Tighten YuE for 16GB VRAM and stop the false Waiting status.

Cap Stage-A segments and use a heavier mmgp profile so long lyrics stop OOMing mid-run, and show that Comfy is actually working.

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

pages/music.vue

@@ -411,14 +411,35 @@ async function refreshStudioQueue() {
     if (queued.value && !jobId.value) {
       queued.value = false
     }
+    if (jobId.value && /waiting in the job queue/i.test(status.value)) {
+      status.value = selectedEngine.value === 'yue'
+        ? 'YuE running on Comfy — Stage A can take 10–20+ minutes'
+        : 'Generating…'
+    }
     return
   }
   if (row.status === 'waiting' || (row.status === 'held' && !row.liveJobId)) {
     queued.value = true
     busy.value = true
-    status.value = row.status === 'held' ? 'Paused in the job queue' : 'Waiting in the job queue…'
+    status.value = row.status === 'held'
+      ? 'Paused in the job queue'
+      : (selectedEngine.value === 'yue'
+        ? 'Waiting for GPU… YuE is heavy; do not spam Generate'
+        : 'Waiting in the job queue…')
     return
   }
+  if (row.status === 'running' && row.liveJobId) {
+    queued.value = false
+    if (!jobId.value || jobId.value !== row.liveJobId) {
+      jobId.value = row.liveJobId
+      listen(row.liveJobId)
+    }
+    if (/waiting/i.test(status.value)) {
+      status.value = selectedEngine.value === 'yue'
+        ? 'YuE running on Comfy — Stage A can take 10–20+ minutes'
+        : 'Generating…'
+    }
+  }
   if (row.status === 'error') {
     queued.value = false
     busy.value = false

server/assets/workflow_yue.json

@@ -5,9 +5,9 @@
       "xcodec_ckpt": "ckpt_00360000.pth",
       "quantization_model": "fp16",
       "use_mmgp": true,
-      "stage1_cache_size": 16384,
+      "stage1_cache_size": 8192,
       "exllamav2_cache_mode": "FP16",
-      "mmgp_profile": 2
+      "mmgp_profile": 4
     },
     "class_type": "YUE_Stage_A_Loader",
     "_meta": {
@@ -23,7 +23,7 @@
       "repetition_penalty": 1.1,
       "prompt_start_time": 0,
       "prompt_end_time": 30,
-      "max_new_tokens": 3000,
+      "max_new_tokens": 2000,
       "use_dual_tracks_prompt": false,
       "use_audio_prompt": false,
       "offload_model": true,
@@ -38,8 +38,8 @@
   "16": {
     "inputs": {
       "stage_B_repo": "m-a-p/YuE-s2-1B-general",
-      "stage2_cache_size": 8192,
+      "stage2_cache_size": 4096,
-      "stage2_batch_size": 2,
+      "stage2_batch_size": 1,
       "exllamav2_cache_mode": "FP16",
       "use_mmgp": true,
       "info": ["8", 1]

server/utils/musicChain.ts

@@ -32,8 +32,8 @@ function sleep(ms: number) {
 
 function classifyError(message: string) {
   const lower = message.toLowerCase()
-  if (lower.includes('out of memory') || (lower.includes('cuda') && lower.includes('alloc')) || lower.includes('vram')) {
+  if (lower.includes('out of memory') || lower.includes('exceed allowed memory') || (lower.includes('cuda') && lower.includes('alloc')) || lower.includes('vram')) {
-    return 'ComfyUI VRAM allocation failed. Try fewer steps or a shorter duration.'
+    return 'ComfyUI ran out of VRAM. For YuE use shorter lyrics (fewer sections) or ACE-Step instead.'
   }
   if (lower.includes('econnrefused') || lower.includes('unreachable') || lower.includes('fetch failed')) {
     return 'ComfyUI host connection dropped. Confirm the desktop instance is running.'
@@ -245,11 +245,18 @@ function watchMusicJob(job: Job): Promise<void> {
           return
         }
         const entry = (job.promptId ? history?.[job.promptId] : undefined) as {
-          status?: { status_str?: string; completed?: boolean }
+          status?: {
+            status_str?: string
+            completed?: boolean
+            messages?: Array<[string, Record<string, unknown>]>
+          }
         } | undefined
         const status = entry?.status?.status_str
-        if (status === 'error') await fail('ComfyUI reported an execution error')
+        if (status === 'error') {
-        else if (status === 'interrupted') await fail('Job interrupted.')
+          const err = entry?.status?.messages?.find(([type]) => type === 'execution_error')?.[1]
+          const detail = String(err?.exception_message || err?.message || '').trim()
+          await fail(detail || 'ComfyUI reported an execution error')
+        } else if (status === 'interrupted') await fail('Job interrupted.')
       } catch {
         // history can miss while the graph is still running
       }
@@ -382,14 +389,19 @@ export async function startMusicJob(params: MusicJobParams) {
       try {
         const queued = await queuePrompt(graph, job.clientId, { timeoutMs: 25_000 })
         if (queued?.prompt_id) job.promptId = queued.prompt_id
-        emitJob(job, { type: 'status', message: 'Job queued on ComfyUI', progress: 8 })
+        const runningMessage = params.engine === 'yue'
+          ? 'YuE Stage A on Comfy — 16GB job, often 10–20+ minutes. Leave it alone.'
+          : 'Job queued on ComfyUI'
+        emitJob(job, { type: 'status', message: runningMessage, progress: 8 })
       } catch (queueError: unknown) {
         const statusCode = Number((queueError as { statusCode?: number; status?: number })?.statusCode
           || (queueError as { status?: number })?.status || 0)
         if (statusCode >= 400) throw queueError
         emitJob(job, {
           type: 'status',
-          message: 'Waiting for ComfyUI to pick up the job...',
+          message: params.engine === 'yue'
+            ? 'YuE submitted — waiting for Comfy to start Stage A…'
+            : 'Waiting for ComfyUI to pick up the job...',
           progress: 8
         })
       }

server/utils/musicWorkflow.ts

@@ -116,7 +116,9 @@ function buildYueWorkflow(params: MusicWorkflowParams): WorkflowGraph {
   const prefix = params.filenamePrefix || musicFilenamePrefix()
   const tags = params.tags
   const lyrics = params.lyrics || '[instrumental]\n'
-  const segments = Math.max(1, Math.min(10, lyrics.split(/\[[^\]]+\]/g).filter(part => part.trim()).length || 2))
+  // 5080 16GB OOMs past ~3 long Stage-A segments on fp16+mmgp. Cap hard.
+  const sections = lyrics.split(/\[[^\]]+\]/g).filter(part => part.trim()).length || 2
+  const segments = Math.max(1, Math.min(2, sections))
   for (const node of Object.values(graph)) {
     if (node.class_type === 'YUE_Stage_A_Loader') {
       node.inputs.stage_A_repo = yueStageARepo()
@@ -128,8 +130,9 @@ function buildYueWorkflow(params: MusicWorkflowParams): WorkflowGraph {
       node.inputs.genres_prompt = tags.replace(/,/g, ' ').replace(/\s+/g, ' ').trim()
       node.inputs.lyrics_prompt = lyrics
       node.inputs.seed = params.seed
-      node.inputs.prompt_end_time = params.duration
+      node.inputs.prompt_end_time = Math.min(params.duration, 45)
       node.inputs.run_n_segment = segments
+      node.inputs.max_new_tokens = 2000
     }
   }
   applyPrefix(graph, prefix)