantEvict)
+		}
+		// Simulate the eviction + load actually happening.
+		for _, e := range st.evict {
+			eff.states[e] = process.StateStopped
+		}
+		eff.states[model] = process.StateReady
+		s.OnSwapDone(SwapDone{ModelID: model})
+		s.OnServeDone(ServeDoneEvent{ModelID: model})
+	}
+
+	step("qwen36", nil)                // cold load, nothing else running
+	step("qwen35", []string{"qwen36"}) // evict qwen36
+	step("sdxl", []string{"qwen35"})   // evict qwen35
+	step("qwen36", []string{"sdxl"})   // RELOAD qwen36, evict sdxl
+
+	want := []string{"qwen36", "qwen35", "sdxl", "qwen36"}
+	if got := servedOrder(eff); !sameOrder(got, want) {
+		t.Fatalf("serve order=%v want %v", got, want)
+	}
+}
+
+func sameOrder(a, b []string) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	for i := range a {
+		if a[i] != b[i] {
+			return false
+		}
+	}
+	return true
+}
+
+func TestSerial_SwapError_FailsCallerAndAdvances(t *testing.T) {
+	eff := newFakeEffects()
+	eff.states["a"] = process.StateStopped
+	eff.states["b"] = process.StateStopped
+	s := newSerial(eff)
+
+	s.OnRequest(req("a"))
+	s.OnRequest(req("b")) // queued behind a
+
+	// a's load fails: its caller is errored and b proceeds.
+	s.OnSwapDone(SwapDone{ModelID: "a", Err: errors.New("boom")})
+	if eff.errored("a") != 1 {
+		t.Fatalf("errored(a)=%d want 1", eff.errored("a"))
+	}
+	if got := eff.startsFor("b"); got != 1 {
+		t.Fatalf("StartSwap(b)=%d want 1 after a's load failed", got)
+	}
+}
+
+// TestSerial_GrantServeFalse_Advances verifies that when the active request's
+// caller has disconnected by serve time, the queue advances to the next request.
+func TestSerial_GrantServeFalse_Advances(t *testing.T) {
+	eff := newFakeEffects()
+	eff.states["a"] = process.StateStopped
+	eff.states["b"] = process.StateStopped
+	eff.serveResult["a"] = false // a's caller is gone by grant time
+	s := newSerial(eff)
+
+	s.OnRequest(req("a"))
+	s.OnRequest(req("b")) // queued
+
+	eff.states["a"] = process.StateReady
+	s.OnSwapDone(SwapDone{ModelID: "a"}) // grant fails -> advance to b
+
+	if got := eff.served("a"); got != 0 {
+		t.Errorf("served(a)=%d want 0 (caller gone)", got)
+	}
+	if got := eff.startsFor("b"); got != 1 {
+		t.Fatalf("StartSwap(b)=%d want 1 (advanced after gone caller)", got)
+	}
+}
+
+func TestSerial_OnCancel_QueuedRequest(t *testing.T) {
+	eff := newFakeEffects()
+	eff.states["a"] = process.StateStopped
+	eff.states["b"] = process.StateStopped
+	s := newSerial(eff)
+
+	s.OnRequest(reqCh("a")) // starts loading a
+	cancelled := reqCh("b")
+	s.OnRequest(cancelled) // queued behind a
+	if len(s.queued) != 1 {
+		t.Fatalf("queued=%d want 1", len(s.queued))
+	}
+
+	s.OnCancel(cancelled)
+	if len(s.queued) != 0 {
+		t.Fatalf("queued=%d want 0 after cancel", len(s.queued))
+	}
+
+	// a completes; b is gone, so nothing starts for it.
+	eff.states["a"] = process.StateReady
+	s.OnSwapDone(SwapDone{ModelID: "a"})
+	s.OnServeDone(ServeDoneEvent{ModelID: "a"})
+	if got := eff.startsFor("b"); got != 0 {
+		t.Errorf("StartSwap(b)=%d want 0 (cancelled before its turn)", got)
+	}
+}
+
+func TestSerial_OnShutdown_FailsQueuedAndActiveLoad(t *testing.T) {
+	eff := newFakeEffects()
+	eff.states["a"] = process.StateStopped
+	eff.states["b"] = process.StateStopped
+	eff.states["c"] = process.StateStopped
+	s := newSerial(eff)
+
+	s.OnRequest(req("a")) // active (loading)
+	s.OnRequest(req("b")) // queued
+	s.OnRequest(req("c")) // queued
+
+	s.OnShutdown(errors.New("shutting down"))
+
+	if got := eff.errored(""); got != 3 {
+		t.Errorf("error grants=%d want 3 (active load + 2 queued)", got)
+	}
+	if len(s.queued) != 0 {
+		t.Errorf("queued=%d want 0 after shutdown", len(s.queued))
+	}
+}
+
+// TestSerial_OnUnload_WhileServing verifies that unloading the model that is
+// actively serving does not strand the queue: OnUnload stops the process but
+// leaves the active request to end via OnServeDone, which then advances.
+func TestSerial_OnUnload_WhileServing(t *testing.T) {
+	eff := newFakeEffects()
+	eff.states["a"] = process.StateReady
+	eff.states["b"] = process.StateStopped
+	s := newSerial(eff)
+
+	s.OnRequest(req("a")) // served immediately (a ready)
+	s.OnRequest(req("b")) // queued behind a
+	if got := eff.served("a"); got != 1 {
+		t.Fatalf("served(a)=%d want 1", got)
+	}
+
+	// Unload a while it is serving: the process is stopped, but the queue must
+	// not advance yet — the active serve is still outstanding.
+	s.OnUnload([]string{"a"}, time.Second)
+	if len(eff.stops) != 1 || !sameSet(eff.stops[0].ids, []string{"a"}) {
+		t.Errorf("StopProcesses=%+v want one call stopping [a]", eff.stops)
+	}
+	if got := eff.startsFor("b"); got != 0 {
+		t.Fatalf("StartSwap(b)=%d want 0 before the serving request ends", got)
+	}
+
+	// The killed handler returns -> OnServeDone advances to b.
+	eff.states["a"] = process.StateStopped
+	s.OnServeDone(ServeDoneEvent{ModelID: "a"})
+	if got := eff.startsFor("b"); got != 1 {
+		t.Fatalf("StartSwap(b)=%d want 1 after the serving request ended", got)
+	}
+}
+
+func TestSerial_OnUnload_DropsQueuedAndStops(t *testing.T) {
+	eff := newFakeEffects()
+	eff.states["a"] = process.StateStopped
+	eff.states["b"] = process.StateStopped
+	s := newSerial(eff)
+
+	s.OnRequest(req("a")) // active (loading a)
+	s.OnRequest(req("b")) // queued
+
+	// Unload a: its active load is failed and a is stopped.
+	s.OnUnload([]string{"a"}, time.Second)
+
+	if eff.errored("a") != 1 {
+		t.Errorf("errored(a)=%d want 1 (active load failed)", eff.errored("a"))
+	}
+	if len(eff.stops) != 1 || !sameSet(eff.stops[0].ids, []string{"a"}) {
+		t.Errorf("StopProcesses=%+v want one call stopping [a]", eff.stops)
+	}
+	// b was queued and not unloaded; with a's load cancelled it now starts.
+	if got := eff.startsFor("b"); got != 1 {
+		t.Errorf("StartSwap(b)=%d want 1 after unload advanced the queue", got)
+	}
+}

internal/server/api.go

@@ -88,6 +88,12 @@ func renderCapabilities(caps config.ModelCapConfig) (arch map[string]any, capsMa
 		if contains(caps.In, "image") && contains(caps.Out, "image") {
 			capsMap["image_to_image"] = true
 		}
+		if contains(caps.In, "text") && contains(caps.Out, "video") {
+			capsMap["video_generation"] = true
+		}
+		if contains(caps.In, "image") && contains(caps.Out, "video") {
+			capsMap["image_to_video"] = true
+		}
 	}
 
 	if caps.Tools {

internal/server/captures.go

@@ -37,7 +37,8 @@ const (
 )
 
 // captureFieldsByPath overrides the default capture mask for routes carrying
-// large binary payloads (audio/image) where storing the full body is wasteful.
+// large binary payloads (audio/image/video) where storing the full body is
+// wasteful.
 var captureFieldsByPath = map[string]captureFields{
 	"/v1/audio/speech":         captureReqAll | captureRespHeaders,
 	"/v1/audio/voices":         captureReqHeaders | captureRespAll,
@@ -46,6 +47,9 @@ var captureFieldsByPath = map[string]captureFields{
 	"/v1/images/edits":         captureReqHeaders | captureRespHeaders,
 	"/sdapi/v1/txt2img":        captureReqAll | captureRespHeaders,
 	"/sdapi/v1/img2img":        captureReqHeaders | captureRespHeaders,
+	// video: the request multipart can carry a conditioning frame and the
+	// response body IS the clip — headers only, both directions.
+	"/v1/videos/sync": captureReqHeaders | captureRespHeaders,
 }
 
 // captureFieldsFor returns the capture mask for a request path. Unlisted routes

internal/server/server.go

@@ -80,6 +80,15 @@ var modelPostJSONRoutes = []string{
 var modelPostFormRoutes = []string{
 	"/v1/audio/transcriptions",
 	"/v1/images/edits",
+
+	// video generation: ONLY the blocking sync shape (vLLM-Omni
+	// /v1/videos/sync — the response body is the clip). The async
+	// OpenAI-style POST /v1/videos is deliberately NOT dispatched: its
+	// poll/download companions (GET /v1/videos/{id}[/content]) carry no
+	// model field to route by, so registering creation alone would start
+	// unretrievable upstream jobs. Extraction is content-type driven, so
+	// JSON bodies dispatch here too.
+	"/v1/videos/sync",
 }
 
 // modelGetRoutes are model-dispatched GET endpoints (the model arrives as a

internal/server/server_test.go

@@ -1,9 +1,11 @@
 package server
 
 import (
+	"bytes"
 	"context"
 	"encoding/json"
 	"io"
+	"mime/multipart"
 	"net/http"
 	"net/http/httptest"
 	"strings"
@@ -339,3 +341,56 @@ func TestServer_LogStream_UnknownID_Returns400(t *testing.T) {
 		t.Errorf("status=%d want 400", w.Code)
 	}
 }
+
+func TestServer_VideoRoutesDispatch(t *testing.T) {
+	s := newTestServer(
+		newStubRouter([]string{"videogen-model"}, "video response"),
+		newStubRouter(nil, ""),
+	)
+
+	// Multipart form on /v1/videos/sync.
+	var buf bytes.Buffer
+	mw := multipart.NewWriter(&buf)
+	mw.WriteField("model", "videogen-model")
+	mw.WriteField("prompt", "a cat")
+	mw.Close()
+	req := httptest.NewRequest(http.MethodPost, "/v1/videos/sync", &buf)
+	req.Header.Set("Content-Type", mw.FormDataContentType())
+	w := httptest.NewRecorder()
+	s.ServeHTTP(w, req)
+	if w.Code != http.StatusOK || w.Body.String() != "video response" {
+		t.Fatalf("/v1/videos/sync: status=%d body=%q", w.Code, w.Body.String())
+	}
+
+	// JSON bodies dispatch too (extraction is content-type driven).
+	jsonReq := httptest.NewRequest(http.MethodPost, "/v1/videos/sync",
+		strings.NewReader(`{"model":"videogen-model","prompt":"a cat"}`))
+	jsonReq.Header.Set("Content-Type", "application/json")
+	w = httptest.NewRecorder()
+	s.ServeHTTP(w, jsonReq)
+	if w.Code != http.StatusOK || w.Body.String() != "video response" {
+		t.Fatalf("/v1/videos/sync json: status=%d body=%q", w.Code, w.Body.String())
+	}
+
+	// Unknown model on the video route still 404s.
+	nopeReq := httptest.NewRequest(http.MethodPost, "/v1/videos/sync",
+		strings.NewReader(`{"model":"nope","prompt":"a cat"}`))
+	nopeReq.Header.Set("Content-Type", "application/json")
+	w = httptest.NewRecorder()
+	s.ServeHTTP(w, nopeReq)
+	if w.Code != http.StatusNotFound {
+		t.Fatalf("/v1/videos/sync unknown model: status=%d want 404", w.Code)
+	}
+
+	// The async job-creation route is NOT dispatched (no poll routes to
+	// complete the flow) — a stray OpenAI-videos client gets a clean 404
+	// instead of an unretrievable upstream job.
+	asyncReq := httptest.NewRequest(http.MethodPost, "/v1/videos",
+		strings.NewReader(`{"model":"videogen-model","prompt":"a cat"}`))
+	asyncReq.Header.Set("Content-Type", "application/json")
+	w = httptest.NewRecorder()
+	s.ServeHTTP(w, asyncReq)
+	if w.Code != http.StatusNotFound {
+		t.Fatalf("POST /v1/videos: status=%d want 404 (async family not routed)", w.Code)
+	}
+}

ui-svelte/src/components/playground/VideoInterface.svelte

@@ -0,0 +1,359 @@
+<script lang="ts">
+  import { models } from "../../stores/api";
+  import { persistentStore } from "../../stores/persistent";
+  import { generateVideoSync } from "../../lib/videoApi";
+  import { playgroundStores } from "../../stores/playgroundActivity";
+  import ModelSelector from "./ModelSelector.svelte";
+  import ExpandableTextarea from "./ExpandableTextarea.svelte";
+
+  const selectedModelStore = persistentStore<string>("playground-video-model", "");
+  const selectedSizeStore = persistentStore<string>("playground-video-size", "");
+  const negativePromptStore = persistentStore<string>("playground-video-negative-prompt", "");
+  const numFramesStore = persistentStore<number>("playground-video-num-frames", 0);
+  const fpsStore = persistentStore<number>("playground-video-fps", 0);
+  const stepsStore = persistentStore<number>("playground-video-steps", 0);
+  const guidanceStore = persistentStore<number>("playground-video-guidance", 0);
+  const seedStore = persistentStore<number>("playground-video-seed", -1);
+
+  let prompt = $state("");
+  let isGenerating = $state(false);
+  let videoUrl = $state<string | null>(null);
+  let videoMime = $state("video/mp4");
+  let videoBytes = $state(0);
+  let elapsedSeconds = $state(0);
+  let error = $state<string | null>(null);
+  let abortController = $state<AbortController | null>(null);
+  let showSettings = $state(false);
+  let initImage = $state<File | null>(null);
+  let initImagePreview = $state<string | null>(null);
+  let fileInput = $state<HTMLInputElement | null>(null);
+  let timer: ReturnType<typeof setInterval> | null = null;
+
+  let hasModels = $derived($models.some((m) => !m.unlisted));
+
+  $effect(() => {
+    playgroundStores.videoGenerating.set(isGenerating);
+  });
+
+  function onInitImageChange(event: Event) {
+    const input = event.target as HTMLInputElement;
+    const file = input.files?.[0] ?? null;
+    setInitImage(file);
+  }
+
+  function setInitImage(file: File | null) {
+    if (initImagePreview) URL.revokeObjectURL(initImagePreview);
+    initImage = file;
+    initImagePreview = file ? URL.createObjectURL(file) : null;
+  }
+
+  function clearInitImage() {
+    setInitImage(null);
+    if (fileInput) fileInput.value = "";
+  }
+
+  async function generate() {
+    const trimmedPrompt = prompt.trim();
+    if (!trimmedPrompt || !$selectedModelStore || isGenerating) return;
+
+    isGenerating = true;
+    error = null;
+    elapsedSeconds = 0;
+    abortController = new AbortController();
+    timer = setInterval(() => (elapsedSeconds += 1), 1000);
+
+    try {
+      const result = await generateVideoSync(
+        $selectedModelStore,
+        trimmedPrompt,
+        {
+          negativePrompt: $negativePromptStore || undefined,
+          size: $selectedSizeStore || undefined,
+          numFrames: $numFramesStore,
+          fps: $fpsStore,
+          steps: $stepsStore,
+          guidanceScale: $guidanceStore,
+          seed: $seedStore,
+          initImage,
+        },
+        abortController.signal
+      );
+      if (videoUrl) URL.revokeObjectURL(videoUrl);
+      videoUrl = result.url;
+      videoMime = result.mime;
+      videoBytes = result.sizeBytes;
+    } catch (err) {
+      if (err instanceof Error && err.name === "AbortError") {
+        // User cancelled
+      } else {
+        error = err instanceof Error ? err.message : "An error occurred";
+      }
+    } finally {
+      isGenerating = false;
+      abortController = null;
+      if (timer) clearInterval(timer);
+      timer = null;
+    }
+  }
+
+  function cancelGeneration() {
+    abortController?.abort();
+  }
+
+  function clearVideo() {
+    if (videoUrl) URL.revokeObjectURL(videoUrl);
+    videoUrl = null;
+    error = null;
+    prompt = "";
+  }
+
+  function downloadVideo() {
+    if (!videoUrl) return;
+    const ext = videoMime.includes("webm") ? "webm" : "mp4";
+    const link = document.createElement("a");
+    link.href = videoUrl;
+    link.download = `generated-video-${Date.now()}.${ext}`;
+    document.body.appendChild(link);
+    link.click();
+    document.body.removeChild(link);
+  }
+
+  function formatBytes(n: number): string {
+    if (n > 1024 * 1024) return `${(n / 1024 / 1024).toFixed(1)} MB`;
+    if (n > 1024) return `${(n / 1024).toFixed(0)} KB`;
+    return `${n} B`;
+  }
+
+  function formatElapsed(s: number): string {
+    const m = Math.floor(s / 60);
+    return m > 0 ? `${m}m ${s % 60}s` : `${s}s`;
+  }
+
+  function handleKeyDown(event: KeyboardEvent) {
+    if (event.key === "Enter" && !event.shiftKey) {
+      event.preventDefault();
+      generate();
+    }
+  }
+</script>
+
+<div class="flex flex-col h-full">
+  <!-- Model selector and options -->
+  <div class="shrink-0 flex flex-wrap gap-2 mb-4">
+    <ModelSelector
+      bind:value={$selectedModelStore}
+      placeholder="Select a video model..."
+      disabled={isGenerating}
+      capabilities={["video_generation", "image_to_video"]}
+      matchAny={true}
+    />
+
+    <select
+      class="px-3 py-2 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+      bind:value={$selectedSizeStore}
+      disabled={isGenerating}
+    >
+      <option value="">Model default size</option>
+      <optgroup label="Landscape">
+        <option value="1280x704">1280x704</option>
+        <option value="832x480">832x480</option>
+      </optgroup>
+      <optgroup label="Portrait">
+        <option value="704x1280">704x1280</option>
+        <option value="480x832">480x832</option>
+      </optgroup>
+      <optgroup label="Square">
+        <option value="704x704">704x704</option>
+      </optgroup>
+    </select>
+
+    <button
+      class="px-3 py-2 rounded border border-gray-200 dark:border-white/10 bg-surface hover:bg-secondary-hover transition-colors"
+      onclick={() => (showSettings = !showSettings)}
+    >
+      {showSettings ? "Hide Settings" : "Settings"}
+    </button>
+  </div>
+
+  <!-- Settings panel -->
+  {#if showSettings}
+    <div class="shrink-0 mb-4 p-4 rounded border border-gray-200 dark:border-white/10 bg-surface">
+      <div class="grid grid-cols-2 md:grid-cols-5 gap-3 mb-3">
+        <label class="flex flex-col gap-1">
+          <span class="text-xs text-txtsecondary">Frames (0 = default)</span>
+          <input
+            type="number"
+            class="px-2 py-1 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+            bind:value={$numFramesStore}
+            min="0"
+            max="1000"
+          />
+        </label>
+        <label class="flex flex-col gap-1">
+          <span class="text-xs text-txtsecondary">FPS (0 = default)</span>
+          <input
+            type="number"
+            class="px-2 py-1 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+            bind:value={$fpsStore}
+            min="0"
+            max="60"
+          />
+        </label>
+        <label class="flex flex-col gap-1">
+          <span class="text-xs text-txtsecondary">Steps (0 = default)</span>
+          <input
+            type="number"
+            class="px-2 py-1 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+            bind:value={$stepsStore}
+            min="0"
+            max="150"
+          />
+        </label>
+        <label class="flex flex-col gap-1">
+          <span class="text-xs text-txtsecondary">Guidance (0 = default)</span>
+          <input
+            type="number"
+            class="px-2 py-1 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+            bind:value={$guidanceStore}
+            min="0"
+            max="30"
+            step="0.5"
+          />
+        </label>
+        <label class="flex flex-col gap-1">
+          <span class="text-xs text-txtsecondary">Seed (-1 = random)</span>
+          <input
+            type="number"
+            class="px-2 py-1 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+            bind:value={$seedStore}
+            min="-1"
+          />
+        </label>
+      </div>
+
+      <label class="flex flex-col gap-1">
+        <span class="text-xs text-txtsecondary">Negative Prompt</span>
+        <textarea
+          class="px-2 py-1 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary resize-y text-sm"
+          bind:value={$negativePromptStore}
+          rows="2"
+          placeholder="Elements to avoid..."
+        ></textarea>
+      </label>
+    </div>
+  {/if}
+
+  <!-- Empty state for no models configured -->
+  {#if !hasModels}
+    <div class="flex-1 flex items-center justify-center text-txtsecondary">
+      <p>No models configured. Add models to your configuration to generate videos.</p>
+    </div>
+  {:else}
+    <!-- Video display area -->
+    <div
+      class="flex-1 overflow-auto mb-4 flex items-center justify-center bg-surface border border-gray-200 dark:border-white/10 rounded"
+    >
+      {#if isGenerating}
+        <div class="text-center text-txtsecondary">
+          <div
+            class="inline-block w-8 h-8 border-4 border-primary border-t-transparent rounded-full animate-spin mb-2"
+          ></div>
+          <p>Generating video… {formatElapsed(elapsedSeconds)}</p>
+          <p class="text-xs mt-1">
+            Video generation takes minutes — a cold model load takes longer still.
+          </p>
+        </div>
+      {:else if error}
+        <div class="text-center text-red-500 p-4">
+          <p class="font-medium">Error</p>
+          <p class="text-sm mt-1 break-all">{error}</p>
+        </div>
+      {:else if videoUrl}
+        <div class="relative max-w-full max-h-full flex items-center justify-center">
+          <!-- svelte-ignore a11y_media_has_caption -->
+          <video src={videoUrl} class="max-w-full max-h-full object-contain" controls autoplay loop></video>
+          <button
+            class="absolute bottom-2 right-2 p-2 bg-black/60 hover:bg-black/80 text-white rounded-full transition-colors"
+            onclick={downloadVideo}
+            aria-label="Download video"
+            title="Download ({formatBytes(videoBytes)})"
+          >
+            <svg class="w-5 h-5" fill="none" stroke="currentColor" viewBox="0 0 24 24">
+              <path
+                stroke-linecap="round"
+                stroke-linejoin="round"
+                stroke-width="2"
+                d="M4 16v1a3 3 0 003 3h10a3 3 0 003-3v-1m-4-4l-4 4m0 0l-4-4m4 4V4"
+              ></path>
+            </svg>
+          </button>
+        </div>
+      {:else}
+        <div class="text-center text-txtsecondary">
+          <p>Enter a prompt below to generate a short video clip</p>
+          <p class="text-xs mt-1">Attach an image to animate it (image-to-video)</p>
+        </div>
+      {/if}
+    </div>
+
+    <!-- Init image chip -->
+    {#if initImagePreview}
+      <div class="shrink-0 flex items-center gap-2 mb-2">
+        <img src={initImagePreview} alt="Conditioning frame" class="h-12 rounded border border-gray-200 dark:border-white/10" />
+        <span class="text-xs text-txtsecondary">Animating this image (image-to-video)</span>
+        <button
+          class="px-1.5 py-0.5 text-xs rounded border border-gray-200 dark:border-white/10 hover:bg-red-500 hover:text-white hover:border-red-500 transition-colors"
+          onclick={clearInitImage}
+          disabled={isGenerating}
+        >
+          Remove
+        </button>
+      </div>
+    {/if}
+
+    <!-- Prompt input area -->
+    <div class="shrink-0 flex flex-col md:flex-row gap-2">
+      <ExpandableTextarea
+        bind:value={prompt}
+        placeholder="Describe the video you want to generate..."
+        rows={3}
+        onkeydown={handleKeyDown}
+        disabled={isGenerating || !$selectedModelStore}
+      />
+      <div class="flex flex-row md:flex-col gap-2">
+        {#if isGenerating}
+          <button class="btn bg-red-500 hover:bg-red-600 text-white flex-1 md:flex-none" onclick={cancelGeneration}>
+            Cancel
+          </button>
+        {:else}
+          <button
+            class="btn bg-primary text-btn-primary-text hover:opacity-90 flex-1 md:flex-none"
+            onclick={generate}
+            disabled={!prompt.trim() || !$selectedModelStore}
+          >
+            Generate
+          </button>
+          <label
+            class="btn flex-1 md:flex-none text-center cursor-pointer {isGenerating ? 'opacity-50 pointer-events-none' : ''}"
+          >
+            Image…
+            <input
+              type="file"
+              accept="image/*"
+              class="hidden"
+              bind:this={fileInput}
+              onchange={onInitImageChange}
+            />
+          </label>
+          <button
+            class="btn flex-1 md:flex-none"
+            onclick={clearVideo}
+            disabled={!videoUrl && !error && !prompt.trim()}
+          >
+            Clear
+          </button>
+        {/if}
+      </div>
+    </div>
+  {/if}
+</div>

ui-svelte/src/lib/types.ts

@@ -8,6 +8,8 @@ export interface ModelCapabilities {
   audio_speech?: boolean;
   image_generation?: boolean;
   image_to_image?: boolean;
+  video_generation?: boolean;
+  image_to_video?: boolean;
   function_calling?: boolean;
   reranker?: boolean;
 }

ui-svelte/src/lib/videoApi.ts

@@ -0,0 +1,57 @@
+export interface VideoGenerationOptions {
+  negativePrompt?: string;
+  size?: string; // "WxH"
+  numFrames?: number;
+  fps?: number;
+  steps?: number;
+  guidanceScale?: number;
+  seed?: number; // -1 = random (omitted)
+  initImage?: File | null; // image-to-video conditioning frame
+}
+
+// POST /v1/videos/sync (multipart; llama-swap routes by the `model` field).
+// The response body IS the encoded clip — returns an object URL for <video>.
+// Callers must URL.revokeObjectURL() the result when done with it.
+export async function generateVideoSync(
+  model: string,
+  prompt: string,
+  options: VideoGenerationOptions = {},
+  signal?: AbortSignal
+): Promise<{ url: string; mime: string; sizeBytes: number }> {
+  const form = new FormData();
+  form.set("model", model);
+  form.set("prompt", prompt);
+  if (options.negativePrompt) form.set("negative_prompt", options.negativePrompt);
+  if (options.size) {
+    const [w, h] = options.size.split("x").map(Number);
+    if (w > 0 && h > 0) {
+      // Both conventions: vLLM-Omni reads width/height, OpenAI-shaped
+      // upstreams read size; they can never disagree.
+      form.set("width", String(w));
+      form.set("height", String(h));
+      form.set("size", options.size);
+    }
+  }
+  if (options.numFrames && options.numFrames > 0) form.set("num_frames", String(options.numFrames));
+  if (options.fps && options.fps > 0) form.set("fps", String(options.fps));
+  if (options.steps && options.steps > 0) form.set("num_inference_steps", String(options.steps));
+  if (options.guidanceScale && options.guidanceScale > 0)
+    form.set("guidance_scale", String(options.guidanceScale));
+  if (options.seed !== undefined && options.seed >= 0) form.set("seed", String(options.seed));
+  if (options.initImage) form.set("input_reference", options.initImage, options.initImage.name);
+
+  const response = await fetch("/v1/videos/sync", {
+    method: "POST",
+    body: form,
+    signal,
+  });
+
+  if (!response.ok) {
+    const errorText = await response.text();
+    throw new Error(`Video API error: ${response.status} - ${errorText}`);
+  }
+
+  const blob = await response.blob();
+  const mime = response.headers.get("Content-Type") || "video/mp4";
+  return { url: URL.createObjectURL(blob), mime, sizeBytes: blob.size };
+}

ui-svelte/src/routes/Playground.svelte

@@ -2,12 +2,13 @@
   import { persistentStore } from "../stores/persistent";
   import ChatInterface from "../components/playground/ChatInterface.svelte";
   import ImageInterface from "../components/playground/ImageInterface.svelte";
+  import VideoInterface from "../components/playground/VideoInterface.svelte";
   import AudioInterface from "../components/playground/AudioInterface.svelte";
   import SpeechInterface from "../components/playground/SpeechInterface.svelte";
   import RerankInterface from "../components/playground/RerankInterface.svelte";
   import ConcurrencyInterface from "../components/playground/ConcurrencyInterface.svelte";
 
-  type Tab = "chat" | "images" | "speech" | "audio" | "rerank" | "concurrency";
+  type Tab = "chat" | "images" | "video" | "speech" | "audio" | "rerank" | "concurrency";
 
   const selectedTabStore = persistentStore<Tab>("playground-selected-tab", "chat");
   let mobileMenuOpen = $state(false);
@@ -15,6 +16,7 @@
   const tabs: { id: Tab; label: string }[] = [
     { id: "chat", label: "Chat" },
     { id: "images", label: "Images" },
+    { id: "video", label: "Video" },
     { id: "speech", label: "Speech" },
     { id: "audio", label: "Transcription" },
     { id: "rerank", label: "Rerank" },
@@ -92,6 +94,9 @@
     <div class="h-full" class:tab-hidden={$selectedTabStore !== "images"}>
       <ImageInterface />
     </div>
+    <div class="h-full" class:tab-hidden={$selectedTabStore !== "video"}>
+      <VideoInterface />
+    </div>
     <div class="h-full" class:tab-hidden={$selectedTabStore !== "speech"}>
       <SpeechInterface />
     </div>

ui-svelte/src/stores/playgroundActivity.ts

@@ -2,18 +2,20 @@ import { writable, derived } from "svelte/store";
 
 const chatStreaming = writable(false);
 const imageGenerating = writable(false);
+const videoGenerating = writable(false);
 const speechGenerating = writable(false);
 const audioTranscribing = writable(false);
 const rerankLoading = writable(false);
 
 export const playgroundActivity = derived(
-  [chatStreaming, imageGenerating, speechGenerating, audioTranscribing, rerankLoading],
+  [chatStreaming, imageGenerating, videoGenerating, speechGenerating, audioTranscribing, rerankLoading],
-  ([$chat, $image, $speech, $audio, $rerank]) => $chat || $image || $speech || $audio || $rerank
+  ([$chat, $image, $video, $speech, $audio, $rerank]) => $chat || $image || $video || $speech || $audio || $rerank
 );
 
 export const playgroundStores = {
   chatStreaming,
   imageGenerating,
+  videoGenerating,
   speechGenerating,
   audioTranscribing,
   rerankLoading,