Add a Sleep GPU button that stops Comfy to free VRAM.

Co-authored-by: Cursor <cursoragent@cursor.com>

components/SettingsModal.vue

@@ -10,6 +10,36 @@
           <button type="button" class="rounded-full border border-white/10 px-3 py-1 text-sm text-zinc-400 hover:text-white" @click="emit('close')">Close</button>
         </div>
 
+        <section class="mt-6">
+          <h3 class="text-sm font-semibold uppercase tracking-wider text-zinc-500">GPU</h3>
+          <div class="mt-2 space-y-3 rounded-2xl border border-white/10 bg-zinc-950/50 px-4 py-3">
+            <p class="text-sm text-zinc-400">
+              Stops Comfy on the desktop GPU so it drops out of VRAM. The PC stays on. Generate or Wake GPU will start it again.
+            </p>
+            <div class="flex flex-wrap items-center gap-2">
+              <button
+                type="button"
+                class="rounded-xl border px-3 py-2 text-sm disabled:cursor-not-allowed disabled:opacity-60"
+                :class="gpuSleeping ? 'border-amber-300/40 text-amber-100' : 'border-white/15 text-zinc-200 hover:border-amber-300/50 hover:text-white'"
+                :disabled="gpuSleeping || gpuWaking"
+                @click="sleepGpu"
+              >
+                {{ gpuSleeping ? 'Stopping GPU…' : 'Sleep GPU' }}
+              </button>
+              <button
+                type="button"
+                class="rounded-xl border px-3 py-2 text-sm disabled:cursor-not-allowed disabled:opacity-60"
+                :class="gpuWaking ? 'border-amber-300/40 text-amber-100' : 'border-amber-400/40 text-amber-200 hover:border-amber-300 hover:text-amber-50'"
+                :disabled="gpuWaking || gpuSleeping"
+                @click="wakeGpu"
+              >
+                {{ gpuWaking ? 'Starting GPU…' : 'Wake GPU' }}
+              </button>
+            </div>
+            <p v-if="gpuMessage" class="text-xs" :class="gpuError ? 'text-red-300' : 'text-zinc-500'">{{ gpuMessage }}</p>
+          </div>
+        </section>
+
         <section class="mt-6">
           <h3 class="text-sm font-semibold uppercase tracking-wider text-zinc-500">Storage</h3>
           <div class="mt-2 rounded-2xl border border-white/10 bg-zinc-950/50 px-4 py-3">
@@ -256,6 +286,10 @@ const loginNext = ref('')
 const drafts = reactive<Record<string, { name: string; protect: boolean; password: string; currentPassword: string }>>({})
 const storageLoading = ref(true)
 const storageError = ref('')
+const gpuSleeping = ref(false)
+const gpuWaking = ref(false)
+const gpuError = ref(false)
+const gpuMessage = ref('')
 const storage = ref<{
   libraryBytes: number
   breakdown: { clips: number; stills: number; tmp: number; other: number }
@@ -392,6 +426,38 @@ async function loadStorage() {
   }
 }
 
+async function sleepGpu() {
+  if (gpuSleeping.value || gpuWaking.value) return
+  gpuSleeping.value = true
+  gpuError.value = false
+  gpuMessage.value = 'Stopping Comfy on the GPU…'
+  try {
+    await $fetch('/api/comfy/sleep', { method: 'POST' })
+    gpuMessage.value = 'Comfy is stopped. VRAM is free until you generate or wake the GPU.'
+  } catch (error: any) {
+    gpuError.value = true
+    gpuMessage.value = error?.data?.statusMessage || error?.statusMessage || 'Could not stop Comfy.'
+  } finally {
+    gpuSleeping.value = false
+  }
+}
+
+async function wakeGpu() {
+  if (gpuWaking.value || gpuSleeping.value) return
+  gpuWaking.value = true
+  gpuError.value = false
+  gpuMessage.value = 'Starting Comfy on the GPU…'
+  try {
+    await $fetch('/api/comfy/wake', { method: 'POST' })
+    gpuMessage.value = 'Wake sent. The first job can still take a couple of minutes if Comfy is cold.'
+  } catch (error: any) {
+    gpuError.value = true
+    gpuMessage.value = error?.data?.statusMessage || error?.statusMessage || 'Could not wake the GPU.'
+  } finally {
+    gpuWaking.value = false
+  }
+}
+
 async function loadLoraKeys() {
   loraRowsLoading.value = true
   try {

pages/index.vue

@@ -30,6 +30,17 @@
             >
               {{ gpuWaking ? 'Starting GPU…' : 'Wake GPU' }}
             </button>
+            <button
+              v-if="showSleepGpu"
+              type="button"
+              class="rounded-full border px-3 py-1 hover:border-amber-300/50 hover:text-white disabled:cursor-not-allowed disabled:opacity-80"
+              :class="gpuSleeping ? 'border-amber-300/40 text-amber-100' : 'border-white/15 text-zinc-300'"
+              :disabled="gpuSleeping || rigBusy"
+              :title="gpuWakeError || (rigBusy ? 'Cancel the running job first' : (gpuSleeping ? 'Stopping Comfy on the GPU' : 'Stop Comfy on the GPU to free VRAM. Does not shut the PC down.'))"
+              @click="sleepGpu"
+            >
+              {{ gpuSleeping ? 'Stopping GPU…' : 'Sleep GPU' }}
+            </button>
           </div>
           <span class="hidden md:block text-zinc-400">{{ userName }}</span>
           <button type="button" class="rounded-full border border-white/10 px-3 py-1 text-zinc-300 hover:border-amber-300/50 hover:text-white" @click="settingsOpen = true">
@@ -2453,6 +2464,7 @@ const comfyOk = ref(false)
 const comfyBusy = ref(false)
 const comfyAsleep = ref(false)
 const gpuWaking = ref(false)
+const gpuSleeping = ref(false)
 const gpuWakeError = ref('')
 const userName = ref('')
 const authEnabled = ref(false)
@@ -2726,8 +2738,17 @@ const liveMaxStep = computed(() => (
 const rigBusy = computed(() => (
   comfyBusy.value || imageComfyBusy.value || videoBusy.value || editBusy.value
 ))
-const showWakeGpu = computed(() => gpuWaking.value || (!comfyOk.value && !imageComfyOk.value))
+const showWakeGpu = computed(() => gpuWaking.value || (!gpuSleeping.value && !comfyOk.value && !imageComfyOk.value))
+const showSleepGpu = computed(() => gpuSleeping.value || ((comfyOk.value || imageComfyOk.value) && !gpuWaking.value))
 const rigBadge = computed(() => {
+  if (gpuSleeping.value) {
+    return {
+      label: 'Stopping…',
+      className: 'border-amber-400/40 text-amber-200',
+      dot: 'bg-amber-400 animate-pulse',
+      title: 'Stopping Comfy on the GPU'
+    }
+  }
   if (gpuWaking.value) {
     return {
       label: 'Starting…',
@@ -2857,6 +2878,7 @@ const recommendHint = computed(() => {
   return 'Send this still and draft to Beast Qwen VL.'
 })
 const gpuSleepHint = computed(() => {
+  if (gpuSleeping.value) return 'Stopping Comfy on the GPU to free VRAM. The PC stays on.'
   if (comfyOk.value || imageComfyOk.value) return ''
   if (gpuWaking.value) return 'Starting Comfy on the GPU. Generate will wait in the job queue until it answers.'
   if (comfyAsleep.value) return 'GPU is asleep. Generate will wake it — the first job can take a few minutes. Or click Wake GPU to pre-warm.'
@@ -3528,7 +3550,7 @@ async function pollComfyHealth() {
 }
 
 async function pokeGpu() {
-  if (gpuWaking.value) return
+  if (gpuWaking.value || gpuSleeping.value) return
   gpuWaking.value = true
   gpuWakeError.value = ''
   try {
@@ -3550,6 +3572,24 @@ async function pokeGpu() {
   }
 }
 
+async function sleepGpu() {
+  if (gpuSleeping.value || gpuWaking.value || rigBusy.value) return
+  gpuSleeping.value = true
+  gpuWakeError.value = ''
+  try {
+    await $fetch('/api/comfy/sleep', { method: 'POST' })
+    comfyOk.value = false
+    imageComfyOk.value = false
+    imageBeastOk.value = false
+    comfyAsleep.value = true
+    await pollComfyHealth()
+  } catch (error: any) {
+    gpuWakeError.value = error?.data?.statusMessage || error?.statusMessage || error?.message || 'Could not stop Comfy.'
+  } finally {
+    gpuSleeping.value = false
+  }
+}
+
 onMounted(async () => {
   try {
     withSound.value = localStorage.getItem('aigen-generate-sound') !== 'false'

server/api/comfy/sleep.post.ts

@@ -0,0 +1,22 @@
+import { listJobs } from '~/server/utils/jobs'
+import { listOwnersWithStudioQueues, listStudioJobs } from '~/server/utils/studioQueue'
+import { requestComfySleep } from '~/server/utils/comfyLifecycle'
+
+export default defineEventHandler(async () => {
+  const live = listJobs().filter(job => job.status === 'queued' || job.status === 'uploading' || job.status === 'running')
+  if (live.length) {
+    throw createError({
+      statusCode: 409,
+      statusMessage: 'A job is still running. Cancel it first, then sleep the GPU.'
+    })
+  }
+  for (const owner of listOwnersWithStudioQueues()) {
+    if (listStudioJobs(owner).some(job => job.status === 'running')) {
+      throw createError({
+        statusCode: 409,
+        statusMessage: 'A job is still running. Cancel it first, then sleep the GPU.'
+      })
+    }
+  }
+  return await requestComfySleep()
+})

server/utils/comfyLifecycle.ts

@@ -81,7 +81,7 @@ export async function fetchQueue() {
   }
 }
 
-async function controlRequest(path: string, method = 'GET') {
+async function controlRequest(path: string, method = 'GET', timeoutMs = 5000) {
   const { controlUrl, controlToken } = settings()
   if (!controlUrl) return null
   try {
@@ -91,10 +91,9 @@ async function controlRequest(path: string, method = 'GET') {
         Accept: 'application/json',
         ...(controlToken ? { Authorization: `Bearer ${controlToken}` } : {})
       },
-      signal: AbortSignal.timeout(5000)
+      signal: AbortSignal.timeout(timeoutMs)
     })
-    if (!res.ok) return { ok: false, status: res.status }
+    const body = await res.json().catch(() => null) as {
-    return await res.json() as {
       ok?: boolean
       http?: boolean
       process?: boolean
@@ -108,9 +107,23 @@ async function controlRequest(path: string, method = 'GET') {
       port?: number | null
       proxyPort?: number | null
       stopped?: boolean
+      killed?: unknown
       error?: string
+      running?: number
+      pending?: number
       status?: number
+    } | null
+    if (!res.ok) {
+      return {
+        ok: false as const,
+        status: res.status,
+        error: body?.error,
+        running: body?.running,
+        pending: body?.pending,
+        killed: body?.killed
+      }
     }
+    return body || { ok: true }
   } catch {
     return null
   }
@@ -151,6 +164,43 @@ export async function requestComfyWake() {
   }
 }
 
+export async function requestComfySleep() {
+  const { controlUrl } = settings()
+  if (!controlUrl) {
+    throw createError({
+      statusCode: 503,
+      statusMessage: 'COMFY_CONTROL_URL is not set. The studio cannot poke the Beast host agent.'
+    })
+  }
+  const remote = await controlRequest('/stop', 'POST', 25_000)
+  if (!remote) {
+    throw createError({
+      statusCode: 503,
+      statusMessage: 'Beast host agent is unreachable. Confirm it is running on the desktop.'
+    })
+  }
+  if (remote.status === 409 || remote.error === 'queue-busy') {
+    const running = Number(remote.running || 0)
+    const pending = Number(remote.pending || 0)
+    throw createError({
+      statusCode: 409,
+      statusMessage: `Comfy is still working (${running} running, ${pending} queued). Cancel the job first, then sleep the GPU.`
+    })
+  }
+  if (remote.ok === false) {
+    throw createError({
+      statusCode: Number(remote.status) || 502,
+      statusMessage: 'Beast host agent refused to stop Comfy.'
+    })
+  }
+  return {
+    ok: true,
+    stopped: true,
+    asleep: true,
+    killed: remote.killed ?? null
+  }
+}
+
 async function findLocalComfyProcess() {
   const platform = process.platform
   try {