Coordinate shared GPU access across studio instances

server/utils/promptComfy.ts

@@ -1,4 +1,5 @@
-import { comfyInputFilename } from '~/server/utils/comfy'
+import { assertSharedGpu, sharedGpuHeaders, sharedGpuConfigured } from '~/server/utils/sharedGpu'
+import { comfyInputFilename, getComfyHost } from '~/server/utils/comfy'
 import { getSidecarImageHost } from '~/server/utils/imageComfy'
 import { ensureSidecarReady } from '~/server/utils/imageComfyLifecycle'
 import { buildVisionPromptWorkflow } from '~/server/utils/promptWorkflow'
@@ -19,6 +20,7 @@ const jobs = new Map<string, PromptJob>()
 const MAX_JOBS = 20
 
 export function getPromptComfyHost() {
+  if (sharedGpuConfigured()) return getComfyHost()
   return getSidecarImageHost()
 }
 
@@ -31,6 +33,12 @@ async function promptComfyFetch(path: string, init?: RequestInit) {
   if (!host) {
     throw createError({ statusCode: 503, statusMessage: 'Qwen VL is not configured. Set COMFY_HOST or IMAGE_COMFY_HOST to Beast.' })
   }
+  if (init?.method && !['GET', 'HEAD'].includes(init.method.toUpperCase())) {
+    await assertSharedGpu()
+    const headers = new Headers(init.headers)
+    for (const [key, value] of Object.entries(sharedGpuHeaders())) headers.set(key, value)
+    init = { ...init, headers }
+  }
   try {
     return await fetch(`${host}${path}`, init)
   } catch (error) {
@@ -260,3 +268,7 @@ export function failPromptJob(job: PromptJob, error: unknown) {
   job.error = err.statusMessage || err.message || 'Prompt recommend failed'
   job.message = job.error
 }
+
+export function hasActivePromptJobs() {
+  return [...jobs.values()].some(job => job.status === 'queued' || job.status === 'running')
+}